# Built with Claude: Life Sciences: Project Gallery - **Event:** [Built with Claude: Life Sciences](https://cerebralvalley.ai/e/built-with-claude-life-sciences) - **When:** Jul 7 at 12:00 PM – Jul 14 at 12:00 AM (EDT) - **Where:** Online - **Hosts:** [Anthropic](https://cerebralvalley.ai/u/anthropicai), [Cerebral Valley](https://cerebralvalley.ai/u/cv) - **Projects:** 299 (7 placed) - **Page:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery ## Projects ### 1. Splice with Claude I asked which RNA-splicing changes in liver cancer actually affect patient survival — and whether isoform switching correlated with survival rather than gene expression. Using only public data, I screened 23,909 splicing events in TCGA-LIHC against overall survival (382 significant, FDR<0.05), correlated splicing with matched protein abundance, and validated 18 survival-associated genes at the protein level in an independent cohort (CHCC-HBV deep proteome). I independently re-quantified tumor-vs-normal splicing with two orthogonal pipelines (SUPPA2 and LeafCutter), which agree on 90 genes (p=1.9×10⁻¹³). Running survival on the exact differentially-spliced events showed that 8 of 18 survival-defining genes, including the lead gene FN1, are prognostic through isoform switch alone, with no expression change: biology a conventional differential-expression analysis would miss. FN1's survival-linked exon is the IIICS/V oncofetal-fibronectin cassette, a splice-switching-ASO target hypothesis. Why it matters: I'm a wet-lab RNA-splicing PI. This is a multi-cohort, dual-pipeline proteogenomic + survival study that would normally need a bioinformatician, a splicing-methods specialist, and a biostatistician over months. With Claude Science I ran the whole thing myself, in days, including catching and correcting my own false leads and validating across two cohorts. The expertise was always mine; what changed is I no longer depend on a team to act on it and/ or move the field forward. - **Team:** [SAFIYA KHURSHID](https://cerebralvalley.ai/u/soprich2) - **GitHub:** https://github.com/mohsinahad/liver-cancer-splicing-survival - **Demo video:** https://youtu.be/b9uC9Rmsp2s - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=276 ### 2. Todd Appleby Vidalign is a local, GPU-first web workbench that puts behavior video and the rest of an experiment's data — sensor traces, operant/Med-PC hardware logs, human annotations — onto one shared master clock, then turns the aligned footage into quantitative, exportable measures of what an animal did and when. It runs a full pipeline, each stage usable on its own: import and align heterogeneous, independently-clocked data on a single timeline (by wall-clock, by dragging, or by auto-aligning a sensor against the animal's motion); track the animal with SAM3 (a text prompt or one click) and distill a fast per-project detector that sweeps whole sessions for presence; embed and sort tracked figures across recordings with DINOv3; matte them into clean cutouts; cluster movement and posture into a labeled ethogram without hand-scoring; and model event-locked structure with CASTLE-style focused latents and press-aligned CEBRA. Everything runs on your own machine — a FastAPI backend and React frontend— a project is just a folder with a SQLite database, and source media is referenced in place and never modified. What we found - Promptable video segmentation is now good enough to track a lab animal with essentially zero setup. On real top-down IR footage from a 3-hour operant self-administration session, SAM3 held a full mask on a freely-moving mouse, through an approach to the active lever and away again, from a single click, at maximum confidence, with no training. That single capability removes the biggest manual bottleneck in behavior work. You can bootstrap a whole-session detector from the researcher's own segmentation, with no external labels — if you get the negatives right. The key insight was a failure mode: harvesting "empty-looking" frames as negatives quietly teaches a still or sleeping animal as background. Training only on verified detections plus mined false positives fixes it. Once video and the operant log share one master clock, event-locked representation learning becomes possible: press-aligned CEBRA over mask-focused visual features yields an embedding organized by time relative to a lever press, evaluated with a held-out recording. Just as important, we kept nonlinear projection (UMAP) out of the scientific labeling path — labels come from the feature space, so a visualization choice can never change a result. Behavioral neuroscience is throttled by two unglamorous problems: behavior is still scored by hand, frame by frame, and the data that explains it lives in separate devices on clocks that were never synchronized. vidalign collapses both into one reproducible, local workbench, and puts modern vision and representation methods (SAM3, DINOv3, CEBRA) within reach of labs that could never assemble that stack themselves. The payoff is fewer hours spent aligning, labeling keypoints, and scoring, and a straight path from raw recordings to analyzable, event-aligned behavior. - **Team:** [Todd Appleby](https://cerebralvalley.ai/u/tappleby) - **GitHub:** https://github.com/toddappleby/vidalign - **Demo video:** https://www.youtube.com/watch?v=4H9AsClf0GA - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=277 ### 3. Genome Therapy Sequence-to-expression "oracles" (Enformer, Borzoi, AlphaGenome, Evo2) and the generative-DNA designers built on them all aim to predict how cis-regulatory DNA sets a gene's expression. But they are rarely scored on the quantity that matters for regulatory therapeutics: the change in expression (Δ) a real perturbation produces, measured against experimental ground truth and against a trivial baseline. CIS-Δ is a reproducible benchmark that does exactly that. It scores a model's predicted Δ for a (cell-context, cis-intervention) pair against experimental truth — single-cell CRISPRi/CRISPRa, eQTL fine-mapping, reporter MPRA — across seven scoring axes, anchored on 90,955 K562 CRISPRi pairs. We ship a one-file submission format, a scoring harness with gene-clustered bootstrap CIs and DeLong tests, an auto-generated leaderboard, and a self-test that round-trips committed truth to one part in a million. The finding is a carefully-bounded negative result. On distal-enhancer CRISPRi, a naive distance-to-TSS baseline out-ranks Borzoi and Enformer (AUROC 0.87), because 71–88% of real regulatory links fall outside the models' receptive field. On natural-variation eQTLs the oracles do clear the distance prior (AlphaGenome, significant after FDR correction) — but statistical fine-mapping still out-ranks them. And cell-type specificity can't be demonstrated at all: the eQTL ground truth is 93–96% shared across tissues, so that axis is under-powered by construction. Even the 2026 state of the art does not close the gap. Why it matters: before trusting a sequence model to design a regulatory edit, you need to know it can score one better than distance-to-TSS. CIS-Δ is the instrument that checks — and a living leaderboard the field can submit to. - **Team:** [Qihao Duan](https://cerebralvalley.ai/u/Qihao) - **GitHub:** https://github.com/Qihao-Duan/PerturbExpression.git - **Demo video:** https://youtu.be/2IDTpwpvsTU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=278 ### 4. geoscope I built GEOscope, a discovery layer for NCBI GEO’s 288,904 public functional-genomics studies. GEO metadata is largely free text, so relevant datasets are often missed when researchers and submitters use different vocabulary. GEOscope processes the full corpus into canonical study records, normalizes organisms, sex, and assay concepts, generates embeddings, and indexes everything in Elasticsearch. Search combines BM25, semantic retrieval, live NCBI candidates, and Claude reranking. The same search powers a side-by-side website and a three-tool MCP server for Claude and other agents. I found that semantic retrieval can surface highly relevant studies absent from the first 100 candidates displayed by GEO’s literal keyword search. I also found that embeddings alone were insufficient: normalized metadata and hybrid retrieval were essential for precise, trustworthy discovery. This matters because researchers can find reusable public datasets using natural-language scientific questions instead of guessing the exact terminology used by submitters. This helps valuable existing data get discovered and reused. - **Team:** [kevin w](https://cerebralvalley.ai/u/coffeetime) - **GitHub:** https://github.com/kevinwuhoo/GEOscope https://geoscope.kevinformatics.com/ - **Demo video:** https://www.youtube.com/watch?v=_bppapP2Ons - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=279 ### 5. James Hu Bayleaf is an AI-assisted decision gate for genomics quality control. The context needed to trust a run is usually scattered across tools that do not talk to each other. Bayleaf brings that context together and follows the run end to end, from composing the pipeline to reviewing the results and preserving a record of why each decision was made. Operators build pipelines as graphs of typed tool cards that compile to runnable Nextflow. Claude can help author a step using a vetted, cited catalog, but pipeline design and execution remain separate and approval-gated. When results come back, deterministic rules assign each sample a verdict: proceed, hold, rerun, or escalate. Each finding is tied to the metric and runbook threshold that triggered it. Missing or unmeasured QC is treated as a signal, not a silent pass. Claude explains decisions and answers questions with citations, helping the operator move faster. But the verdict always comes from the rules, never the model. Underneath, an append-only event log and content-hashed artifacts make every decision reproducible. The log is the source of truth, while the database is a rebuildable projection. Bayleaf is a decision-support tool, not a clinical system. Thresholds are configurable policy, and confidence values are heuristics rather than calibrated probabilities. - **Team:** [James Hu](https://cerebralvalley.ai/u/jchu0) - **GitHub:** https://github.com/jchu0/bayleaf - **Demo video:** https://youtu.be/vlOZRzLoP7Y - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=280 ### 6. Ali Chegini **Project description** This repository is the **computational side** of a PhD thesis (Schimmer lab, U of T Medical Biophysics): *Identifying Variable Dependencies in the Mitochondrial Proteome in AML.* The goal is to find genes that are **variably essential** — lethal in only a *subset* of AML myeloid cell lines rather than pan-essential — from DepMap CRISPR gene-effect screens, then follow the biology. **Built:** a variability engine (apps/DepMapVariablyEssential/) that scores selective essentiality across several algorithms (Static, Dynamic, the PCA/RPCA/PPCA families, Likelihood1D, plus ML classifiers), with a posterior decision-boundary framework and GO/MitoCarta3/Reactome pathway-GSEA. A core methodological point: raw variance ≠ selective essentiality — pan-essential genes swinging between "death and more death" look variable but aren't real vulnerabilities; a true hit must *cross* the essential/non-essential boundary. The pipeline was also made deterministic (exact SVD, seeded PPCA, sorted intersections). **Found:** the leading hits converge on **mitochondrial translation** — the mitoribosome and mitochondrial aminoacyl-tRNA synthetases (mt-aaRSs) — with **RARS2** as the flagship gene and YARS2/FARS2/EARS2/NARS2 as the validation panel. **Why it matters:** selective dependencies are candidate AML therapeutic targets that spare cells lacking the vulnerability; wet-lab validation follows. - **Team:** [Ali Chegini](https://cerebralvalley.ai/u/AliChegini) - **GitHub:** https://github.com/AliChegini97/Built-with-Claude-Life-Sciences.git https://github.com/AliChegini97/aars-cyto-mito-similarity.git - **Demo video:** https://youtu.be/qwAC-gHnGkI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=281 ### 7. FibrosisExplorer I built an interactive fibrosis enhancer combining direct perturbation CRISPR screen data (perturbseq) and evolutionary signals - **Team:** [Yasir Suhail](https://cerebralvalley.ai/u/ys_) - **GitHub:** https://yasirs.github.io/gladstone-hackathon/ - **Demo video:** https://www.youtube.com/watch?v=dXZiWpcys_E - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=283 ### 8. RWE-Agent What we built. We created a Real-World Evidence Methodologist — an AI specialist that acts as a built-in epidemiologist and causal-inference expert, sitting alongside any life-sciences analysis that relies on observational data (patient registries, electronic health records, insurance claims, biobanks, or clinical-trial data being reused for a new question). Why it matters. In life sciences, the hard part of learning from real-world data is almost never running the statistics — it's the design decisions made before the analysis: who counts as a patient, how a disease or drug exposure is actually defined in messy data, and when the clock starts for each person. Get those wrong and you get confident, published-looking answers that are simply artifacts of bias — a drug looks protective when it isn't, a biomarker looks predictive when it's really just tracking how sick someone was to begin with. This specialist exists to stop that failure at the source. What it does. Instead of jumping straight to a result, it first asks how comfortable you are with these methods and adjusts its language to match — full technical depth for a biostatistician, plain guidance for a bench scientist or clinician — then walks you through the study design interactively, turning each choice into a simple picture you approve: a cohort diagram showing exactly who's included or excluded and why, side-by-side cards comparing candidate definitions of a condition or outcome, and timelines that visually expose traps like "immortal time" and selection bias. Only once the design is sound does it move to any computation, and it flags every assumption and hidden pitfall along the way. Whom it impacts: It lowers the barrier for the many researchers, clinicians, and translational scientists who have valuable real-world datasets but aren't trained epidemiologists — letting them ask rigorous causal questions without accidentally fooling themselves — while giving trained methodologists a faster, visual, reproducible way to specify and document a defensible study. The net effect is fewer spurious findings, more trustworthy evidence, and study designs that are transparent and reviewable before a single result is generated. - **Team:** [Nayan Chaudhary](https://cerebralvalley.ai/u/nayanchaudhary) - **GitHub:** https://github.com/nayanchaudhary/claude-science-rweAgent - **Demo video:** https://youtu.be/l-CmxNZDHK4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=289 ### 9. NUDGE NUDGE — an "honesty engine" for AI-driven science. Point it at a differentiable model of a biological system; it tells you which parameters you can actually trust, and designs the experiment that resolves the ones you can't. Why we built it: AI can now fit any model and hand you confident numbers — including confident wrong ones. In a drug trial that's dangerous: when amyloid plaque falls, you can't tell if the drug did it or the body did. We wanted a tool that reports what it measured, not what it guessed. What we found: a capable AI agent can already derive this math by hand — so the real value isn't capability, it's trust: NUDGE gives audited, reproducible, guarded answers. It refuses to over-abstain (won't throw out a sloppy-but-predictive model), refuses to over-claim (won't fake a number on an unidentifiable design), designs the trial that separates a drug's effect from natural clearance (~259× sharper, same scans), and every figure ships with the code to reproduce it byte-for-byte. - **Team:** [Nicholas Roy](https://cerebralvalley.ai/u/nicholasroy) - **GitHub:** https://github.com/NicholasEhsanRoy/NUDGE - **Demo video:** https://youtu.be/GM831aIhIBU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=282 ### 10. The Regulators We built a reproducible discovery-to-target pipeline using a genome-scale CRISPRi Perturb-seq atlas of primary human CD4+ T cells (Marson 2025), scoring every perturbation along two axes: T-cell activation and Th1/Th2 polarization. We benchmarked each axis against the authors' published regulator coefficients before ranking any candidates. The polarization axis showed strong concordance (Spearman ρ = 0.72; AUROC = 0.92 for known regulators), with directional agreement across 21 established Th1/Th2 regulators; the activation axis showed weak concordance (ρ = 0.13) and was therefore retained as exploratory only. The validated polarization axis recovered established immunoregulatory targets in the expected direction, including IFNAR1, the target of approved anifrolumab, and TNFRSF9, a therapeutically pursued immunoregulatory target. The central finding is a trade-off between novelty and tractability: the most reproducible potentially novel hits are intracellular scaffolds and chromatin factors that are difficult to drug. TATDN2 emerged as our strongest potentially novel lead: its knockdown produced a reproducible Th2-skewing phenotype, and it is annotated as a metal-dependent nuclease with a potentially tractable catalytic pocket. These results generate testable target hypotheses for allergic and atopic disease and for Th1-driven autoimmune conditions such as MS, IBD, and RA, while leaving the indication-specific therapeutic direction to be established experimentally. - **Team:** [Xiang Niu](https://cerebralvalley.ai/u/steveniu) - **GitHub:** https://github.com/stevexniu/tcell-perturbseq-target-discovery - **Demo video:** https://www.loom.com/share/91bf703330484310af4af853092f8586 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=284 ### 11. Jedidiah Carlson Biological research is increasingly reliant on complex computational pipelines, often involving multiple steps of QCing, transforming, processing, ingesting, and summarizing data on the journey from raw biological material to scientific insights. For any one of these tasks, there could be dozens of published software tools that have been developed to parse data for subsequent analysis. A ubiquitous problem for computational biologists is simply choosing which tool to use for their particular dataset and research questions. Though independent benchmarking studies are routinely published to compare different software tools at a high level, these benchmarks often focus on broad measures of output data quality and/or computational performance using heavily vetted "gold standard" datasets. In reality, it is difficult to make judgement calls about which tools is best suited for specific data, research questions, or practical constraints. Moreover, the pace at which new tools are published makes it difficult for researchers to stay on top of (much less implement and test) emerging methods. BioBakeoff is a collection of AI agent skills that creates sandboxes for on-the-fly benchmarking of various tools for common bioinformatic tasks. Triggering the agent prompts users to provide the following: 1. Sample data 2. Relevant research question(s) being pursued by the analysis 3. Any practical constraints (compute resources, timeline, cost, etc.) With this information, BioBakeoff then spins up a customized Docker orchestrator that runs a set of candidate tools (without the user ever needing to deal with complex installations/dependencies or tool-specific outputs), aggregates the results using standardized benchmarking statistics, and produces a brief custom report detailing how the tools performed relative to one another, along with suggestions and caveats to help inform researchers' decision about which tool to proceed with. ### Notes Currently BioBakeoff focuses on variant calling as a proof-of-concept, but I will soon be extending it to other genomics analysis tasks. I have intentionally kept this as a low-level implementation rather than getting fancy with a GUI, etc. Ideally, this agent can be hooked into existing platforms and workflows. - **Team:** [Jedidiah Carlson](https://cerebralvalley.ai/u/jedmsp) - **GitHub:** https://github.com/carjed/biobakeoff - **Demo video:** https://github.com/carjed/biobakeoff/tree/main/demo - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=285 ### 12. PHANTOM What I built / investigated A degree-bias assessment for protein–protein interaction (PPI) ranking. The central question: when a PPI method ranks candidate partners for a protein complex, how much of that ranking reflects genuine, specific binding versus a confound — the fact that well-studied, highly-connected “hub” proteins tend to score high for everything? I attacked this on two levels: A controlled complex example — a 29-protein network spanning three complexes (Arp2/3, CCT/TRiC, MCM), with a naive network score and a degree-corrected re-ranking, to show the mechanism cleanly. Real data vs. a published model — the same degree assessment applied to two experimental interactomes (BioPlex 3.0, 293T and HCT116; score = CompPASS-Plus pInt) and to a sequence-only deep predictor (D-SCRIPT human_v1, run from the published model over its own benchmark pairs; score = interaction probability). The one label-free, directly comparable statistic across all of them is the Spearman correlation between score and node degree. What I found The direction of the degree relationship flips between experimental and computational scores. BioPlex top scores are negatively correlated with degree (−0.097 / −0.110) — they favor specific, low-degree pairs. D-SCRIPT is strongly positive (+0.659, far beyond a degree-preserving null, p ≈ 0.001) — its top scores are enriched for hubs. Most of the predictor’s apparent skill on its benchmark is a sampling artifact. Benchmark positives have higher endpoint degree (6.6) than the randomly-sampled negatives (2.8), so a biology-free degree-only baseline recovers the labels at AUROC 0.905. D-SCRIPT reaches 0.944 — a biology premium of only +0.039 AUROC over degree alone. In the controlled example, ranking inverts biology: the hub ACTB (degree 23) ranks #1 while the specific true member MCM6 (degree 6) is buried at #29. Degree correction promotes the seven Arp2/3 members and the seven CCT/TRiC members to the top as confirmed targets and demotes the frequent-flyer hubs to “hypothesis only.” Why it matters A low-degree, top-ranked pair in experimental data is plausibly a true, specific interaction. The identical position in a computational predictor evaluated on a standard random-negative benchmark instead tends to reward high-degree hubs — so top-ranked computational predictions cannot be read as “specific” the way experimental ones can. For anyone using PPI predictors to nominate targets, a large fraction of the ranking signal can be study/sampling bias rather than binding, and a degree control is a cheap, label-free way to see how much. The framing is deliberately conservative: the assessment measures what survives a degree control, not ground-truth correctness — it reports a provable “% artifact” and a small non-zero “survives-assessment” premium, and does not claim “% real.” Scope caveat: D-SCRIPT was assessed on a seed-42 subset of 858 pairs (612 proteins), not the full 52,725-pair test set (~8 h of GPU embedding); the subset preserves the degree structure, so extending it is a compute question, not a method change. - **Team:** [Cathy Fung](https://cerebralvalley.ai/u/Sin) - **GitHub:** https://github.com/madnessfish/phantom - **Demo video:** https://youtu.be/BkIxWZOsxIc - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=286 ### 13. Abraham Most locally advanced rectal cancers are mismatch-repair proficient and have no molecularly-guided treatment. We asked whether drugs that already exist could help, guided by the biology that makes these tumors resist chemoradiation. Anchoring to the chemoradiation-resistance gene program (GSE209746), two independent engines — one by drug target, one by measured kinase binding — nominated 287 clinically-available drugs. A permutation test confirmed the anchor is non-arbitrary (p=0.001 vs 1,000 random gene sets). FDA-label safety filtering distilled this to two ranked shortlists of 44 drugs, ready for patient-derived organoid testing. Claude drove the full scientific method: planning, building both engines, integrating eleven public databases, reasoning through methodology, drafting the manuscript, then acting as a critical peer reviewer that exposed weak claims — which we revised into a stronger, defensible result. One researcher, the complete arc of science. - **Team:** [Abraham Lin](https://cerebralvalley.ai/u/ALin7) - **GitHub:** https://github.com/ALin-Orbits/Claude_Science_Hackathon - **Demo video:** https://youtu.be/PGftJQQX3z8 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=288 ### 14. CTRA - CD4+ T Cell Regulator Atlas I built an integrated, honestly-scored prioritization of gene-regulatory "master regulators" in primary human CD4+ T cells from the Marson/Pritchard 2025 genome-scale CRISPRi Perturb-seq screen (~22M cells, 4 donors, 3 stimulation states). Eight functional-genomics evidence layers shipped with the screen — trans-effect breadth, on-target CRISPRi QC, cross-guide and cross-donor reproducibility, gene-regulatory-network module membership, autoimmune-disease enrichment, polarization/aging regulator role, and K562 cross-cell-type specificity — are joined into ONE master table of 33,983 perturbed-gene × culture-condition rows, with a transparent additive regulator_score (0–9) that rewards convergent evidence. Because the paper's core thesis is that active regulators change with stimulation state, the table is stratified by Rest/Stim8hr/Stim48hr, never pooled. What I found: the ranking recovers known CD4+ biology unsupervised — top regulators are BATF (9), STAT3/STAT6/GATA3/RASA2 (8), and the broadest trans-effect hubs are the proximal TCR machinery (CD3E, LAT, ZAP70, PLCG1; >5,000 downstream genes each). Context-specificity is quantified with a model-free reimplementation of the cz-benchmarks Perturbation Expression Prediction metric: the same perturbation's effect transfers across stimulation states only modestly (median Spearman rho ~ 0.20, ~20x a random baseline) and degrades with context distance — a direct confirmation of the paper's thesis and a floor any model (including scLDM.CD4) must beat. Autoimmune-disease-enriched GRN modules are most numerous at Stim8hr, several of them novel programs. Why it matters: this turns a 22-million-cell screen into a triage-ready shortlist of context-specific, disease-linked CD4+ regulators — with honest scoring (a documented heuristic, not a black box) and 36 automated raw-vs-built integrity checks that all pass, so every number is reproducible and auditable. It's a template for turning any genome-scale Perturb-seq resource into a prioritized, disease-anchored regulator map. - **Team:** [JIN KIM](https://cerebralvalley.ai/u/Jinky) - **GitHub:** https://github.com/jink-ucla/Claude_Hackathon - **Demo video:** https://youtu.be/sAe4PzEq3rE?si=AER5FTJo6CDBHCGh - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=297 ### 15. scRegen We are always looking to find inspirations to regain a more regenerative and youthful states in medicine so we can treat patients with disease and injury by optimizing their own healing abilities. One place to find that is to look for inspiration in the natural development of organs. Thus, we first reconstruct by looking into the existing dataset of a fully multiomic covered time course organogenesis Single cell dataset. Here, for example, we use lung as the primary focus where we will expand into other organs when data are more flushed out and possible. After the compilation, we reconstruct the gene regulatory network throughout the time course and look for key master regulators that mediates critical selfie changes throughout time. After that, We lock in between the lineage specific progenitors, in this case in lung, is the lung distal progenitors that are multipotent and also the mature eighty two cell that is the facultative stem cells that in injury context can have limited regenerative capacity. By analyzing the differences between the progenitors and the matured facultative stem cell groups, we find critical changes in gene regulatory network and dose master regulator and perform transcriptional factor interventions via virtual knockout or overexpression. By iterating all kinds of combo, we optimize for intervention combo that can maximize scoring that approach progenitor states the most. The package that has been involved here is a cell oracle, Scenic Plus, and also Pertformer. We also tested a currently popular way to reactivate and rejuvenate in vivo that has been, um, moving into clinical trials where a partial or interrupted reprogramming using, uh, Yamanaka factors or as it's known as the OSKM. To see if the exposure of those can have an effect in bringing mature eighty two cells back to a distal progenitor fade, and we benchmark that against the new combination that we can mine by the development reconstructed datasets, gene regulatory network predictions. - **Team:** [Jimmy Cai](https://cerebralvalley.ai/u/jimmychai) - **GitHub:** https://github.com/thejimmychai/scElixir_Claude_Hackerthon - **Demo video:** https://youtu.be/vYEiUnVVBLQ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=287 ### 16. JAI SOGANI Keystone is a Claude-powered discovery engine for life sciences that verifies evidence before it reasons. Across four real disease programs — immunology (CD4+ T-cell/Th2), glioblastoma, brain hemorrhage, and diabetes — built on real literature, real retraction records, and real Perturb-seq data from Gladstone Institutes, a coordinated cell of Claude agents (planner, evidence-quality checker, contradiction miner, adversarial reviewer, principal investigator) proposes and ranks competing hypotheses. But every claim is gated by a deterministic integrity check first — nothing reaches a conclusion unverified. Scientists can also feed in their own reference list, their own bench data (plate-reader CSVs), or any gene symbol, and get the same live integrity-and-reasoning pipeline run on their own work, not a curated demo. What we found: A gated, reviewer-checked pipeline of 8 agents admits zero unverified claims into its final recommendation; an ungated swarm of 300 agents running the identical task cites unverified sources at real, measured cost — proof that more agents isn't better science, verification is. Our load-bearing citation classifier agrees with human annotators at 0.82, inside the human-to-human agreement band (0.69–0.75) on the same task. And the system caught a real, still-uncorrected problem in the wild: a foundational glioblastoma paper is retracted, and a widely used cell line (U-87MG) is independently flagged as misidentified — Keystone surfaces both automatically before letting any hypothesis build on them. Why it matters: The bottleneck in AI for science was never generating hypotheses — it's knowing what not to trust. Scientists routinely build years of work on literature that later turns out compromised, and most AI research tools make this worse by confidently citing sources they never checked. Keystone inverts that default: every recommendation ships with its own falsification condition, every claim traces to a source you can open, and the AI is built to downgrade its own confidence — live, in front of you — the moment the evidence doesn't hold up. - **Team:** [Jai Sogani](https://cerebralvalley.ai/u/jaisogani) - **GitHub:** https://github.com/jaisogani-ai/Keystone-Life-Science-System.git - **Demo video:** https://drive.google.com/file/d/1vPhQUBwmG26PW9ZZdCJeV_nhf_eeGUsj/view?usp=drivesdk - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=290 ### 17. Ultimate Robotics Manual labeling is the bottleneck of biosignal ML — every hour of EMG needs a human marking what the muscles were doing. We built a tool that removes the human: a first-person camera watches the hands, a pose model turns the video into timestamped gesture labels, and any instrument recording on the same clock — EMG in our lab, ECG or IMU in yours — inherits those labels for free. The camera is a teacher, not the product. During the event the loop closed: a pretrained hand model labeled 320 frames of our own footage (zero human keypoints), a small pose model fine-tuned on them and outgrew its teacher — 71% vs 46% — and now labels its own next dataset. On a consumer AMD GPU each fine-tune takes minutes, so labeling improves between recordings. Standalone, sensor-agnostic, open source: a session folder in, two CSVs out. - **Team:** [Kseniia Kyrylyshena](https://cerebralvalley.ai/u/paladin) - **GitHub:** https://github.com/ultimaterobotics/camera-autolabeler - **Demo video:** https://www.youtube.com/watch?v=lHY-8xWjjfw - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=291 ### 18. Perturb2Target Genome-scale screens and GWAS both tell you which genes matter in disease — but not what to do about them. A knockdown that reverses pathology is a drug lead; one that worsens it is a warning. Perturb2Target closes that gap by integrating evidence across biological scale, from population genetics to protein structure, to repurpose existing drugs for underserved autoimmune diseases like multiple sclerosis. We built an agentic AI pipeline that reasons over four layers. First, population data: we anchored candidates in GWAS through Open Targets credible-set fine-mapping and CD4⁺ T-cell cis-eQTLs, yielding 239 genetically-supported targets. Second, integrated single-cell atlases: we harmonized 1.9 million CD4⁺ T cells spanning ~20 inflammatory diseases into one batch-corrected manifold, and derived a contamination-filtered MS disease-direction from cross-atlas replication. Third, a CD4⁺ CRISPRi Perturb-seq screen scored, for every knockdown, whether its transcriptional effect reverses the MS direction — supplying the therapeutic sign that genetics alone cannot. Fourth, AlphaFold2 drug–protein interaction modelling assessed structural druggability of the leads. The pipeline is orchestrated by agentic AI, which chains these heterogeneous datasets into a single ranked, directional, genetics-anchored, druggable shortlist. Our headline finding: IL2RA emerges as a promising MS target — genetically supported (L2G 0.74), reversing the disease signature on knockdown, a validated immunotherapy node (the IL-2 receptor axis) and a cognate ligand. By unifying genetics, cell-atlas context, functional perturbation, and structure, Perturb2Target nominates not just which genes, but which way to push them — and whether they can be drugged. - **Team:** [Yizhou Yu](https://cerebralvalley.ai/u/yizhouyu), [Oishi Deb](https://cerebralvalley.ai/u/odeb) - **GitHub:** https://perturb2target.streamlit.app/ AND https://github.com/odeb1/Perturb2Target - **Demo video:** https://youtu.be/5rbH33VO1NU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=292 ### 19. cocaine I create a dry-wet-agent loop operator system which is the loop Engineering . I think it is - **Team:** [Wenyi Yang](https://cerebralvalley.ai/u/Cocaine) - **GitHub:** https://github.com/cocaineeeeee/dry-wet-agent-loop-operator-system - **Demo video:** https://claude.ai/code/artifact/a2598364-956a-433b-a49a-cf4e41b65198 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=293 ### 20. Akul Sharma What I built and investigated: A complete biomarker discovery pipeline for post-traumatic epilepsy (PTE) in a longitudinal rodent TBI model, covering: rigorous QC (scanner/site effects, sham reference stability, tractography quality), exploratory analysis establishing the general injury phenotype, region- and whole-tract-level statistical and machine learning analysis, and the core methodological contribution: a localized, multimodal normative modeling framework. The pipeline addresses a genuinely hard problem: PTE and TBI-only animals both undergo substantial injury and recovery, so any PTE-specific signal is small and easily masked. What we found: QC revealed real site/scanner effects requiring harmonization, and confirmed the sham group as a stable normative reference despite noisy individual longitudinal metrics. Whole-tract averaging diluted focal injury signal (AUC 0.90 localized vs. 0.63 averaged), and classification collapsed to chance. Our normative framework, representing each tract as a 20-segment × 9-metric profile and modeling deviation from the sham distribution, recovered injury detection far more sensitively, with framework structure. Applied to thalamic tracts, it revealed tract-specific metric signatures (diffusivity-led vs. fiber-fraction-led) and generated concrete, testable hypotheses about injury propagation. Why it matters: This is a full pipeline from data quality through interpretable, localized biomarker discovery and can generalize to other tracts, metrics, and injury or disease models beyond this specific dataset. Beyond the present study, the framework may be more broadly extended to other preclinical rodent imaging studies for discovering biomarkers or evaluating therapeutic interventions. - **Team:** [Akul Sharma](https://cerebralvalley.ai/u/chronicpain) - **GitHub:** https://github.com/akul0119/tbi-pte-pipeline - **Demo video:** https://youtu.be/HQWBUX3_ak0 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=294 ### 21. J-solo-claude-science What I built and investigated. GWAS have catalogued tens of thousands of trait-associated loci, but naming the causal effector gene — usually acting through a non-coding regulatory variant far from the gene — remains unsolved at scale. The field's default evidence, cis-QTL colocalization, has a ceiling that is larger than commonly stated. I set out to (i) measure that gap rigorously and (ii) build a sequence-based line of evidence that adds what colocalization cannot: a directional enhancer→gene link available where colocalization is silent. Over the past week this ran end to end: I assembled a harmonized five-source, 10M-row multi-biobank credible-set substrate (All-of-Us, UKB-WGS, FinnGen R13, BBJ molecular-QTL, UKB metabolomics); harvested published colocalization results from UKB-PPP, FinnGen R13 and BBJ; derived 6.7M FDR-controlled signed enhancer→gene pairs from the DBNascent nascent-RNA atlas (validated to Pearson = 1.0 against published pairs, plus a clean-room mouse reimplementation); embedded ~1.2M regulatory windows through a frozen Nucleotide Transformer v2; and trained a family of signed-E2G models culminating in CATE, a convolution-attention two-tower encoder. What I found. Three results. (1) The gap is real and large: on my own substrate only 1–2% of confident QTL signals share a causal lead with a GWAS signal despite ~90% being feasibly close, and on a 1,160-locus benchmark cis-eQTL colocalization names the curated effector at only 24.5% of loci (44.7% silent, 30.9% wrong — 75.6% unresolved), corroborated by three flagship resources. (2) A DBNascent-derived sign predicts causal direction (Spearman ρ = −0.224, p = 5×10⁻⁹; blood×K562 odds ratio 5.4) — weak but robust, and unique among E2G methods. (3) The model is complementary to distance: it recovers 39 effector genes distance misses (18 at coloc-silent loci, including GLP1R, MYC, SNCA, VDR), and a methodological finding — link AUROC plateaus at ~0.61 across gradient-boosted trees, a pooled MLP, and cross-attention — localizes the ceiling to pooling away sequence positions, motivating the per-token CATE model. Why it matters. Effector-gene assignment is the bottleneck between a GWAS hit and a drug target: get the gene wrong and every downstream decision is wrong too. Quantifying exactly where the standard tool fails, and adding a sequence-only, directional signal available at the silent majority of loci, gives a more honest and more complete locus-to-gene map. - **Team:** [Jerome Irudayanathan](https://cerebralvalley.ai/u/madd-ebb) - **GitHub:** https://github.com/jerome-f/Claude_Life_Sciences_Hackathon_Project.git; https://jerome-f.github.io/Claude_Life_Sciences_Hackathon_Project/ - **Demo video:** https://youtu.be/fY6m1YjTCGA - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=295 ### 22. Quincy Labs Lab Memory turns a scattered pile of a lab's files (papers, notebooks, protocols, ClinicalTrials.gov exports, CSVs, notes) into a queryable, cited memory. Drop files in, and Claude answers questions grounded only in those files — with a citation to the exact source for every claim, a reasoning chain you can verify, an auto-built knowledge graph, and the sharp follow-up questions worth asking next. lab-memory.vercel.app - **Team:** [Henry Ogedegbe Jr](https://cerebralvalley.ai/u/danchou) - **GitHub:** https://github.com/zetsuchan/lab-memory - **Demo video:** https://youtu.be/rhzN4V_eu8s - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=296 ### 23. Accelerate gpu-sceptre is an independent, from-scratch GPU implementation of SCEPTRE (Barry et al., 2021), the calibrated conditional-resampling test for single-cell CRISPR (Perturb-seq) screens. We re-expressed both expensive stages — the per-gene negative-binomial GLM precompute and the per-pair resampling kernel — as batched float64 tensor operations, so the entire screen runs on a single GPU. What we found: on SCEPTRE's own example dataset (526 genes × 45,919 cells), a single free Colab T4 ran the analysis ~29× faster than SCEPTRE on one CPU core and ~23× faster than on two — SCEPTRE parallelizes at only ~63% efficiency, so matching one GPU takes ~30+ cores. Results match an independent double-precision reference to machine precision (~1e-15 kernel, ~1e-10 precompute); null p-values stay calibrated and all planted knockdowns are recovered. Why it matters: SCEPTRE's statistical rigor normally requires a CPU cluster. gpu-sceptre puts the gold-standard method on hardware any lab or student already has for free — and every number is reproducible from committed scripts. Saves a lot of time hence accelerate the research. - **Team:** [Shrishti Shah](https://cerebralvalley.ai/u/Shrishti18) - **GitHub:** https://github.com/Shrishti18/gpu-sceptre - **Demo video:** https://youtu.be/8uAsuCAQXfc - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=298 ### 24. AntibodyForge abdev-suite — Project Description WHAT I BUILT AND INVESTIGATED I built abdev-suite, a pluggable in-silico platform for antibody developability assessment. Given a VH/VL sequence — or a stored portfolio of candidates — it models the Fv structure, runs a set of analyzer modules that each score one developability trait, and serves the results through a REST API and a single-page web interface: a sortable and filterable candidates table, a per-candidate detail view that places every score on its reference distribution, a VH/VL sequence map of CDRs and liability sites, and synchronised dual 3D structure views. The pipeline models the Fv with ABodyBuilder2, protonates and repairs it with PDBFixer, and then runs modules of three kinds under one uniform contract: - Structural computation — a histidine scan (phswitch) that predicts each position's pKa with PROPKA3, flags pH-switch-window hits, and classifies each as CDR or framework. - Rule-based scanning — a chemical-liability panel (liabilities) that flags PTM and degradation motifs (N-glycosylation, deamidation, isomerization, oxidation, Asp-Pro clip, free cysteine), split framework-vs-CDR and weighted by solvent exposure. - Data-driven scoring — sequence-similarity modules (solubility, affinity, gh_pssm) calibrated on a curated functional-antibody library. Heavy computation runs on a durable job queue with background workers: ESM-2 latent scoring (plm), explicit-solvent molecular-dynamics stability (gromacs), and constant-pH molecular dynamics (cphmd), with live GROMACS/Amber runners wired to GPU hosts. Results persist per candidate, so a portfolio accumulates structural, MD, and sequence-level evidence over time. The scientific thread underneath the platform is a structural pH-switch design study: engineering histidine-based, pH-dependent switches into an antibody framework by scanning positions for the pKa-versus-burial "sweet spot," triaged quickly from structure and validated against constant-pH MD. WHAT I FOUND The most significant finding concerns the tool itself, which I characterised honestly rather than overstating: - The mechanistic modules are trustworthy. Rule-based liability scanning (grounded in documented degradation chemistry) and the structural histidine-pKa switch scan rest on chemistry and physics rather than fit-to-library, and are the signals worth acting on. - The similarity modules are heuristics, not predictors. affinity and solubility interpolate a reference library by sequence similarity; similarity to a library of other binders does not causally predict a new antibody's absolute affinity or yield against its own antigen, so a single "predicted" number is false precision. The honest form is nearest-neighbour retrieval with distance gating — "your closest known antibodies and their measured values" — not a headline value. - gh_pssm is honest only within its scaffold. Its H3 model is length-anchored, so off-scaffold sequences must read N/A rather than a low score. - The PLM latent score is a real signal presented poorly — an ESM-2 latent novelty/naturalness measure whose raw Mahalanobis value is uninterpretable and unactionable as shown. - A methodological point: framing "kNN versus PLM latent space" as a choice is a category error — the representation (a protein-language-model embedding) and the algorithm (k-nearest-neighbours) are orthogonal. The strong form is retrieval on PLM embeddings with distance gating; dimensionality-reduction plots (UMAP, t-SNE) are non-metric and exploratory only; and nothing here is a genuine predictor until held-out validation shows it generalises beyond its training distribution. Along the way I also resolved real correctness issues. The modeled structures are raw-numbered rather than IMGT-numbered, so an IMGT-range CDR classifier mis-tagged boundary residues; I unified all CDR handling — sequence map, GH-PSSM, 3D viewer, phswitch, and liabilities — onto a single sequence-based definition. I also removed UI elements that implied more than the method supported, such as a solvent-dominated MD "potential" number and a hotspot ranking that failed to recover the constant-pH-MD-validated switches. WHY IT MATTERS In-silico developability screening is crowded with tools that present weakly validated, correlational scores with unwarranted confidence. A tool that is explicit about which signals are mechanistic versus correlational — and that refuses to extrapolate beyond its reference set — is more trustworthy for real triage decisions than one that hides that distinction behind a polished number. The mechanistic core connects to a genuine engineering goal: pH-dependent switches enable antigen release for recycling antibodies and conditionally active binders, and pairing a fast structural triage with constant-pH MD validation is a defensible workflow rather than a single opaque score. Finally, the pluggable architecture — a uniform module contract, reference- anchored display, and a durable job queue — makes the platform a reusable substrate: better-validated modules can be dropped in, and the honest framing (scope warnings, retrieval rather than prediction, N/A when out of scope) lives in the presentation layer itself. As much as it is a screening tool, it is a case study in building computational-biology software honestly — calibrating the interface's confidence to the method's actual epistemic standing. - **Team:** [Hung-Pin Peng](https://cerebralvalley.ai/u/hppeng) - **GitHub:** https://github.com/px172/abdev-suite - **Demo video:** https://www.youtube.com/watch?v=ZViqkndR9Tw - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=299 ### 25. SP 1. Project name ||The Tempo of Transformation — Post-Translational Timing Gates How Driver Mutations Spread Through the Colorectal Cancer Network|| Short variant : ||Tempo of Transformation: PTM Timing Gates CRC Damage|| 2. Tagline When a PTM is modelled as timing rather than a switch, it changes how far a driver mutation's damage travels through the cancer network. 3. What it does The project tests a modelling choice that cancer systems biology usually makes without noticing: a driver mutation and a post-translational modification (PTM) are both encoded the same way — a node forced ON or OFF. In a published, independently re-verified 96-node Boolean model of cancer signalling (Fumiã & Martins 2013), we encode the *same* β-catenin destruction-complex lesion two ways — as a node clamp (the standard approach) and as a PTM-edge deletion on a separated fast timescale (the biologically faithful approach) — and measure how far a seeded perturbation spreads as driver mutations accumulate in the canonical colorectal order. Under synchronous updating the two encodings are indistinguishable — a documented negative control (Δ(η-slope) = −0.0020 per driver). Once fast (phospho/degron) and slow (transcriptional) processes run on separate clocks, the PTM-edge encoding propagates the perturbation substantially further: Δ(AUC) = +0.309 (95% CI [0.3058, 0.3115], p < 0.0001) at the APC+KRAS stage. A pre-registered timescale-knockout control (τ=1) removes 98.2% of the effect — proof that the phenomenon *is* the timescale separation, not the lesion itself. We report the finding with its limits intact. The effect exceeds a degree- and sign-preserving topology null (p = 0.042) but sits at the 94th percentile of a random-edge null (p = 0.058), so it is a property of a *class* of coupling-preserving edges rather than the β-catenin degron uniquely; the network stays firmly ordered (η < 1) throughout; and external tests in static CPTAC phosphoproteomics and DepMap dependencies are honest nulls or weak, underpowered probes, not validation. The contribution is a ||novel multi-timescale encoding method plus an honest attempt to test it|| — not a claim to have found a colorectal-cancer drug target. 4. Inspiration The study began as a "butterfly effect" question: does a small change early in tumorigenesis ripple out to reshape everything downstream? We retired that framing deliberately, because in this setting it is a category error. A butterfly (Lyapunov) exponent describes sensitive dependence on *initial conditions* in a chaotic system. A mutation is not a change of initial condition — it is a change of the *rule*, a structural (bifurcation) perturbation of the network itself. And eukaryotic regulatory networks are not chaotic: they sit in an ordered or near-critical regime (Kauffman 1969; Shmulevich, Kauffman & Aldana 2005). The honest discrete-dynamics analogue of "a small change spreading" is therefore ||Boolean damage-spreading|| and the ||Derrida annealed map|| (Derrida & Pomeau 1986), applied to a network whose *rules* are edited by the mutation and the PTM. Turning a vague metaphor into a rigorous, measurable estimand is what made the study both tractable and novel — and it sharpened the real question: not *whether* damage spreads, but whether the post-translational layer changes the *pace* at which it spreads. 5. How we built it A six-stage pipeline, entirely in the Claude Science workbench: 1. ||Verified model sourcing (not reconstructed).|| The Fumiã & Martins (2013) update rules live only in the paper's supplementary Text S1, so we sourced them two independent ways — the EuropePMC supplementary ZIP and an independent GitHub sympy re-encoding — confirmed the network closes exactly (96 nodes, 249 edges, zero dangling references), and validated the simulator against the paper's three published normal-cell attractors (proliferative period-7 cycle, quiescent fixed point, apoptotic fixed point). 2. ||LOF/GOF driver encoding.|| Drivers edit the rules, not generic toggles: APC (loss-of-function, node clamped 0), KRAS (gain-of-function, Ras clamped 1 and decoupled from EGFR/RTK to respect anti-EGFR resistance biology), TP53 (LOF), SMAD4 (LOF) — accumulated in Fearon–Vogelstein order. 3. ||PTM-edge layer (the crux).|| The β-catenin degron lesion is encoded as deletion of the single inhibitory edge GSK-3/APC ⊣ β-catenin, which keeps β-catenin coupled to p53 (σ_βcat = sgn(−p53 + 1)), and contrasted against the standard node clamp that severs that coupling. 4. ||Multi-timescale simulator.|| A custom pure-numpy engine partitions the network into 33 fast (phospho/signalling/degron) and 57 slow (transcriptional/cyclin) nodes plus 6 inputs; slow nodes update on a separated clock (τ_slow = 50); the τ=1 knockout collapses the separation as the pre-registered mechanism test. 5. ||Modal sweep.|| The full damage-spreading sweep — primary estimand + three null models + synchronous negative control — ran on Modal (BYOC, 32-core CPU container, pure-numpy worker) in ~22 min: M = 500 null/bootstrap realizations, 15,000 samples per timescale cell, 2,000 bootstrap draws. 6. ||CPTAC / DepMap validation.|| Cross-validated against CPTAC colon phosphoproteomics (cBioPortal `coad_cptac_2019`, phospho matrix 97 × 106) and probed against DepMap 23Q2 CRC dependencies; CrossRef verified every citation; all statistics (empirical two-sided p-values, bootstrap CIs, Benjamini–Hochberg FDR) were pre-registered before the sweep. 6. Challenges we ran into - ||The synchronous-null → multi-timescale-signal reframe.|| Our first comparison of the two PTM encodings under synchronous updating found nothing (Δ(η-slope) = −0.0020 per driver, negligible). Instead of abandoning the hypothesis, we recognised that a synchronous clock is exactly what erases a timescale effect — and made the timescale separation the treatment. The "null" became the documented negative control, and the effect appeared (Δ(AUC) = +0.309) once fast and slow clocks were separated. - ||The count-vs-combination scoring bug we caught and fixed.|| The DepMap external score initially keyed criticality on raw *driver count*, which mis-scored APC+TP53 lines as near-critical and inflated an apparent signal to p = 0.038 with 4/7 effectors surviving FDR. But the sweep showed the effect requires *both* the APC-driven degron lesion *and* an intact p53 seed (it collapses once TP53 loss clamps p53). Scoring lines mechanism-faithfully (APC-mutant AND TP53-wild-type) dropped the result to p = 0.20 — the honest number we report, superseding our own inflated version. - ||Data-access chains.|| The base model existed only as paper supplementary text; CPTAC came through the cBioPortal REST API; DepMap through figshare. We also caught and corrected a data-hygiene error — an early "CTNNB1 S45F in 4 patients" assumption traced to a *different* dataset, not this CPTAC cohort (where the Wnt degron is disabled overwhelmingly through APC loss, 74.5%). - ||Ordered, not critical.|| The network stays firmly in the ordered regime (η < 1 at every driver load), not at the edge of chaos. This was a pre-registered contingency (Park et al. 2023, PRX Life), so rather than force "criticality" language we reframed the result as a shift of *position on the order↔criticality axis* — precisely at the intermediate +APC/+KRAS loads where the degron lesion is active. 7. Accomplishments we're proud of - ||Cleared a real novelty gate.|| We gated the idea against actual prior art — Chu, Lee & Cho (2015), criticality-versus-driver-load in CRC, which we cite as the platform, and the Cho-lab attractor-reversion program, whose vocabulary and target-nomination framing we deliberately avoided — and pre-registered a narrow surviving wedge before computing the headline result. - ||A clean mechanism proof.|| The τ=1 timescale knockout removes 98.2% of the effect with the *same* lesion and the *same* seed — direct evidence that the phenomenon is the timing, not the mutation. - ||Honest null reporting.|| Every null model (topology p = 0.042; PTM-location p = 0.058; driver-order rank 8/24), the ordered-regime reframe, the CPTAC identifiability nulls, and the weak DepMap probe are reported whether or not they favour the hypothesis — including superseding our own tempting-but-wrong count-based DepMap result. - ||Full reproducibility.|| Verified model provenance (dual-sourced rules + attractor validation), a locked pre-registration, a pure-numpy simulator, a reproducible Modal sweep, and every figure and dataset saved as a project artifact. 8. What we learned - ||Timing can be the mechanism, not a detail.|| Modelling a PTM as a fast-timescale edge rewiring rather than a node switch changes a mutation's reach by a large, highly significant margin (Δ(AUC) = +0.309, p < 0.0001) that the standard node encoding literally cannot see. For a perturbation spreading through a network, *when* a coupling acts can matter as much as *whether* it acts. - ||Pre-registration and null models earn their keep.|| Locking the estimand, the mechanism control, and three null models before the sweep is what let us separate a genuine multi-timescale phenomenon from degree/wiring artefacts — and what kept us honest when the DepMap scoring bug produced a seductive significant result. - ||Static data has hard identifiability limits.|| Δ(AUC) is a statement about *dynamics over time*. A single-timepoint bulk snapshot (CPTAC) can replicate a *direct* mutation→PTM link (TP53-R273 → Ser315, p = 0.018) but cannot resolve a *dynamic* multi-timescale coupling (Tests 2 and 3 null, p = 0.83 and p = 0.70). The data type bounds the claim — a limit we pre-registered and confirmed rather than papered over. 9. What's next - ||Time-resolved perturbation phosphoproteomics.|| The one experiment that could test Δ(AUC) directly — a dynamic estimand — is time-course phospho data following a defined perturbation, which does not exist for this cohort. That is the decisive next measurement. - ||Multi-model generalisation.|| The null-model class argument implies the effect is a property of coupling-preserving edges in *any* multi-timescale network, not the Fumiã–Martins model or the β-catenin degron uniquely. Sweeping additional base models and PTM families would test that argument head-on. - ||The tractable-kinase PTM-edge-writer angle.|| Because the effect is carried by a *class* of coupling-preserving edges, the actionable question becomes which druggable kinases and enzymes *write* those edges — reframing "which node to hit" as "which timing to retune." This is a hypothesis-generation direction, not a target claim. - **Team:** [PARAMESWAR SAHU](https://cerebralvalley.ai/u/s-parameswar) - **GitHub:** https://github.com/Sandbox-commission/Hackathon-claude-built-with-science-The-Tempo-of-Transformation - **Demo video:** https://github.com/Sandbox-commission/Hackathon-claude-built-with-science-The-Tempo-of-Transformation/blob/main/video/tempo_of_transformation_demo.mp4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=300 ### 26. Noozra Two open-source tools, built with Claude, that catch expensive biology failures early — on public data, as decision-support (never a diagnosis). The off-target copilot screens a therapeutic protein against its human look-alikes before the wet lab: anti-myostatin is 89.9% identical to GDF-11, so an antibody would hit both. It finds a selective epitope (identity to GDF-11 drops 89.9%→60%), checks whether animal models will engage it, and flags developability and immunogenicity liabilities — and it retrospectively reproduces 10/10 documented antibody cross-reactions. The clinical suite re-reads old genetic reports: it re-checks variants against current ClinVar, triages missense calls, flags drug–gene mismatches (CPIC), prioritizes genes from patient phenotypes (HPO), and reads gene constraint (gnomAD). All twelve tools are driven by two Claude agents over MCP. Claude Code built the repo autonomously; five Claude agents researched the field in parallel; Claude Science ran the analyses and drew the figures. It never says "safe." 76 tests · one repo · MIT. - **Team:** [Gus Vernon](https://cerebralvalley.ai/u/Noozra) - **GitHub:** https://github.com/nochallenge/life-sciences-safety - **Demo video:** https://youtu.be/-Qm6wUcS5q8 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=2 ### 27. Cardio21 PYRIN-PLAQUE Digital Twin investigates whether MEFV/pyrin inflammasome biology can be separated from the saturated NLRP3 atherosclerosis story and mapped to human plaque cell states using only public data. We built an auditable Claude Science workflow that combines a 57-claim literature evidence graph, verified public datasets, single-cell plaque module scoring, bulk plaque triangulation, a dimensionless pyroptosis ODE, caveat-aware priority scoring, and reviewer-agent audits. The project deliberately separates a pyrin-specific regulatory backbone from NLRP3 and excludes shared downstream genes such as PYCARD, CASP1, GSDMD, IL1B, and IL18 from discriminative scoring. In GSE159677 single-cell plaque data, MEFV was sparse overall but enriched in macrophage/myeloid cells, and PYRIN_BACKBONE scored highest in the myeloid compartment while NLRP3 localized differently. In GSE120521 bulk plaque data, PYRIN_BACKBONE rose in unstable plaque, but this signal largely collapsed after myeloid-marker adjustment. The ODE model similarly suggested that generic inflammatory priming may dominate pyrin-specific threshold effects. The result is not a validated drug target or clinical claim. It is a transparent, reproducible, caveat-aware hypothesis engine that prioritizes the MEFV/pyrin threshold axis for future experimental validation while preserving confounded and weak findings instead of hiding them. - **Team:** [Ibrahim Halil Tanboga](https://cerebralvalley.ai/u/Ibrahim_halil) - **GitHub:** https://github.com/ihtanboga/pyrin-plaque-digital-twin - **Demo video:** https://www.youtube.com/watch?v=wCPon-HImi8 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=3 ### 28. INTEGRITY INTEGRITY is a forensic scanner for life-science manuscripts. It validates every biological entity a paper names - gene symbols, cell lines, antibody RRIDs - against authoritative registries (HGNC, Cellosaurus, SciCrunch) and reports only what those registries actually say. It never reasons about biology and never accuses an author: every finding is labeled REVIEW, never ERROR. The core insight: when Excel silently converts a gene symbol like SEPT2 into a date, it does not store the text "2-Sep" - it stores a typed date cell (data_type='d'). That evidence is destroyed the moment the sheet is exported to CSV and was never in the manuscript PDF, so no model reading the paper can recover it. INTEGRITY reads the raw spreadsheet bytes instead. Run on the real 2016 Ziemann supplementary file - the paper that first documented this corruption - it finds exactly 1213 corrupted cells across 128 distinct registry lookups, and refuses to flag registry-vindicated symbols (Oct-4 resolves to POU5F1 and SEPT2 to SEPTIN2, both PASS). Only genuine corruption is flagged. When a registry is unreachable it fails safe with a distinct exit code rather than reporting a false clean. It ships as three surfaces: a CLI, a live web app with streaming results (https://integrity-lemon.vercel.app), and a Claude Code Stop hook that blocks an agent from ending its turn if a registry disagrees with a biological claim it just made. Why it matters: 30.9% of papers carrying supplementary Excel gene lists were corrupted by 2021 (up from 19.6% in 2016), the errors are silent and cheap to catch, and today nobody catches them. - **Team:** [Javier De Jesus](https://cerebralvalley.ai/u/javierdejesusda) - **GitHub:** https://github.com/javierdejesusda/integrity - **Demo video:** https://youtu.be/7y2HNXwHlIk - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=4 ### 29. Oocytezero OocyteZero asks a concrete question behind a disease that affects up to 1 in 8 women: can today's DNA-reading AI models name the single causal base at an unsolved PMOS (polycystic ovary syndrome) genetic locus? It's personal — someone I love lives with PMOS — and it's genuinely unsolved: most of the disease's heritable signal sits in non-coding DNA no one has decoded. Working solo with Claude Code and Claude Science, I built an in-silico saturation-mutagenesis benchmark on 589 measured variants at the DENND1A locus and ran an escalating ladder of 12 model classes — from frozen Enformer/Borzoi to AlphaGenome to the reporter-allele model Malinois. Every one reads regulatory activity, but none resolves the causal allele. Two positive controls prove the pipeline works and the negative is real; a mechanistic diagnosis — tissue gap + data gap + an eQTL-silent locus — explains why. Claude Science independently recomputed every number and caught my own overclaims. It's a rigorous, reproducible NO — and a finding others can build on: a reusable, model- and locus-agnostic benchmark, a bound that stops the field from over-calling causal variants, and the exact wet-lab experiment (a matched-tissue steroidogenic MPRA) that could flip it. Come travel the DNA strand with me → oocytezero.vercel.app - **Team:** [Jose Contreras](https://cerebralvalley.ai/u/Pepe) - **GitHub:** https://github.com/josemacontrerasp-arch/oocytezero - **Demo video:** https://youtu.be/TtpO0E8MmXI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=5 ### 30. Historify Most people's medical history is a pile of PDF lab reports — one per clinic, each in its own layout, none of them talking to each other. A single value tells you almost nothing; the trend across all of them tells you everything, but nobody ever gets to see it. Historify turns those scattered PDFs into one structured, longitudinal health record. Upload a report and Claude reads it page by page, extracting each marker with a confidence score, while a deterministic backend normalizes units and reconciles results across labs — reconstructing a trajectory that no single exam, and no single clinic, could ever show. Low-confidence readings are flagged and linked to the exact source line: the AI is allowed to be uncertain, and the person stays in control. Any marker can be traced over years, or exported as a one-click report to share with a doctor. Built from scratch during the hackathon, Historify's real finding is a design principle: the hardest part of a health record isn't intelligence — it's trust, and trust comes from traceability. Claude reads, the backend calculates. It never diagnoses. - **Team:** [David Castro](https://cerebralvalley.ai/u/davidcasr) - **GitHub:** https://github.com/davidcasr/historify-backend - **Demo video:** https://youtu.be/B31jwHD5m3g - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=6 ### 31. The chief dawgs We investigated whether a genome-scale CD4+ T-cell Perturb-seq screen contains reproducible regulatory architecture beyond ranked hits, statistical noise and connectivity hubs We built a quality-audited regulator × gene × condition operator from a 1.8 TB public CRISPRi atlas using a laptop-scale pipeline. Spectral denoising reduced the operator to 92 empirically supported signal directions, and consensus community detection revealed structure 259 standard deviations above a matched null The strongest result was annotation-blind recovery of mitochondrial Complex I, identified only after clustering, at BH-FDR 1.4 × 10⁻⁷. A second SAGA-centered module was supported by convergent evidence across CORUM, tensor factors and an independent K562 screen, while cross-cell-type analysis across CD4, K562 and RPE1 separated universal from T-cell-specific regulation We also tested the predictive boundary of this structure. Real and shuffled regulator features performed identically for leave-regulator-out prediction, showing that recoverable structure does not imply inductive predictability The result is an open, reproducible analysis and manuscript that others can build on - **Team:** [Santiago Silva](https://cerebralvalley.ai/u/cuentadesanti) - **GitHub:** https://github.com/cuentadesanti/cd4-perturbseq-regulator-atlas - **Demo video:** https://youtu.be/3tu_3JmHh14 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=7 ### 32. PaperTrail ▎ PaperTrail is a provenance-grade verification platform for clinical-evidence claims. Given a claim like "lecanemab caused brain swelling in 21.3% of patients," it retrieves the primary source (PubMed / ClinicalTrials.gov), extracts the actual finding, recomputes the effect size deterministically, and flags exactly where the claim diverges — with a trust score and a citation trail where every flagged span maps to a verbatim substring of the cached source. ▎ ▎ The core design decision: deterministic math decides every verdict; Claude only writes grounded language and routing hints — it is never in the numeric or scoring path. The grounding guarantee is enforced as a code invariant (an audit ran 500k+ adversarial fuzz iterations and could not produce an unsourced span). ▎ ▎ Three named-user tools sit on top: Verify a claim, a Lab Notebook that turns dictated bench notes into a grounded, reproducible experiment record (nothing it can't quote is kept), and a Trial Matcher that assesses a de-identified patient against ClinicalTrials.gov criterion-by-criterion. A Mixture-of-Agents pipeline composes 20+ engines (MiniCheck, cross-source aggregation, PyMARE pooling, a discrepancy auditor, STORM debate) into one deterministic verdict. ▎ ▎ What I found (honest): on a single-source clinical benchmark the deterministic + audit path beats a plain Claude call 95% vs 90%; on contested multi-source claims it ties at 100% but stays 87% accurate when the API is down (Claude-alone: 0%). I then audited the entire statistical core (~18 modules — meta-analysis, GRADE, survival, network-meta, fragility, pharmacovigilance) against the biostatistics literature and fixed real correctness bugs: a confidence-interval false-positive on the verify path, a 2× trial-sequential error, a fragility-index bug that inverted "fragile" and "robust," and negation-parity bugs that flipped support↔refute. ▎ ▎ Why it matters: a translational researcher or medical-affairs reviewer must produce and defend numbers. A hallucinating LLM can't be trusted for that. PaperTrail's verdicts are deterministic and every claim traces back to its primary source — provenance, not opinion. - **Team:** [arhan subasi](https://cerebralvalley.ai/u/arhansuba) - **GitHub:** https://github.com/caelum0x/papertrail - **Demo video:** https://youtu.be/K1JPQiUIjWE - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=8 ### 33. Discoverybytes Co-localization of signaling molecules to a stable cell-cell interface is a key prerequisite for signaling in systems such T cell activation. We had previously built a model that indicated bifunctional antagonists, such as bispecific antibodies, might have unexpected activating function in such systems. We revisited the concept to do the following. 1. Confirm the math behind the model and build a user-friendly interface to visualize activation / inhibition 2. Generate a table of receptor proteins that span the cell membrane once, and are involved in signaling. These are candidate targets for modulation via this co-localization mechanism of action. 3. Expand the candidate target data to include data for interaction partners, disease associations (including rare/orphan), therapeutic objective (i.e. activate/inhibit), highest drug clinical stage for any targeting the receptor. We ranked pairs by a composite Opportunity Score, then narrowed to three rare/orphan diseases with clean ligand-replacement logic: CD70-CD27 (severe combined immunodeficiency), DLL4-NOTCH1 (Adams-Oliver syndrome), and EDA-EDAR (hypohidrotic ectodermal dysplasia). DLL4-NOTCH1 lacks a disease-modifying option, making it our lead despite highest mechanistic risk: Notch activation requires active endocytic force, not proximity alone. Molecular design concept would tether a nanobody against EFNB2, an endothelial protein coupled to Notch signaling, to one against NOTCH1. - **Team:** [John Proudfoot](https://cerebralvalley.ai/u/discoverybytes) - **GitHub:** https://github.com/discoverybytes/built-with-claude-life-sciences - **Demo video:** https://youtu.be/BSuJnv4Ps38 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=9 ### 34. Mickey Pentecost Traumatic brain injury and inherited genetic risk both raise the odds of Alzheimer's, but no one had shown they act through the same mechanism. Working only from public data, I pre-specified two opposing gene programs in microglia ( an inflammatory "accelerator" and a pro-resolving "brake") and tested whether injury and inherited risk converge on the same one. They do. Across nine analyses spanning human AD, mouse TBI, chronic traumatic encephalopathy, single-nucleus and bulk RNA, spatial transcriptomics, and brain proteomics, the accelerator program is elevated in both conditions and concentrates spatially at the injury lesion and around amyloid plaques. Partitioned heritability shows 31% of inherited AD risk falls on microglial regulatory DNA (1.5% of the genome, 21-fold enrichment) and loads specifically onto the accelerator arm while avoiding the brake, a directional genetic test made possible by pre-specifying both modules. Critically, inherited risk enters upstream (APOE, TREM2), while the downstream effector genes are installed by injury or disease itself: TBI installs environmentally what AD risk alleles predispose to genetically. Both arms are read out at a single receptor, CD44, whose elevation reproduces across four independent modalities and both species. Why it matters: the work turns an unexplained epidemiological link into a defined molecular circuit with ranked, testable therapeutic predictions, and it does so entirely from existing public datasets reanalyzed through one harmonized pipeline, making it a reusable resource rather than a one-off result. - **Team:** [Mickey Pentecost](https://cerebralvalley.ai/u/MickeyP) - **GitHub:** https://github.com/mickeypentecost/microglial-tbi-ad-axis/blob/main/SUBMISSION.md - **Demo video:** https://youtu.be/LplLZ6PIfbU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=10 ### 35. TrustCheck Problem. A biologist generates new data and wants to build on the published literature, increasingly with an AI scientist running the analysis. That raises two questions: are the findings they build on reproducible in their data, and is the AI scientist reliable? Solution. The biologist runs TrustCheck: they name a finding, point it at their cohorts, and it computes, live on the raw data, whether it holds. The verdict is that number, not the AI scientist's judgment, and the AI scientist works blind, never told which finding is supposed to reproduce, so it cannot just confirm what is expected. When a finding fails, two follow-up checks, power and method, explain why. Example. We tested two discoveries from our 2023 Nature Medicine paper on single-cell data from immunotherapy-treated liver (HCC) tumors, against two independent cohorts. One reproduced in all three. The other, testable in only one of the two, reproduced in HCC but not in melanoma, where 100% power and 0 of 20 analyses rule out both a small cohort and a bad pipeline. TrustCheck does not call the finding false: an effect as large as the original would have shown up here, but a smaller one cannot be excluded. - **Team:** [Assaf Magen](https://cerebralvalley.ai/u/Magen), [Chunyi Liu](https://cerebralvalley.ai/u/cliu238jhu) - **GitHub:** https://github.com/asmagen/trustcheck - **Demo video:** https://youtu.be/Jh5vnVy_AaU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=11 ### 36. Kai Specimen Casebook is auditable data reconciliation for biodiversity collections. Herbarium and museum data managers digitize specimens whose sources disagree — a label, a field ledger, and a legacy database may record three different collection dates. Conventional pipelines flatten these into a single value and discard the reasoning, silently turning uncertainty into fact. Specimen Casebook uses Claude's structured extraction to read every source field by field, returning a verbatim supporting quote and a source id for each claim. A deterministic policy layer re-checks each quote as an exact substring, drops any that were altered or invented, and refuses to auto-resolve a field when two grounded values disagree — marking it CONFLICTING and routing it to human review. The output is a Darwin Core CSV that carries unresolved disagreements in `informationWithheld` instead of hiding them, plus a full provenance JSON that lets a reviewer reconstruct every decision. Claude proposes; the policy layer decides what can ship. The primary metric — unsupported resolution rate — holds at 0% across adversarial tests and live runs against claude-opus-4-8. Built for a named herbarium collections data manager, it is working software, not a demo. - **Team:** [Keigo Takahashi](https://cerebralvalley.ai/u/ktak5) - **GitHub:** https://github.com/keigotak/specimen-casebook - **Demo video:** https://youtu.be/LVumKu5Eh-0?si=rhJSeq6jSAdX7Hyc - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=12 ### 37. enTrust What it is. A trust-first interpreter for non-coding, regulatory DNA variants — the frontier where variant-of-uncertain-significance (VUS) calls mostly fail. Coding-variant tools are mature; a variant in regulatory DNA usually gets no principled call at all*, because its effect isn't "changes a protein," it's "changes how much a gene is expressed." I fine-tune a DNA language model on the Deng et al. 2024 cortical MPRA (*Science*, adh0559), score a single-base change's effect via in-silico saturation mutagenesis, annotate the mechanism via transcription-factor motif gain/loss, and — the core contribution — ground every prediction in independent evidence (held-out MPRA, GTEx eQTLs, an independent organoid-context model, TSS proximity) to return a calibrated confidence with an auditable evidence chain**. When the model and evidence agree, confidence is high; when they conflict, the tool surfaces it instead of hiding it behind a number. What I found. 1. Optimizing the allelic-difference directly — a shared-weight "siamese" model trained on measured skew — beats the standard subtract-endpoints proxy by **+47%** on a bidirectional Caduceus backbone (variant-effect Δ-Pearson 0.19 → 0.28**, emVar AUC → 0.67**), the project's biggest gain. 2. A stacking meta-learner that fuses the primary model with an independent organoid model beats a single-feature calibrator (emVar AUC 0.623 vs 0.610**), the independent model being the lever. 3. I reproduce Deng's headline emVar set end-to-end (**163 ≈ 164**), and I report honest negatives — reverse-complement averaging, lower calibration τ, bigger HyenaDNA, and a frozen Enformer feature (tested to full 15k-variant coverage) — each logged with its reason, including one case where I caught our own thin-coverage artifact (a promising feature weight that collapsed to zero under full coverage). Why it matters. The deliverable isn't a bare score — it's a calibrated confidence and an auditable evidence chain for a class of variants curators currently can't call. A confident-but-wrong regulatory call is dangerous; a calibrated "uncertain — here's the conflicting evidence" is usable. The tool works on any variant and outlives the hackathon as a variant-curation aid. - **Team:** [Sujit Shelar](https://cerebralvalley.ai/u/Sujit_Shelar) - **GitHub:** https://github.com/shelarsujit/Regulatory_Variant_Interpreter - **Demo video:** https://www.youtube.com/watch?v=5oHtrrASPSM - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=13 ### 38. CellVerdict CellVerdict is a multi-method consensus annotator and trust layer for single-cell RNA-seq data. Cell-type annotation is the linchpin of every single-cell analysis — a wrong cluster label silently poisons downstream differential expression, cell-cell communication, and trajectory results while the UMAP still looks clean. Existing annotators each give an answer, often with unearned confidence, and they disagree with one another; a non-expert can't tell which to trust. Instead of adding one more annotator, we built a council. Every cluster is annotated independently by five voices — a marker-database matcher (multiple curated reference DBs), a CellTypist ML classifier, a prior-studies voter that web-searches primary literature for the specific tissue and votes with clickable, citation-ranked PMIDs, and a grounded Claude annotator — on top of a deterministic evidence engine that computes doublet, stress, ambient-contamination, and marker-specificity signals. A separate Claude adjudicator then reconciles all the votes against the computed evidence into one final label, a green/yellow/red trust score, and natural-language reasoning that names which DBs, which model, and which papers drove the decision. What we found: the fusion catches failures a single tool confidently mislabels — planted doublets and dying-cell states get flagged red instead of forced into a clean label — and that bad-label risk propagates downstream, so CellVerdict lists exactly which cell-cell communication interactions rest on a low-trust annotation. Give it an unlabeled dataset and it annotates; give it a labeled one and it audits. Every verdict is glassbox: click any node in the flow canvas to see the votes and the raw numbers behind them. Why it matters: it turns a two-day, subjective, cluster-by-cluster eyeballing task into a two-minute, reproducible, citable trust report — and makes the risk in an annotation visible instead of hidden. - **Team:** [junseong kim](https://cerebralvalley.ai/u/junseong) - **GitHub:** https://github.com/JunseongKim1104/CellVerdict - **Demo video:** https://youtu.be/fn3CIIhXQ-w?si=il5XzOCRS8V1riNn - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=14 ### 39. Kinga Zielinska We set out to build a machine-learning classifier that predicts a clinically meaningful vaginal-microbiome state from a baseline metagenomic profile, using only public data. Because intervention-response labels proved unrecoverable from public deposits — probiotic/live-biotherapeutic response had zero read-joinable labels, and preterm-birth cohorts are dominated by V3–V4 16S that resolves L. iners from L. crispatus poorly — we selected the one abundant, per-sample, self-supervised target the data could support: next-day community-state-type (CST) transition. The cohort was PRJEB37731 (Danish daily shotgun metagenomics, 40 subjects, ~967 consecutive day-pairs), reached through a documented funnel of 90 public datasets → 30 mined publications → 4 candidate targets → 8 feasibility-vetted cohorts → 1. The classifier was engineered for rigor: 184 CLR-transformed species plus covariates, leave-one-subject-out cross-validation, honest baselines (prevalence, persistence, current-state), subject-clustered GEE for effect estimates, and a synthetic-data power analysis. The primary result was an honest negative — the full microbiome composition adds essentially nothing beyond the current-state label (transition AUROC 0.66, matched by a covariate baseline), and next-day onset of dysbiosis is unpredictable (AUROC 0.51–0.55; up to 5 days of memory does not help). We then tested the textbook claim that L. iners is a "gateway to dysbiosis." iners strongly predicts community movement (adjusted OR 5.87, 95% CI 3.15–10.94, p=2.5×10⁻⁸), but that movement is predominantly recovery toward L. crispatus (0.57) rather than descent to dysbiosis (0.24). L. iners marks mobility, not decline — a novel refinement of the field's model. Why it matters: BV recurs in >50% of women within 6–12 months, making treatment response fundamentally a prediction problem. This is a first-of-its-kind, fully leakage-safe attempt to predict near-term vaginal-microbiome dynamics from public data — it honestly maps what is and isn't predictable at current data scale, and delivers a reproducible pipeline ready for a future cohort with linked intervention outcomes. - **Team:** [Kinga Zielinska](https://cerebralvalley.ai/u/Kinzie) - **GitHub:** https://github.com/Kizielins/Anthropic_hackathon_Claude_Science - **Demo video:** https://youtu.be/rWXCYf2jW0w - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=15 ### 40. Nick Lamb When two well-run papers disagree, it is rarely because one of them is wrong. Usually the reason is a methodological difference buried in the methods. StudyDiff is a contradiction explorer for bench scientists: give it two papers and it explains why they reach different conclusions, grounding every claim in the source text. Claude (claude-sonnet-5), through forced tool-use, turns each paper into a fixed study-card schema, returning a verbatim supporting quote for every field. A deterministic OpenGATE grounding check then verifies it: every quote must be a real substring of the source, and every number must trace back. There is no LLM-as-judge. Ungrounded fields are downgraded to "not reported" before they can be cited. StudyDiff compares only the verified evidence, ranks the design dimensions that most plausibly drive the disagreement, and suggests what evidence would resolve it. It also ships as an MCP server, so Claude can call the engine as a tool with the same guarantees. It never invents a confidence score and never picks a winner. If a paper doesn't report something, it says so. Most literature tools help you read a paper. StudyDiff explains why two papers conflict: an explanation, not a summary. - **Team:** [Nick Lamb](https://cerebralvalley.ai/u/pharmatools) - **GitHub:** https://github.com/nickjlamb/studydiff https://studydiff.pharmatools.ai https://www.npmjs.com/package/studydiff-mcp - **Demo video:** https://youtu.be/BwegvGp51-4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=16 ### 41. Ash0723 RegScope- Now available at https://auronideep777.github.io/RegScope/ https://github.com/auronideep777/RegScope The short version: RegScope is a small suite of browser-based tools that do one connected thing: read a piece of DNA, predict how active it is as a regulatory element, and then work out what a single-base change would do to it, not just at the level of "which transcription-factor motif did you break," but across the structural language of the genome as well: G-quadruplexes, i-Motifs, Z-DNA, R-loops, cruciforms, triplex/H-DNA, DNA shape, duplex stability, and nucleosome positioning. It grew out of the Gladstone / Built-with-Claude research question, train a model that reads DNA to predict regulatory activity, then ask what a single base does, using the Deng et al. developing-cortex MPRA dataset (Science, adh0559). The result is four linked pieces plus a landing page, all of which run entirely in a web browser with nothing to install. Why it needed to exist: If you sit at a bench and you have a non-coding variant you care about, the honest state of the art is fragmentation. You go to one web server to ask about G-quadruplexes, a different one for R-loops, a third for Z-DNA and repeats, a fourth for DNA shape, a fifth for duplex melting, a sixth for nucleosomes, each with its own input format, its own output, and no shared frame of reference. Separately, if you want a data-driven prediction of regulatory activity you reach for a heavyweight deep-learning model (Enformer, Sei, Basenji-style) that needs a GPU, a genome build, and real engineering effort, and that tells you almost nothing about why a base matters structurally. Nobody had put the two halves together in a way a non-specialist could actually use. RegulatoryScope's argument is simple: a regulatory variant's effect is a sum of grammar the field usually studies in isolation, so the tool should let you see all of it at once, the trained activity model, the structural decomposition, and the per-mutation breakdown, reproducibly, transparently, and without a cluster. How it works, four tools, one pipeline: A Home page ties the suite together. From there: ● The MPRA Dashboard. A sequence-to-function deep-learning model trained on the cortex MPRA: a reverse-complement-equivariant, residual-dilated convolutional tower with squeeze-excite and attention pooling, run as a 3-seed ensemble. It predicts regulatory activity from raw DNA, flags silencers, scores single-base variant effects, runs saturation mutagenesis, and interprets learned motifs against JASPAR via TOMTOM. Every number is reported on held-out chromosomes. ● The structural feature engine (11 families). For any window of DNA it scores G-quadruplex (G4Hunter), i-motif, DNA shape (MGW/Roll/ProT/HelT/EP), duplex destabilization (nearest-neighbour ΔG / SIDD), Z-DNA, R-loop-forming potential, cruciform (inverted repeats), triplex/H-DNA (mirror repeats), nucleosome positioning, plus transcription-factor motifs. Each is a fast, transparent model with a documented path to its gold-standard tool. ● The decomposition & rigor layer (Additional Data). Six per-family dashboards plus a combined atlas that ask, with real statistics, how much each structural axis adds to explaining variant effect on top of GC content and TF motifs, nested R², partial correlation controlling for GC, a GC-matched control, and a bootstrap confidence interval. It reports negatives as loudly as positives. ● SDM Studio. A site-directed-mutagenesis explorer: enumerate every single-base change (or paste a specific list), get an exportable table of all mutations, and click any one to see its predicted impact across all families plus TF motifs, with the reference context and a citation for every method used. ● RegScope scanner. A base-resolution view of a whole sequence, structure tracks, an in-silico mutagenesis map, CpG islands, ORFs and TF binding, for orienting yourself before you drill into a single variant. The through-line is that all five share the same structural engine and the same honest reporting style, so a number in the scanner, the SDM table, and the decomposition atlas all mean the same thing. How I tested it: Two kinds of claim need two kinds of test, and it matters not to blur them. Biological accuracy of the structural engine was measured with positive/negative controls: for each discrete family I embedded a canonical motif (a known G4, a C-rich i-motif, a CG/CA Z-former, a G-rich R-loop initiator, an inverted repeat, a homopurine mirror repeat) into random background, generated matched motif-free negatives, and computed the area under the ROC curve (AUROC) of the raw score separating them (n = 60 positives / 60 negatives each). I spot-checked G4Hunter against five textbook G-quadruplexes, and validated the duplex-stability model exactly against published SantaLucia nearest-neighbour values. Predictive accuracy of the model is quoted directly from the dashboard's held-out-chromosome results, these are the project's own reported numbers, and I present them as reported rather than re-deriving them, while noting how they compare to the baselines shipped alongside. Software reliability was tested by loading all six pages headlessly and driving the interactive tools through edge cases, empty input, invalid characters, a too-short sequence, an oversized sequence hitting the safety cap, a custom mutation list mixing four notations and deliberate errors, sorting, filtering, family toggling, and every export path, while watching for a single uncaught JavaScript error. On the testing itself. Every control set was generated programmatically, with canonical motifs implanted at random positions inside random-composition backgrounds, and the AUROC was computed as the rank-based probability that a positive outscores a negative. The full suite was re-run to confirm the scores are deterministic, and the reliability pass deliberately drove each tool through its failure modes as well as its normal path, treating any single uncaught JavaScript error as a failure. Honest limitations ● The structural engine ships with fast proxy models. G4/i-motif/R-loop/repeat detection is excellent on controls; DNA shape and nucleosome are directionally right but not validated against DNAshapeR/NuPoP. Every panel says which gold-standard tool to swap in. ● The decomposition dashboards' extended-family numbers were computed on a synthetic demo cohort to exercise the pipeline end-to-end; they are badged as such and become real the moment the pipeline is pointed at your own MPRA/caQTL table. The RegNetDNA metrics and provenance, by contrast, are real and held-out. ● Variant-effect discrimination (AUROC ~0.62) is modest, but that is largely the biology: only about 4% of the library's variants have a significant effect, so there is little signal to find, and the honest move is the confidence-weighted ensemble rather than an inflated score. ● This is a research and hypothesis-generation instrument, not a clinical classifier. Bottom line RegulatoryScope does something the field hasn't packaged before: it reads DNA, predicts regulatory activity with a model that is honest about its held-out performance, and then explains a single-base change through the full structural grammar of the genome, with statistics, citations, and exports, entirely in a browser. The predictive model is strong for its task (Spearman ~0.60, AUROC ~0.80), the structural engine is excellent on controls (AUROC 0.96 mean; thermodynamics exact), and the software is reliable (zero errors across every scenario tested). Its honesty about what its features don't do is, in a field full of overclaiming, one of its most useful features. Datasets used: Developing human cortex MPRA (training, validation, variant library, and all held-out metrics). Deng C et al., Science 2024, both the primary tissue and organoid arms. The activity table used here is byte-identical to Supplement S1 and the variant library to Supplement S2; splits are by chromosome so held-out chromosomes never appear in training. JASPAR 2022 core vertebrate motif profiles, used for the transcription-factor scans and for TOMTOM motif matching. Non-B DB predicted non-B DNA motif tracks, the reference resource the cruciform/triplex and Z-DNA proxies defer to. WTC11 iPSC data, used only as an independent multi-cell negative-control transfer, deliberately a clean negative. Structural-engine control sequences (canonical G-quadruplexes such as c-MYC Pu27, VEGF, BCL2, KIT, and the human telomere repeat) are drawn from the standard G-quadruplex literature; the nearest-neighbour thermodynamic parameters are from SantaLucia 1998. The complete writeup has been uploaded on Github - **Team:** [Auroni Deep](https://cerebralvalley.ai/u/ASH0723) - **GitHub:** https://github.com/auronideep777/RegScope - **Demo video:** https://youtu.be/TuI5c0yHcIc - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=17 ### 42. Marissa Famularo Verastar is a morning literature harness built by and for a busy vascular surgeon who can't read the hundreds of papers published in her field each month, yet can't afford to miss them. The insight: frontier models are superb at reading papers, but can't be trusted on numbers. Verastar wraps Claude Opus in a deterministic verifier: every statistic is located in the source text and gated before it renders, showing what it proved, not what a model asserted. A number that can't be found is never shown; verified numbers appear inline. Papers link to their PDFs and full text, saved work forms a searchable library, and connections tie active projects to new literature. Everything saved becomes a plain markdown file in a folder the clinician owns. Because that memory is an open folder, any agent she trusts can read it -- the seed of a research companion whole teams could build on. Verastar exists to bring doctors who haven't used AI into the fold, and give busy clinicians who never had time for research the chance to change their field. Built end-to-end with Claude Code, including 181 adversarial false-verify tests, by a physician who never coded before late 2025. - **Team:** [Marissa Famularo](https://cerebralvalley.ai/u/caivory03) - **GitHub:** https://github.com/MarissaFamularo/verastar - **Demo video:** https://youtu.be/RmT7JjTTqeM - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=18 ### 43. Claudia & Claude(ius) Right now, 57 million people are living with dementia — one more begins to forget every three seconds. The best-evidenced tool for helping them relearn what matters — a granddaughter's name, "my walker is by the door" — is Spaced Retrieval, and it still works exactly the way it did in 1989: a stopwatch, a stack of index cards, and a clinician's afternoon. The practice belongs at home, daily. But at home, the caregiver becomes the tester of their own mother — the largest analog trial failed on exactly that strain — and the data dies on paper. The field's biggest questions have stayed open for twenty years because home practice is invisible. Keepsake makes it visible. One decision everything hangs on: the device is the therapist; the family is the companion. Claude shapes a daughter's rambling description into a protocol-valid memory target — rejecting its own first draft, on screen. A pure, property-tested engine runs the expanding-interval session on a calm kiosk. On a miss, the screen delivers the errorless correction — a daughter never has to tell her mother she's wrong. Every session ends on a win. And every probe becomes a structured, protocol-faithful trial log. Then the jaw-drop: ask a research question over those logs, and Claude chooses and runs its own analyses — the tool trail shown on screen — and writes an honest n=1 study report. At n=1 it tunes one patient's practice. At n=1000, the identical schema answers questions the field has argued over since the 1990s. Every install is a study site. The design itself is evidence that this works: our own literature review overturned it twice (FSRS and the cueing hierarchy are in the documented graveyard), and Monika Stroińska of Centrum Montessori Senior — the Polish practitioner-trainer organisation in the Camp/Brush lineage — validated the protocol mid-build, before the first screen shipped. - **Team:** [Klaudia Kromolowska](https://cerebralvalley.ai/u/user_1898095096) - **GitHub:** https://github.com/KlaudiaKromolowska/KeepSake - **Demo video:** https://canva.link/eccln14j2rls93r - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=19 ### 44. FACET Scientists in different roles ask the same question and need different answers. Ask about KRAS and a medicinal chemist wants crystal structures and warhead chemistry; a pathologist wants mutation frequencies and IHC localization by tumor type; a computational biologist wants ADC constructs, linkers, and DAR comparisons; a cell biologist wants mechanism and cell-line selectivity. Generic search averages all four into one page, so each person re-digs through material irrelevant to their work. FACET conditions retrieval on the asker's role at three points: planning (which connectors run, and how), extraction (which findings survive a discipline relevance filter), and synthesis (how the dossier is structured and voiced). Role changes the evidence, not just the wording. On the same KRAS query, the four personas draw on different connector sets. Two decisions make it trustworthy. Retrieval runs through deterministic connectors (ChEMBL, PDB, PubChem, UniProt, ClinVar, PubMed, STRING, Reactome, GTEx, AlphaFold, Open Targets, and more), not LLM tool-calling — keeping sources credible and cost near-zero. And retrieved values are sacrosanct: each is tied to a numbered evidence ID, and when a connector returns nothing, FACET says so rather than inventing it. In our demo the medicinal chemist refuses to cite PDB accession codes the evidence block didn't surface, flagging them as established knowledge instead. That refusal-to-fabricate is the feature: hallucination has almost no surface to occur on. FACET isn't a search replacement — it's a lens on top of one, plugging into the retrieval stack biopharma teams already run so the same corpus answers four people correctly instead of one approximately. Future roadmap: per-persona learned rerankers (evidence ordered by what each discipline actually cites), click-through citation traceability to every connector call and timestamp for a regulated-grade audit trail, and cross-persona collaboration on one shared query. FACET conditions the search stack a lab already runs rather than replacing it. - **Team:** [Keerthi Harikrishnan](https://cerebralvalley.ai/u/iKeerthii) - **GitHub:** https://github.com/drkeerthiharikrishnan-h/FACET - **Demo video:** https://youtu.be/NMDtOpBxia8 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=20 ### 45. FlexScience Spatial perturbation transcriptomics aims to measure how a genetically edited cell affects surrounding tissue. I investigated whether these effects form a point-spread-function-like spatial response, bringing a mathematical idea from electrical engineering into an unfamiliar biomedical field. Using a public mouse-brain Spatial Perturb-seq dataset containing 229,775 cells, I found a consequential replication problem. Treating 24,453–31,189 source-recipient pairs as independent produced 16 of 20 nominal program-level hits. Keeping the biological comparisons and point estimates fixed, but computing uncertainty at the 78 or 101 independent perturbation and control sources, increased standard errors by a median of 5.5× and eliminated every hit: 0/20. Wild-cluster bootstrap, source-label permutation, equal-source weighting, non-overlap analysis, alternative adjustments, and leave-one-mouse-out checks agreed. Two Lrrk2 contrasts remain power-limited, method-dependent leads rather than findings. I packaged the central check as a reusable command-line audit that produces naive, source-aware, wild-bootstrap, and equal-source results from a source-recipient table. This matters because neighbour-level pseudoreplication can create false confidence across spatial screens. The resulting design guidance is concrete: count independent sources, balance them across animals, and power experiments at the source level. The broader demonstration is that Claude can help an outsider enter a new research field and contribute rigorously without lowering the standard of evidence. - **Team:** [Tianxiang Dai](https://cerebralvalley.ai/u/txdai) - **GitHub:** https://github.com/txdai/Perturbation-PSF - **Demo video:** https://www.youtube.com/watch?v=R0H_EdXbDcA - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=21 ### 46. JARVIS CuraVar is a "show your work" assistant for classifying genetic variants. Claude reads the evidence and proposes which ACMG/AMP rules apply — each with a citation — then deterministic code does the final scoring, so the AI never touches the number and the same evidence always gives the same answer. Every step is written to a tamper-proof, hash-chained ledger that can be re-verified on demand. Built for the Built with Claude: Life Sciences hackathon. Research decision-support only — the final call stays with a qualified professional. - **Team:** [Husnain Arif](https://cerebralvalley.ai/u/hunain_arif) - **GitHub:** https://github.com/arifhusnain478-cpu/curavar-web.git , https://curavar-web.vercel.app/ - **Demo video:** https://youtu.be/E7D9S8CIZsU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=22 ### 47. Strata Strata is a living meta-analysis like tool for the teams who depend on evidence. A meta-analysis can pull scattered trial data into a single defensible answer, but it goes stale the day it publishes, and refreshing one by hand is slow and expensive. Strata keeps that answer current and lets you trace every number back to its source. You ask a research question, and Strata pulls the matching trials, reads the results out of each one, and gives a first read on how much to trust each. A person stays in the loop: Strata surfaces low-confidence extractions and the trust judgments that carry weight for a human to confirm before anything is pooled, rather than deciding those alone. It then combines what passes into a single answer: the size of the effect, how confident you can be, a forest plot, and a plain-English summary, with every figure clickable back to the sentence it came from. When a new trial reports, it re-runs and tells you whether the answer changed. This works because of one rule: Claude reads, code does the maths. Claude finds and structures the numbers; a statistics library does all the pooling; and plain sanity checks flag anything that doesn't add up rather than folding it in. When the data is too thin to answer, Strata says so instead of inventing a confident number. Claude drives the whole workflow through an MCP server, with a full web platform and CLI over the same core. - **Team:** [S](https://cerebralvalley.ai/u/SimonR) - **GitHub:** https://github.com/simonrab/Strata - **Demo video:** https://www.loom.com/share/bd31b1253cf44b93af0a7cfb87a6f146 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=23 ### 48. e11 Clinical datasets that could move research forward sit unused because two questions are hard to answer together: is this use consented, and is the data de-identified enough to share? Consent tools track permission but do no de-identification math; de-identification engines compute risk but are blind to consent. Lupta joins them. Claude reads the free-text consent document, checks a proposed use against it, discovers quasi-identifiers, and reasons about re-identification — while a deterministic engine owns every number, so no figure in the decision memo is ever produced by the model. It emits an auditable memo: a verdict, the risk findings tagged to the tool that computed them, and every legal assertion cited to a gazette-verified provision of India's DPDP Act — ahead of the May 2027 enforcement date. Built for the researcher trying to get a dataset safely out of the drawer. - **Team:** [Abhishek Srivastava](https://cerebralvalley.ai/u/abhiis) - **GitHub:** https://github.com/abhiis/lupta - **Demo video:** https://youtu.be/S2FxFKZGBpA - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=24 ### 49. IMMUNA In one week, I turned a biological question into a journal-ready paper and reproducible biomarker. IMMUNA-CD4 is a compact 17-gene measure of immune aging in naive CD4 T cells. I developed it using several cohorts and validated it in an independent population. It tracks age and is also linked to immune health after accounting for age. I then tested the biomarker in a genome-scale CRISPRi screen. This created a map of genes that may affect the measured aging signal. The final project includes a complete paper, figures, Source Data, analysis code, provenance, and a verified reproducible build. - **Team:** [Fridolin Haugg](https://cerebralvalley.ai/u/frido) - **GitHub:** https://github.com/frido22/immuna-cd4-paper - **Demo video:** https://www.youtube.com/watch?v=oLNe4ucfsAg - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=25 ### 50. Demir Ege Ortac I ran my tool on published, peer-reviewed data from the Wilbrecht lab at Berkeley (DANDI:001340, dopamine in the nucleus accumbens during reward learning). 8 of 15 reasonable preprocessing pipelines REVERSE the sign of the result. All 16 are significant. The choice that decides it is the bleaching correction, and a researcher would probably never register it as a choice at all. fiberqc is a quality-control tool for fiber photometry, the technique used across systems neuroscience to record dopamine release in behaving animals. THE PROBLEM Before a raw recording becomes a result, it passes through a chain of preprocessing decisions: filter cutoff, bleaching correction, motion regression, normalization. Each one is defensible. Each one is somewhat arbitrary. The field's own reference primer (Simpson, Akam, Patriarchi et al., Neuron 2024) states plainly that there is no systematic comparison of these choices and no established best practice. So results get published without anyone checking whether they depend on a filter setting that was picked without thinking about it. WHAT IT DOES fiberqc runs the same analysis through every reasonable pipeline at once. You declare the pipeline you actually ran. It locates your result in that space, tells you how many defensible alternatives disagree with it, and names the choice that decides the outcome. THE FINDING The eight pipelines using high-pass bleaching correction find a significant decrease (d = -0.33). The eight using a double-exponential fit find a significant increase (d = +0.50). Nothing else flips the sign: not the filter cutoff, not motion correction, not normalization. A researcher would draw opposite biological conclusions depending on that one decision. WHAT THE TOOL FOUND IN ITSELF Running on independently published data exposed three serious bugs in fiberqc. Each is fixed, and each fix ships with a regression test that fails without it. 1. The verdict counted p-values but ignored sign, so a result where half the pipelines found a significant increase and half a significant decrease was labelled ROBUST. A QC tool that marks a dangerous result as safe is the worst failure it can have. 2. Non-uniform sampling was accepted silently. A 140-second LED-off gap was being filtered as if it were continuous signal, and that step discontinuity inverts the measured response. 3. Only the t-statistic was reported. But t = d * sqrt(n), so a trivially small effect over hundreds of trials looked overwhelming (Chen et al., 2017). Magnitude now leads, and the t-statistic is shown as the evidence for the effect rather than as the effect. Three independent labs (Akam/Blanco-Pozo, Lerner/GuPPy, Wilbrecht/DANDI). pyPhotometry, TDT and NWB formats. 29 tests, green on Python 3.10 to 3.12. API docs, CI, MIT. - **Team:** [Demir Ege Ortac](https://cerebralvalley.ai/u/demiregeortac) - **GitHub:** https://github.com/demiregeortac666/fiberqc - **Demo video:** https://youtu.be/NAhpafIWrg4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=26 ### 51. CanCell I built an interactive analysis that separates cell-state composition from cell-intrinsic gene regulation in 1,714 acute-leukemia RNA-seq samples. Using an integrated 35-state bone marrow and thymus reference, we mapped AML, B-ALL, and T-ALL onto normal development and validated the results with matched single-cell data. We found that fusion-defined leukemias are often explained by differentiation state, while several mutations retain strong expression signals beyond composition. This helps distinguish where leukemic cells sit in development from what they express abnormally. - **Team:** [Onur Karakaslar](https://cerebralvalley.ai/u/onurr) - **GitHub:** https://eonurk.github.io/acute-leukemia-map/ - **Demo video:** https://www.youtube.com/watch?v=W3gs1cbZuAA - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=27 ### 52. Ruth-Anne Pai I took one disease from idea to an investable program in a week and then turned that workflow into open AI agents for any disease. It began with eosinophilic esophagitis (EoE), a chronic food-driven condition I live with. From public omics, I mined EoE datasets into a ranked, druggability-filtered target landscape (surfacing CCL26 and POSTN), designed de novo protein binders, and carried it through a business plan, scientific plan, and regulatory strategy. Knowing the burden of life-long symptom maintenance firsthand, I initiated Project Tolera, an antigen-specific program to reprogram the immune system: because EoE is a food peptide presented on MHC-II to a pathogenic Th2 T cell, targeting that triad could induce durable tolerance instead of lifelong suppression. My engine nominated 832 tolerizing epitopes and modeled lead pMHC-II complexes. Seeing the pipeline generalize, I packaged every step into reusable skills, a Therapeutic Program Architect, and a Manuscript Architect, then re-ran it on esophageal cancer as a test. I also built the Patient Organization Navigator: a specialist agent and three skills that help patient advocates and rare-disease nonprofits map where their disease stands, what's been done, where the gaps are, and where their funding and capacity sit. - **Team:** [Ruth-Anne Pai](https://cerebralvalley.ai/u/Rpai) - **Demo video:** https://youtu.be/GV9aKAsQ9I8 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=28 ### 53. OmniGate I built OmniGate, a tool that answers questions about real single-cell perturbation data in plain English, and shows its work. "Analyze" takes a question, writes the analysis in Python, runs it in a sandbox against the actual cells, and returns the figure, the statistics, and the exact code that produced them. "Predict" covers experiments nobody has run: it finds the closest perturbations that were measured, blends their real signatures into an estimate, then audits itself by hiding each known perturbation, predicting it blind, and scoring the result against the truth it withheld. Why it matters is because single-cell analysis is locked behind knowing scanpy, and most tools that unlock it for non-specialists do so by letting a language model narrate plausible-sounding results. OmniGate never lets the model produce a number. The model writes code, the sandbox runs it, and the answer is whatever the data says. - **Team:** [Kian Khooban](https://cerebralvalley.ai/u/kiankhooban) - **GitHub:** https://github.com/kiankhooban/omnigate - **Demo video:** https://www.youtube.com/watch?v=aCHPVNXEHtQ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=29 ### 54. Tuna Birgun omicau (Omics Audit) is a reproducible, leakage-safe command-line tool, with an optional no-code web interface, that audits a multi-omic dataset before anyone trusts a predictive result. Given one matrix per omic layer (RNA, protein, methylation, and so on) plus a clinical outcome, it aligns the layers, locks them with a value-level SHA-256 provenance hash, and tests for the failure modes that make omics machine learning non-reproducible: information leakage, batch confounding, target-linked missingness, and modality redundancy. It then benchmarks whether combining layers beats the best single layer, under group-aware cross-validation that keeps every subject's samples on one side of the split, and runs three shuffled-label control baselines that gate the result — when a control scores above chance, omicau flags leakage and refuses to certify the score. It writes a one-page executive verdict for clinicians and a research tab with confidence intervals, calibration, and per-layer diagnostics, alongside a DOME methods block and a model card; it runs fully offline, installs from PyPI, and was built with Claude Code. I validated omicau on eight public datasets spanning human, mouse, zebrafish, yeast, and Arabidopsis, across classification, regression, and survival, and checked every result against the published literature. On TCGA-BRCA (PAM50 subtype, RNA+CNV, n=493) it reached AUROC 0.953 with controls at chance and certified the result, while its leave-one-layer-out test put the gain of fusion over RNA alone at +0.006 with a 95% confidence interval spanning zero, so it reported that one layer suffices; on an Arabidopsis Botrytis-infection set (n=12) it reached a perfect AUROC of 1.000, yet the shuffled-target control also beat chance (0.89), so it declined to certify — the correct call when the sample is far smaller than the feature count. This matters because leakage, batch effects, and redundant layers are a documented cause of irreproducible omics machine learning; integration methods (MOFA, MOGONET) and reporting checklists (DOME, TRIPOD+AI) exist, but no tool audits a dataset before modeling for these failures and then declines to certify a score it cannot trust. omicau fills that gap: it tells a researcher — a genomics postdoc, a lab, a clinic — when not to trust their own model, backed by a provenance hash anyone can recompute. - **Team:** [Tuna Birgun](https://cerebralvalley.ai/u/tunabirgun) - **GitHub:** https://github.com/tunabirgun/omicau - **Demo video:** https://youtu.be/qQGXTUGV8Gs?si=OKJq9BSNWrfELgsP - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=30 ### 55. Phebe I created PKU Commons, a framework for developing and maintaining a phenylalanine estimating app usable for both managing the dietary requirements of PKU as well as clinical and research purposes. I'm a PKU mom. PKU Commons is intended to systematize the "real world evidence" represented by dietary phenylalanine inputs and logging usually conducted by parents. As AI begins to create actionable new drug development for PKU, it is critical that we speed up the data feedback cycle, both to increase safety and accelerate the cycle of cure discovery. - **Team:** [Nina Kilbride](https://cerebralvalley.ai/u/phebepku) - **GitHub:** https://cureledger.github.io/pku-commons/index.html - **Demo video:** https://youtu.be/6g_epRqbXO8 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=31 ### 56. Concinnity We built a directional readout of organelle dynamics. Standard analysis measures how much of an organelle is present - it can't separate a cell building an organelle up from one blocking its clearance, and the clearance genes (selective-autophagy receptors like BNIP3) aren't in conventional organelle signatures. We score the building and clearing programs separately and subtract: one signed number whose sign is the biology. What we found: reading direction, not amount, predicts mitochondrial vulnerability in cancer. Across 1,066 DepMap cell lines, higher mito-biogenesis direction means stronger genetic dependency on that machinery (ρ=−0.35, p=1.7×10⁻³²) - and it's specific (the same test on ribosomes is null). From expression alone it also ranks drug response: the clinical Complex-I inhibitor IACS-010759 lands in the top 3% of 1,514 drugs, and MitoQ ranks #1. Direction beats amount on the vulnerability endpoint (AUROC 0.67 vs 0.61). Why it matters: OXPHOS-targeting cancer drugs keep stalling on patient selection - no one can tell in advance which tumors depend on their mitochondria. This reads exactly that, from data already collected. Honest bounds: effect sizes are modest, the cancer results are in cell lines, and our one patient-survival test (TCGA-KIRC, 510-tumor cohort) is a clean null at power - reported. All public data. - **Team:** [Taras Nazarov](https://cerebralvalley.ai/u/concinnity) - **GitHub:** https://github.com/different-change/organelle-direction - **Demo video:** https://youtu.be/T2vFfWQJ5Wc - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=32 ### 57. sonrg Project: We benchmarked whether Boltz-2 (an AlphaFold-class structure predictor) can substitute for a binding measurement, testing two complementary questions across two systems: does its confidence score rank affinity (SH2 domains vs. measured KD panels for c-Src and Fyn), and does it detect a damaged interface (MDM2:PMI under site-saturation mutagenesis). What we found: Boltz-2 reliably separates binders from non-binders — free two-peptide competition is the best readout (65% c-Src, 81% Fyn) — and maps which residues matter with anchor-localized, chemically-coherent sensitivity. But it cannot resolve fine affinity rank among true binders (the Fyn 1-log test collapses, residual pointing the wrong way), and the peptide-linker fusion trick does not transfer to a folded protein–protein interface (barnase–barstar fails at every linker length, even femtomolar wild-type). Why it matters: It draws a clear, evidence-backed boundary for using structure-prediction confidence as an affinity proxy — sound for binder/non-binder triage and interface-hotspot reading, unreliable for quantitative ranking — so users know which question the score can actually answer before they trust it. - **Team:** [Ryeogang Son](https://cerebralvalley.ai/u/pulzang) - **GitHub:** https://drive.google.com/file/d/1qDgJ2B9fEnjECtoG8DJhE80ZFUiTcuw8/view?usp=sharing - **Demo video:** https://youtu.be/EoF65RKxpH8 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=33 ### 58. TFModulators Conserved gene co-expression & regulatory networks across vertebrates — abstract One of the holy grails of biology is to understand how 1 genome can make all the different cell types, each with a unique regulatory network. While these questions are starting to be answered in model organisms such as the mouse and human, it's unclear if these regulatory networks are conserved across species in the same cell types. If these networks are conserved across species, then gene regulatory networks found in model organisms are conserved and can be used to predict genotype-to-phenotype relationships in other mammals or vertebrates. However, if the same transcription factors are not regulating the same genes in the same cell types, then data must be collected in distant organisms. Using both bulk-RNASeq and scRNASeq, we found that a conserved core — but not all — of the gene co-expression networks is preserved across 26 vertebrate species spanning humans to zebrafish, representing ~430 million years of evolution. Starting from a well-powered human reference (GTEx, 30 tissues), we discovered 27 tissue-defining co-expression modules and tested their preservation across the vertebrate tree. The modules split cleanly into two classes: a core set of programs — neuronal-synaptic (brain), meiotic (testis), oogenic (ovary), contractile (muscle) and immune — that is conserved essentially everywhere, and a labile set — pancreatic, adipose and lung — that is not. Strikingly, conservation is governed by which program a network belongs to, not by how long ago two species diverged (preservation vs. divergence time Spearman rho = -0.24, n.s.; Blomberg's K << 1 for the conserved core), implying that these core programs are held in place by deep stabilizing selection acting across all lineages at once rather than eroding gradually with evolutionary distance. Within each conserved module, gene membership is preserved while the fine-scale hub wiring drifts — a pattern we confirmed at single-cell resolution (Tabula Sapiens / Muris / Microcebus for human, mouse and lemur, plus a zebrafish outgroup): the identity of the master regulators is the conserved axis (regulator-identity conservation 0.42), while their downstream target sets are the labile one (0.17). The conserved regulator core reaches all the way to teleost fish — MEF2 (muscle), SPI1/PU.1 and IKZF1 (immune), and EBF3/PBX (neuron) recur as top regulators in zebrafish (permutation p <= 0.016) — which brackets the intervening marsupial and afrotherian lineages. We discovered transcription factors that regulate different genes, depending on the presence of other transcription factors — combinatorial "cocktails" recovered purely from co-expression (e.g. hepatocyte HNF1A+HNF4A+HNF1B, immune IKZF1xSPI1, testis FOXM1xE2F2). Critically, single-cell data showed that the same TF pair can act as a genuine co-regulator in one species or cell type and as a cell-composition artifact in another (e.g. IKZF1xIRF4 is a genuine plasma-cell cocktail only in lemur), demonstrating that combinatorial regulatory logic is itself context-dependent across species. These results argue that gene regulatory networks discovered in model organisms are, for the conserved core, transferable — master-regulator identity can be used to predict genotype-to-phenotype relationships across mammals and vertebrates — while a minority of labile, clade-variable programs (notably the KRAB-zinc-finger-rich pituitary and adipose modules) will require data collected directly in the target organisms. This work represents a limited pilot; additional data in more distantly related and undersampled vertebrates (marsupials such as koala and Tasmanian devil, and afrotherians) is needed to determine precisely which networks remain conserved under which master regulators. I also worked on another project, it's a mobile app for clinical trial matching, see the github - **Team:** [Brandi Cantarel](https://cerebralvalley.ai/u/bcantarel) - **GitHub:** https://github.com/bcantarel/claudescience_hackaton_20260713 - **Demo video:** https://youtu.be/-HLpOqsZe4g - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=34 ### 59. CoronaryAtlas The research question. When a patient's troponin rises, the heart is injured — but why? Is it a Type I MI, a true infarction from an atherothrombotic plaque rupture in a coronary artery, or a Type II MI, where the heart is starved of oxygen by something else entirely (sepsis, anemia, a fast arrhythmia, respiratory failure)? The two demand very different treatment — one may go to the cath lab, the other needs the underlying stressor fixed — yet troponin and even the ECG frequently cannot tell them apart. There is no validated blood test that cleanly separates rupture from demand. Our question: across the entire molecular cascade of atherothrombotic MI, which biomarker — known or overlooked — could actually make that distinction? What we built. To answer it systematically rather than by hunch, we constructed a ground-up, interactive atlas of the Type I MI pathway. Starting from raw sources, we mined and normalized 1,948 molecules from PubMed (2,716 articles), ClinicalTrials.gov (974 trials), Open Targets (177 genes), the GWAS Catalog (209 genes), and omics repositories, placing each into a 9-step mechanistic cascade with mechanism, full provenance (PMIDs/trials/omics/genetics), and druggability. Every candidate is then scored against a 7-driver Type-II confounder panel, for plaque-rupture responsiveness, a direct Type-I-vs-Type-II differential (only where head-to-head evidence exists), assay feasibility, and evidence strength — combined into a composite Type-I Discrimination Index. A Next.js app makes all of it explorable: pathway browser, per-molecule evidence pages, a discrimination view, a weight-tunable diagnostic-utility explorer, assay panels, and a dashboard with a rupture-vs-demand "signature map," a linked cross-filter, and a discrimination leaderboard. What we found. The answer is a gap. The best-discriminating markers are the familiar injury markers (CK-MB, cardiac troponin I, copeptin, cardiac myosin-binding protein C), but of the 1,948 molecules, only 6 have any direct, head-to-head Type-I-vs-Type-II evidence in the literature — every other candidate is scored from indirect signal or flagged as insufficient. That scarcity, visible as the near-empty "ideal Type-I-specific" corner of the signature map, is the result: the field knows extensively whether troponin rises in MI and almost nothing about which kind of MI it distinguishes. Existing classifiers largely detect the demand/Type-II axis or shared necrosis; there is no validated circulating marker of the plaque-rupture event itself — and the atlas surfaces exactly the under-studied rupture-axis candidates (chymase, MMP-2, thrombus-specific signals) that deserve testing. Why it matters. Type II MI is common, frequently misclassified, and treated differently from Type I, yet no biomarker cleanly separates the two. The atlas doesn't claim to discover new markers — it turns a scattered literature into a systematically-derived, interpretable feature prior: a two-axis (rupture vs demand) map that names both the current best options and the specific evidence gaps a targeted study or an adjudicated-cohort model (High-STEACS, APACE, SWEDEHEART, or a MIMIC-IV pilot) would need to close. - **Team:** [Narayana Sarma Singam](https://cerebralvalley.ai/u/Singamnv) - **GitHub:** https://github.com/singamnv/t1-mi-pathway-atlas - **Demo video:** https://www.youtube.com/watch?v=OgdaWenwbDo - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=35 ### 60. BiotechOS BiotechOS — an operating system for running a drug-discovery pipeline. What we built: A system that reads a drug program's real inbox — CRO emails, quotes, invoices, contracts, assay reports — and turns every message into a concrete action a human just approves. A classifier routes each email to a specialized processor that extracts exactly what that document type needs; nothing changes state until a human signs off. On top sits a molecule layer (registry gate → Molecule Database with live pass/fail against a Target Product Profile → deep-dive → structure modeling to design new analogs) and a memory layer: a time-stamped record of every fact, a self-linking entity graph, and an immutable decision log, all partitioned by program. What we found: Most of a preclinical program's operational load — procurement, finance, legal, data QC, molecule tracking — is coordination that frontier AI plus domain-specific software can absorb, with the human kept as the approver, not the operator. Why it matters: AI and outsourcing have collapsed the cost of doing the science, but not the headcount needed to coordinate it. BiotechOS attacks that coordination tax directly, so one person can run a whole pipeline, not just a project. - **Team:** [Sourav Bandyopadhyay](https://cerebralvalley.ai/u/sza) - **GitHub:** https://github.com/souravUCSF/BiotechOS/ - **Demo video:** https://drive.google.com/file/d/15HOwhutplWfe9OYXJeBQVS7wyM7P__vL/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=36 ### 61. CREative To understand how Alzheimer’s disease reshapes the brain’s regulatory landscape cell type by cell type, I used Claude Science to analyze Droplet Paired-Tag data from 10 donors across the hippocampus and cerebellum. This method measures RNA together with two histone marks in the same nucleus: H3K27ac, associated with active enhancers, and H3K27me3, associated with Polycomb repression. The analysis produced a two-mark cis-regulatory atlas and identified coordinated pathway-level remodeling in specific cell types. The main finding was a region-specific glial chromatin switch in the AD-vulnerable hippocampus. Reactive astrocytes gained H3K27ac at inflammatory programs, including NF-κB, IL-1, and IL-6 signaling, while homeostatic astrocytes lost enhancer activity. At the same time, homeostatic microglia lost H3K27me3, suggesting de-repression or priming before full activation. These results support a testable model in which AD glial reactivity is an ordered, chromatin-encoded state transition, potentially initiated by local pathology. By integrating chromatin change, CRE–gene links, RNA expression, and AD genetics, the analysis also prioritized candidate targets, led by SORL1 in hippocampal microglia, and motivated a three-aim proposal to test where the switch occurs, what drives it, and why it matters. - **Team:** [Xia Han](https://cerebralvalley.ai/u/xhan) - **GitHub:** https://github.com/Xiaxhan/AD_CUTtag_claude_science/ - **Demo video:** https://youtu.be/XL4lIuoHrvU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=37 ### 62. Basecase I built Basecase, an AI-assisted genomic research workspace that connects natural-language questions directly to an interactive genome browser. The system handles disease, gene, variant, regulatory, conservation, transcript, genomic-range, and VCF-based queries using grounded biological evidence. Researchers need more than prose: answers become useful when genes, variants, ranges, and annotations remain inspectable and actionable in the browser. Basecase therefore supports linked loci, multiple selected regions, live reasoning summaries, evidence sources, precise variant anchors, and agent-generated annotation tracks. This matters because it creates a faster, more transparent workflow for moving from a biological question to evidence that can be independently examined. - **Team:** [Kevin Hu](https://cerebralvalley.ai/u/khu) - **GitHub:** https://github.com/kevinhu/basecase - **Demo video:** https://www.youtube.com/watch?v=qo9Ggrqrm3g - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=38 ### 63. InsideOut Bio Every approved atopic dermatitis (AD) drug blocks the type-2 pathway from the outside: injectable antibodies against extracellular cytokines. The frontier is inside the cell, at the transcription factors that run the allergic program, because degrading a master node promises deeper, oral suppression. STAT6 is the proof of concept, but no one had systematically asked which other intracellular regulators would collapse the program if shut down. I screened a genome-scale CD4⁺ T-cell Perturb-seq atlas (~34,000 knockout signatures), scoring every perturbation for how strongly it phenocopies the STAT6 and GATA3 "Th2-collapse" signature. The pipeline recovers both anchors at rank 1 which is a built-in positive control and surfaced an unexpected finding: in circulating cells, STAT6 and GATA3 knockouts anti-correlate, meaning GATA3 maintains the type-2 program STAT6-independently. This is corroborated three ways (cosine screen, an independent supervised classifier that ranks GATA3 #4/6,923 but STAT6 #4,809, and prior literature(Ouyang 2000, Yamashita 2004), and it reproduces in independent lesional-skin scRNA-seq (GSE147424). I then expanded to every AD immune axis a T cell can represent (JAK-STAT, Th22/AHR, Th1, activation-tone, AD-GWAS TFs), cross-referenced the known AD drug universe, and layered on degrader-tractability, human-genetics (allergic-disease GWAS), and literature-novelty annotation. Deliverables: a druggability-annotated multi-axis target map, a reusable Th2-collapse classifier, and a fully rerunnable pipeline others can build on. Why it matters: it converts a single validated drug target (STAT6) into a ranked, evidence-annotated landscape of the next ones with IL4R, ORAI1, SMARCA4, SMAD4 and a TYK2-axis lead (IFNAR1/anifrolumab) surfacing as candidates, and it delivers a classifier that scores any new perturbation for Th2-collapse potential. - **Team:** [Sayane Shome](https://cerebralvalley.ai/u/Say181) - **GitHub:** https://github.com/sayaneshome/atopic-dermatitis/ - **Demo video:** https://viddler.com/qrJ2d5 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=39 ### 64. Bench to Biobank Autoimmune drug discovery often fails when a gene that looks important in a cell experiment doesn't hold up in real people. We built a reproducible pipeline that takes drug-target hypotheses from the Marson-lab genome-scale CD4+ T-cell Perturb-seq screen all the way to individual-level human genomes in All of Us. Every regulator passes through five independent evidence layers: bench causality (CRISPRi knock-down that calms T-cell activation) → druggability → human GWAS → biobank PheWAS (FinnGen + GWAS Catalog, cross-ancestry check in BioBank Japan) → individual-level genomes in All of Us. From ~4,000 regulators, 2,771 pass the bench filter and 86 are druggable and disease-linked. What we found: the method works (PTPN22 and TYK2 controls replicate); IL23R validates for inflammatory bowel disease in the right disease (OR 0.73, P 4.4x10^-5); the three novel lupus receptor hypotheses return honest, well- controlled results; and the pipeline catches its own false positives with a nearest-gene stress test (e.g. our flagship "hidden hit" maps to a famous neighbor, not our gene). It also surfaced ATP2B1 as an SLE-protective lead, then honestly down-weighted it when its genetics turned out cardiometabolic. Why it matters: the discrete result is not "three lupus targets" — it's a reproducible bench-to-biobank pipeline that confirms a real target, gives honest nulls, and catches its own false positives. It's packaged as an open, config- driven tool (bench2biobank) that runs on any perturbation screen. Built entirely in Claude Science. - **Team:** [Yukang Jiang](https://cerebralvalley.ai/u/Kanny) - **GitHub:** https://github.com/kannyjyk/bench2biobank - **Demo video:** https://youtu.be/9pDIZY7dqPI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=40 ### 65. Protocol To Data protocol-to-data turns a clinical trial protocol into a validated, analyzable synthetic SDTM dataset in one Claude-driven agentic loop. Drop in a protocol (PDF/HTML/text) and Claude reads the prose into a typed study design — arms, visit schedule, endpoints mapped to the right SDTM domains — then deterministic Python generates SDTM-shaped synthetic CSVs, validates them (schema, referential + temporal integrity, physiologic ranges, planned-domain coverage), and on any failure Claude repairs its own design and regenerates. A second Claude agent then injects controlled data-quality defects and hunts them down, scored against ground truth. What I found: the loop generalizes from a toy cardiology sample to a real 179-page oncology protocol (Amgen AMG 510 vs docetaxel) with no code changes — extracting cycle-based visits and self-repairing 8 unproducible oncology domains in one pass. The generated data is biologically responsive: it models docetaxel-induced myelosuppression, and the Copilot's "average NEUT per arm" chart shows the docetaxel arm's neutrophils sitting well below the sotorasib arm — real pharmacology, not random distributions. The detector even flags a grade-4 neutrophil count on its own. Why it matters: clinical data managers and biostatisticians spend days-to-weeks hand-building SDTM test data before an EDC (Medidata/Veeva) is stood up. This collapses that to minutes and produces Databricks-ready, reproducible, PHI-free data they can stress-test analysis and edit-check pipelines against before first patient in — turning sequential setup into parallel work. It ships with a two-tab web UI (pipeline + a memory-safe DuckDB Data Copilot with charts), a Registry Cross-Check against ClinicalTrials.gov, a clean HTTP/MCP API, and a live cloud deployment. - **Team:** [Chetan Tamballa](https://cerebralvalley.ai/u/chetantamballa) - **GitHub:** https://github.com/chetankumart/protocol-to-data — public, tagged v1.0.0. Live demo: https://protocol-to-data.onrender.com - **Demo video:** https://youtu.be/JJXIagmZX3Q - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=41 ### 66. Aging Lab What I built. A deep generative single-cell model of the human immune system that simulates aging and rejuvenation at single-cell resolution, then reasons forward from the simulation to mechanism and to candidate drugs. On an atlas of 1,248,980 immune cells from 981 donors (ages 19–97), I trained an scVI latent space and an OT-CFM flow field whose forward direction is aging and backward direction is rejuvenation — a trajectory you can traverse in silico in both directions. The entire pipeline, from research proposal to submission-ready manuscript, was run with Claude alone through a purpose-built toolchain of 9 Claude Code skills (proposal → analysis plan → methods → figures → manuscript, with an adversarial reviewer loop). What I found. The simulation reveals a reversible loss-of-proteostasis / translation axis as a central mechanistic feature of immune aging. It recovers known hallmarks as positive controls (p16/CDKN2A, GZMK⁺ inflammaging, WRN, the mTOR/PI3K geroprotector axis) and nominates 1,558 new biological-age markers absent from any curated aging database. Closing the loop from simulation → target → drug, a genome-scale CRISPRi screen identified 47 causal reverser genes, and molecular docking against three druggable targets (LSD1/KDM1A, MAO-A, VEGFR1/FLT1) yielded 12 structurally-plausible clinical/approved drug candidates that reverse the cytotoxic-inflammaging arm of immune aging. A key signature replicated in an independent cohort (CD8 T, r = 0.92, different chemistry). Why it matters. Aging is nearly impossible to study experimentally — you can't wait decades for a cohort to age, or for a rejuvenation to happen and then sample it. A generative engine that traverses the aging trajectory virtually turns that intractable experiment into something you can run in silico, in both directions, and mine for interventions. This work demonstrates a complete, reproducible loop — virtual simulation → mechanism → causal target → concrete drug hypothesis — carried end-to-end by an AI agent. The results are computational hypotheses, not validated geroprotectors, but they define a new, testable, low-cost front end for anti-aging drug discovery. - **Team:** [Sungwon Jeon](https://cerebralvalley.ai/u/jsungwon) - **GitHub:** https://github.com/jsungwon/anti-aging-drug - **Demo video:** https://youtu.be/Q3VdnJbTPlA - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=42 ### 67. Meow Immune checkpoint blockade (ICB) reactivates T cells to kill tumors and has transformed cancer treatment, but many patients don't respond. Most work on ICB response has focused on CD8⁺ killer T cells, but emerging evidence shows the CD4⁺ helper T cell is also important. Focusing on CD4⁺ helper T cells, I asked which gene-regulatory pathways push them into a non-responding state, and whether those pathways are druggable. I compared CD4 T cells from responders and non-responders in single-cell RNA-seq of 32 melanoma patients to define a non-responder signature, then integrated a genome-scale Perturb-seq atlas of ~2,600 gene knockdowns in healthy-donor CD4 T cells. Because each knockdown is a direct perturbation, the model learns which regulators causally push cells toward responding or non-responding, not merely what correlates. A model trained on the resting-condition perturbation data reproduced the patient signature well, and the top knockdowns cluster into coherent pathways. The interferon pathway was the highest-ranked controller of the response state, and its downstream genes are upregulated in non-responders — so inhibiting interferon should push cells toward responding. Critically, interferon-receptor antibodies and JAK inhibitors are already-approved interferon-pathway inhibitors, turning the analysis into a concrete, testable hypothesis to sensitize non-responders. - **Team:** [Shanshan Yin](https://cerebralvalley.ai/u/Shirlay) - **GitHub:** https://github.com/shanshan-yin/Claude-Science-T-cell-pertubseq-melanoma-ICB https://raw.githack.com/shanshan-yin/Claude-Science-T-cell-pertubseq-melanoma-ICB/main/websites/index.html - **Demo video:** https://youtu.be/htTqtNN_LVY - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=43 ### 68. Team SyL In this project, we established an automated pipeline that seamlessly generates explicit-solvent molecular dynamics (MD) ensembles directly from SMILES inputs, extracts the full distributions of 3D descriptors, and enables interactive exploration of the resulting data. Applying this framework to 37 benchmark molecules spanning the entire flexibility spectrum revealed that increasing molecular flexibility strongly correlates with a larger individual frame variance in 3D descriptors, particularly 3D-PSA (Spearman ρ ≈ 0.5). Furthermore, we demonstrated that for highly flexible molecules, descriptor values derived from a single static structure frequently fail to represent the broader ensemble distribution. These empirical findings substantiate a fundamental, first-principles limitation of static single-conformation descriptors in predicting the physicochemical properties of large, flexible molecules, underscores the absolute necessity of ensemble-based computational approaches. - **Team:** [Sangyeon Lee](https://cerebralvalley.ai/u/syl) - **Demo video:** https://youtu.be/N0GSxMwF-Uc - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=44 ### 69. Anuj Dev Singh I tested a specific, high-stakes question for disease genetics: can today's best sequence-to-function deep-learning models (AlphaGenome, Borzoi, and the DeepSEA/Basset lineage) predict the DIRECTION — up or down — of a noncoding variant's regulatory effect in a defined cell type? This is the quantity a variant-interpretation pipeline actually needs, yet it is rarely benchmarked directly. I assembled a chromosome-split benchmark of noncoding variants with laboratory-measured effect direction across nine cell contexts (6,377 variant×context measurements), plus a native human-microglia chromatin-accessibility QTL map from 95 donors as an on-target test. On the fairest possible test — accessibility direction in the native cell type — AlphaGenome (0.537), Borzoi (0.551), motif-PWM floor baselines, and models trained in-distribution on the caQTL data itself were ALL statistically indistinguishable from chance (majority-class 0.565). I then found the mechanism. In the same 95 donors, a variant's effect on chromatin accessibility and its effect on gene expression are statistically INDEPENDENT (sign concordance 0.506, 95% CI 0.455–0.556, p=0.87). A model that reads one functional layer cannot recover the direction of a layer it does not read. The Alzheimer's risk variant rs6733839 (upstream of BIN1) anchors it: the risk allele opens chromatin AND raises BIN1 expression, yet represses episomal enhancer activity — three "directions" at once. A Boltz-2 co-fold shows a tighter MEF2A–DNA interface at the risk allele (+12.5% contacts). Why it matters: this is a well-powered NEGATIVE result with a mechanism. It tells the field that directional predictions from single-layer sequence models should not be trusted for variant interpretation — and explains precisely why. I release the harmonized benchmark, the paired-layer independence statistic, and all per-variant scores as a reusable community resource. - **Team:** [Anuj dev Singh](https://cerebralvalley.ai/u/anujdevsingh) - **GitHub:** https://github.com/anujdevsingh/regulatory-direction-independence - **Demo video:** https://youtu.be/pgvrCK41Xu8?si=5bRzg5qvzTGyiBhV - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=45 ### 70. team holly Glioma Copilot helps a neuro-oncology clinician and, through them, the patient - decide whether a clinical trial fits this specific glioma patient. It is not a trial finder: clinicians already know the trials. The hard, still-manual part is fit. Eligibility criteria and patient charts are both free text, so after filtering, a clinician verifies every criterion by hand , since the 2021 WHO reclassification, a stale diagnosis label can silently misroute eligibility. I built a copilot that reads both sides and produces a per-criterion fit table , met, not-met, or unknown, each source-cited , turning unknowns into a workup checklist. Its principle: Claude is never the source of truth. Diagnosis runs a deterministic WHO CNS5 rule engine; drugs are grounded in RxNorm/ChEMBL; the literature layer is generated with Claude Science and audited against PubMed (23/23, zero hallucinated). A three-agent loop drafts, an Opus verifier rewrites over-claims, and an independent Opus auditor re-derives eligibility blind and challenges the table. What we found: with real de-identified TCGA patients against real strict trials, almost nothing is a clean "eligible" , so the tool refuses to fake a green match and shows honest unknowns instead. And it does what no trial tool does ,it brings the patient into the decision with a plain-language explanation, a curated FAQ, and a shared-decision handout. It shows the clinician why a trial fits, or doesn't - **Team:** [Holly Tang](https://cerebralvalley.ai/u/Holly99) - **GitHub:** https://github.com/tangxiya-star/glioma-copilot - **Demo video:** https://youtu.be/OQuaID7qxZ4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=46 ### 71. Tanja Hann Most drugs for autoimmune and inflammatory disease work by blocking a cytokine everywhere, all the time. That is effective, but it broadly dampens immunity and raises infection risk. A more precise strategy is to target the immune response only when it is active. Resting and stimulated T cells run very different genetic programs, so a gene that controls a disease-relevant cytokine only in the stimulated state is a far cleaner target: interfering with it should calm an over-active response while sparing resting immune function. I built a reproducible pipeline that finds these context-specific regulators and then asks the practical question that decides whether a target matters: can it actually be drugged, and how? Starting from the genome-scale CD4+ T-cell Perturb-seq screen of Zhu, Dann et al. 2025 (every gene knocked down, one at a time, read out at rest and under stimulation), I tested a perturbation × condition interaction for each gene across 30 cytokines, rather than comparing the two states separately. A gene counts only if its knockdown moves a cytokine under stimulation, does essentially nothing at rest, and passes that interaction test genome-wide. I scored each hit for selectivity (rewarding focused single-cytokine regulators over whole-cell master switches) and applied a cross-donor reproducibility and on-target filter. This generated a high-confidence shortlist of 128 context-specific regulators, both suppressors (anti-inflammatory leads) and inducers (leads for boosting immunity, for example in cancer). The part I'm most excited about is the druggability layer, which draws directly on my background in RNA therapeutics. Target discovery has historically branded much of the genome "undruggable" because it lacks a small-molecule pocket or a cell-surface handle. But those genes are often perfectly targetable at the RNA level with ASOs or siRNA. So I scored all 128 genes on two independent axes: protein-level druggability (small molecule / antibody) and RNA-level druggability (siRNA / ASO), the RNA axis informed by transcript accessibility, predicted stability, and subcellular localization, the same considerations that govern real oligonucleotide design. Plotting both axes turns a binary "druggable/undruggable" call into a map with a concrete modality for nearly every gene: of 128, only a handful score poorly on both, and a distinct RNA-preferred group emerges where an oligonucleotide strategy is the natural path. Why it matters: the output is a ranked, modality-annotated shortlist of new candidate targets for tuning specific cytokines in activated T cells. It is a resource others can build on, and a pipeline retargetable to any cytokine, cell type, or disease. By treating the RNA axis as a first-class route rather than a fallback, it deliberately keeps in play the regulators traditional discovery throws away, which is exactly where modern RNA medicine is strongest. - **Team:** [Tanja Hann](https://cerebralvalley.ai/u/thann) - **GitHub:** https://tanja-hann.github.io/Context-Specific_Cytokine_Regulators/ - **Demo video:** https://youtu.be/7SkWnNmmvUI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=47 ### 72. Brakepoint Brakepoint is a genome-scale discovery engine for the next generation of cancer-immunotherapy drug targets. The best checkpoint drugs work by cutting the "brakes" off a patient's own T cells — but only a handful of those brakes have ever been drugged, and most patients still don't respond. Brakepoint goes looking for the rest, across the whole genome. WHAT I BUILT. A fully reproducible pipeline that reads a public 2,638,736-cell, genome-wide human CD4+ T-cell CRISPRi Perturb-seq screen (Marson lab, Gladstone Institutes; 12,449 gene knockdowns) and, for every knockdown, asks two questions at once: how hard did switching the gene off hit the cell (a coverage-equalized energy-distance effect size), and which way did it push the cell — toward a stronger "fighter" (effector) state or a weaker (exhausted) one (a signed direction axis: a 16-gene effector program minus a 13-gene exhaustion program). WHAT I FOUND. At 2.6M cells, ~97.5% of knockdowns clear significance, so p-values can no longer rank targets — and ranking by raw effect size points straight at essential machinery (18 of the top 20 hits are genes the T cell needs to survive, including its own TCR-signaling core). Adding the signed axis fixes this: it separates a candidate brake from essential machinery. With zero prior hints, Brakepoint rediscovered CBLB — a brake already advancing through clinical trials (NX-1607, HST-1011) — then surfaced four more candidates: CD5, DGKA, SMAD3, and UBASH3A, each scored across seven independent lines of evidence (causal effect, direction, donor consistency, screen fitness, target tractability, immune genetics, and clinical precedent). WHY IT MATTERS. Recovering a target the industry already drugs — blind, from raw data — validates the method; the other four are concrete, testable hypotheses for the next experiment. Brakepoint is a blueprint for AI-native target discovery: one person, one week, on public data, with every result a versioned artifact carrying its exact code and reasoning trail. LINKS. Live project + interactive explorer: https://duanchengchen-oss.github.io/brakepoint/deliverables/ | Open-source code (MIT): https://github.com/duanchengchen-oss/brakepoint | Methods deep-dive: https://github.com/duanchengchen-oss/brakepoint/blob/main/pipeline/METHODS.md - **Team:** [Chengchen Duan](https://cerebralvalley.ai/u/Sam_Duan) - **GitHub:** https://duanchengchen-oss.github.io/brakepoint/deliverables/ - **Demo video:** https://youtu.be/Jksufatf51A - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=48 ### 73. Team Jack Living Protocol Copilot A grounded, safety-first copilot for wet-lab protocols What I built Living Protocol Copilot is a web application that transforms static laboratory SOPs into structured, risk-aware “living protocols.” It does two main things. First, it analyzes a protocol step by step, highlights high-risk procedures, and identifies every important detail that the SOP leaves unspecified. Rather than guessing or completing missing information, the system explicitly marks those gaps for the user. Second, it provides evidence-grounded troubleshooting for failed experiments. For a symptom such as “bands are weak and smeared,” the system returns a ranked list of possible causes. Each cause is linked to both a specific step in the user’s own SOP and a published source from a curated whitelist of 42 trusted references, including Thermo Fisher Scientific, Bio-Rad, Abcam, and New England Biolabs. Users can paste a protocol, upload a PDF, or photograph a printed protocol using their phone. The application also includes a grounded conversational agent for asking follow-up questions about the protocol. The system is built with Next.js and FastAPI. All Claude calls are handled server-side, using Sonnet 5 as the primary model, Opus 4.8 as a fallback, and Haiku for scope gating. Responses are grounded through the Claude Citations API. What I found The hardest part was not getting Claude to answer. It was getting Claude to refuse when the evidence was insufficient. A general-purpose assistant may confidently invent a missing reagent concentration or suggest a plausible-sounding experimental cause. In a wet-lab environment, however, a fabricated number can become a safety risk and undermine experimental reproducibility. For that reason, the entire system is built around four red lines: 1. Never invent biological or procedural information. 2. Never declare a procedure “safe.” 3. Remain grounded in evidence or remain silent. 4. Express uncertainty clearly, treating “insufficient information” as a valid first-class answer. In a representative live run, the protocol parser identified 15 unspecified details and filled in none of them. The troubleshooting system generated 10 candidate causes, all supported by citations. Eight were linked directly to a specific SOP step, and every external source was verified against the curated whitelist. Why it matters Failed experiments are expensive, time-consuming, and often difficult to diagnose. The explanation may be buried in scattered vendor manuals, troubleshooting PDFs, or the undocumented knowledge of an experienced researcher. There is strong evidence that this problem is real: manufacturers routinely publish symptom-indexed troubleshooting tables. Eight of the 42 sources in this project are dedicated troubleshooting guides. These documents exist because experimental procedures do not always work as expected. Living Protocol Copilot explores a more responsible way to apply large language models in high-stakes scientific workflows: cite real evidence, expose missing information, communicate uncertainty, and openly admit when the answer is unknown. This principle extends beyond wet-lab research. In any high-stakes domain, a confident but incorrect answer may be far more costly than saying, “I don’t know.” - **Team:** [JACK NG](https://cerebralvalley.ai/u/jacksen) - **GitHub:** https://github.com/jacksen-ng/living-protocol-copilot - **Demo video:** https://youtu.be/bYIiml9bOeo - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=49 ### 74. DoctorDean What I built. Lazarus is an agent that takes a dead research repository that is open, cited, and no longer runnable, and turns it back into a callable, containerized software component, from nothing but a GitHub URL. It reads the repo and the paper to write its own goal and a falsifiable success test, then runs a build → run → read-the-traceback → repair loop in a disposable Docker sandbox (CPU, remote, or GPU), pinning dependencies to the repo's commit era and fixing real source bugs along the way. Every revival emits the same "contract" in the form of a pinned container, a clean API/CLI, and a smoke test. This way revived tools from any language or era compose into pipelines, and the fixes go back upstream as pull requests. It's a real Python package on the Claude Agent SDK, published to PyPI (pip install lazarus-bio), with a docs site, a zero-setup Colab notebook, and CI. What I found. I revived six abandoned methods spanning wildly different dead stacks: MaSIF, ScanNet, dMaSIF (GPU), fpocket (2010-era C), Basset (2016 Lua Torch7, genomics), and DiffDock (2023 diffusion docking), two of them from a bare URL with the agent writing its own plan. It didn't just make them run: it fixed a 15-year-old undefined-behaviour bug in fpocket's C, patched a CUDA path in dMaSIF that unlocked GPU execution the original forced onto CPU, and caught a silent soft-masking bug in Basset that a naive run would have shipped as a wrong result. It reproduced published benchmarks as a trust layer (MaSIF 0.82 vs 0.85; Basset AUROC 0.894 vs 0.895; DiffDock top-1 ≈0.375 vs the reported ~0.40), and when DiffDock's shipped example wouldn't honestly clear its own 2 Å success bar, the agent refused to fake a pass and reproduced the paper's aggregate accuracy instead. Composed together, the revived tools ran a live binder-triage pipeline on PD-L1 and correctly concluded "a flat protein–protein interface, an antibody target, not a small-molecule one." Why it matters. Computational science has a reproducibility crisis that is really a software-decay crisis: a huge fraction of published methods stop running within a few years, and the cost of resurrecting them is days of dependency hell and niche debugging across compilers, drivers, and framework eras. This cost is high enough that people abandon and re-invent instead. Lazarus turns that days-long expert task into something anyone can trigger with a link and a pip install, and it does it honestly by reproducing the paper's numbers rather than just producing output that looks plausible. That's the difference between "the code ran" and "you can trust the result." For small, budget-constrained ML-for-biology teams, lowering that activation energy and keeping methods alive via upstream PRs is the whole point. - **Placement:** Finalist - **Team:** [Dean Sherry](https://cerebralvalley.ai/u/DoctorDean) - **GitHub:** https://github.com/DoctorDean/lazarus https://pypi.org/project/lazarus-bio/ - **Demo video:** https://youtu.be/zSC8miJpTtk - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=50 ### 75. Claudician I built Clinical Digital Twin, an open-source R system (Shiny/plotly dashboard, plumber API, SQLite, and a Telegram bot powered by Claude) that gives clinicians in elder-care institutions a personalized, interpretable model of each resident's fall risk. An engineered pooled-logistic-regression digital twin ingests simulated wearable data (heart rate, blood pressure, steps, sedentary time) and clinical history to predict fall probability at 24 hours and 7 days, and answers counterfactual "what-if" questions in real time (e.g., what happens to risk if mobility increases or a medication is deprescribed). I found that a single, transparent model can serve two very different interfaces (a visual dashboard and a natural-language Telegram bot) without sacrificing interpretability, and that grounding Claude's replies in real model output rather than free generation keeps the bot clinically trustworthy. We also surfaced a modeling caveat worth reporting: under our synthetic data-generating process, static risk factors are only identifiable through their effect on sensor trends, not directly; a reminder that interpretability claims need statistical scrutiny, not just a simple model. This matters because falls are a leading cause of injury death in older adults, and institutional fall-prevention often fails not from data scarcity but from data never reaching a clinician in time. - **Team:** [Daniel Gallardo](https://cerebralvalley.ai/u/dgalgom) - **GitHub:** https://github.com/dgalgom/clinical-digital-twin - **Demo video:** https://youtu.be/u8PsLidXqMY - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=51 ### 76. Past the Wall Orphan Finder: reliability-aware target prediction for natural products, drug repurposing, and antimicrobial resistance. Roughly half of all approved drugs come from natural products, yet for most natural products we still do not know which protein they act on. They are "orphan" molecules: demonstrably active, but with no known target, so they never advance. The standard way to predict a target is by chemical similarity, assuming a new molecule hits the same protein as a known drug it resembles. That approach fails silently for exactly the molecules that matter most, because orphan molecules, by definition, resemble nothing already characterized. We built Orphan Finder to measure and cross that boundary. Across 695,133 natural products, we quantified where similarity-based prediction collapses: below a Tanimoto similarity of 0.5, and 63% of natural-product space lies past that "analog wall." Orphan Finder always knows which side of the wall a prediction is on. Where a close analog exists it uses it; where none does, it stops guessing and switches to structure, docking the molecule into candidate protein pockets against real positive and negative controls. The entire pipeline is orchestrated with Claude Science, which assembles the compound libraries (COCONUT, ChEMBL, the Broad repurposing set), reads protein structures, dispatches GPU docking on Modal, and ranks the results with a calibrated confidence. Three findings show it works. First, an orphan natural product and a structurally unrelated repurposing-library compound (Tanimoto 0.28, no shared scaffold) both dock into the same pocket of M. tuberculosis InhA, the enzyme the frontline drug isoniazid targets, both engaging the catalytic Tyr158 and beating a matched decoy set, with an independent structure predictor confirming the poses. Second, because naming a target is not enough for an antibiotic, we added a mutation-aware layer that predicts whether binding survives resistance mutations; validated against known resistance, it then nominated seven orphan natural products predicted to hold their binding against DHFR precisely where the frontline drug fails. Third, the same engine generalizes to a second application: pointed at gut-microbe metabolites, it first re-discovered known microbe-to-host signaling (niacin to HCAR2, secondary bile acids to FXR and TGR5) as a positive control, then flagged a family of nine microbial carotenoid pigments all converging on the human xenobiotic receptor PXR, the master regulator of drug clearance, with every one docking within 1 kcal/mol of the canonical agonist rifampicin. Why it matters: antimicrobial resistance and undruggable natural products are two of the largest bottlenecks in finding new medicines. A target-prediction tool that reports its own reliability, and that asks whether a target will still be druggable after resistance emerges, turns a flood of orphan molecules into a ranked, structure-backed, resistance-aware set of testable hypotheses. Every result is a computational nomination for wet-lab follow-up, and everything is fully open and reproducible from the repository. - **Team:** [James Martin](https://cerebralvalley.ai/u/jmart24225) - **GitHub:** https://github.com/DocMartin95/analog-wall-target-prediction - **Demo video:** https://github.com/DocMartin95/analog-wall-target-prediction/raw/main/demo/Orphan_Finder_demo.mp4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=52 ### 77. OPUS-Bio OPUS-ET-AGENT: agent-conducted cryo-electron tomography Cryo-electron tomography (cryo-ET) is the only technique that resolves macromolecules at near-molecular detail inside intact cells. But turning raw tilt-series into in-cell 3D maps normally takes a week of hand-driving several specialist packages (WARP, AreTomo2, PyTOM, OPUS-ET, M), with expert judgment gating nearly every step — which is why in-cell structural biology remains labor-intensive and low-throughput. In this project, we put Claude Code in the driver's seat as a supervised-autonomy "conductor." It discovers the toolchain, configures and submits cluster jobs, runs QC at each stage, and then stops at each scientific checkpoint ("gate") to hand the decision to a human — backed by the evidence it has just computed. We defined five gates that guard processing quality: alignment QC, particle-picking QC, compositional-state selection, resolution, and joint refinement. On a real dataset (EMPIAR-10988), the agent drove two molecular species to high resolution at once: a ribosome to 7.76 Å and the far rarer fatty-acid synthase (FAS) to 13.88 Å. It did this by importing both into a single refinement population, so multi-particle refinement solves the shared tilt-series geometry using every particle from both species — the abundant ribosomes anchor the model while the sparse FAS rides along, gaining roughly 12 Å. Both maps were then placed back at every particle's pose inside the tomogram, reconstructing molecular sociology straight off the data. The key challenge: template matching is run for recall, so its picks are noisy and heavily over-picked. Claude Code leverages OPUS-ET's per-particle heterogeneity analysis to separate signal from junk, clustering each species' candidates into compositional states where the genuine high-resolution population stands out. For the sparse FAS, most of its 30 latent states are noise; only the clean D3 barrels — identifiable by their three-fold central pore — emerge as the real population. That per-particle separation of true species from a noisy template-matching set is what lets a rare complex resolve at all. Why it matters Operationally: a weeks-long, expert-only workflow becomes reproducible, test-covered (187 tests), and agent-driven. The agent explores, runs experiments, and presents the evidence, while the scientist keeps command of every call that matters. Along the way it even caught two silent, crash-free scientific bugs — a CTF pixel-size mis-scaling and a missing-wedge sign error — that all-green tests would never have surfaced. Biologically: placing every molecule back at its true position turns isolated structures into a map of who sits next to whom — molecular sociology. The crowding and neighbor relationships revealed here can never be seen in purified single-particle structures. By determining and interpreting how molecules pack in space, subtler in-cell mechanisms become discoverable. Where we're headed OPUS-ET already resolves each particle's conformational and compositional state. Next, by drawing on Claude Code's visualization abilities, we can integrate per-particle state directly into the in-cell map — labeling each placed molecule not just by position but by the functional state it occupies there. This would let OPUS-ET-AGENT deliver state-resolved molecular sociology: far richer biological context than structure or position alone, and a route to molecular-level mechanisms you can only see when you know both where a molecule is and what it's doing. - **Team:** [Zhenwei Luo](https://cerebralvalley.ai/u/alncat) - **GitHub:** https://github.com/alncat/opus-et-agent - **Demo video:** https://youtu.be/Kf_MkRYBOCo - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=53 ### 78. Sourabh Mosaic is an MCP server for preclinical target diligence. Our user — a comp-bio at a 10–30 person biotech, or a solo diligence consultant — gets asked "which of these five targets deserves the next six months of budget?" and answers it by hand: PubMed, ChEMBL, Lens, DepMap, ClinicalTrials.gov, three days to two weeks per target. The real problem isn't scattered data; it's that they deliver the verdict without being able to see their own blind spots. Open Targets shows data but takes no position; Causaly is enterprise-priced; a raw LLM invents confident answers with no provenance. Mosaic runs inside their existing Claude session, composes across sources via MCP without touching proprietary data, and returns a coverage-honest verdict — one that states confidence per dimension and explicitly flags where its coverage is too thin to judge. For a budget decision, a system that tells you where it's blind beats one that just sounds sure. - **Team:** [Sourabh Nyalkalkar](https://cerebralvalley.ai/u/sourabhn) - **GitHub:** https://github.com/sourabhnk/mosaic-mcp - **Demo video:** https://youtu.be/2WTIgS9viGQ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=54 ### 79. ai星星 What we investigated. We started from an asset rather than a disease: a genome-scale CD4⁺ T-cell Perturb-seq dataset (a CRISPR knockout screen read out in single cells). The question we set was inverted from the usual pipeline — instead of picking a disease and hunting for data, we asked which disease would let this perturbation asset reveal a high-impact drug target? We chose Sézary syndrome (SS), a leukemic cutaneous T-cell lymphoma driven by a circulating malignant CD4⁺ T cell, because it satisfied three criteria: public clinical single-cell transcriptomes are available, blood CD4⁺ T cells are central to the pathology, and the CD4⁺ T cell itself is the therapeutic target. What we built. On a 7 GB laptop, we integrated five public SS scRNA-seq datasets (≈180,000 cells) with Harmony, isolated the malignant Sézary T-cell population (63,127 cells) with a signature score, and defined the SS-specific expression profile — reproducing the known diagnostic hallmarks (loss of CD7; gain of KIR3DL2, TOX, GATA3). We then defined a reversal score: for each of ~34,000 knockouts we compared its transcriptional effect to the SS disease signature by weighted cosine similarity and flipped the sign, so a positive score means the knockout pushes cells from the disease state back toward healthy (connectivity-map logic). What we found. IL2RB (CD122) is the single largest disease-axis mover among 7,822 high-confidence knockouts — reversing the disease signature ~2× more than the targets of drugs already approved for SS (bexarotene/RXR, HDAC inhibitors). This target is invisible to standard analysis. The reversal score is essentially uncorrelated with a gene's own expression change in SS (Spearman ρ = +0.055) — so IL2RB would never surface from differential-expression or enrichment analysis. It only appears through functional perturbation. A route to drug it. IL2RB is a membrane protein with no small-molecule pocket, and existing IL-2-receptor antibodies act by cell depletion, not signal blockade. Because Perturb-seq measures loss-of-function, the faithful pharmacological equivalent is degradation. Screening degradation machinery in SS T cells, only RNF43/RNF149/RNF167 are expressed in CD4⁺ T cells; RNF149 stands out — expressed in 42% of malignant cells and ~8× restricted to blood/lymphoid tissue, offering potential T-cell selectivity and reduced off-target toxicity. Why it matters. SS has high unmet need and no IL2RB-targeted, degradation-based therapy in the competitive landscape. Our conclusion — degradation of IL2RB via RNF149 would be a promising therapeutic strategy for SS — is a reproducible, computationally derived hypothesis that demonstrates a generalizable discovery pattern: pairing a disease's clinical transcriptome with a functional perturbation screen surfaces targets that expression-based analysis alone cannot. - **Team:** [Seishin Sakai](https://cerebralvalley.ai/u/Seisei) - **GitHub:** https://github.com/seishin0125/BuiltWithClaude_LifeSceince_Hackathon/blob/main/research_writeup.md - **Demo video:** https://youtu.be/0poE8tlOu_g - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=55 ### 80. SALUS Zero SALUS Zero is a clinical training simulator and bedside decision companion for medicine where nothing is available. Most of the world's children are treated with no CT, no ultrasound, no surgeon — only a clock and one doctor. It serves that world through two doors. The scored night: Claude narrates a deteriorating pediatric case while code owns everything that must be true — the stage physiology, every vital and lab, and a 0–100 debrief computed in auditable arithmetic, never by the model. Its signature constraint: no CT — you decide with what the clinic actually has. The consult companion: a narrow, reference-grounded appendicitis assistant for a clinician facing a real child — it starts from your inventory, computes PAS and Alvarado in code, forces the mimics into every assessment, and refuses to prescribe doses. Appendicitis is only the pilot: any domain, surgical or medical, plugs in as expert-authored data. A chatbot is built to agree; this engine is built to push back. In its first days, six residents, an associate professor, and a district-hospital general practitioner ran it. Built solo in six days with Claude Code by a practicing pediatric surgery resident. Open source, AGPL-3.0, live at saluszero.com. - **Team:** [Sahin Parlak](https://cerebralvalley.ai/u/hannibal66) - **GitHub:** https://github.com/sahinparlak/salus-zero - **Demo video:** https://www.youtube.com/watch?v=nViBs7w1IeA - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=56 ### 81. moonicu NeoVUS is a point-of-care tool for NICU clinicians who hit a variant of uncertain significance (VUS) in a neonatal genomic report. Existing tools either rank variants across a whole exome (Exomiser) or serve genetics professionals (VarSome) — none give a neonatologist an actionable, auditable interpretation of ONE already-flagged variant. What I built: enter a variant as written on the report (c./p. HGVS, rsID, or coordinate) plus the baby's phenotypes, and NeoVUS returns candidate diseases re-ranked by phenotype match, a neonatal checklist of symptoms with cited work-up and follow-up actions, the variant mapped onto its protein domain (UniProt/AlphaFold), and a plain-language summary — where every claim links back to its source database. The report tracks its own citations and flags any unsupported claim, so the clinician verifies rather than trusts a black box. What I found: on ClinVar variants reclassified across the VUS boundary, NeoVUS's transparent in-silico evidence agreed with the eventual reclassification 84.8% of the time, at 100% pathogenic precision — validated on open data, with no ClinVar label used as a feature. Why it matters: rapid genome sequencing is reaching the NICU faster than neonatologists have tools to interpret it at the bedside, and NeoVUS puts a transparent, evidence-traceable first read in their hands. The tool is actively being upgraded: a calibrated, ACMG-aware v2 is already in development (live at https://neovus2.streamlit.app, code at https://github.com/moonicu/neovus2) — it adds a provisional point-based ACMG classification, calibrated PP3/BP4 thresholds, PVS1/PS1/PM5/PM1 evidence drawn from gnomAD gene constraint and same-residue ClinVar, and clinician-supplied family/functional evidence (e.g. a confirmed de novo) that re-classifies the variant live at the bedside. Open-source (MIT), reproducible, open public data only. - **Team:** [Moonyeon Oh](https://cerebralvalley.ai/u/moonicu) - **GitHub:** https://github.com/moonicu/neovus - **Demo video:** https://youtube.com/watch?v=DMBYsHMh2VA - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=57 ### 82. Sharab A Journals now require that Methods declare RRIDs for antibodies, authentication for cell lines, and animal-rigor reporting. Methods Verifier is the pre-submission check: paste your Methods + Key-Resources and get a per-resource verdict — PASS / FAIL / NEEDS-VERIFICATION — each cited to a public record. Claude does the two jobs a lookup can't. It extracts a typed resource inventory out of messy, hyphen-stripped prose, and it reasons about whether each antibody was validated with a genetic knockout control — the specificity gold standard. Deterministic gates then issue every identity verdict against ground truth a model cannot fabricate: cell lines against the ICLAC register, antibodies against the Antibody Registry, software against SciCrunch RRIDs. A missing datum → it abstains, cited. Every finding is tagged registry (deterministic) or model judgment (Claude), so the two are never blurred. Why it must exist: over the entire 594-line ICLAC register, a frontier model names a contaminated line from memory only 18% of the time — confidently wrong on 230. Give it the register and extraction catches 92%; on 47 real open-access papers, 89% with 0 false alarms on 14 clean controls, every FAIL cited to an ICLAC ID + CVCL. - **Team:** [Sharab Ahmad](https://cerebralvalley.ai/u/Moenv) - **GitHub:** https://github.com/Moe1117/verdict - **Demo video:** https://youtu.be/HQW0DZOr84U - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=58 ### 83. Ritz What I built. Verdikt is an AI agent that turns one input — a drug, gene, or aging-related target — into a transparent, confidence-scored decision brief. It reads across five live public biomedical databases (Open Targets, ClinicalTrials.gov, ChEMBL, PubMed, openFDA), then uses Claude as a rubric-driven judge to weigh the evidence, reconcile agreeing versus conflicting findings, and calibrate a 0–100 confidence score — with every claim traced back to its source. It's built with Claude Code as a Python/FastAPI engine and a single-page web app, and it runs live. What I found. Drug discovery doesn't fail because scientists lack data — it fails on which evidence to trust. A landmark analysis found only a small fraction of preclinical findings could be independently reproduced: evidence is abundant, but reliability is uneven. Building Verdikt showed the problem is tractable. By gathering the evidence deterministically first and having the model reason only over what it finds, the system produces honest, calibrated confidence — it stays low when the evidence is thin and surfaces genuine contradictions, instead of hallucinating certainty. Why it matters. A single go/no-go decision can cost billions of dollars and years of work, yet today it rests on weeks of manual, error-prone evidence review. Verdikt is the decision-intelligence layer between scientific evidence and billion-dollar investment decisions — making the evidence transparent, traceable, and honest, so scientists can trust the evidence, not the AI. - **Team:** [Ritu Chakrawarty](https://cerebralvalley.ai/u/RituChakrawarty) - **Demo video:** https://youtu.be/tlJJYXLS54Y - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=59 ### 84. BibimbapKiller What if a model could say the one word every drug-discovery pipeline needs but no other tool can utter — "unreachable"? Cell-state-reachability reframes perturbation biology around a new question: not "what will this gene knockout do?" but "can we actually steer a cell from state A to state B — and should you believe the answer?" We surveyed 91 perturbation-prediction methods; none return a feasibility verdict. Ours does. Built on public Perturb-seq screens (Replogle K562/RPE1, Norman combinatorial doubles), it produces a map of reachable cell-state space plus the falsification machinery that flags which regions you're allowed to trust: additivity bounds for multi-gene edits, weak-instrument confidence intervals, and negative-control and construct-validity checks. The payoff is targets a supervised ranker structurally can't nominate — like ZAP70: small-molecule-tractable, genetically associated, yet undrugged. The stance is AI-as-instrument, not oracle. The deliverable isn't a gene list; it's understanding of the data's structure earned inside a falsification loop — fitting for a field whose own 2025 benchmark shows deep models don't beat linear baselines. - **Team:** [Justin Kim](https://cerebralvalley.ai/u/MasalaKimchi) - **GitHub:** https://github.com/MasalaKimchi/cell-state-reachability - **Demo video:** https://www.youtube.com/watch?v=GJbxLxYUBMo - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=60 ### 85. Panoscope Panoscope is a grounded interpretation layer over jazzPanda's spatial-transcriptomics output. A wet-lab biologist opens a cluster and asks in plain language, for example "why is this still Stromal when COL1A1 is missing?", and Panoscope answers with a cell-type call, a confidence band, the driving markers with their real jazzPanda numbers, and a live-fetched PubMed citation for the biology. It never invents a marker, a number, or a reference. Two design choices make it trustworthy. A confident floor: a deterministic engine computes every confidence band from jazzPanda's spatial signal and enforces the panel-absence rule, so a canonical marker that was never on the 280-gene panel is not treated as evidence against a cell type. An open ceiling: literature reasoning runs live through Claude behind a grounding gate that rejects any number or citation it cannot trace to source, and a test suite fails the build if the agent states anything it cannot cite. When a biologist overrides a call, Panoscope keeps their decision and cross-checks the literature, saving a versioned, git-tracked lab note with both the agreement and the dissent visible. The same discipline extends past marker calls. A second workflow interprets gene-set enrichment on the panel, carrying a mandatory panel-coverage caveat (for example, only 8 of a 200-gene program are actually measured) that deterministically weakens the confidence when a "pathway" really rests on a handful of immune genes bleeding in from neighbouring cells. Each analysis method is a self-contained skill file, so supporting a new bioinformatics tool means writing a new grounded interpretation contract, not a new app. Why it matters: annotation confidence is exactly where biologists get stuck, and LLMs are exactly where fabrication creeps in. Panoscope is a reusable pattern for Claude Science: put Claude on top of any real bioinformatics output and make it structurally unable to state a fact it cannot cite. Every one of the 223 citations it ships resolves to a real paper. - **Team:** [Xinyi Jin](https://cerebralvalley.ai/u/melody_jin) - **GitHub:** https://github.com/MelodyJIN-Y/Panoscope - **Demo video:** https://youtu.be/djmImgY6RU8 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=61 ### 86. The CD4 Perturbation Predictability Audit Team Everyone builds bigger models to predict what a gene knockdown does. I asked the prior question: how much is even predictable? Using a genome-scale CRISPRi Perturb-seq screen in primary human CD4⁺ T cells (~22M cells, 4 donors), I ran seven pre-registered investigations, each gated on CPU to fail cheaply before spending GPU. Six failed — and the failures are the finding. Under honest, reliability-ceiling-calibrated measurement, the per-perturbation signal sits at the noise floor: only ~17% of knockdowns have an effect you can trust (SNR > 3), and reaching reliability would need ~12× more sequencing depth. The one method that worked — a causal do-operator (C2 +0.118/+0.162) — beats its twin only within-distribution; on external causal data it's correlational, not causal. I wrapped all seven probes into a reliability-ceiling-calibrated predictability audit: a model-agnostic pre-flight scorecard, anchored by a positive control that proves it detects signal when present. Why it matters: perturbation models exist to prioritize expensive bench experiments. This audit tells a lab what its data can and can't support — before a single experiment runs. - **Team:** [Lukas Weidener](https://cerebralvalley.ai/u/LSW) - **GitHub:** https://github.com/VibeCodingScientist/CD4-Perturbation-Predictability-Audit - **Demo video:** https://youtu.be/AO0A4UhhMZE - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=62 ### 87. HISTORA HISTORA is a non-diagnostic, agentic clinical-research navigator built with Claude Code for a clinical/translational researcher. Testing a hypothesis like "does periodontal inflammation share mechanisms with cardiometabolic disease?" normally means weeks of chart review — and often the data isn't even there. HISTORA turns a fragmented public corpus (NHANES) into a research-ready cohort in seconds: an eligibility funnel (20,905 → 442), per-field completeness, and — the differentiator — it states what the data cannot answer (a missing datum is a collection flag, never imputed). It grounds one falsifiable, non-diagnostic hypothesis in genetics it runs live: IL-6R → coronary disease is genetically causal (replicated across FinnGen and BioBank Japan), circulating CRP is not, and the Alzheimer's axis stays exploratory. Operating its own pipeline under a reviewer agent, it caught, retracted, and fixed a bug in its own flagship number (+0.705 → +0.553), with a regression test. It's reproducible, tested, Apache-licensed software a lab can run without us in the room. Why it matters: the mouth is a cheap, modifiable window into the inflammation behind heart disease and diabetes — HISTORA is the instrument to study that link honestly. - **Team:** [Matias Molinas](https://cerebralvalley.ai/u/matiasmolinas) - **GitHub:** https://github.com/matiasmolinas/dental-analysis - **Demo video:** https://youtu.be/qNmgm4Bs4Dw - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=63 ### 88. Xiao Background & motivation Mesothelioma is an aggressive, asbestos-linked cancer of the pleura with poor prognosis and few effective therapies. It is particularly important to study in Australia, which has one of the highest incidences of the disease in the world, driven by its historical use of asbestos, while only hundreds of public omics datasets now profile it across modalities and this information is scattered and rarely integrated. The goal is to use Claude to automate retrieval and integration of public multi-omics datasets (genomics, transcriptomics, epigenetics), then systematically identify survival-associated features that are reproducible across cohorts, biologically meaningful, and have potential for clinical translation — such as cancer-vaccine target discovery. Project description I developed a new feature-selection workflow. A reliability-first method for high-dimensional, small-sample omics survival data: repeated event-stratified split screening + epistasis-hub detection + bootstrap stability LASSO-Cox, combined so that only features reproducing across resamples survive (released as the reusable omicsfs library). It was applied to mesothelioma on the MESOMICS cohort (120 patients × ~25,000 multi-omic features: expression, DNA methylation, LOH, CNV, driver alterations. I benchmarked it against commonly used methods. Head-to-head against SIS, LASSO-Cox, Random Survival Forest and XGBoost, and DeepSurv under one common evaluation: cross-cohort survival transfer to external cohorts, KEGG pathway enrichment, single-cell cell-type expression, and a curated-literature check. Panels of prognostic features for mesothelioma were selected to take forward. The workflow yields reproducible, cross-cohort-validated biomarker panels — a shortlist of candidate features for further experimental validation. What found. The in-house consensus panel generalizes best across cohorts (expression-surrogate C-index ~0.67; native multi-omics transfer to TCGA ~0.72), edging established methods — though at n=120 the bootstrap confidence intervals overlap, so this is competitive and best-generalizing. A key methodological result: internal cross-validation is optimistic. Biologically, the prognostic signal sits largely outside canonical driver pathways, and several top targets are immune-compartment genes (T-cell, mast, neutrophil, dendritic). Why it matters. Methodologically, this is a template for a defensible biomarker benchmark — verified comparators only, honest confidence intervals, generalization prized over internal fit. Translationally, it delivers a short, druggable, immuno-oncology-relevant target shortlist for Mesothelioma, an asbestos-linked cancer with few effective therapies and growing use of immunotherapy. Methodologically, it can be used to select features for other cancers. - **Team:** [Xiao Zhong](https://cerebralvalley.ai/u/xiaozhong) - **GitHub:** https://github.com/Xiao-Zhong/multi-omics-feature-selection - **Demo video:** https://www.youtube.com/watch?v=z_EsnO1XFAo - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=64 ### 89. Wayfinder Wayfinder is a hypothesis referee. Literature-based discovery (LBD) overproduces hypotheses and tests almost none — the critique that has dogged the field for decades. Wayfinder answers it: it pairs LBD generation with a deterministic referee that tests each machine-generated gene→program→disease hypothesis against a genome-scale CRISPRi Perturb-seq resource, and returns a verdict with a receipt at every hop — including a confident, receipt-backed "no." A failed knockdown comes back untested, not a false negative: an artifact caught, not a discovery missed. The whole loop ran inside Claude Science, whose independent reviewer model flagged and cut the project's own overclaims. On the Marson lab's CD4+ T-cell Perturb-seq data, 22,039 hypotheses were culled to a small set of receipt-backed candidates; the standout — NAB2→Th1/Th2→atopic eczema — is a connection the literature had never drawn (the eczema label is a genetic-association nomination). Measured as a predictor on held-out post-2016 links, the method returns a null at its primary metric — reported straight. Why it matters: it makes machine-generated hypotheses triage-able and trustworthy — every claim carries an experimental receipt, and the system is willing to abstain and to refute. - **Team:** [Dayanjan Wijesinghe](https://cerebralvalley.ai/u/Dayanjan) - **GitHub:** https://github.com/dayanjan/wayfinder - **Demo video:** https://www.youtube.com/watch?v=MbgojaAFfz0 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=65 ### 90. Ken Kamiya morphoDE asks a deliberately simple question: where two cell types meet, the boundary bulges outward in some places and caves inward in others—do the cells on a bulge express different genes than the cells in a dent? We tested it across 11 published 3D spatial-transcriptomics datasets and 126 interface–strata, from a human lymph-node tumor to an Arabidopsis leaf. The correspondence holds almost everywhere: 125 of 126 interface–strata carry signal that survives a geometry-breaking null (which shuffles the shape while keeping the biology fixed), signal appears in all 11 datasets, and 16 reach a perfect 12/12 in species sharing no common ancestry. On the human tumor the difference is 93% within a single cell type—convex bulges carry immune and cholesterol programs, concave dents carry hypoxia and invasion (PTHLH, KRT17). Why it matters: interface shape and cell state are linked as a general property of tissue architecture, not a tumor quirk—giving a geometry-readable spatial map of cell state, from cancer to plants. - **Team:** [ken kamiya](https://cerebralvalley.ai/u/kkami1115) - **GitHub:** https://github.com/kkami1115/morphoDE - **Demo video:** https://youtu.be/h1vk-xAkmjY - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=66 ### 91. preCaution Early in a routine lab protocol, piranha acid is prepared, used, and its spent waste poured into an acid waste carboy. A step later, someone rinses sodium azide buffer glassware into that same carboy. The two are never combined in any single instruction. They meet in the vessel, because the waste goes to a shared container, and the azide reacting with the carried-over acid releases hydrazoic acid, which is toxic and potentially explosive. No ingredient-screening tool catches this, because catching it means reading the procedure as a sequence of vessels and events, not a list of chemicals. That is what preCaution does. It reads a scientist's free-text protocol and produces a short, sourced safety brief for that specific procedure, not a stack of data sheets to wade through. It flags hazards that come from the procedure itself: reagents combined in one step, or waste streams that quietly converge across steps through a shared vessel. It also flags where a step leaves something its chemicals' own safety data calls for unstated, like ventilation or waste handling, as a tentative prompt to confirm rather than a verdict. Every hazard claim traces to a live lookup, PubChem for GHS hazard data and a NOAA CAMEO-sourced table for reactive-group interactions, never a model's memory. The brief itself is composed by code with zero model calls, so there is no generation step where an ungrounded claim could slip in. The single claim that rests on Claude's reading of the protocol, that the azide enters that specific carboy, is marked unverified on screen. It is the only claim marked that way, so the marker carries weight. What I found building it: the safety data that exists in structured, queryable form is calibrated for industrial and hazmat chemistry and thins out fast for biological reagents. preCaution is honest about that boundary in its own output rather than papering over it. The person this is written for is usually a newcomer. In its first case study of an academic laboratory, following the 2010 Texas Tech University explosion that seriously injured a graduate student, the U.S. Chemical Safety Board found systemic failures in hazard evaluation and in oversight by the principal investigators and the institution: a critical exposure limit existed only as a verbal instruction, and none of the students the Board interviewed knew it. Those are precisely the gaps a newcomer is least equipped to catch alone. Existing lab-safety tools are systems of record for the institution. preCaution is a reasoning tool for the person at the bench, about to run a protocol they did not write. - **Team:** [Yuvraj Gupta](https://cerebralvalley.ai/u/theyuvrajgupta) - **GitHub:** https://github.com/theyuvrajgupta/PreCaution - **Demo video:** https://www.youtube.com/watch?v=D09dpqCn_s0 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=67 ### 92. Alex Maier We tested a timely scientific question: Does the brain (neocortex) fundamentally compute predictions (so-called predictive coding)? Specifically, do neurons signal every kind of surprise with *one* mechanism or *many*? We made use of open science (a free Allen Institute OpenScope dataset from a public archive). - **Team:** [Alex Maier](https://cerebralvalley.ai/u/alexmaier) - **GitHub:** https://github.com/maierav/ai_oscp_neuro - **Demo video:** https://youtu.be/72kfgGo-8uM - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=68 ### 93. Team Coincidence Proteins are spelled in a twenty-letter alphabet, and every one of those letters is an ordinary Roman letter. So English words — HEALTH, ELVIS, VEGAN — fall out of human protein sequences by pure chance. The question isn't whether they exist. It's whether they land anywhere that matters. **The finding.** This is not an absence of a result. Words *avoid* the parts of a protein that matter, in a specific and measurable direction, and the cause is letter frequency. Right on a protein's active site — the tiny pocket where it does its chemistry — you find only 82% of the words chance would put there. Across 10,000 composition-preserving shuffles of the reviewed human proteome, not one placed as few words on active and binding residues as reality does: a depletion of −7.02σ, holding across all 27 languages tested. The cause is compositional and unmystical: functional sites are built from cysteine, histidine and aspartate; common words from A, L, S and E. Same alphabet, opposite frequencies — so the stretches that spell words and the stretches that do chemistry are drawn from nearly disjoint pools. Words also show no thematic match to their host genes: HEART sits in a ubiquitin hydrolase, LIVER in a vesicle protein. The pipeline goes further than the headline. Words are traced through Ensembl orthologs across a billion years — six survive intact all the way back to yeast, sitting in proteins the cell cannot live without. They're mapped to their genes' true GRCh38 coordinates (the gene-rich X spells 95 words; the gene-poor Y, just 8). And they're tracked through ClinVar variants and splice isoforms, where a documented disease mutation writes a word the reference proteome never spells. **You can walk it yourself via a standalone developed game.** The explorer is a single offline HTML file: type your name, or any word, and travel the protein it hides in — residue by residue through the real sequence, with the functional sites marked. Type SLUG and you land on the catalytic selenocysteine of glutathione peroxidase, the very atom the enzyme uses to protect your cells from damage. Type CARAVAN and you're in a proteasome subunit, unchanged for a billion years, back to yeast. Type JACOB and you learn it can never exist in any protein — J and O aren't amino acids. Five of our letters aren't, and between them they rule out 53% of English. **Every objection is tested, not asserted.** Counts are compared against composition-preserving permutations run within each protein, so uneven annotation cannot manufacture the signal; the depletion survives restriction to structurally ordered residues (AlphaFold pLDDT ≥ 70); and featured words are ranked by a hashed, tamper-checked scoring rule, so nothing can be quietly tuned after the fact. The full threats-to-validity analysis is in the manuscript. **Why it matters.** First, it's a null result reported plainly, with every threat to validity tested rather than buried — confirming that a coincidence really is a coincidence, rigorously, is a finding worth publishing. Second, it's useful: a lexicon of strings that look like biological signal and provably carry none is a calibrated negative control for short-linear-motif discovery, a field chronically vulnerable to false positives. Third, it's an educational tool: sequence, function and chance become something a non-scientist can hold in ten seconds. Genetics rarely offers a handle this tangible — a word, sitting on a binding groove, in a protein that is really inside you. Everything is open — pipeline, result tables, manuscript, and the explorer. Reproducible from public data with fixed seeds; every figure regenerated from tables. - **Team:** [Thomas Stranzl](https://cerebralvalley.ai/u/Tomsn) - **GitHub:** https://github.com/tstdk/words-in-proteome - **Demo video:** https://youtu.be/5utw-nm4AyY - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=69 ### 94. CBMR Multimodal **What we investigated.** ML leaderboards present ranked models with narrow error bars, which reads as evidence of predictive strength. But a February 2026 audit of the Therapeutics Data Commons (TDC) ADMET benchmark (Receptor.AI, bioRxiv 2026.02.26.708193) found that most top-ranked entries fail on unavailable code, non-reproducible environments, or data leakage, and that deliberately overfitting a model to the open test set moves it up the rankings. On a benchmark with a fully open test set, climbing and cheating can be the same act. We asked: can an AI agent iteratively improve ADMET models *honestly*, and what does honest iteration reveal about the benchmark itself? **What we built.** An agentic pipeline, driven by Claude Code as a research agent, that does two things in one loop. It **climbs**: an agent proposes model designs, trains them, and selects among them to improve validation performance. And it **audits**: the same loop checks every model and data split for leakage and reports its own results with honest uncertainty. Honesty is enforced structurally, not by good intentions. Four roles are separated so the component that touches the test set cannot influence the one that designs models: a proposer (designs models on validation only), an executor (trains and scores on validation), a deterministic arbiter (promotes only on non-overlapping confidence intervals; keeps simpler models on ties), and a walled auditor (the only component allowed to touch test-set molecules, for leakage checks, and it returns only a summary verdict with no labels). The test set is reachable only by the auditor and by a single final-evaluation step, run once after a model is locked. The iteration loop has no code path to the test set. The human is inside the loop but subject to the same accounting as the agent: repeated testing widens the reported uncertainty, and the tool refuses high-complexity model tiers on endpoints too small to support them. **What we found**, across 10 of the 22 ADMET endpoints: - Honest iteration climbs where there is signal and plateaus where there is not. In 3 of 10 endpoints models measurably separated, all the same physically motivated move (physicochemical descriptors beating substructural fingerprints on descriptor-driven properties). In the other 7 the models were statistically indistinguishable, and the loop reported that instead of manufacturing a winner. - On small endpoints, the leaderboard ranks noise. The test-set bootstrap interval on small imbalanced endpoints is wide enough to span most of the leaderboard, so the fine-grained rankings there are not supported by the sampling error. - A documented leakage bug drifted silently across data versions (patched on two endpoints, still live on a third), and a two-axis leakage check proved necessary: on one endpoint, real duplicates that structural similarity missed were caught only by a desalted-identity check, while on another, a structural false positive was correctly cleared by the identity axis. **Why it matters.** Our contribution is not a better ADMET model. It is a demonstration that an AI agent can do the iterative modeling work without the test-set overfitting that compromises the benchmark, and can audit the benchmark's integrity as it goes, producing results honest about what the data can and cannot support. The same pipeline, run as new models are submitted, is a standing integrity check on a leaderboard rather than another entry gaming it. The full pipeline, the 10-endpoint findings, and the cached data reproduce from a fresh clone on a standard laptop. - **Team:** [Dewei Hu](https://cerebralvalley.ai/u/deweihu) - **GitHub:** https://github.com/deweihu96/admet-honest-audit - **Demo video:** https://www.youtube.com/watch?v=BN4r4DUT9cg - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=70 ### 95. Han Wei Tan Most "AI made a discovery" claims collapse the moment someone asks for a control. I built the opposite: a self-falsifying engine that predicts the missing members of protein complexes from CRISPR genetics and refuses to trust itself. The engine mines co-essentiality across two independent genome-wide CRISPR screens (Broad DepMap, 1,100 cell lines; Sanger Project Score, 324), then forces every prediction through independent physical-interaction evidence (BioPlex, hu.MAP, CFdb) and a confound firewall that rejects genetic-context artifacts. Benchmarked honestly: leave-one-out recall@10 is about 28% (roughly 475x chance), AUROC about 0.69 in both screens, and it independently re-derives dark-gene functions only published in 2022 to 2025. Why it matters, shown in a live test: I asked Claude Opus 4.8 with no rules for a protein's complex partners. It confidently returned ten, and its built-in reviewer approved all ten. Under the engine's rules, only one clears every check. The honest breakdown is four clear false positives, five real members held back as unproven, and one you can fully trust. Same model in both halves, so the rules, not a smarter AI, earn the trust. On Gladstone's Krogan head and neck cancer interactome, a hand-seeded KEAP1-CUL3-NRF2 pathway recovers UBE2M as a positive control (known biology, labeled as such; the automated pipeline's top CUL3 hit is ARIH1, and I say so). I then reproduced the work in Claude Science, whose independent reviewer caught me calling a correlation "replicated" when it sat below my own 0.3 threshold. The safety net fired on my own analysis. It is a reusable rigor layer, not a one-off result: a scientist can run `python scripts/complete.py --seed BRCA1,BRCA2,PALB2` and get BARD1 (BRCA1's obligate partner) as the sole trustworthy hit, with every candidate labeled PASS, WEAK, or NO-PHYSICAL. The repo also ships an honest map of ten attempts, most of which collapsed under proper controls, because a negative established rigorously beats a positive established loosely. Built by Claude Code, audited by Claude Science. Not a hype reel. A trust layer. - **Team:** [Han Wei Tan](https://cerebralvalley.ai/u/hanwei) - **GitHub:** https://github.com/tanhanwei/science - **Demo video:** https://youtu.be/KwLw49yvooQ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=71 ### 96. The Gene Team PlaqueQTL builds a cell-type-resolved eQTL resource for atherosclerotic plaque entirely from public single-cell RNA-seq — no new sequencing, no genotyping array. The key move: donor genotypes are called directly from the RNA reads of published plaque single-cell/single-nucleus datasets using Monopogen, recovering 3,812,318 high-quality variants across 57 donors and all 22 autosomes with zero missingness. Using these de-novo genotypes, we mapped cell-type-specific cis-eQTLs genome-wide (16,902 genes; 118,286 gene×cell-type tests) in diseased human plaque. The method validates cleanly: across 9 cell types it recovers established GTEx artery eGenes in the expected cell type. It then does what bulk tissue cannot — it localizes the known coronary-artery-disease gene GUCY1A1 specifically to the fibroblast compartment (rs7671138, p=8.9×10⁻⁸), and it reveals a distinct class of regulation exemplified by CFDP1: a gene expressed across every cell type whose genetic effect fires only in the smooth-muscle-derived mesenchymal lineage — cell-type specificity invisible to bulk artery. Why it matters: GWAS has found hundreds of CAD loci but rarely resolves the gene and the cell type through which a variant acts, and the standard vascular reference (GTEx) is bulk, non-diseased tissue. PlaqueQTL shows that the thousands of public single-cell datasets are themselves a latent statistical-genetics resource — every one carries recoverable genotypes — offering a low-cost, high-yield route to disease-context, cell-type-resolved regulatory maps. - **Team:** [James Hilser](https://cerebralvalley.ai/u/JRH) - **GitHub:** https://github.com/thegeneteam/PlaqueQTL - **Demo video:** https://www.youtube.com/watch?v=_d9BmmTHEOg - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=72 ### 97. leegun OnLabel is a self-verifying OTC medication-safety assistant for US consumers. Ask a chatbot "Can I take Tylenol and DayQuil together?" and it's often right on average — but for a medication-safety decision, being right on average isn't enough: it buries the answer in hedged prose, makes you do the math, and recalls numbers from memory that shift when you rephrase. OnLabel separates language from judgment. Claude writes the answer and extracts the products the user named — but Claude never makes the safety call. Every product resolves to its FDA-grounded active ingredients, and a pure-code verifier builds an ingredient ledger to check active-ingredient duplication (the hidden acetaminophen in Tylenol + DayQuil), cumulative dose vs. the FDA daily maximum, and drug-class overlap. The verdict is composition plus arithmetic — reproducible every time, and every number traces to an FDA label line you can audit. Why it matters: acetaminophen overdose is the leading cause of acute liver failure in the US, and the overlaps that cause it are exactly what generic AI answers miss. Built by a pharmacist who sees patients arrive with wrong AI advice daily. - **Team:** [gun lee](https://cerebralvalley.ai/u/dlrjsdl200) - **GitHub:** https://github.com/dlrjsdl200-byte/onlabel - **Demo video:** https://www.youtube.com/watch?v=JXQKidfVldw - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=73 ### 98. Faith Ogundimu Diffuse midline glioma is a paediatric brainstem cancer that is almost uniformly fatal: about ninety percent of children die within two years, and the first systemic drug, approved in 2025, does not cure. Its biology is largely non-coding, which is exactly where variant prediction is least trustworthy. NCypher is a triage layer, not another predictor. In the matched fetal-OPC cell context it couples a regulatory-activity score, the mechanism (the motif and the exact base that breaks), and an honest confidence flag into one agent call, promoting a variant only when independent lines of evidence agree. Across 10,869 somatic non-coding variants from an H3 K27M cohort it returns a two-sided finding, not a manufactured driver: somatic mutation does not drive the super-enhancers this tumour depends on, yet those enhancers sit on markedly more conserved OPC regulatory sequence (matched-permutation p<0.001), with NPAS3 leading the shortlist. The engine is validated where it should work (progenitor caQTLs, AUROC 0.69) and honest where it should not. Six independent stress tests, each re-derived by a separate reviewer agent, harden it. NCypher ships as a rich MCP tool, an Agent Skill with a built-in sceptic, a reproducibility gate, and a downloadable regulatory map. - **Placement:** Finalist - **Team:** [Faith Ogundimu](https://cerebralvalley.ai/u/faith-ogun) - **GitHub:** https://github.com/faith-ogun/ncypher - **Demo video:** https://youtu.be/0EQfVwYXFG8 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=74 ### 99. Stanza I built Stanza to chase a frustrating question: when a resistance mutation kills a cancer drug, can you get to a redesigned molecule in minutes instead of years? You give it a target and a mutation, or you just hand it a research paper. Claude reads the whole PDF and pulls out the design brief, and the part I care about most is that it shows you the exact sentence behind every field and leaves things blank instead of making them up. On an EGFR resistance paper it correctly worked out that the warhead shouldn't hit the mutated residue at all, it should hit a different cysteine deeper in the pocket. That is a genuinely non-obvious call, and watching Claude reason its way there in real time was the moment the whole thing clicked for me. From there Stanza builds the wild-type and mutant structures, has Claude design covalent molecules for that pocket, docks them, ranks them, and feeds the results back so each round is smarter than the last. What I did not expect was how much of the work turned out to be catching myself being wrong. A selectivity number that looked impressive was an artifact. A confidence score disagreed with the harder structural check on four molecules out of five. I found four separate places where my own headline numbers were misleading, and I fixed or surfaced every one. So Stanza does not pretend to be an oracle. It is a triage filter that shows its work, puts an error bar on every score, and says "not resolved" when a result sits inside the noise. To check the engine wasn't fooling itself, I ran it on a case with a known answer, BCR-ABL T315I, and it scored the defeated drug as defeated and the surviving drug as surviving. Why it matters: resistance quietly retires real, approved drugs, and the first pass of designing around it is slow and expensive. Stanza compresses that pass into minutes. And in a field where one confident hallucination can send a lab down a dead end for months, a tool that is honest about what it knows and what it doesn't is worth more than one that just sounds sure. - **Team:** [Ayush Kumar](https://cerebralvalley.ai/u/ayush00git) - **GitHub:** https://github.com/ayush00git/stanza - **Demo video:** https://www.youtube.com/watch?v=o6uQK9lP_V4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=75 ### 100. Michael Rothrock What I built: I built BioSurface Audit, a deterministic tool that verifies the claims in an AI‑assisted single‑cell / Perturb‑seq analysis. It takes an AI assistant's prose write‑up, turns it into explicit structured claims, and checks each one across several independent verification surfaces. The model drafts the claims and suggests missing evidence, but a deterministic engine decides what the data supports. This is a clear trust boundary. Each claim ends up with a reproducible status tied to the exact evidence behind it, plus two summary numbers: Trust Coverage (how much of the required evidence was actually checked) and Corroboration (the share of checkable claims the independent surfaces agree on). What I found: I ran it on the published Marson data and a second Replogle K562 dataset to check that it generalizes. It catches the kinds of mistakes an AI assistant tends to make: a reversed fold‑change direction, or a target "nominated" from expression alone with no pathway or network support. In a blind test it caught all five errors I planted, each tied to the exact contradicting row, e.g. a claim that RASA2 knockdown lowers CXCL8 when the data shows it going up. The output is byte‑for‑byte reproducible, so it can gate a CI pipeline. Why it matters: The lesson we are learning in software development is that LLMs generate far more than human reviewers can digest. We need to augment them with deterministic tests that allow us to make guarantees about output, paired with stochastic evaluation that suggests where our limited human attention is best spent. For this domain, my starting point is that trust isn't any one engine's verdict, it's where independent lines of evidence converge. BioSurface makes that triangulation mechanical: it reports what it can't check rather than guessing, and every result is reproducible. It's the single‑cell instance of the Verification Surface framework I've been developing (https://michael.roth.rocks/research), so the same idea extends well beyond biology. - **Team:** [Michael Rothrock](https://cerebralvalley.ai/u/mrothroc) - **GitHub:** https://github.com/mrothroc/biosurface - **Demo video:** https://youtu.be/gDmDqymTXY4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=76 ### 101. Richard Zhu What I investigated: I used the EvE Bio "pharmome" — a single-laboratory dataset measuring 1,397 approved drugs against 256 human targets (kinases, GPCRs, nuclear receptors) under one consistent framework — to ask two questions heterogeneous public data cannot answer cleanly. First, do modern computational methods (the Boltz-2 affinity predictor and molecular docking) recover the comparative quantities that drive drug-discovery decisions? Second, mined as a screen, does the dataset reveal credible drug-repurposing hypotheses? What I found: Boltz-2 (a state of the art AI model for predicting binding affinity) recovers not just potency but genuine cross-target selectivity (ρ = 0.78 after removing drug and target baselines), and partially recovers resistance-mutation effects. In other words, Boltz-2 is able to figure out, out of a set of proteins, which will actually selectively be bound by a particular drug, as well as how this binding changes when mutations occur in the protein (a very common scenario in cancer that eventually leads to drug mutation). I find that Boltz-2 successfully predicts changes in binding affinity for mutations that physically reshape the drug-binding pocket, failing for allosteric ones. A memorization control showed this resistance signal is the least explained by training-data overlap. Mining the EvE Bio dataset also revealed two exposure-validated, docking-supported repurposing hypotheses that extend to whole drug classes: HIV protease inhibitors like saquinavir binding to the metabolic enzyme LRH-1 to explore metabolic effects of these drugs, and BTK inhibitors like zanubrutinib binding to the immune protein LTB4R. Moreover, we present an additional drug-specific repurposing hypothesis of the drug nintedanib being used against the drug-resistant PDGFRα-D842V mutant (common in many gastrointestinal stromal tumors). Why it matters: Systematically collected, internally consistent drug-target binding datasets like the EvE Bio pharmome can effectively act as computational benchmarks to help us test biological AI models on key drug discovery questions like selectivity and mutation effect prediction, as well as help us develop novel drug repurposing hypotheses. - **Team:** [Richard Zhu](https://cerebralvalley.ai/u/rzhu) - **GitHub:** https://drive.google.com/file/d/1DoD1_PmfH6m0xiP-tqjPNXh7L231qX9V/view?usp=sharing - **Demo video:** https://drive.google.com/file/d/1ra17whuqbH6suVVOwfgvs3NgUyH7SmDC/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=77 ### 102. Sabih We mapped the causal control architecture of a human immune cell and used it to nominate druggable, disease-relevant targets. From the Marson-lab genome-scale CRISPRi Perturb-seq atlas (22M primary human CD4+ T cells; rest / 8h / 48h activation) we built the causal trans-regulatory network for every validated knockdown -> differentially-expressed gene is one edge, and characterised its architecture. Finding: the network is hub-dominated and sparse-but-pleiotropic (the top 5% of regulators drive ~78% of all trans-effects). The first causal confirmation of the Pritchard-lab 2026 topology prediction. Across activation, it is shape-invariant but identity-labile; the concentration (Gini ~ 0.92) is pinned across all states while up to 59% of the top hubs are replaced. The cell keeps the shape of its control while swapping out who is in control. Every result survives power, edge-definition, knockdown-efficiency and detectability confounds, and replicates in a different cell type (K562). Why it matters: the hubs that switch on specifically with activation are ~2x enriched for monogenic-disease genes and are druggable (ZAP70, ITK, LCK, PTPRC, IL12RB2). The candidate state-specific control points for immune modulation. All results are reproduced from public data with code, a streaming notebook, and figures on GitHub. - **Team:** [Syed Sabih ur Rehman](https://cerebralvalley.ai/u/Sabih) - **GitHub:** https://github.com/Sneakypeat/HumanCD4CoDEGNet - **Demo video:** https://youtu.be/unAUdmUt2Zg - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=78 ### 103. Team Negaverse Negaverse is a Python toolkit for generating biologically informed negative datasets for protein–protein interaction models. The same approach can later extend to protein–ligand, protein–RNA, and protein–DNA tasks. Most datasets create negatives by randomly pairing proteins. The problem is that some of these pairs may be real but undiscovered interactions, which introduces false labels and teaches models the wrong biology. Negaverse reduces that risk through a staged pipeline. It first removes known interactions using BioGRID, IntAct, and high-confidence STRING data. It then scores the remaining pairs using network topology and biological and biophysical rules. Only the most uncertain cases are sent to an LLM for literature-based reasoning. Each result includes confidence, difficulty, reasoning, and full provenance. To test the system, we hid 1,000 real interactions inside the candidate pool. Naive hard-negative mining selected roughly three out of four of them, while Negaverse’s default configuration selected almost none. Across HuRI and DRYAD, Negaverse reduced hidden false negatives by about 10× compared with random sampling. Every run also produces an interactive report explaining why each pair was selected. A D-SCRIPT model trained on Negaverse negatives already generalizes better to unseen proteins than one trained on random negatives. The result is more credible and more challenging training data for drug discovery, with fewer false negatives and less wasted experimental validation. - **Team:** [Igor Gerasimov](https://cerebralvalley.ai/u/geravant), [Bailin Zhang](https://cerebralvalley.ai/u/lucyzhang1116) - **GitHub:** https://github.com/Geravant/negaverse - **Demo video:** https://youtu.be/4RHxGCo-5es - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=79 ### 104. Throughline This is Throughline, a workbench built for patients and families suffering from a rare or chronic illness. For chronically ill individuals, the vast majority of time is spent sifting through clinical documents and biological information, trying to find where to go next. This information is often scattered and researching as an individual can be extremely challenging. Throughline compiles all of a patient's information in one place and facilitates the research process for the patient themselves. This information can then be easily translated to doctors during appointments and empower patients to understand more fully what next steps can be taken. This tool is not meant to replace clinical advice, instead it provides an avenue for the patient to play a greater role in advocating for themselves and their treatment. - **Team:** [Linus Hansen](https://cerebralvalley.ai/u/linushansen) - **GitHub:** https://github.com/linushansen12/Chronic-Illness-Workbench/tree/main - **Demo video:** https://youtu.be/5pHpuJCqqgk - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=80 ### 105. Paul Gueguen SpatialScribe is a self-serve copilot for imaging-based spatial transcriptomics (Xenium, CosMx, MERSCOPE, and whole-transcriptome Atera). A wet-lab scientist points it at a raw run folder and gets back a QC'd, annotated, spatially-explored tissue map plus a re-runnable report - in plain language, with no terminal, no R, and no waiting for a bioinformatician. It's a React + FastAPI + deck.gl app over a single shared analysis engine of 40 capabilities, drivable four ways: guided rails, one click ("Run full analysis"), one headless command, or a grounded Claude copilot that runs the real analysis and can recolour the map, subcluster a population, or load a whole new section from plain English. The thesis is honesty. Most tools label every cell. SpatialScribe refuses to: it abstains when the evidence isn't there, tells you up front which cell types your panel can and cannot resolve, and reports one 0-1 Annotation Quality Index. Three findings I didn't expect: A low "confidently typable" score is usually a DEPTH verdict, not a panel defect. On the bundled 5K-panel section, restricting to the panel's genes costs 1 of 10 cell types; thinning to the section's real ~50 counts/cell costs 5. Buying a bigger panel would not have helped. Reference agreement is not annotation quality. Transferring a lab's own 28-type skin reference lifted agreement with their existing RCTD from 0.23 to 0.70 while the quality index collapsed from 0.22 to 0.01 - those fine types simply aren't separable on a 495-gene panel. Coarsening the reference first buys both (agreement 0.65-0.75, AQI 0.28). Per-cell confidence heuristics barely rank correct cells (within-type AUC ~0.54, near a coin flip). Only cross-method ensemble agreement does (~0.77). So the app declares which regime it's in instead of pretending its greying is trustworthy. Why it matters: spatial runs are piling up far faster than bioinformaticians can read them, and the failure mode that actually hurts a biologist is a confident wrong label. The quality index is validated against expert ground truth (Spearman ~0.78, 95% CI 0.66-0.86, across 11 sections and 5 tissues) - and it is an index, explicitly not a probability of being correct. - **Team:** [Paul Gueguen](https://cerebralvalley.ai/u/pgueguen) - **GitHub:** https://github.com/p-gueguen/spatial-scribe - **Demo video:** https://youtu.be/1wR6y0JIw3k - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=81 ### 106. genetic_drift I used the genome-scale CD4+ T-cell Perturb-seq atlas from Marson/Pritchard labs as an empirical genotype-phenotype map to ask how the architecture of the T-cell gene regulatory network shaped which Neanderthal variants survived in modern humans. Introgression frequencies came from the Li et al. 2024 IBDmix map across 1000 Genomes populations. I set out expecting a clean enrichment story, i.e. that strong T-cell regulators would carry excess archaic ancestry, and I got the opposite. After a properly matched null (constraint, gene length, local gene density, and a directly-measured "introgression desert" covariate replacing a gene-density proxy), the regulator effect is indistinguishable from zero at every resolution tested. Network topology metrics were directionally consistent with hub constraint but statistically redundant with the desert, which inverts the usual reading: the desert is not a nuisance confound but the integral of the genotype-phenotype map: the stabilizing-selection imprint of 50,000 years, consistent with theoretical work. Why it matters: it is a worked example of network-conditioned mutation-effect theory tested on real data, and an honest one: several hypotheses (context-dependence, three-way viral convergence, epistatic co-introgression) were tested and did not survive their controls, and the manuscript says so. The negative results constrain the field's expectations, and the 151 leads are a concrete, testable hypothesis set for follow-up. - **Team:** [Alexandre Gouy](https://cerebralvalley.ai/u/agouy) - **GitHub:** https://github.com/agouy/introgression_landscape_tcell - **Demo video:** https://drive.google.com/file/d/1oxzzolsnMObP-IgEjZTg6kSPG-Jg_Oe8/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=82 ### 107. Team Dhruvik PaedScale — a starting pediatric dose from adult pharmacokinetics, with a cited rationale Most drugs are never trialed in children, so when a kid needs one that has no pediatric label, clinicians extrapolate the dose by hand, usually by scaling the adult dose down by body weight. The problem is that a young child's liver and kidneys aren't finished maturing, so they clear drugs more slowly than their weight suggests. Scale linearly, and you overdose the child. That gap between "weight-proportional" and "what the body actually does" is where many pediatric medication errors occur. I built PaedScale to close that gap in a defensible way. You give it a drug, the child's weight, and their age, and it returns a starting-dose estimate built on the Anderson–Holford model: allometric scaling (weight^0.75) multiplied by an organ-maturation curve for whichever pathway actually eliminates that drug (a specific CYP enzyme, renal filtration, glucuronidation, etc.). The split I cared about was keeping the math and the judgment separate — Python does the pharmacokinetic arithmetic deterministically, and Claude does the part a calculator can't: figuring out which elimination pathway a given drug maps to, pulling the adult PK numbers live from PubMed and openFDA, checking the result against real dosing guidelines, and writing out the reasoning. It's a multi-agent setup — an orchestrator agent driving a cheaper retrieval subagent — and the same retrieval tools are exposed over MCP. The rule I held to throughout is cite-or-abstain. Every number that drives the estimate has to trace back to a source; if the data can't be retrieved, the engine refuses to invent a maturation curve, and the whole result drops to a "grade D — insufficient data" instead of producing a confident-looking guess. Each answer is graded A–D by evidence strength, with a guideline concordance ratio and flags for clinically important factors (narrow therapeutic index drugs that require blood-level monitoring, active metabolites, oral bioavailability gaps). What I found building it: the approach genuinely works for drugs whose pediatric dosing is really about matching adult exposure — I validated it across ~20 drugs (midazolam, vancomycin, morphine, the aminoglycosides, several others), and the estimates landed within guideline range for the ones with a guideline to check against. Just as useful was learning where it breaks. Concentration-dependent aminoglycosides get underdosed by the exposure-matching math, so the agent has to guideline-anchor and flag that mismatch rather than trust its own number. And some drugs (digoxin, acetaminophen) are empirically de-linked from adult PK altogether, so I deliberately excluded them — forcing a single-pathway extrapolation there would just produce a confident wrong answer, which is exactly the failure mode this project exists to avoid. Why it matters: This isn't meant to replace a prescriber, and I've been careful to frame it as decision support, not prescribing. But the honest reality is that a neonatologist facing an off-label drug at 3 am is doing this extrapolation by hand from sparse references anyway. PaedScale gives them a starting point that's transparent about its own uncertainty — you can see the pathway it assumed, the sources it used, and how far it sits from the nearest guideline — instead of a black-box number. In a domain where the tolerance for a silent error is basically zero, making the reasoning auditable is the whole point. Live at paedscale.by-dhruvik.in. - **Team:** [Dhruvik Pandya](https://cerebralvalley.ai/u/dhruvik_pandya) - **GitHub:** https://github.com/heyiamnotacoder/dosagecalc - **Demo video:** https://youtu.be/DSgwuBE4Rpo - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=83 ### 108. molecular composites Cadence is an AI copilot for the Learn phase of the synthetic-biology Design–Build–Test–Learn (DBTL) cycle. Four Claude-powered agents run the loop end to end; the Learn agent ingests Test-stage datasets, flags undefined, duplicated, and anomalous data tokens, and scores each with a reproducible confidence index that feeds fixes back into the next Design. We validated it against a real study — Singhal, Allen, Gaddes, Baer & Demirel, "Biomanufacturability of a Squid Ring Teeth Protein Library via Orthogonal High-Throughput Screening" — whose thesis is that predictable biomanufacturing "runs through models learned from library-scale data." Given the per-bin summary behind the paper's Figures 5–6, the Learn agent reproduced its headline conclusions from the summary alone: FACS brightness saturates as a yield predictor (Cliff's δ +0.46 Low→Medium, −0.03 Medium→High), growth is not the bottleneck, and codon-adaptation index carries the residual signal (bin-mean CAI vs mCherry R² = 0.91). On a schema-faithful stand-in it also caught three data-integrity defects unprompted — a malformed column, an exact duplicate (r = 1.000), and a zero-variance feature. The entire system — orchestration, bug fixes, the live run, and this write-up — was built with Claude (Cowork) in one session. - **Team:** [Melik Demirel](https://cerebralvalley.ai/u/MelikDemirel) - **GitHub:** https://github.com/melikcdemirel/claude - **Demo video:** https://youtu.be/QCd-Tl8T0KI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=84 ### 109. Vitalii Babenko ClinVal Validator is built for one named user: the clinical-AI researcher who has trained a medical-imaging model and needs to know whether the reported AUC is real or an artifact of how the data was split. It ingests a model's predictions.csv (predictions plus patient/scanner/site metadata) and runs three deterministic checks: patient-level leakage (do patient IDs cross the train/test boundary?), a split-quality audit (re-score the same predictions on a patient-disjoint subset), and per-subgroup degradation (AUC by scanner and site). It then hands the resulting JSON to Claude, which writes either a GREEN/YELLOW/RED risk report or a journal-reviewer critique with a corrected, leakage-safe Methods paragraph the author can paste into their manuscript. One architectural rule is enforced end to end: Python computes every number; Claude only explains it. The validators run fully offline; the reporter is the only code that touches the network, and it may cite only figures already present in the JSON. It never computes a metric. What I found: on the real RSNA Screening Mammography cohort, a naive row-level random split — the default most people reach for — leaves 7,511 of 7,524 test patients (99.8%) also present in training, because each patient contributes four to five images. Nothing is wrong with RSNA. The mistake is splitting by image instead of by patient, and it is completely invisible unless someone counts. Why it matters: a leaked split doesn't merely inflate a number — it makes the number unanswerable, because it no longer measures what a clinical reader assumes it measures. The detectors are verified against six ground-truth-labeled synthetic scenarios, so the tool's own claims are checked, not asserted. MIT, no PHI, synthetic/public-schema data only. - **Team:** [Vitalii Babenko](https://cerebralvalley.ai/u/vbabenko2191) - **GitHub:** https://github.com/vbabenko97/clinval-validator - **Demo video:** https://youtu.be/ybS2RcKtzbY - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=85 ### 110. AxeCap The user I built for is a genetic counselor holding a variant of uncertain significance in a regulatory region. Roughly 40% of what a clinical exome returns is uncertain, and the noncoding ones tend to stay that way forever, because the ACMG classification rules were written for coding variants and mostly do not apply. VUS Copilot takes a variant, or a whole VCF straight from a sequencing pipeline, and checks it against five real sources at once: RegulomeDB, Ensembl VEP, Open Targets with ClinVar, gnomAD, and a Myint et al. 2020 MPRA, which is an actual wet lab measurement rather than a prediction. Deterministic code ranks the cohort and shows every point it awarded. Claude reconciles the evidence and explains it in plain language. The interesting part is what happens when the sources disagree. In our demo cohort the CFTR 3849+10kb variant is Pathogenic by ClinVar practice guideline, and yet RegulomeDB ranks it a 4 and Ensembl VEP sees no regulatory feature at all. The predictors miss a known pathogenic variant. Going the other way, variants the wet lab confirms are held out of the worklist by ACMG BA1 once you notice that 40 to 87% of people carry them. All of it is auditable. A reproducible grounding benchmark ships with 64 fabricated claims the audit has to catch, and it names the 7 it still misses. - **Team:** [Krishna Venkatesh](https://cerebralvalley.ai/u/KrishVenky) - **GitHub:** https://github.com/KrishVenky/vus-copilot - **Demo video:** https://youtu.be/MM6rcVh-tcQ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=86 ### 111. Wen Group Cell-Type-Resolved Chromatin Effects of AD/PD Variants: DNA Language Model vs. ChromBPNet. We asked whether a general DNA language model can rival the specialized CNN ChromBPNet at predicting how non-coding variants affect chromatin accessibility. We fine-tuned our model, NucEL, on the same brain DNase-seq and GM12878 ATAC data and ran ChromBPNet's own endpoints. NucEL beat zero-shot ChromBPNet at accessibility classification in all four brain cell types and on GM12878 (0.967 vs 0.899), and independently reproduced ChromBPNet's variant-effect atlas (ρ=0.50; CR1 the shared top AD hit). A 1×→50× scaling curve shows the gap is data-limited, not architectural — suggesting general sequence models can match purpose-built regulatory CNNs given enough data. - **Team:** [Ke Ding](https://cerebralvalley.ai/u/FreakingPotato) - **GitHub:** https://github.com/FreakingPotato/NucEL_Chromatin_Effects/tree/main - **Demo video:** https://youtu.be/NeMU1IttMME - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=87 ### 112. ShiftScope (solo) ShiftScope compares any two single-cell populations, measures how far apart they are, finds the genes driving the difference, and ranks which genes to test next. It is condition-agnostic, so the same tool handles perturbation, disease, development, or cell-type comparisons. Its main use runs on Alex Marson's genome-scale CD4+ T-cell CRISPR screen. ShiftScope ranks hits by how strong the effect is and how little the gene has been studied, using a live PubMed count so novelty is grounded, not guessed. Claude then gives a keep-or-skip verdict. Famous regulators like VAV1 and CD45 drop; strong but overlooked genes, including chromatin-complex members with almost no papers, rise as candidates worth testing. It recovers correct biology across four domains and two real screens. Why it matters: a screen yields thousands of hits but a lab can test only a few. These are hypothesis-generating nominations that point wet-lab effort toward high-value, overlooked biology. - **Team:** [Tina Shen](https://cerebralvalley.ai/u/Tina233) - **GitHub:** https://github.com/aenorhabditis6/claude_science_hackathon/tree/main - **Demo video:** https://youtu.be/eL98YqIpexQ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=88 ### 113. TransBench TransBench turns a clinician's bedside observation, written in plain words, into a grounded, testable, bench-ready computational experiment. It ships as an MCP connector for Claude Science. You paste a de-identified observation - for example, "49M, type 2 diabetes with persistent postprandial hyperglycemia despite maximal metformin and confirmed adherence; elevated fasting glucagon; blunted GLP-1 response to mixed-meal testing." An eight-agent pipeline decomposes it into biological mechanisms, writes up to three falsifiable hypotheses, retrieves real published abstracts for each (PubMed, ClinicalTrials.gov, and Europe PMC), and grades every hypothesis as supported or refuted against those sources. Three rigor gates then reject anything that does not hold up: entailment (does the cited source actually support the claim?), grounding (drop any claim with no resolvable citation), and novelty (demote textbook facts so they are never sold as discoveries). For any hypothesis that is both a genuine open question and grounded in evidence, it designs one runnable single-cell experiment on a public dataset whose accession it content-verifies against the real record, and it emits a paste-ready prompt that Claude Science executes into a figure. What matters most: when the evidence is not there, TransBench ships nothing instead of inventing a plausible answer. The refusal is the feature. What I found across real, de-identified runs: over five domains (lupus, type 2 diabetes, resistant hypertension, melanoma, and rheumatoid arthritis) it shipped experiments for three and correctly declined for two because no hypothesis cleared the evidence bar; it caught two proposed datasets that did not match their claimed content and replaced them; and it fabricated zero citations. For the Gladstone example, a lupus / low-regulatory-T-cell observation produced a PTPN2-in-Tregs experiment grounded in 41 real citations, pointed at Gladstone's own dataset GSE278572 (Marson lab Treg/Teff Perturb-CITE-seq), which Claude Science then executed into a real composite figure with an FDR-corrected statistics table. That full run, with a sha256 manifest of every file, is committed in the repo under golden-test-results/. Why it matters: it is the difference between a confident chatbot answer and a defensible research lead. Every claim is tied to a real, resolvable paper or it is dropped; every dataset is verified before it is proposed; and the whole pipeline replays byte-identical with no API key (golden mode), so anyone can confirm that nothing shown is mocked up. TransBench is a research hypothesis-generation tool only - not clinical, diagnostic, or prescribing advice. - **Team:** [Kayomarz Darukhanawalla](https://cerebralvalley.ai/u/kayomarz97) - **GitHub:** https://github.com/kayomarz97/TransBench - **Demo video:** https://youtu.be/RFRhDaPUonE - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=89 ### 114. SecondLook SecondLook is a trust layer for an AI basal cell carcinoma detector, built for a dermatopathologist, Mohs surgeon, or dermatologist reading skin-pathology slides under real clinic load, where long days and fatigue nudge the hardest calls. AI support on a cancer-or-not call only helps if the doctor can trust it, and a confident wrong answer is worse than none. I trained EfficientNet-B4 detectors for basal cell carcinoma and built an app that scores regions of a slide, shows a Grad-CAM heatmap of where the model looked, and sorts each score into positive, uncertain, or negative. On a confident positive, Claude reads the heatmap and attention metrics and returns VERIFIED, FLAGGED, or DEFER. Claude never diagnoses and never changes the detector's call; it checks whether the model looked at the right place. The doctor keeps the final say. What I found: strong test scores cannot separate a trustworthy model from an unreliable one. Two models with near-identical scores can be looking at tissue or at a blank corner, and Claude tells them apart. It runs today as a live, open-source app, turning a detector a doctor could not trust into one they can. - **Team:** [Edward Yi](https://cerebralvalley.ai/u/aiedwardyi) - **GitHub:** https://github.com/aiedwardyi/SecondLook - **Demo video:** https://youtu.be/v3zw5bIx1yw - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=90 ### 115. A Class-I Act HLA class I molecules sit at the center of a trade-off: the same allele that helps T cells kill a tumor may also raise autoimmune risk. We built engageability — a new, validated, sequence-derived measure of how permissively an HLA allele's TCR-facing surface engages diverse T-cell receptors — and used it to show that immune "potency" is not one number but several independent axes. When public data proved structurally unable to answer the germline cancer-vs-autoimmunity trade-off (which we prove with an identifiability argument and a blind power analysis), we followed the statistical power to somatic immune escape across ~50,000 tumors. There the pipeline recovers the known B2M loss-of-function control and surfaces a novel, reproducible finding: a locus-wide HLA-B loss/retention hierarchy under HLA loss-of-heterozygosity, with HLA-B*58:01 the one allele individually singled out as preferentially lost (FDR 0.003) — pointing to the alleles tumors keep as escape-durable targets for TCR-based therapy. - **Team:** [Rebecca Nadler](https://cerebralvalley.ai/u/rbnadler) - **GitHub:** https://github.com/rebeccanadler/hla-engageability - **Demo video:** https://youtu.be/6zx3Qa12fLk - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=91 ### 116. Tastify Citera is a trust layer for AI-generated clinical regulatory documents. Large language models can write documents that sound compliant, but in clinical research that can get a study suspended. Citera checks each requirement and proves, with evidence, whether a consent form really meets the regulation. Every quote is span-verified against the source document, so nothing is made up. It also has a Verification Loop: when a fix is proposed, Citera verifies it and repeats until the document is "Submission Ready", with a full audit trail. This matters because it lets reviewers trust AI instead of just believing it. - **Team:** [Husein Indra Kusuma](https://cerebralvalley.ai/u/huseindra) - **GitHub:** https://github.com/huseindra/citera - **Demo video:** https://www.loom.com/share/e2491ddc212a4769b16d6c6653982bcf - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=92 ### 117. Spinal injury Globally, over 15 million people are living with spinal cord injury. We built a cross-species contrast to surface CNS-regeneration targets the field has overlooked: spinal-cord and appendage regeneration in three regeneration-competent species against mammalian injury, compared at the human-ortholog level. From a 1,034-dataset catalog we derived a 587-gene "competent-not-in-mammal" signature, scored 123 candidate targets, and annotated each for druggability. A from-scratch physics and control-theory engine adds what a differential-expression list cannot: causal direction, and evidence that no single perturbation reaches a four-way cocktail ceiling — so the robust targets are convergent downstream effectors (19 of 73 druggable now). The shipped ranking is what survived our own engine attacking it twice. Its first pick (MATK #1) was a near-zero-baseline artifact; a Claude self-audit caught it and re-ranked to COL9A1, which then held across five orthogonal axes — all directional and nominal, hypothesis not proof. It matters because it is reproducible, uses public data and APIs only, and every number opens to an artifact others can build on: a prioritized, falsifiable hypothesis set. - **Team:** [Tien-chen Lin](https://cerebralvalley.ai/u/tclin) - **GitHub:** https://github.com/darkbreaker0/ClaudeScienceHackthon2026_SpinalInjury - **Demo video:** https://youtu.be/e1BheOL0ylw - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=93 ### 118. Tim Lantin Aging research is rate-limited by long iteration loops; it may take years or decades to know if an intervention has any effect. However, recent research provides evidence that astronauts “exhibit many hallmarks of aging…often on accelerated timelines”, which raises the interesting possibility of spaceflight-as-an-accelerated-aging model. Here, Claude and I investigate three questions: 1. To what extent do the effects of spaceflight resemble the aged human transcriptome? 2. Are these effects reproducible enough to be an accelerated aging model? 3. What existing drugs might reverse these effects? We reframed open human spaceflight blood transcriptomics as an accelerated model of immune aging. Using OPEN NASA OSDR data only, we found the conserved spaceflight signature most resembles human immune aging (inflammaging), replicating at the pathway level across three independent missions (Inspiration4, Ax-1, JAXA) — different platforms, biofluids, 3 to >120 days. The telomere axis doesn't transfer — reported straight. Reversing the signature against LINCS and cross-filtering against lifespan-extending geroprotectors yields a dual-evidence drug shortlist (taxifolin, everolimus, curcumin) — hypothesis-generating, not clinical advice. Astronauts age fast and recover on a known clock, making spaceflight data a testbed for geroprotectors — the actionable step existing work stops short of. - **Team:** [Tim Lantin](https://cerebralvalley.ai/u/timlantin) - **GitHub:** https://github.com/timlantin/spaceflight-aging - **Demo video:** https://www.youtube.com/watch?v=3BysATOBFtE - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=94 ### 119. Pls let me win What I built. Perturb-seq Copilot — a no-code web app that takes a wet-lab immunologist from a processed single-cell CRISPR Perturb-seq count matrix to a ranked shortlist of high-impact gene perturbations and grounded, testable hypotheses in about two minutes. You upload a file (or click "Try the example") and it runs the full pipeline — quality control, donor-aware pseudobulk differential expression, pathway and transcription-factor activity, and an E-distance effect-size ranking — then produces a grounded Claude interpretation and a one-click, self-contained HTML report. It was built end to end with Claude Code: the scientific pipeline, the Streamlit interface, the Dockerized deployment, and an interpretation layer that sends Claude only computed facts and passes a grounding check. What I found. On real published T-cell CRISPR screens, the biology validates. In Shifrut 2018 (primary human T cells), CBLB — a known brake on T-cell activation — ranks #1 by effect size with a coherent chemokine/effector program; in Datlinger 2017 (Jurkat CROP-seq), the top hits are the core TCR-activation transcription factors: EGR1, ETS1, NFATC1, NR4A1, and ZAP70. But the most important result is methodological. When there are too few donors, gene-level significance cannot honestly be established — so the tool says so: it labels the run exploratory, refuses to promote any target to a "confirmed regulator," and in one case Claude flagged that its own #1-ranked hit was dominated by a guide-construct artifact rather than real biology. What I built. Perturb-seq Copilot — a no-code web app that takes a wet-lab immunologist from a raw single-cell CRISPR Perturb-seq count matrix to a ranked shortlist of high-impact perturbations and grounded, testable hypotheses in about two minutes. Built end to end with Claude Code: QC, donor-aware pseudobulk differential expression, E-distance ranking, a grounded Claude interpretation, and a self-contained report. What I found. The biology validates on real T-cell screens — CBLB tops Shifrut 2018, and the core TCR transcription factors (EGR1, ETS1, ZAP70) top Datlinger 2017. More importantly, when donors are too few the tool refuses to overclaim: it labels the run exploratory and even flags when a top hit is a technical artifact. Why it matters. It puts rigorous, honesty-first Perturb-seq analysis in the hands of the biologist who generated the data — every output a testable hypothesis, never a false discovery. - **Team:** [Alnur Kengesbay](https://cerebralvalley.ai/u/Alnur) - **GitHub:** https://github.com/a1murt/perturb-seq-copilot - **Demo video:** https://youtu.be/UuLHZ57qyok - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=95 ### 120. Builder_QXY Target selection is one of the earliest and most expensive decisions in drug discovery. TraceTarget is a drug target prioritization tool that scores gene targets across several key dimensions—human genetics, clinical precedence, in vivo validation, disease biology, tractability, and safety. It presents them as a single radar scorecard, but every axis opens onto the evidence beneath it, and every number links back to its source. A future version will let scientists bring in their own internal data to prioritize targets against the same transparent framework. - **Team:** [Qiao Xuanyuan](https://cerebralvalley.ai/u/QiaoX) - **GitHub:** https://github.com/qiaoxy0/tracetarget - **Demo video:** https://www.loom.com/share/ad351b4a34844cf0a0efbee76867b8ee - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=96 ### 121. eraya Most immunogenomic models predict T-cell receptor–peptide binding. KSHAMA inverts this: given a primary human CD4⁺ T cell and a candidate genetic perturbation (knockout, knockdown, or CRISPR activation), it predicts whether the perturbation drives commitment to a stable, suppressive FOXP3⁺ regulatory (Treg) fate, whether that state is durable, and it ranks perturbations for wet-lab prioritization. Anchored on the biology recognized by the 2025 Nobel Prize (peripheral tolerance, Tregs, FOXP3) and its lethal failure mode, IPEX, we (i) map 292 recent papers of the subfield, (ii) construct a signed, directed regulatory network of 48 circuit genes whose edges are 82% supported by STRING v12.0, (iii) identify the highest-leverage “trigger” nodes by network centrality, and (iv) simulate how single perturbations cascade toward stable Treg or effector/autoimmune fate using a continuous-logical dynamical model. The simulation reproduces the IPEX collapse (FOXP3 loss de-represses the T-bet effector axis) and recovers RBPJ—a repressor identified by a 2025 genome-wide CRISPR screen—as the top-ranked trigger. All mechanistic claims were audited by an independent reviewer agent against primary literature. We enumerate live datasets (genome-scale Perturb-seq in primary human CD4⁺ T cells), baseline models with hyperparameters, benchmarking harnesses, and four therapeutic modalities through which ranked perturbations become interventions. - **Team:** [Pardeep Singh](https://cerebralvalley.ai/u/Gen4) - **GitHub:** https://drive.google.com/file/d/1JotFmmx9_kxLJF5wfFS2JDII9d4S57ao/view?usp=sharing - **Demo video:** https://youtu.be/FqZ76cAaz3o - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=97 ### 122. CBBL DeepPro-v2 is a multi-view, multi-backbone deep learning model for prokaryotic promoter prediction across 23 species. It aggregates five complementary genomic foundation model experts through a learnable anchor–prototype fusion module with a soft-MoE aggregator, together with a differentiable motif-grammar side branch. Training follows a three-stage pipeline: task-adaptive pretraining of each backbone on the pooled 23-species corpus, followed by per-species two-stage 5-fold training, and finally ensemble voting. Findings: DeepPro-v2 attains a mean test MCC of approximately 0.81 across the 23 species, exceeding strong genomic foundation model baselines (GROVER, HyenaDNA, DNABERT2-CAMP, msBERT) retrained under a faithful, paper-matched reproduction protocol. Beyond headline accuracy, we subject the model to an end-to-end analysis and interpretability campaign spanning statistical significance testing (Friedman–Nemenyi and pairwise Wilcoxon), probabilistic calibration (ROC/PR and reliability curves), per-view contribution analysis, representation-space separability (UMAP), motif discovery, error stratification, and a full 23×23 cross-species transfer matrix. This campaign also yields honest negative results: the soft-MoE gate does not aggressively prune experts—itself evidence that the five views are non-redundant—and an anchor-prototype probe revealed no latent class structure, which we report as a negative finding. Most strikingly, the standalone differentiable motif-grammar branch spontaneously recovered the sigma-70 −10 box (TATAAT; best per-expert correlation r = 0.67) from binary promoter/non-promoter labels alone, without any motif-level supervision. Why it matters: Accurate and interpretable promoter identification underpins genome annotation and synthetic biology circuit design. By pairing fair, paper-matched benchmarking with a comprehensive, reproducible analysis and interpretability suite—rather than reporting a single accuracy number—we provide a rigorous reference for this task and make transparent both where the model succeeds and where it does not. All of the work presented here was built during a one-week hackathon: "v2" denotes an in-week iteration of our own model rather than an extension of any pre-existing project. - **Team:** [Johnson V](https://cerebralvalley.ai/u/Awakumori) - **GitHub:** https://github.com/Phlogistic-Rain/DeepPro-v2 - **Demo video:** https://youtu.be/pkxclL3HM0E - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=98 ### 123. chroma-code regulatr — cell-type-aware non-coding variant prioritization What we built. regulatr is a web tool that answers a question standard variant annotation can't: is this non-coding variant sitting in active regulatory DNA in the cell type that actually matters — and which gene does it likely control? Most disease-associated genetic variation lies outside protein-coding genes, where its effect is regulatory: it changes whether and where a gene is switched on. But regulatory DNA is cell-type-specific — an enhancer variant only matters in a cell type where that enhancer is open. So "is this variant functional?" is the wrong question; "is it functional in this cell type?" is the right one. regulatr operationalizes exactly that. You give it a single ENCODE chromatin-accessibility experiment (ATAC-seq), which fixes the cell type. It scans that experiment's peak file genome-wide to find where regulatory DNA is open, pulls every ClinVar variant in a chosen locus, flags which variants fall inside open chromatin, maps each to its nearest gene via Ensembl, infers the gene's strongest disease association from Open Targets, and ranks the variants by a transparent composite score (accessibility strength, clinical significance, and gene–disease score, equally weighted, in-peak variants first). It adds ordered pathway enrichment over the most-accessible genes (g:Profiler), an embedded igv.js genome browser, and JSON/CSV export. The entire backend is Python standard library only — zero third-party dependencies — and pulls all data live from ENCODE, ClinVar, Ensembl, Open Targets, and g:Profiler. What we found. The tool makes the cell-type dependence of non-coding variation visible. Loading the same locus under two different experiments — say a CD8 T-cell line — surfaces different accessible regions and a different set of variants landing in open chromatin. The same variant can be "in open regulatory DNA" in one cell type and inert in another, which is precisely the biological signal that a cell-type-agnostic annotation would erase. Building it also surfaced real data-integration subtleties worth handling correctly: ENCODE peak BEDs use inconsistent schemas, so the signal column must be resolved per file rather than assumed; gnomAD genome frequencies (not exome) are the right subset for non-coding variants; and gnomAD's own region API can't serve wide windows, so ClinVar's embedded frequencies are the reliable source. Why it matters. Interpreting non-coding variants is one of the central bottlenecks in human genetics. regulatr turns a genome-browser hunch into a ranked, exportable, provenance-tracked list of cell-type-specific mechanistic hypotheses — the candidates worth testing at the bench first. It is deliberately honest about scope: accessibility locates open regulatory DNA but never calls variants; nearest-gene mapping is a first approximation, not chromatin-contact evidence; and an in-peak variant is a hypothesis, not proof. Every result is coordinate-checked to hg38 and traceable back to its ENCODE source file and lab. By combining public accessibility, clinical, and target-disease data behind one accession-driven query, regulatr lowers the barrier to asking the right, cell-type-aware question — for any of the hundreds of biosamples ENCODE already provides. - **Team:** [Tarun Naithani](https://cerebralvalley.ai/u/Chaiwala) - **GitHub:** https://github.com/btarun13/Regulatr-app - **Demo video:** https://youtu.be/7Di5r7MKUfY - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=99 ### 124. Human Phenotype Project (HPP) I built the Claude Human Phenotype Project (HPP) Research Harness, a paper-calibrated, local-first research environment for private biomedical cohorts. Instead of treating a lab’s papers, code, schemas, and notes as disconnected files, it reconstructs the research paths behind them—datasets, cohorts, joins, methods, limitations, and provenance—and turns that knowledge into an executable workflow. I demonstrated it on the Human Phenotype Project, a deeply multimodal longitudinal cohort. The harness maps prior work, surfaces conflicts and gaps, identifies feasible underexplored questions, and produces reviewable research packages containing data routes, analysis plans, code, uncertainty, reports, and reusable lessons. The central finding was that the bottleneck is often not a lack of research ideas, but fragmented institutional knowledge and the high cost of making each idea analysis-ready. Research Harness lowers that startup cost while preserving scientific uncertainty, privacy, and traceability. HPP is the proof, but the approach can generalize to other private biomedical cohorts. - **Team:** [Hagai Rossman](https://cerebralvalley.ai/u/hrossman) - **GitHub:** https://github.com/hrossman/claude-hpp-research-harness - **Demo video:** https://www.youtube.com/watch?v=Peg6tlgo1AI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=100 ### 125. ParadoxMap I set out to test whether the IL-23/IL-17 therapeutic paradox — the same pathway helping in one organ and harming in another — has a measurable tissue-intrinsic signature in public single-cell atlases, using a preregistered, blinded framework across gut, psoriatic skin, and HS. The scientific result is honest and bounded: the cross-tissue comparison is not computable from current public data — psoriasis and HS are absent from the public single-cell disease vocabulary entirely — and the within-tissue P comparison is underpowered rather than null. What survives is reference-free and real: a large IL23R-negative effector reserve, and tissue-specific co-activation (gut IL-23+IL-1, skin IL-23+TCR). The deeper output is methodological. Working with Claude Science, I logged nine data-integrity defects, none caught by any automated check — each one about what a number counts, not whether it computes. Deciding what is true, it turns out, is still the human's job. That discipline is the transferable result. - **Team:** [Jingyu Xie](https://cerebralvalley.ai/u/ValerieTse) - **GitHub:** https://github.com/ValerieTse/ParadoxMap_ClaudeScienceHackathon - **Demo video:** https://youtu.be/S3pV6QazTSM - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=101 ### 126. PORTRAIT PORTRAIT turns population health data into a per-patient passport - calibrated position, honest uncertainty, and what genuinely stands out - and refuses when the data cannot support a description. Built and validated during the event on public NHANES. The describability-and-calibration gate holds per-slice coverage in [0.85–0.95] across all six age×sex groups on 2,247 held-out adults from unseen survey years, where a conventional split under-covers every group and abstention halves. A pre-registered confirmatory test found profile coherence adds mortality signal beyond twelve markers plus age and sex (ΔC≈+0.005, permutation p<0.001, survives frailty adjustment) - small, real, and against our stated null. Eight phase-gated sessions; every hypothesis git-tagged before running; every reported number traced to a manifest; a leakage guard separating an eligible target (AUC 0.48) from covert self-prediction (0.998). The interface is grounded in a 73-requirement specification drawn from UX systematic reviews and device standards (IEC 62366, WCAG 2.2 AA). What it means. A description tool that states its precondition and abstains when unmet is safer than one that always answers. PORTRAIT ships that behaviour with the negative results that earned it. - **Team:** [Daniela Oehring](https://cerebralvalley.ai/u/danielaoehring) - **GitHub:** https://github.com/doehring-gh/PORTRAIT.git - **Demo video:** https://youtu.be/xbHdHCdteXA - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=102 ### 127. EvolvingScholar DNN EvolvingScholar is an AI "research trainee" that learns the way a real medical trainee does — mentor-guided and improving across a *sequence* of gene–disease projects. Given a gene and a disease, the Scholar generates research questions, reviews the literature, analyzes synthetic EMR data to test gene–disease associations, and proposes testable hypotheses. At checkpoints it "meets its PI" and receives structured, rubric-based feedback that drives its growth before the next project. The ultimate goal is to evaluate and compare AI Scholar Agent evolving trajectory growth, rather than autonomous co-scientist. The system is built as three nested loops: **Loop A** does the research for one gene–disease pair; **Loop B** injects structured PI-mentor feedback; **Loop C** — the actual research contribution — rewrites the Scholar between projects. The core design bet is that **growth is not prompt accumulation**: cross-project learning happens only through updates to external, typed, versioned artifacts (the concept model, question-design logic, method repertoire, and earned EPAs), never through an ever-swelling prompt. This means "the Scholar evolved" is a recorded git diff, competency is read out on the clinical Entrustable Professional Activity (EPA) entrustment ladder (observe → independent → supervise), and the evolution measurement is free of the context-length confound that grow-the-prompt systems suffer. **What we found (working pilot, v0.1.0):** two scholars — one on the Claude Agent SDK with rich tools, one on the raw Messages API with minimal tools — each completed a full A→B→C run on the first project (TTR / hereditary transthyretin amyloidosis). The API scholar reached entrustment level 2 with 4 earned EPAs on run 1. Both share one review-and-evolve harness but keep separate, endowment-coupled experience stores with no cross-pollination. **Why it matters:** the AI Scholar is a controllable model system for a developmental process — how research questioning and disease conceptualization evolve — that is knowledge-entangled and nearly impossible to isolate in humans. Longer term it's a first step toward a self-taught AI clinical expert, the EPA based AI application in medicine, and a testbed for how AI can serve in medical education and research. PS: slides https://dlmp-niu.github.io/EvolvingScholar/demo/talk/index.public.html - **Team:** [Zhiyv Niu](https://cerebralvalley.ai/u/nealniu_mc) - **GitHub:** https://github.com/DLMP-Niu/EvolvingScholar - **Demo video:** https://youtu.be/WqVg4UOW2Ac - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=103 ### 128. Fareed Khan Vellum is a glass-box agentic ICD-10 coder. Five Claude Opus roles read a hospital discharge summary, resolve each finding to a current code through live terminology tools, challenge one another, and sign off a chart. What makes it a product rather than a black box is that every code carries a receipt: the sentence the model quoted, and a deterministic verdict, with no model and no network, on whether that sentence exists in the note. The motivating finding is that a fluent model sometimes invents its citation, quoting a sentence that was never written, and a more capable model makes the fabrication more convincing. A code whose citation cannot be found is never accepted on the system's own authority. I evaluated it in two stages. On ninety real inpatient notes, self-hosted open models hallucinate most of their codes, and even with retrieval plateau near 0.16 exact F1. On four identical synthetic notes, the deployed pipeline reaches 0.93 to the best open configuration's 0.57, and it is the only one that grounds every code and fabricates none. The limits are stated as plainly as the numbers. - **Team:** [Fareed Khan](https://cerebralvalley.ai/u/fareedkhandev) - **GitHub:** https://github.com/FareedKhan-dev/vellum - **Demo video:** https://www.youtube.com/watch?v=FBezih99aBI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=104 ### 129. Basepair Basepair watches for one specific kind of preventable mistake. A patient's genetics can make a drug useless or dangerous. Clopidogrel does nothing for a CYP2C19 poor metabolizer. Some HLA-B carriers can die from a normal dose of abacavir. The hospital usually has the genetic result sitting in a file somewhere, but when a doctor prescribes months later, nobody re-checks it, and the patient gets the wrong drug. So I built the thing that re-checks. Basepair listens to prescribing events from the EHR, matches the patient's genotype against 2,115 published CPIC and DPWG rules, and catches the ones that conflict. Each conflict becomes a short, cited alert that a pharmacist clears in one click before the drug goes out. It runs as a durable task that waits for that human decision and records the whole exchange as FHIR. The part I care about most is that the model never makes the call. A plain lookup finds the rule. Claude only reads the messy inputs and checks each alert against the guideline's actual words. The bar is a published citation. - **Team:** [Tyler Gibbs](https://cerebralvalley.ai/u/tylergibbs) - **GitHub:** https://github.com/tylergibbs1/basepair - **Demo video:** https://youtu.be/5FBv_NA6fbQ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=105 ### 130. Zeyu YANG Disease predisposition arises from complex, multivariate gene regulation, yet conventional approaches typically nominate a single target gene and fail to capture the underlying regulatory network. Two complementary genome-scale resources now make the full network tractable: genome-wide CRISPRi Perturb-seq, which reveals how knocking down any individual gene reshapes the entire transcriptional program landscape, and human loss-of-function (LoF) genetics, which quantifies how a gene's inactivation affects a disease trait. Inspired by a recent Nature framework (Ota et al.), we integrate these to reconstruct, for a chosen disease, a directed network linking regulator genes to gene-expression programs and ultimately to the trait. We demonstrate it on rheumatoid arthritis (RA): non-negative projection of CD4⁺ T-cell Perturb-seq onto RA patient–derived programs, followed by a program-to-trait regression against RA LoF burden, yields the complete regulator→program→trait map. The network recovers the Th17 program—a hallmark pathogenic CD4⁺ T-cell state in RA—as a positive control, and additionally surfaces the SOCS2/CISH cytokine-signalling-feedback program (P19) and its candidate upstream regulators as a novel, testable lead. Finally, we release a reproducible, disease-swappable pipeline, letting the same framework be applied to any trait with matched Perturb-seq and LoF data. - **Team:** [ZEYU YANG](https://cerebralvalley.ai/u/TYR00) - **GitHub:** https://zeyuyang-0420.github.io/PLoTM/ https://github.com/Zeyuyang-0420/PLoTM - **Demo video:** https://www.youtube.com/watch?v=TjZziQq_UfU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=106 ### 131. Octopus Octopus is a wet lab operating system. It turns a written protocol into a live, voice-guided run and an auditable report, so the experiment record is created while the work happens, not reconstructed afterward. It’s built for the wet-lab researcher at the bench, where hands are busy and fidelity leaks between what a protocol says, what was actually done, and what was observed. Everything lives in one app. Deterministic, tested code owns every number, the model never computes a volume, it reads and explains it. A voice copilot guides the run hands-free, and every result cites its source, all flowing into a lab memory you can search. With Octopus you can generate and scale a protocol, run it by voice while capturing notes, photos, measurements and deviations, get a fully sourced Planned / Performed / Observed / Interpreted report, and then ask your whole lab questions in plain language. - **Team:** [Clement Djezvedjian](https://cerebralvalley.ai/u/ClemDjez) - **GitHub:** https://github.com/clemdjez-byte/claude-hackaton - **Demo video:** https://youtu.be/rL8hDNQmuOs - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=107 ### 132. Mightytreefolk Why it matters. It is frequently cited that roughly a third to half of all proteins bind a metal, and the metal is frequently what makes an enzyme an enzyme, the catalytic center itself, and often a small molecule target (hemoglobin's iron is why carbon monoxide is lethal). Yet thousands of protein families remain completely uncharacterized. Homology-based annotation is blind to these orphans by construction. MetalloHunter demonstrates that you can take genuinely dark proteins and, without any homolog, identify which ones carry metal sites and assign them a known functional character, here, dark DUFs that display the coordination chemistry of known Ca²⁺-dependent hydrolases. Rigor was checked against a blinded 76-structure control package (metal-prior sensitivity 0.78 at specificity 0.92; a separate 29-structure benchmark reached 0.74 at 100% specificity), with the only false positives being the trypsin catalytic-triad decoy we planted deliberately. What we built. MetalloHunter is a four-stage discovery pipeline that scans AlphaFold structures of functionally uncharacterized protein families (Pfam "Domains of Unknown Function," DUFs) for metal-binding sites, classifies their coordination geometry and chemistry, and then physically stress-tests the top candidates on a GPU, placing the predicted metal, co-folding the holo structure with Chai-1, and measuring whether the coordination sphere actually reconstructs. The premise: a cluster of Cys/His/Glu/Asp residues in an apo model is only a hypothesis that a metal binds; a genuine site locks the metal into place when physics acts on it, while an adventitious cluster falls apart. That contrast is the validation signal. What we found. Sweeping 600 DUF families (408 with usable AlphaFold models, 1,809 proteins, 4,820 candidate sites), the pipeline surfaced 1,072 proteins with a metal-pocket signal and a tight top tier of 35 with the strongest dual signal (ideal-polyhedron geometry and a catalytic open-coordination vertex). Running these through AllMetal3D (metal-identity CNN) then Chai-1 (holo co-folding) produced physically validated hits, most notably two previously-uncharacterized proteins whose Ca²⁺ sites fully reconstruct (PF16480 and PF13961): both are carboxylate-anchored Ca²⁺ with an open coordination vertex, the coordination signature of the catalytic-Ca hydrolase class (phospholipase A₂, staphylococcal nuclease). We also confirmed the discriminator works in both directions, large multidomain proteins where the scan caught adventitious clusters correctly fail to reconstruct, and a Cys-rich family (PF25205) resolved to a structural zinc-ribbon fold rather than a catalytic site (verified not to be disulfides, the SG–SG distances are 3.8–4.1 Å, far above the 2.05 Å of a disulfide bond). - **Team:** [Peter Thompson](https://cerebralvalley.ai/u/Mightytreefolk) - **GitHub:** https://github.com/mightytreefolk/MetalloHunter - **Demo video:** https://youtu.be/BO6G3VlNQww - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=108 ### 133. Disruptors Hippocampus AI is the memory and reasoning engine for scientific discovery. Every experiment generates valuable knowledge, but that knowledge becomes fragmented across ELNs, protocols, GitHub, notebooks, sequencing pipelines, papers, and conversations. As researchers leave and projects evolve, context is lost, experiments are repeated, and discoveries remain disconnected. Hippocampus continuously captures knowledge from a laboratory's existing tools, builds a living knowledge graph of experiments, reagents, protocols, analyses, hypotheses, and literature, and exposes that memory to Claude Science through an MCP server. Claude reasons over the lab's complete institutional knowledge, not just public scientific literature. Instead of waiting for prompts, Hippocampus proactively detects contradictions, identifies repeated work, surfaces relevant prior experiments and new publications, and recommends the highest-value next experiment. Every conclusion is backed by traceable evidence from the laboratory's own records, making hallucinated citations structurally impossible. Rather than replacing existing workflows, Hippocampus becomes the intelligence layer that turns every experiment into knowledge and every future experiment into a better-informed decision. - **Placement:** Finalist - **Team:** [Chirag H](https://cerebralvalley.ai/u/chirag01) - **GitHub:** https://github.com/soul0101/Hippocampus-AI - **Demo video:** https://www.youtube.com/watch?v=cD-bKsvqiAo - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=109 ### 134. Traxome Traxome is a deterministic biomedical interpreter: it turns any symptom, gene, variant, or drug into a fully-sourced evidence dossier for patients, clinicians, and researchers. A single search unifies 20+ public databases — ClinVar, gnomAD, dbNSFP predictors, Open Targets, CPIC, the DEA schedule, openFDA, ClinicalTrials.gov, and more ; and every field traces back to its source. Its core contribution is a Provable Evidence Standard: any result can be frozen into a content-hashed Evidence Capsule that is byte-for-byte reproducible, re-verifiable offline, and diffable over time to flag reclassifications. All ten evidence tools are exposed to Claude over a remote MCP server, giving an AI agent grounded, deterministic, citable biomedical evidence instead of hallucination. This matters because clinical genomics runs on trust: interpretations must be transparent, reproducible, and auditable; exactly what black-box models and paywalled tools fail to deliver. - **Team:** [Kenneth Kusima](https://cerebralvalley.ai/u/KennethKusima) - **GitHub:** https://github.com/kkusima/Traxome - **Demo video:** https://www.youtube.com/watch?v=ViOzwMpICoI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=110 ### 135. PerturbGate PerturbGate is an open-data, reproducible pipeline that turns a genome-scale primary human CD4 T-cell Perturb-seq screen into evidence-gated mechanism hypotheses — not another ranking of hits. Each of 924 knockdowns is scored not by "does the cell change?" but by "does it reverse a donor-paired juvenile-arthritis disease direction?", then put through gates for technical validity, disease direction, an external patient cohort, and translatability. The result is an honest triad: PAK2, a strong reproducible cellular hit, was rejected as non-directional; RIPK1 is an unsupported comparator; RICTOR reverses the inflamed T-cell programme (+0.161 internal, +0.165 in an independent JIA cohort, 6/6 leave-one-donor-out) but is retained only as a mechanism hypothesis with an explicit modality gap - never a validated target. Every number resolves to a committed frozen table; negatives and corrected results are first-class outputs. It recomputes on a laptop in minutes. - **Team:** [Margarita Soloshenko](https://cerebralvalley.ai/u/Margosolo) - **GitHub:** https://github.com/MargoSolo/perturbgate-perturbseq - **Demo video:** https://drive.google.com/file/d/1QVZjKWSkjHtPjq03FXxN32L6DgIl_A4c/view?usp=drivesdk - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=111 ### 136. Cogworks We built An Agentic Interface Over Existing Wet Lab Equipment Most lab equipment today is very difficult for researchers to use. Every piece of equipment comes with its own domain-specific language. Many times researchers have to spend months just to learn that language before they can actually start using the robot, and for every workflow they have to go and code that specific workflow into the machine. This makes lab automation equipment almost unusable. In fact, there is a whole class of jobs that has spun up just because of this, called automation consultants. Their only job is to talk to scientists, understand what protocols they want to run, and code them into the robotics. They charge around $150,000 per year. Our product can do the same work with the $200 Claude subscription while removing the conversation overhead between scientists and consultants. - **Team:** [dhruv agarwal](https://cerebralvalley.ai/u/firstfly) - **GitHub:** https://github.com/sunnnybala/cogworks-labverify - **Demo video:** https://www.youtube.com/watch?v=oBeHY43MH68 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=112 ### 137. Natioan taiwan university hospital Cancer center What we built: A genome-scale CRISPRi Perturb-seq screen in primary human CD4⁺ T cells turned into a working target-discovery system — not just an analysis. A FastAPI scoring engine for researcher and physician converts raw differential-expression data into 39-field "target cards" (effect size, knockdown confidence, cross-donor robustness, druggability, safety liability, human genetic support), feeding a readiness engine that outputs one of four calls (advance/validate/watchlist/deprioritize) per gene, served through a React portal and a live CSV-upload path so anyone can score their own screen with the same engine. What we found: Two independently-motivated ranking paths — a portal-readiness funnel (302 advance-ready targets) and a publication-decision funnel (39 context-specific, druggable targets) — converge on a 5-gene "Core-5" intersection with an independent 15-gene primary-outcome shortlist. Those calls hold up against data we never used to build them: 55/55 Open Targets disease associations, STRING-confirmed interactors at ≥700 confidence, and replication in an independent public CRISPRa HIV screen (GEO GSE318876). Why it matters: The hard part of a screen like this isn't generating hits, it's not fooling yourself about which ones are real. We enforced that architecturally: only 4 scores and 7 red-flags can move a target's call, every safety/genetic overlay is descriptive-only and regression-locked so it can never silently override that call, and missing data is shown as "unknown," never imputed as zero — known golden-standard genes (ZAP70 correctly grade-4, MED12 correctly flagged as broad-effect). That's what turns one expensive screen into a reusable, falsifiable pipeline instead of a one-time hit list. - **Team:** [Hsiangwei Huang](https://cerebralvalley.ai/u/Ericeric777777) - **GitHub:** https://cd4-target-discovery.netlify.app/ and https://github.com/erichuang777777/GWT_perturbseq_analysis_2025 - **Demo video:** https://youtu.be/blPKLObkapo - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=113 ### 138. Evolve Past Biology Evolution Arena is an interactive 3D WebGL survival game that turns evolutionary biology into something you can experience rather than simply read about. You control a small organism in a living ecosystem and must gather food, manage energy and warmth, evade predators, cross rivers, survive storms and nights, find shelter, and reproduce. At the end of each successful generation, you choose an inheritable mutation from faster limbs and camouflage to larger eyes, armor, swimming fins, glide wings, toxin resistance, stronger jaws, or more efficient digestion. Every adaptation has both a benefit and a cost, and different environments reward different bodies. Faster limbs consume more energy. Armor reduces damage but slows movement. Better digestion extracts more energy but reduces maximum speed. The habitats a player spends time in also influence which evolutionary opportunities are presented to the next generation, creating lineages that visibly and mechanically diverge over time. The central insight behind the project is that evolution becomes much more intuitive when adaptation is experienced as a series of constrained tradeoffs under changing selection pressures. There is no universally optimal organism: success depends on the environment, accumulated traits, available resources, predators, and the player's behavior. That matters because concepts such as natural selection, fitness landscapes, ecological niches, inheritance, adaptation, and evolutionary tradeoffs are often taught abstractly. Evolution Arena makes them tangible, visual, and playable. - **Team:** [JJ Ben-Joseph](https://cerebralvalley.ai/u/JJBJ) - **GitHub:** https://github.com/tensorspace-ai/evolution-arena - **Demo video:** https://www.youtube.com/watch?v=UWRvIPVu_Ws - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=114 ### 139. DMV Petri Dish Synthetic lethality is one of the cleanest routes to targeted cancer therapy, yet it remains understudied, and most cancer-screening data sits siloed behind deep expertise. We made over 10 datasets (over 20GB of drug data) explorable by anyone, with quick tooling to screen for potential drugs by cancer-type. We grounded our work by replicating our lung synthetic-lethal study (doi.org/10.21203/rs.3.rs-9559070/v1 – preprint) which used RB1/CDKN2A/PTEN/TP53 as driver gene mutations and found relevant synthetic lethal pairs (original dataset was DepMap, but we now have other datasets that show this generalization as well). The value of this project is that preclinical cancer researchers can massively look at an interpretation of cancer therapeutics using the synthetic lethality model, and get access to new drug targets and mechanism interpretations. Because this mechanism requires two genes, a driver gene and targeted gene – only big datasets are good a finding this non-obvious but very useful class of potential therapeutics. - **Team:** [Matthew Zamora](https://cerebralvalley.ai/u/MoCoMakers) - **GitHub:** https://github.com/MoCoMakers/tumor-preclinical-discovery - **Demo video:** https://www.youtube.com/watch?v=swFN98a-rZg - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=115 ### 140. Anees Ahmed MMC - the Mechanistic Model Compiler AI models that predict what happens when you switch off a gene have a problem: the good ones are black boxes, and they still barely beat a linear baseline (Ahlmann-Eltze, Nature Methods 2025). So we asked a different question. Not only that, can Claude predict better? But can Claude build a model that a biologist can actually open up, poke, and know when to trust? What we built: MMC is a Claude-driven loop that reads the newest human T-cell atlas (Zhu 2025, 22 million cells). Claude proposes a wiring diagram of how genes control one another, compiles it into a runnable simulation, tests it against real experiments, reads where it fails, reasons about why it fails, and rewrites its own wiring. The output is not a number. It's a circuit you can interrogate. The moment to watch. In our Th2 circuit, knockdown of GATA3 → IL-5 drops by 5.0. Knock down GATA3 and STAT6 together → IL5 drops by 5.0 again. Not more. Because STAT6 acts through GATA3, hitting both changes nothing. The additive baseline says −5.4. It's wrong, and it can't tell you why. That's mechanistic reasoning a black box structurally cannot do. What we found, and it's the point. Claude proposed a new hypothesis: STK11 represses chemokines. It looked good. It was grounded, the data genuinely support that edge, and it passes the same interpretability checks that textbook biology passes. But it didn't work. Across 76 hypotheses, in every regime we tested, single-gene knockdowns and combinatorial double-knockouts grounded mechanisms were never once incorporated into a model that beat a simple linear baseline (0 of 76; 95% CI [0, 4.8%]). MMC's own held-out test refused to certify STK11. We killed our own best result. Why does it matter? The scary failure mode of AI in science isn't hallucination. It's the opposite: hypotheses that are grounded, interpretable, and pass every check we usually rely on and still don't predict. Interpretability didn't catch it. Plausibility didn't catch it. Only the held-out performance against a strong baseline did. We're shipping the method, an interrogable T-cell circuit, and a map of exactly when mechanistic modelling helps and when it doesn't, a referee for a field where prediction is either stuck or bought with proprietary data. Who this is for, Biologists get a T-cell circuit they can open, poke, and argue with, and a map of exactly which regulatory questions mechanistic models can answer today, so they stop spending months where the data can't support them. AI builders get the harder lesson: the model's hypotheses can be grounded, interpretable, and pass every sanity check you have and still be worthless. Only held-out performance against a strong baseline catches it. Ship that gate, or ship a plausible lie. And next, the loop is general. Point it at any system with a simulator and intervention data, and now we know exactly which regimes are worth pointing it at, because we measured where mechanism earns its keep and where it doesn't. - **Team:** [Anees Ahmed Mahaboob Ali](https://cerebralvalley.ai/u/aneesahmed) - **GitHub:** https://github.com/ahmedanees-m/mmc - **Demo video:** https://youtu.be/28YprpGq8Hs - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=116 ### 141. RAFLab/histo.fyi I built a well curated dataset and then used it to answer 8 key scientific questions. They were mainly around devising interaction footprints and understanding about the important relationships with them around the role of sequence and structure in generating the interaction geometry. I also used what we found to ask questions about the underlying biology of cross-reactivity and analysed structure predictions by State of the Art models to understand if they are a proxy for experimentally determined structures (not yet!). One of the most important findings is that the prediction methods memorize rather than generalise and have a much less diverse interaction footprint than the experimentally determined structures and assays. We will do follow-on work to validate results and publish them, and more work with Claude Science to prioritise data gathering and see if we can teach the structure prediction methods how to predict better. - **Team:** [Chris Thorpe](https://cerebralvalley.ai/u/drchristhorpe) - **Demo video:** https://www.youtube.com/watch?v=YsJpH8RoBX0 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=117 ### 142. Odola Labs I created a workbench where you can name a protein and specify an objective in simple terms, and the agent will then manage the entire campaign. After determining whether your target has a measurable fitness landscape, it conducts a propose-measure-feedback cycle against it. In every round, it folds actual assay data into a Gaussian-process surrogate over ESM-2 embeddings and makes decisions about whether to investigate unknown combinations or take advantage of existing hits. The workbench finishes by converting this shortlist into an order that is ready to be copied and sent to a cloud lab like Ginkgo or Adaptyv directly through the tool. As a demo, my agent achieves a best measured fitness of 4.662 in 96 measurements against 149,361 genotypes on GB1's four-site landscape, where each fitness value is derived from the Wu 2016 deep mutational scan as a real prior measurement. It also avoids a reciprocal-sign-epistasis trap that stops greedy hill-climbing cold. The workbench agent takes the judgment seat a scientist normally occupies, so it picks which scorers fit the target, weighs their disagreement, and decides what the shortlist should prioritize instead of just returning raw numbers or ranked lists. Every candidate arrives carrying a rationale that can be interrogated. Using this tool, scientists can spend more time making scientific and design decisions, rather than piecing together various tools and squinting at scores. For live runs, I added a key to the app so that our GPU usage was contained (odola@27) - **Team:** [Chinmay Pala](https://cerebralvalley.ai/u/chinmaypala) - **GitHub:** https://github.com/ChinmayPala/claude_hackathon - **Demo video:** https://youtu.be/z8qoX-Jc9Ng - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=118 ### 143. e-nios I carried out a research study, written up as a manuscript, on one of the hackathon's suggested datasets: a genome-scale atlas of what happens inside human immune cells when each gene is switched off. I asked whether it can be taken from patients all the way to a drug molecule, and built a computational tool from the answer. The method takes a patient's disease signature and finds genes whose silencing pushes cells the opposite way; only those surviving a battery of evidence checks count as targets. They go into molecular design, where molecules are judged not on binding strength but on how cleanly they hit the intended protein while sparing its close relatives — the usual source of side effects. It also enables precision medicine: grouped by disease mechanism rather than diagnosis, patients reveal a drug class whose effect is strong in one subgroup and disappears when the cohort is pooled — precisely the stratification a clinical trial would need in order to see it. Across several immune diseases it rediscovers approved drugs unprompted, proposes new targets, and designs molecules predicted to be more selective than an approved drug. CRUCIBLE, the tool, runs the pipeline end to end for any disease with adequate patient data and withholds a result rather than produce one the data cannot justify. It matters because it delivers a framework to turn a descriptive atlas into a reusable engine for precisely targeted therapies in immune and autoimmune disease. - **Team:** [Eleftherios Pilalis](https://cerebralvalley.ai/u/epilalis) - **GitHub:** https://github.com/epil/crucible - **Demo video:** https://www.youtube.com/watch?v=E4Lli8AtkyM - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=119 ### 144. Sequence Intel SeqIntel turns a patent name into a validated sequence dossier, in minutes instead of hours. Biotech IP researchers face a slow, manual wall. A single AAV patent lists dozens of sequences, and matching each SEQ ID NO to its real sequence by hand takes 4 to 6 hours. Errors slip through. SeqIntel does it in one query. Type a variant name like PHP.​eB, a name that lives in patent prose rather than any lookup field, and get the real sequence, its owner, a freedom-to-operate read, the experimental structure, and a novelty check. Every result cited. Every result traceable. And it runs both directions. Paste a sequence you have engineered, and SeqIntel shows you who already claims it, before you file or commercialize. Built from scratch this week. I'd built a tool for our biologists before, and that is how I learned the data sources and the real bottleneck. A genome researcher team I met confirmed the same wall in their lab. The edge isn't a bigger database. It is automation the big tools don't offer, and an assistant that knows when to stop rather than invent an answer. Built with Claude. For the people who do this by hand today. Live app: https://seqintel.bio · Code: https://github.com/coder-RT/SeqIntel - **Team:** [Rachit Thariani](https://cerebralvalley.ai/u/rachit2311) - **GitHub:** https://github.com/coder-RT/SeqIntel - **Demo video:** https://youtu.be/lwP2FuMdDmI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=120 ### 145. Jimmy Ouellet KatalistSeek — a grounded reasoning layer over fragmented enzyme/reaction databases. Curated knowledge databases about a biochemical pathway or enzyme are scattered all over the internet — PubChem, Rhea, UniProt, COCONUT, AlphaFold, ChEBI. The questions a researcher actually cares about are: how do I make this molecule, and which enzyme catalyses the step? and what else might this enzyme accept? Answering either today means hours of manual cross-referencing to ensure grounded structure and honest provenance. KatalistSeek is a tool that will hopefully fill that gap by providing grounded and a provenance-tagged reasoning layer that collapses that work into minutes. Two modes share one core — make a molecule (target → natural-product gate → grounded producing reactions → candidate enzymes → homologue family → binding-pocket fit) and substrate scope of an enzyme (enzyme → its ortholog family, how the substrate drifts across species, its literature, its 3-D pocket, and a discovery step that proposes candidate promiscuous substrates from structurally-similar natural products — separating those already documented in a reaction from purely structural look-alikes). What makes the output trustworthy enough to act on is one non-negotiable rule: no EC number, accession, or catalytic residue is ever emitted unless a live database call returned it in the current session. Every biological claim carries a provenance handle back to the exact record that produced it; when retrieval fails, the tool says so and stops — it never backfills from memory. On naringenin it recovers chalcone isomerase from a live call and ranks the fold family so that catalytically inactive fold-homologues fall below the true isomerases — showing that conservation is not activity — while refusing to invent a precursor for a xenobiotic that has none. - **Team:** [Jimmy Ouellet](https://cerebralvalley.ai/u/Jouell3) - **GitHub:** https://github.com/jouell3/KatalistSeek - **Demo video:** https://youtu.be/8hObaubnyAg - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=121 ### 146. K-Vector I asked whether tissue-specific AAV regulatory elements could be identified systematically from public single-cell chromatin data, and whether their mouse-to-human translatability could be predicted before any wet-lab work. To test this, I developed OrthoGate as the research method. It applies the same specificity, size, motif, and cross-species analysis across different cell types and returns a ranked list of candidate regulatory elements. Across seven cell types, the method recovered the expected identity gene as a top result in five, including INS in beta cells, GCG in alpha cells, RYR2 in heart cells, and OLFM4 in intestinal cells. It also rejected broadly active housekeeping genes and recovered known regulatory motifs. Most importantly, the analysis flagged candidates whose associated biology may not translate to mice, including PHGR1, an intestinal gene with no mouse ortholog. These results suggest that tissue-specific AAV regulatory elements can be identified computationally for multiple cell types, and that cross-species screening can meaningfully change which candidates should advance. This matters because it could reduce avoidable lab and animal studies and help prioritize safer, more translatable AAV designs for tissues outside the liver. - **Team:** [Karina Herlambang](https://cerebralvalley.ai/u/kh306) - **GitHub:** https://github.com/kh306/aav-regulatory-pipeline - **Demo video:** https://youtu.be/LSwXKt_fbVc - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=122 ### 147. Kilian Maire Repertoire2Structure (R2S) is a multi agent pipeline, orchestrated by Claude, that turns a raw 10x single cell TCR repertoire into QC'd predicted TCR to peptide MHC structures for its top clonotypes, with honest specificity annotation and skeptical structure QC. A researcher drops a 10x contig CSV. A team of Claude agents then runs it end to end: an intake agent frames the run (data, question, compute route), a specificity annotator matches each clonotype by TCRdist against labeled references and refuses to force a label, a structure strategist reasons about the biological question and routes each group of clonotypes to the right folding tool (Protenix by default, with TCRdock, MHC-Fine, and an AF2 presentation classifier wired in), per tool executors build the folds on Colab or a local GPU, and a skeptical QC agent calibrates every structure against a scrambled peptide control before it will call it trustworthy. Two honesty rules are enforced in the output schemas, not just narrated: specificity is annotation by similarity, never prediction; and a predicted structure never confirms specificity. What we found is an honest, convergent negative. On a preregistered repertoire benchmark, structural confidence reads peptide presentation, not TCR recognition, the same conclusion a parallel AlphaFold 3 study (McMaster et al.) reached the same month with a different engine. So R2S does not oversell a specificitrs: the reusable contribution isthe agentic question to structure system plus the guardrails a working immunologist actually needs, leakage aware annotation, pe a fold is spent, and scramblecalibrated QC that is honest about what a fold can and cannot prove. - **Team:** [Kilian Maire](https://cerebralvalley.ai/u/KilianM) - **GitHub:** https://github.com/KilianMaire/tcr-repertoire2structure - **Demo video:** https://www.youtube.com/watch?v=Ru48ZQThA60 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=123 ### 148. Bordeaux Serpents SurvivalLadder is an open-source screening tool — ten Galaxy tools, a Snakemake pipeline, and a Shiny app, all calling one identical R analysis core — that tests a user-defined geneset, a directional _UP/_DOWN signature, or a single gene against clinical outcome across 33 TCGA projects and 21 published immunotherapy cohorts, or against any expression matrix the user supplies. Each signature is scored per sample (ssGSEA, GSVA, PLAGE, z-score or raw expression) and taken through confounder-adjusted regression: a Cox model for the five survival endpoints (OS, PFI, DSS, DFI, and PFS in the immunotherapy trials), and a logistic model reporting odds ratios for RECIST response. It then closes the loop, running differential expression between the two arms the survival model itself separated and screening the resulting signature against cell-line drug sensitivity. What makes it useful for real research questions is not a longer feature list than the established portals — cBioPortal and TIMER3 are strong, and TIMER3 in particular already scores signatures, adjusts for tumour purity, and fits Cox models. It is the shape of the answer. The portals return one query, one split, one model, one number, and give the researcher no way to see how much of that number was manufactured by those choices. SurvivalLadder refuses to collapse either axis: a score is cut at a ladder of percentiles rather than one arbitrary median or quartile, so an effect is read as a trend across the tails and a result that exists only at one lucky split is exposed as such; and each effect is refitted up a ladder of nested adjustment models — unadjusted, plus tumour purity, plus proliferation, plus both — reported side by side, so that the movement between the rungs is the finding. This is what a single adjusted hazard ratio structurally cannot show: that two confounders may act on the same signature in opposite directions, one masking the effect while the other inflates it, a structure any one adjusted number silently averages away. Around this sit three further things the portals do not attempt — screening many signatures in one run with multiple-testing correction across them, where a portal answers one query at a time and corrects for nothing; treating cell proliferation as an adjustable confounder, which neither competitor offers; and keeping prognosis and benefit-from-treatment strictly apart, since a signature can be strongly prognostic and useless for predicting who responds. And because every run emits a reproducibility manifest — pinned package versions, the random seed, and a SHA-256 of every input file — a result can be re-derived years later, which is precisely what a portal whose methods live in a backend that can change without notice cannot promise. The tool runs locally today, from a clean checkout, which is what makes it defensible. The next step is to make it broadly usable: publishing the ten wrappers to the Galaxy ToolShed and registering the Shiny app as a Galaxy Interactive Tool, so that a researcher with no command line gets the same point-and-click access the existing portals offer — but backed by versioned, inspectable, reproducible code rather than a hidden backend. The ambition is not to replace the portals, but to give the field an exploratory screen whose answers survive being asked twice. - **Team:** [Sergio Oterino Sogo](https://cerebralvalley.ai/u/BordeauxSerpents) - **GitHub:** https://github.com/Sergio-ote/SurvivalLadder - **Demo video:** https://youtu.be/NuESzzVZQC0 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=124 ### 149. Ray Recent progress in mass spectrometers and softare drive protein identification from a single HeLa cell to more than 6000. However, how many false positives there are remains an open problem. In this study we evaluated three DIA-NN 2.6.0 features that are on or encouraged by default and that all inflate identification counts — the refined q-value procedure, multi-species database searching for non-human samples, and match-between-runs (MBR) — asking in each case whether the extra identifications are trustworthy. The core method was a two-species entrapment search: the database held the sample's true species plus a foreign species known to be absent, so any precursor mapping only to the foreign proteome is a ground-truth false positive, giving an empirical FDR independent of the engine's own estimate. Spectral quality was measured directly from the raw Thermo files by mapping each precursor to its apex MS2 scan and counting matched b/y fragment ions at ±15 ppm (mass model validated to 0.06 ppm against DIA-NN's reported precursor m/z). What we found Refined q-value trades peptide confidence for protein depth. Precursors rose +4.7% (36,989 → 38,736) but protein groups fell −5.4% (5,069 → 4,794) — the signature of protein-informed rescoring. The added precursors had ~30% fewer matched fragments (median 7 vs 10), though entrapment FDR stayed controlled (0.51% → 0.37%). Shared precursors kept identical RTs, so refined-q only recalibrates q-values, not peak picking. DIA-NN identification is systematically human-biased for non-human samples. In a mouse cell, 6.6% of precursors mapped only to human (~12% size-adjusted) vs ~0.4% for human HeLa. Across 14 mouse liver single cells the human false-positive rate was 5.05–6.18% (median 5.62%, SD just 0.32%) — reproducing to ±0.3% across independent cells proves it is systematic: weak SCP spectra preferentially match the larger, better-annotated human proteome. MBR's homogeneity assumption does not hold well for single cells. Before MBR, mean pairwise cell overlap was only Jaccard 0.54, and more precursors were cell-unique (28.4%) than universal (23.5%) — the "identified in one ⇒ present in others" premise is true less than half the time. MBR added +3–46% precursors (most in the sparsest cells) and lowered entrapment FDR, but second-pass-only precursors carried the same ~30% fragment deficit (median 7–8 vs 10–12). Post-MBR apparent homogeneity (24%→48% "present in all cells") is manufactured by the imputation itself. Crucially, entrapment FDR certifies the aggregate ID list, not whether a precursor is truly present in the specific cell it was transferred into. Why it matters All three features buy depth by admitting lower-confidence identifications, and all three are most hazardous precisely where SCP is most valuable — measuring real cell-to-cell biological differences. A low FDR, even a low empirical entrapment FDR, certifies the aggregate list, not any single per-cell measurement. Practical guidance: turn refined q-value off when confident peptide-level IDs are the goal (it also yields fewer proteins); for non-human data, don't trust the nominal 1% FDR — run an entrapment check, prefer species-specific databases, scrutinize human-annotated hits; for MBR, report both passes and treat transferred values as imputed for any per-cell biological claim, reserving MBR for building complete quantitative matrices for clustering/integration. The cross-cutting principle: report depth and confidence together, and validate rescored or transferred identifications before building single-cell conclusions on them. - **Team:** [Rui Qiao](https://cerebralvalley.ai/u/rqiaorrr) - **GitHub:** https://github.com/volpato30/lifescience-hackathon/blob/main/single_cell_best_practices.pdf - **Demo video:** https://www.loom.com/share/ca3b199628a9446f8175bfc1073cbf0b - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=125 ### 150. DomFico Molaro is a next-generation molecular dynamics environment built as a VS Code extension. Instead of shipping with a fixed set of visualizations and analyses, it provides an extensible agent framework where Claude writes custom Python analysis directly against your trajectory. Every script is shown before execution, can be inspected or edited, and is saved as a reusable Python workflow, allowing your analysis toolkit to grow through conversation rather than waiting for new software releases. To ensure scientific trust, generated analyses are validated against established molecular dynamics libraries such as MDTraj and MDAnalysis. In our benchmark corpus, Claude-generated radius-of-gyration analysis matched reference implementations to within 1.6 × 10⁻⁸ nm, demonstrating that AI-generated workflows can be both flexible and scientifically reliable. The long-term vision extends beyond a single viewer. Molaro is designed as an agent platform where researchers can create, publish, and share new analyses, visualizations, and workflows. Rather than relying on a fixed feature roadmap, the community can continuously expand the capabilities of the software by building and distributing new agent-powered tools. - **Team:** [Dominic Fico](https://cerebralvalley.ai/u/DomFico) - **GitHub:** https://github.com/DomFico/molaro.git - **Demo video:** https://youtu.be/kAY7Gvz6wlE - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=126 ### 151. Lynx Human accelerated regions (HARs) are widely interpreted as developmental enhancers, and human-specific substitutions are known to alter their enhancer activity and transcription-factor binding — but systematically only in the brain. Whether HARs active outside the brain carry substitutions that change binding at experimentally bound transcription-factor sites, distinct from the GC-biased gene conversion (gBGC) that inflates substitution counts at these GC-rich elements, has not been examined. From a single-cell chromatin-accessibility matrix (312 Zoonomia HARs × 16 tissues) we identified 45 dominantly accessible outside the brain and screened each for human-specific substitutions inside experimentally bound (UniBind ChIP-seq) transcription-factor motifs, re-scoring affinity against the reconstructed primate ancestor and testing each change against a per-locus gBGC null. 19 carry a substitution inside a bound motif, and 8 show an affinity gain exceeding the neutral gBGC expectation (P < 0.05). The strongest, ZOOHAR.189, creates a CTCF motif match at an experimentally bound site near MYLK2 (Δ +8.0 bits); others gain bound TP63, GATA3 and PRDM1 sites in tissue-matched contexts. These position-weight-matrix predictions remain hypotheses pending allele-specific assays, but nominate specific loci and tissues at which human substitutions may have rewired regulatory binding. - **Team:** [Y X](https://cerebralvalley.ai/u/xyz7) - **GitHub:** https://hypnolynx.github.io/har-multitissue-manuscript - **Demo video:** https://vimeo.com/1209675634 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=127 ### 152. The Amplicons We are two medical oncologists in London who code. We spend our clinics telling people with cancer that we have run out of options. This hackathon was a chance to challenge that — with Claude — head-on. The idea starts with a quirk of cancer genomes. Tumours amplify chunks of their own DNA, early, and in the same places across thousands of patients. That repetition signals the tumour needs something in there — but amplification is indiscriminate, so a crowd of innocent passenger genes is dragged along too. Could we target these? Antibody–drug conjugates are the most targeted weapon oncology has: an antibody homes to the tumour cell and delivers chemotherapy directly to it. But they are brittle — the tumour stops making the single antigen you aimed at, and the drug fails. Newer formats carry several arms. So what if we aimed at several passenger proteins on the same amplicon at once? The catch: an extra gene copy rarely means extra protein on the cell surface. There are several steps between DNA and protein, and nobody had mapped which passengers actually make it through. We built that map. Over the week we used Claude Science to assemble terabytes of matched genomes and proteomes across six cancer types (CPTAC), promoter methylation, chromatin accessibility, ~465,000 malignant cells from CELLxGENE, and UniProt, DepMap, GTEx and the Human Protein Atlas. Three findings came out: 1. Whether an amplified gene reaches protein is decided early: methylation and chromatin accessibility matter far more than any property of the gene itself. Those settings belong to the tissue of origin, not the individual tumour, which means a construct built for one lung, breast or bowel cancer should work in the next patient carrying the same amplicon. 2. We can predict which genes will be amplified at protein level.. A model trained on gene properties alone — no protein data of any kind — recovers a gene's transmissibility at ρ = 0.52 (leave-gene-out). It is not just reading chromosomal position: holding out whole arms costs only 0.001, and the rankings agree across all six lineages (Kendall W = 0.97). So the map extends into cancers no one has profiled by proteomics — we did exactly that for glioblastoma, nominating constructs with no GBM protein data at all, and they held up across 390,761 single cells. 3. We can find targets. Our funnel narrows from 6,648 genes to 22 surface antigens on 18 recurrent amplicons, assembled into 10 multi-antigen constructs. In single-cell data, six of the eight we could test show their antigens on the same tumour cell more often than chance allows. Remove EGFR — the actual driver — from the lung 7p set, and the remaining passengers still co-detect at 1.30-fold (95% CI 1.13–1.47). The approach stands on the passengers, which makes it new. Every construct ships with the experiment already written: named amplicon-positive and amplicon-negative cell lines, and an explicit go/no-go at each step. We are aiming for preliminary bench data by September. We have circled this idea for years. Claude Science made it real in a week. - **Team:** [Adam Pennycuick](https://cerebralvalley.ai/u/adampennycuick), [Jie Lam](https://cerebralvalley.ai/u/JamesLam) - **GitHub:** https://github.com/ucl-respiratory/amplicon-targeting - **Demo video:** https://youtu.be/_J201ZUocuU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=128 ### 153. Arya What I built: Ratify convenes diverse AI "labs" (each a small crew of persona-driven agents (a PI, members, and a critic)) running on different Claude models and seeded with a distinct scientific worldview to attack one open research problem, deliberate live, and produce competing executable method specs. It turns each spec into running code and scores it against held-out real data. I used it to ask: if you seed different scientific lenses, do you get different methods and do any of them beat a strong biological baseline when held to real evidence? The data came from the Krogan lab's cross-species coronavirus interactome and tried to predict from sequence alone, which host proteins a held-out viral protein binds, using their real SARS-CoV-2 / RaTG13 AP-MS data. What I found: Two things. First, seeding distinct worldviews produces measurably diverse methods: the labs' predictions were truly independent. Second, a simple baseline of "assume a virus grabs the same host proteins as its close relative" gets you most of the way there and is hard to beat. On the first pass, none of my labs beat it. Only after I handed each lab its own held-out scores and asked it to try again did one method (reasoning from molecular evolution) come out ahead. The tool surfaced an approach an ML-only scientist would have skipped. Why it matters: As more of science is handed to autonomous agents, the risk isn't too few ideas: it's confident, wrong ones. Ratify's answer is structural: nothing counts until it's executable and scored against evidence it never saw. That allows an idea generator to also become an idea filter. It also doubles as a way to test multi-agent setups against a real scientific yardstick, and every run leaves behind a chain of verified scientific reasoning. For a working scientist, it's a room of diverse, evidence-held collaborators on demand for you to interact with. - **Team:** [arya kaul](https://cerebralvalley.ai/u/ak1) - **GitHub:** https://github.com/aryakaul/ratify - **Demo video:** https://youtu.be/2JDCN6KQrkw - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=129 ### 154. Grounded Genomics I built a benchmark and a prompt-level "closed-world grounding" layer that measurably reduces hallucination when an LLM writes investigation summaries for genetic variants of uncertain significance (VUS). The problem: AI tools are increasingly used to summarize evidence for VUS, and when the evidence is thin they fabricate — inventing HGVS coding notation, wrong evidence values, misattributed citations, or clinical Pathogenic/Likely-Pathogenic calls the source pipeline never made. Any of these can push a curator, or a family, toward the wrong conclusion. This is personal for me: it comes out of my family's connection to the RASopathy world. The method: over 163 curated Tier-1 RASopathy variants, I run the same model on the same variant with two prompts — an unconstrained baseline and a grounded arm that injects only verified facts plus an explicit non-fabrication contract. Four dataset-anchored detectors then score every output, each tied to a verifiable property of the source data (e.g. the source contains zero HGVS c. notation, so any c. notation is fabrication by definition). The result: grounding raised the clean-output rate from 0.42 to 0.91 raw / 1.00 adjudicated, and calibration from 0.43 to 1.00, cutting mean fabrications per variant from 1.50 to ~0.11. I also report an honest raw-vs-adjudicated finding: my high-recall clinical detector produced 17 false positives in the grounded arm (hedged/negated language), so I publish both numbers rather than silently tuning the detector — the gap quantifies the detector's own calibration. Everything is reproducible: the repo rebuilds the published results from saved model outputs with no API key, and an interactive Space lets you see each fabrication highlighted inline, grounded vs ungrounded, for any variant. - **Team:** [Fernando Rodrigues](https://cerebralvalley.ai/u/Fernandosr85) - **GitHub:** https://github.com/Fernandosr85/closed-world-grounding | https://huggingface.co/spaces/Fernandosr85/closed-world-grounding - **Demo video:** https://youtu.be/sn2mBaPTQOE - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=130 ### 155. Signal Reversal · T-cell Reversal We asked a simple question of a genome-scale CRISPRi Perturb-seq screen of primary human CD4+ T cells (Marson & Pritchard labs): which single-gene knockdown pushes a T cell's transcriptome opposite to active autoimmune disease — i.e. mimics effective therapy — and can itself be drugged? For each of 11,281 knockdowns and three diseases (RA, psoriasis, MS), a Connectivity-Map-style weighted-GSEA score — calibrated with a permutation null and Benjamini–Hochberg FDR — ranks how well the knockdown reverses that disease's published treatment-response signature (RA 1,713 · MS 1,181 · psoriasis 7 significant). Genes significant in ≥2 diseases (238), a six-filter plausibility cascade, and Open Targets tractability leave a 40-gene shortlist — 7 directly targetable, led by ITK (with INPP5D, KAT6A, MAPK14, CD52) — several with approved-drug precedent. Cross-validating the top candidates against four more diseases (seven total) leaves exactly two universal regulators: TET2 and KAT6B. KAT6B is the striking result — a literature-blind screen independently nominated the KAT6A/B pathway in precisely the diseases (MS, IBD) where mouse genetics had already shown the paralog is causal, and a dual KAT6A/B inhibitor is already in clinical trials. A separate Th1/Th2 cell-state method converges on overlapping targets and validates the approach (AUROC 0.92/0.86; independent human-genetics enrichment p = 4.2×10⁻¹¹). Everything is FDR-controlled, deterministic (reproduced at Pearson r = 1.0000), and honestly caveated — all candidates are computational hypotheses for experimental validation, not confirmed drug targets. Data/security: the deliverable is analysis code plus one self-contained, offline HTML report — no server, no telemetry, no credentials or personal data; it uses only the public S3 Perturb-seq bucket, public GEO series, and the public Open Targets / STRING APIs. - **Team:** [Ze Yu](https://cerebralvalley.ai/u/THEU), [Dinesh Ravindra Raju](https://cerebralvalley.ai/u/Dinesh25) - **GitHub:** https://hackathon-reversing-autoimmunity.netlify.app/ (live report) https://github.com/EZUY/Hackathon_Claude_Bioinfo_260707 (source; reproduce with: bash src/run_pipeline.sh) - **Demo video:** https://youtu.be/1ov9PXek3gM - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=131 ### 156. Markus Sommer Determining whether a cancer patient responds to immunotherapy, detecting contamination in food and water, measuring mitochondrial membrane potential: scientists depend on flow cytometry. Experiments can measure thousands of cells; the bottleneck is analysis, where scientists gate by hand or simply give up on 40-marker spectral panels. The tools that could help are expensive, cloud-based, or need years of programming experience, leaving immunologists with sensitive patient data and limited coding experience stuck. WakaFlockaFlow closes that gap: a free, no-code tool that installs in one command and runs on the lab's own machine. Nothing leaves the building. Point it at raw files and it unmixes spectral signal, names cell populations, scores functional state, pools cohorts onto one map, and compares groups statistically. It exports gating paths to FlowJo as real, editable gates, fitting existing workflows rather than replacing them. Validation runs three ways. On synthetic ground truth it recovers signal, returns zero on a null cohort, and scales properly with effect size. On the Bodenmiller benchmark (real, hand-gated PBMCs) it reproduces the B-cell activation response (pS6 up, p 1.2e-6) and caught a real, documented annotation bug in the data. Live on PyPI today. - **Team:** [Markus Sommer](https://cerebralvalley.ai/u/markusjsommer) - **GitHub:** https://github.com/Markusjsommer/WakaFlockaFlow - **Demo video:** https://www.youtube.com/watch?v=NO2lN-pBV9I - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=132 ### 157. PPInteractors One-sentence summary: I rebuilt the SARS-CoV-2 host interactome from the Krogan lab's raw AP-MS data using a novel updatable confidence score that lets structural and chemical evidence rescue borderline interactions, and found that the strongest biology (centrosome hijack via PCNT) is not druggable while a clinical-translatability filter inverts the compound ranking to surface safer repurposing leads, showing that multi-evidence, clinically-aware scoring turns a raw interaction map into an honest, prioritized set of therapeutic opportunities rather than an overclaimed cure. What I built / investigated I built an end-to-end computational pipeline that turns the Krogan lab's raw SARS-CoV-2 AP-MS data into a clinically-filtered, ranked list of therapeutic opportunities, reconstructing the virus/host interactome from scratch rather than trusting the published hit list. Starting from 22,153 unthresholded bait-prey measurements (27 viral baits × 2,750 human proteins), the pipeline runs through six stages, each handled by a specialist role: data QC, network reconstruction, target biology, structural tractability, compound identification/docking, and a clinical and translational reality check. The methodological centerpiece is an updatable composite confidence score. Instead of a fixed pass/fail cutoff, every interaction carries a continuous posterior probability (log-odds space) that combines the non-redundant MS signals, and crucially lets downstream structural and chemical evidence plug in as extra likelihood terms, so a sub-threshold interaction can be recovered when orthogonal data corroborate it. What I found A better, re-iterable scoring metric. The composite posterior hits ROC-AUC 0.999 / PR-AUC 0.955 vs 0.618 for the best single published metric (MIST). The "updatable" design worked end to end: orf10-ELOB was rescued from the marginal tier into high-confidence on the strength of a solved structure (PDB 9BIE), and external complex-membership evidence rescued 61 marginal pairs into HIGH on grounds fully independent of the AP-MS scoring. The strongest biology is centrosome hijack. Three independent axes (AP-MS confidence, network centrality, and evolutionary positive selection via PAML) converge on PCNT (Pericentrin) as the #1 host target (positive-selection q = 5.4×10⁻²⁷). Several single viral baits each bind all subunits of one human complex (nsp1 to DNA polymerase-α/primase; nsp13/M to the Pericentrin-GCP centrosome; nsp8 to AATF-NGDN-NOL10). The clinical filter inverts the compound ranking. Purely biochemical ranking favors oncology drugs (nucleoside analogs, proteasome inhibitors). Adding a clinical-translatability term collapses those and lifts safer options (COMT inhibitors, an entry-inhibitor probe, fostamatinib). Final verdict on 17 near-term repurposing opportunities: 0 clean GO, 3 CONDITIONAL, 14 NO-GO. Notably, XPO1/selinexor, the only candidate ever tested in a randomized COVID-19 trial, was discontinued for futility and trended toward harm, exactly matching our NO-GO call. Why it matters The project delivers three honest categories of output rather than an overclaimed "cure": (1) near-term repurposing leads that are hypothesis-generating and need dedicated early-phase trials; (2) a differentiated novel-modality target, the CRL2^ZYG11B/orf10 viral neo-interaction, as the one axis with a virus-selective rationale; and (3) fundamental virology (the centrosome-hijack axis) that is publishable science but not a drug target due to essential-gene toxicity. Methodologically, it shows that multi-evidence, updatable scoring beats any single AP-MS metric, the right response to the field's well-documented low cross-lab reproducibility, and that baking a clinical-translatability term into target prioritization prevents pursuing biochemically-attractive but clinically-implausible candidates. Finally, the project does not stop at analysis: it delivers a complete, Cell-format manuscript, ready for internal review ahead of submission, with a full figure set (Figures 1 to 6 plus supplements), STAR Methods, and all underlying data tables, packaging the entire pipeline and its findings into a publication-ready document. - **Team:** [Pedro Miguel Belio Mairal](https://cerebralvalley.ai/u/pedrobelio) - **GitHub:** https://drive.google.com/file/d/1JCBE0c3UAtYuj42gB4qjL-eAmN3t7lOh/view?usp=share_link - **Demo video:** https://drive.google.com/file/d/1bnYumDdkSZp2ANEvG-rMbuZJV-G0pOcF/view?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=133 ### 158. Jaiaaryan WHAT WE BUILT • OCTB (Optical Compatibility Test Bench) — try a camera/lens/light swap in software before buying hardware, and prove the instrument still gives the same answer. • It simulates how the new hardware forms the image, runs the ORIGINAL legacy analysis algorithm unchanged, and checks the decision is preserved (same grade, successful decode, metric in tolerance). • Every value is provenance-tagged; the legacy grader is byte-frozen and hash-verified so its verdict can't drift. WHAT WE FOUND • A hardware swap can SILENTLY change the answer — the failure is invisible until after you've bought the part and re-run real assays. • Example the app makes obvious: a Red-630nm light turns every color patch red and wrecks the grader — caught before a cent is spent. WHY IT MATTERS • Legacy machines from the 1990s–2000s still work and produce value, but their specific camera/lens/lamp goes end-of-life. • You can't just drop in a replacement (software was tuned to the old optics), and ripping out the system means costly re-engineering + revalidation. • Today qualifying a part means buy → install → run assays → compare → repeat: weeks per candidate on a validated instrument. • OCTB does the first pass in software: screens the whole market, ranks the few parts worth physically testing, and builds the evidence report — without touching the trusted legacy system. - **Team:** [Jaiganesh Srinivasan](https://cerebralvalley.ai/u/Jaiaaryan) - **GitHub:** https://github.com/Jaiaaryan/OCTB - **Demo video:** https://youtu.be/IGHZtwAKt3w - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=134 ### 159. Luvia Enormous AI spend goes into building biomedical knowledge graphs — and almost all of it is private or paywalled, rebuilt from scratch behind every company's walls and stale within a year. Meanwhile researchers everywhere are already reading the same literature with LLMs. OpenClaims starts from that waste: if your agent is analyzing a paper anyway, it might as well log what it found, in a structure a crowd can actually query. Every reading makes the next person's question better answered — the opposite of a graph that dies in one lab. The user we built for is any researcher whose agent already reads papers. That agent files what it found — typed (subject, predicate, object) claims with polarity, provenance, and the exact grounding sentence — through an MCP server, and the commons adjudicates. Nobody edits the graph; the graph re-adjudicates itself. Three guarantees make it trustworthy: quoted spans are checked verbatim against the source, so fabrications are rejected at write time by non-LLM code; standing is recomputed on read and never stored, so a wrong claim is outweighed in the open rather than silently deleted; and no model decides what's true — entity resolution is deterministic and contradiction is structural. We seeded it with 494 verified readings across 25 papers on the microglial cGAS–STING–interferon axis. Out of that corpus it flags 11 genuine contradictions (e.g. STING→NLRP3 asserted positive in one paper, negative in another) and composes 25 cross-paper chains stated in no single source — for instance, Huntingtin → (↑cGAS, PMID:32581130) → (↑type-I-interferon, PMID:37095396) → an emergent HTT→interferon link neither paper makes, grounded verbatim in both. These are candidate hypotheses the structure surfaces, not validated findings — and the commons labels them exactly that. That honesty is the point: it tells you where the literature already disagrees and where a testable connection is hiding, across papers that never cite each other. - **Team:** [Elis Samskog](https://cerebralvalley.ai/u/Elis) - **GitHub:** https://github.com/Luvia-AB/openclaims - **Demo video:** https://youtu.be/Rxa8of6R50Y - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=135 ### 160. BALL-RL I built BALL-RL: Bayesian Anchored Latent Learning for Reinforcement Learning, a framework for learning personalized treatment policies when the clinically meaningful patient state is measured only intermittently. Our application focuses on repetitive transcranial magnetic stimulation (rTMS), in which clinicians repeatedly select from protocol, intensity, and laterality options, but validated symptom questionnaires are collected only a few times during a treatment course. BALL-RL uses these sparse questionnaires as clinical anchors for a Bayesian teacher-student decoder that reconstructs a continuous, forward-only severity trajectory from dense clinical records. That trajectory provides both the patient state and the session-level reward needed to train a conservative offline RL policy. I evaluated the complete system in a digital twin with known patient states, optimal actions, and counterfactual outcomes. BALL-RL recovered latent severity accurately and enabled the learned policy to close 97.2% of the performance gap between a naive sparse-questionnaire learner and a finite-sample oracle trained with the true latent reward. The benefit remained when BALL-RL was restricted to the same number of training transitions as the sparse comparator, demonstrating that the improvement came from the reward's denser, less biased temporal decomposition rather than simply from having more observations. I then applied the framework to 1,758 observational rTMS courses from 1,508 patients across 15 facilities, using patient-grouped, facility-held-out evaluation. Both fitted-Q evaluation and instrumented off-policy evaluation favored the learned recommendations over observed clinical decisions. AT-OPE estimated 0.350 latent improvement units per course left unrealized when care departed from the learned policy, equivalent to approximately 51.5% of the cohort’s average decoded course improvement. Finally, I built a clinician-facing synthetic-data demonstration with Claude integrated as an interactive chatbot for interpreting the RL model’s decisions. The interface presents the decoded patient trajectory, uncertainty, recommended protocol, supported alternatives, safety status, and SHAP-based drivers. Clinicians can ask Claude why one protocol was favored, why another was not selected, which patient-state features influenced the decision, or how uncertainty affected the result. Claude answers using a tightly constrained payload of model outputs, translating Q-values, feature attributions, and latent-state estimates into understandable clinical language. It does not select, modify, or rerank treatments: the verified RL policy makes the decision, while Claude helps the clinician understand and interrogate its reasoning. FHIR and CDS Hooks interfaces additionally demonstrate how this explainable decision-support workflow could integrate with an EHR. This matters because sparse outcomes are the rule rather than the exception in health care. Most clinical AI predicts a single outcome from a snapshot, but treatment is sequential: clinicians act, observe an incomplete response, and adapt. Without a reliable session-level state, conventional RL cannot determine which action produced improvement or learn when a treatment should be continued, intensified, switched, or stopped. BALL-RL addresses that missing layer by converting occasional validated measurements and routinely collected clinical records into a dense, uncertainty-aware state while preserving the clinical meaning supplied by the questionnaires. The result is a practical bridge between real-world clinical data and sequential decision optimization. It allows health systems to learn from treatment courses that have already occurred, without assigning exploratory treatments to patients, and produces recommendations that can be inspected, explained, support-constrained, and withheld when safety concerns arise. In rTMS, this creates a path to adapt protocol choices throughout a course rather than applying a largely fixed strategy and waiting weeks to evaluate the results. More broadly, the same problem arises wherever meaningful outcomes are delayed or intermittent, including medication titration, rehabilitation, chronic disease management, intensive care, and behavioral health. BALL-RL provides a general framework for turning sparse endpoints into actionable learning signals, enabling personalized treatment sequences in settings where existing measurement practices previously made them inaccessible. - **Team:** [Jason Gibbons](https://cerebralvalley.ai/u/Jsngibbons) - **GitHub:** https://github.com/jbgibbons093/ball-rl - **Demo video:** https://www.youtube.com/watch?v=AxWox7L4GCU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=136 ### 161. TrialBridge - GlobalTrialBridge.com What I built. TrialBridge is a two-sided clinical-trial feasibility layer for Brazil, built entirely with Claude Code — by a founder who has never written a line of code. A sponsor pastes a ClinicalTrials.gov protocol; Claude parses the eligibility criteria into typed rules with its reasoning shown for every criterion; a direct-standardized estimator turns 163GB of public DataSUS data into state-by-state eligible-population estimates with 95% confidence intervals, rendered as an interactive Brazil map and a ranked site shortlist. Sites answer feasibility privately — no patient record ever leaves their walls. It's live, deployed, access-gated, and ships as an MCP server so the same query can be asked inside Claude. What I found. Two findings. A hackathon survey (N=50, run via Claude + Prolific over MCP) showed 70% of pharma/biotech buyers had never heard of Brazil's Lei 14.874/2024, and one paragraph about it lifted their intent to run trials there by +1.39 points (p<.001). And our own calibration harness found the model calibrates near-perfectly in-distribution but breaks across hospital sites — a limit we report openly. Why it matters. Trials are running out of patients where they've always run. Brazil has 214M people, a new law that cut approval from >12 months to ~90 business days, and costs up to 75% lower — but the market is pricing it on outdated information. TrialBridge closes that gap with provable numbers, at the feasibility stage where the decision is actually made. - **Team:** [Angelo Orru Neto](https://cerebralvalley.ai/u/AON) - **GitHub:** https://github.com/aondaai/trialbridge - **Demo video:** https://cap.link/xy360v373ve4f2c - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=137 ### 162. Metavert Aitiome grades whether an environmental chemical has mechanistically-supported links to Parkinson's or Alzheimer's. Give it a chemical and it reconstructs the OECD-endorsed causal pathway (MIE→KE→AO), grounds each edge in queryable evidence, and grades it on curated evidence — never bioactivity, with calibrated confidence. On a curated adversarial benchmark it recovers 13/13 PD and 12/12 AD known neurotoxicants and rejects 6/6 mitochondria-active decoys with zero errors. The key finding, computed live from our own data: bioactivity is anti-diagnostic — every activity signal separating real neurotoxicants from decoys is at or below chance, while the curated rule is perfect. It ships the discovery limits as an honest map, adds a wet-lab candidate-triage queue, and exposes the whole engine over MCP for agents. It also includes a methods study: adversarial RLM surfaced ~10× more counter-evidence than RAG. - **Team:** [Jon Radoff](https://cerebralvalley.ai/u/jradoff) - **GitHub:** https://github.com/jonradoff/aitiome - **Demo video:** https://www.youtube.com/watch?v=wZ7HyH8nqFM&feature=youtu.be - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=138 ### 163. Conduction Lens Cardiac models are routinely fit to an ECG and their conduction parameters reported, but a prior question is rarely asked: which of those parameters can an ECG actually determine? A model can fit perfectly and still carry no information about a parameter it never constrained. I built a calibrated, amortized characterization of that identifiability. A single neural posterior estimator is trained over seven His-Purkinje and myocardial conduction parameters at fixed anatomy, from a simulated 12-lead ECG, and audited with simulation-based calibration, expected coverage, and TARP. Against a stated waveform noise floor (white Gaussian sigma 0.025 mV per sample per lead), four of the seven carry information: interventricular delay and myocardial velocity are well resolved, RV initial extent and conduction velocity are moderate, and LV extent, branch angle, and branch repulsivity stay diffuse, no tighter than the prior. The parameter the ECG resolves best, interventricular delay, is exactly the timing clinicians program into resynchronization pacemakers, and the honest flip side is that a fitted value for a diffuse parameter is a prior belief, not a measurement. The result is simulated (no patient ECG) and conditional on the forward operator and the noise floor, and the calibration audit caught three of my own errors before they shipped. Anyone fitting conduction models to the ECG can reuse the calibrated identifiability recipe (contraction against a stated floor, audited by SBC, coverage, and TARP) to say which of their own reported parameters are measurements and which are prior beliefs, and the released weights, sweeps, calibration artifacts, and verification ledger make that reuse turnkey. The same pipeline already runs on the public Strocchi cohort, so extending it to new anatomies is a config change, not a rebuild. - **Team:** [Ricardo Garcia Ramirez](https://cerebralvalley.ai/u/ricardogr07) - **GitHub:** https://github.com/ricardogr07/ecg-purkinje-npe - **Demo video:** https://www.loom.com/share/a55d0385302942a3ab46fb01087d79af - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=139 ### 164. Team Glass Box I built HAR-annotator, a glass-box pipeline for ranking Human Accelerated Regions — short stretches of DNA that stayed frozen across mammals for a hundred million years, then changed fast only on the human lineage, making them prime suspects for what reshaped the human brain. There are thousands of them, and which ones matter depends entirely on the question you bring. That's the whole design: the pipeline doesn't hand you one fixed answer, it lets you ask a question and shows its full reasoning. The question I brought was just one instance (which HARs are relevant to neurodevelopment and disease?) and the tool is built so anyone can bring a different one. The pipeline works as a funnel followed by a score. Starting from all 3,257 HARs (from this Nature paper https://www.nature.com/articles/s41586-025-08622-x), three biologically motivated gates narrow the field based on information from 5 datasets: keep what’s genuinely conserved across mammals (mean 241-way phyloP above 1.0), then what sits near a known neurodevelopmental-disease gene (proximity within 1 Mb), and then what overlaps a neuropsychiatric GWAS signal. This funnel 3257 HARs into 363 candidate HARs of interest. Each of the 363 survivors is then scored on seven normalized axes combined into one weighted sum: in my investigation into neurodevelopment and disease relevance, gene link and disease genetics carry the most weight (0.22 each), then constraint (0.18), fetal-brain chromatin activity and developmental timing (0.13 each), and human–chimp acceleration and TF-motif disruption weighted low (0.07 and 0.05) because they're first-order or model-based signals. Every axis's contribution ships as its own column, so each rank decomposes into its reasons: the top hit near ZSWIM6 is carried by conservation and gene link, while a lower hit near RPS23 has a strong gene link but barely any conservation. Nothing is hidden in a model. What I found for my question: at the top is a HAR sitting near ZSWIM6, a brain-development gene. It ranks first because it's deeply conserved, carries the strongest human–chimp acceleration signal in the whole top tier, and sits in open, active chromatin in the developing brain. Just behind it, essentially tied, are HARs near TCF20 and POC1B, and each earns its rank differently: TCF20 is a well-established neurodevelopmental-disorder gene, while POC1B rides mostly on its disease-signal axis. Because the score decomposes, you can see exactly why each one placed. The broader shortlist recovers a who's-who of neurodevelopment genes (ZEB2, TCF4, MEF2C, FOXP2, SOX5) without any of them being hard-coded in, which is a good sign the scoring is tracking real biology. And it's robust: across 20,000 random weightings the ranking barely moves (see the weight_eval folder in my GitHub). Most strikingly, it isn't just resurfacing famous genes, ZSWIM6 ranks first despite having only 27 abstracts on PubMed, so the pipeline is pointing at genuinely under-studied biology rather than re-confirming what's already well known. But the more important result is that the tool answers other questions too. Relax the funnel and the candidate set opens up (577 elements if you drop the rare-disease requirement, 2,757 on conservation alone) and re-scoring reshuffles the podium (TCF20 overtakes ZSWIM6, new genes appear). Change the weights and it re-sorts around what you value. Same evidence, different question, different defensible shortlist. Overall, the same HARs that may have made us human also keep showing up next to genes for autism and schizophrenia which means that the thing that built us may be the same thing that makes us vulnerable. This is a reproducible tool that turns thousands of undifferentiated candidates into an auditable shortlist (and reshapes that shortlist around whatever question a researcher actually has). This can hopefully serve as a hypothesis generator many labs can point at their own biology, not just a one-off analysis of mine. - **Team:** [Natalie Chen](https://cerebralvalley.ai/u/nataliechen) - **GitHub:** https://github.com/nataliegits/har-annotator/ - **Demo video:** https://youtu.be/FZlXeXz5NmA - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=140 ### 165. Hokie What I built Clinical trials fail for predictable reasons. The problem is that teams usually discover those reasons after months of work and millions of dollars. I built TrialPremortem, an autonomous agent that stress-tests a clinical trial before the first patient is enrolled. Give it a draft protocol and it does two things. First, a validated machine-learning model estimates the risk that the trial will fail to recruit enough patients. Then the agent investigates why. It generates hypotheses specific to that trial, searches live biomedical sources including ChEMBL, Open Targets, PubMed, and ClinicalTrials.gov, and gathers evidence for or against each one. Before anything reaches the user, an adversarial critic challenges the findings and a programmatic citation gate checks every factual claim against the actual tool output. The result is not a generic protocol review. It is a ranked list of the most likely ways the trial could fail, the evidence behind each one, and a quantified recommendation for what to change. What I found The core result is a validated metric, not an opinion. I reconstructed the real version history of trials on ClinicalTrials.gov. For each trial, I recovered the original protocol, hid everything that happened afterward, made a prediction using only that first draft, and then checked what actually happened. Across 927 real trials, the model achieved an AUC of 0.67. It flagged 43% of the trials that later failed because of poor recruitment, while flagging only 21% of the trials that succeeded. I also found an outcome-leakage bug involving achieved enrollment. Had I left it in, the reported AUC would have been 0.82. That number was wrong, so I removed the leak and report the honest result instead. One of the more interesting findings came from the ablation study. On the hardest matched cohort, the engineered ML model scored 0.61 AUC. A zero-shot agent reasoning from the protocol scored 0.55. That is why TrialPremortem does not ask an LLM to magically predict trial failure from scratch. The agent uses the validated ML model as a tool, then investigates and explains the result. And on just 43 real failed trials that the system successfully flagged, I estimate a conservative $7.8 million in avoidable cost. Why it matters Most clinical trial protocols are reviewed again and again before launch. They still fail. Roughly 76% of protocols are amended after finalization, with each amendment costing an estimated $141,000 to $535,000. Insufficient accrual remains the leading cause of trial termination, accounting for about 39% of terminated studies. This is not because trial teams lack expertise. A protocol may already pass through five to seven expert reviews. The problem is structural. No individual reviewer carries the base rates of thousands of similar trials in their head. And the feedback loop is painfully slow: design a trial, launch it, wait 12 to 24 months, discover what went wrong, and hope that knowledge somehow makes it back to the next team. Usually, it does not. TrialPremortem closes that loop. It gives trial designers something they have never really had: institutional memory across thousands of past trials, applied to the protocol in front of them before the expensive mistakes happen. And the prediction is backed by a metric that anyone can reproduce using public data. - **Team:** [Ragul Seetharaman](https://cerebralvalley.ai/u/Ragul) - **GitHub:** https://github.com/ruuls/claude_lifesciences - **Demo video:** https://claude.ai/code/artifact/4b76978b-9ad1-4915-93b7-7af1ed625483 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=141 ### 166. vcf2report - gbbarra vcf2report turns a raw exome VCF + the patient's phenotype (HPO terms) into an auditable, clinical-grade ACMG variant report — locally, offline, in seconds. Exome interpretation today is slow and manual: an analyst hand-filters thousands of variants across several databases and matches each candidate to the patient's phenotype one by one. vcf2report runs the whole pipeline — parse ▸ QC ▸ annotate ▸ prioritize ▸ classify ▸ report — and returns a draft laudo for expert sign-off, showing the full ACMG reasoning for every variant: all 28 ACMG/AMP criteria, each with its concrete evidence value, its source (database + accession + date), and whether the deterministic engine decided it or it is honestly deferred to expert/model review. What makes it different from a black-box classifier: • Auditable — you see *why* a variant is Pathogenic; the trio/segregation criteria a single-proband VCF cannot support are marked N/A instead of guessed. • Brazilian population frequencies (ABraOM/SABE) on top of gnomAD, so a variant absent from gnomAD but common in Brazilians is correctly dropped. • ClinGen-calibrated AlphaMissense (PP3/BP4), deterministic PVS1 via the ClinGen SVI tree, and ontology-aware HPO phenotype matching (Lin/Information-Content). • Offline gnomAD: a 29.6M-variant DuckDB/Parquet store classifies a whole exome in ~7 s with the VCF never leaving the machine. Why it matters: validated against 200 real ClinVar variants (100 pathogenic / 100 benign, with ClinVar withheld from the engine), it produced 0 gross pathogenic↔benign flips and 100% precision when decisive. It is deliberately conservative — never dangerously wrong, and right whenever it commits — which is exactly the property a draft-generation aid for a clinician needs. - **Team:** [Gustavo Barra](https://cerebralvalley.ai/u/gbbarra) - **GitHub:** https://github.com/gbbarra/vcf2report - **Demo video:** https://youtu.be/QF3twmeWA2U - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=142 ### 167. GroundTruth What I built. Clinical AI Eval Designer turns a clinical AI's intended use — the model, the patient population, the care setting, and the claim it makes — into a structured, citable validation specification: eight fields (study design, sensor/input validation, performance benchmarks, ground truth, sample size, subgroups, regulatory pathway, post-deployment monitoring), each with a recommendation, a cited rationale, a HIGH/MEDIUM/LOW confidence flag, and the expert sign-off it still needs. It runs as one pipeline, three stages: 1) Retrieval — plain code, no model. It queries ClinicalTrials.gov, openFDA, and the literature before anything is written, so every citation is a re-resolvable identifier (PMID, NCT number, FDA K-number/product code) — verifiable by construction, not recalled from memory. 2) Synthesis — Claude writes the spec. Claude Fable 5 (Opus 4.8 fallback) maps the retrieved records into the eight fields under one hard rule: no invented numbers — cite it, or flag it for the study team to set. 3) Review — a three-persona panel. An optional pass has three Claude-played reviewers (regulator, biostatistician, clinical scientist) critique the spec under the same cite-or-flag discipline. What I found. Trustworthy output takes engineered discipline at two boundaries, not one. First, retrieval has to be deterministic. The moment the model fetches its own evidence, the same input returns slightly different sources run to run — and those differences cascade into everything the spec cites; my ablation runs showed recall failures degrade the whole downstream result. So retrieval is plain code with re-resolvable identifiers, reproducible by construction. Second, deterministic retrieval still isn't safe on its own: hand an unconstrained model even perfect evidence and it produces confident, fabricated benchmarks. So synthesis runs under a constraint layer — cite-or-flag, confidence tiers, and naming-and-rejecting off-target records instead of citing them to look thorough. Guard only one boundary and the spec looks authoritative while being subtly false. Guard both, and weeks of expert literature work compress into a grounded starting point in minutes — while staying honest about what still needs a human. Why it matters. A clinical AI can't reach the patients it's meant to help until it clears the FDA — a slow process with no single playbook that fits every model, where the right validation evidence changes by indication, population, and claim. A model can perform well in the lab and still stall on the way to the clinic, because assembling that evidence is slow, manual, and specific to each case. This tool turns weeks of that groundwork into a rigorous, grounded starting point — and an expert-review conversation — with every claim tied to a source the team can re-check. The principle it demonstrates is the one the demo closes on: verify, don't blindly trust — honesty enforced in how the model is allowed to reason, not a surface guardrail: a "not medical advice" disclaimer or warning banner bolted onto the interface while the model underneath could still fabricate. How I validated it. I built a hand-verified answer key for 10 diverse clinical AI cases (devices, drugs, biologics; regulatory-approved and regulatory-null) and scored the live pipeline on precision and recall — never raw counts. To make retrieval trustworthy I ran an ablation study — varying one component at a time across all 10 cases to confirm each earned its place and to lock the best default — including a disease-aware search that recognizes the condition even when the input phrases it indirectly, and a bridge from that condition to the matching FDA product codes and cleared devices. Every one of the ~470 scored identifiers re-resolves against its live registry today, and across the slate the tool cites real cleared devices where they exist and names-and-rejects off-target records everywhere else — for zero fabricated predicates. (Full method: eval_results/ablation_findings.md.) - **Team:** [Victoria Wang](https://cerebralvalley.ai/u/VictoriaxWang) - **GitHub:** https://github.com/victoriaxxwang/clinical-ai-eval-designer - **Demo video:** https://youtu.be/znmxWjoSaGk - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=143 ### 168. Structureless Heart failure switches off OPLAH—a protective enzyme that clears a toxic metabolite—and no one had mapped where its known activator binds because OPLAH has no solved structure. No binding site, no way to design a better activator. We built a pipeline that finds the site without assuming it. From an AlphaFold model of OPLAH (1,288 aa, no experimental structure), we detected every pocket across the whole protein with two geometry methods (P2Rank, fpocket), then docked 5′-AMP — the activator our team identified earlier in a 1,280-compound screen — completely blind, with no pocket specified. Three independent methods converge on the same site: 17 residues fold around it, with a D-x-G-G-T phosphate motif repeated twice. Then we screened 5,000 compounds against that pocket. The top hit (ZINC4126706) shares almost none of AMP's chemistry yet outranks it, and 379 novel scaffolds beat AMP overall a shortlist to test at the bench, not measured affinities. Why it matters: nothing in the method is specific to OPLAH. Any target with a predicted structure and one known modulator can run the same unbiased search and thousands are stuck exactly where OPLAH was - **Team:** [Jose Esquivel](https://cerebralvalley.ai/u/josan) - **GitHub:** https://github.com/anthony-89/structureless-target-screen - **Demo video:** https://youtu.be/5w6GIWi7tHE - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=144 ### 169. Sarah Traore Myomatlas is an open, image-based decision-support tool for abnormal uterine bleeding, one of the most common and most under-addressed problems in women's health. Heavy menstrual bleeding affects up to one in three women and is a leading cause of anemia, yet when fibroids are the cause, their exact location, which determines the right treatment and whether the uterus can be preserved, is often reported inconsistently and rarely explained to the patient. I built an end-to-end pipeline on the open UMD MRI dataset. A U-Net I trained segments the uterine wall, cavity and each fibroid from a pelvic MRI and measures each lesion's size and its position relative to the cavity and serosa. A Claude agent then reasons strictly from an encoded FIGO and PALM-COEIN guideline to classify each fibroid, show its justification, flag when malignancy exclusion is warranted, and derive management options, including uterus- and fertility-preserving ones. The result is presented as an interactive 3D view, a structured clinical report, and a plain-language patient explanation, with a PDF export. The key finding is that the same scan should not always lead to the same recommendation. When I add patient context, a 34-year-old who wants children versus a 58-year-old with postmenopausal bleeding, the agent's malignancy assessment and its ordering of treatment options change accordingly, from uterus-preserving management to flagging for malignancy exclusion. This matters because it makes the location-to-treatment-to-fertility logic explicit and understandable, for both clinicians and patients, in an area of medicine where women are too often unheard. - **Team:** [Sarah T](https://cerebralvalley.ai/u/CybaroseDev) - **GitHub:** https://github.com/Cybarose/myomatlas - **Demo video:** https://www.youtube.com/watch?v=LOoYQ1xTpTA - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=145 ### 170. T-CTRL Perturbation screens often rank genes by effect size, so a gene that changes many transcripts can be mistaken for one that actually controls cell state. We built T-CTRL, an auditable workflow that computes ISCI, our Immune-State Controllability Index, and asks whether functional direction and cross-donor repeatability add information after effect magnitude is already known. In the Marson CD4+ T-cell Perturb-seq screen, the pre-specified full-sample M→M+C test improved recovery of known regulators by +0.357 AUPRC (95% CI [+0.117,+0.538]). A separate, stricter, fully refit leakage-free out-of-fold estimate was +0.215 ([+0.074,+0.560], permutation p=0.010). These are distinct measurements, and we report both. Stress tests deliberately bound the claim. A broad external functional-regulator set fails (ΔAUPRC −0.281); the CAR-T clinical-prediction analysis is NULL; and scGPT corroboration is NOT-EVALUABLE because required inputs were unavailable. Marson CD4+ passes, Schmidt CD4+ and THP-1 are near-misses, and non-immune K562/RPE1 systems fail. The result is therefore a tested scope map, not a universal controller score or target list. T-CTRL ships as a Python CLI, reusable DatasetSpec framework, executable notebook, interactive demo, claim ledger and provenance-bound evidence package, with 151 tests and 21/21 automated release gates. - **Team:** [Abel Costa](https://cerebralvalley.ai/u/anetoc) - **GitHub:** https://github.com/anetoc/ISCI-hackathon - **Demo video:** https://youtu.be/7Rz4PpmQZuI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=146 ### 171. Aimie Garces Most new drugs fail, and the most expensive mistake is made years before a trial begins: choosing the wrong target (the specific protein a drug is designed to act on) for a disease. This is worst in the brain and nervous system, where success rates are the lowest in medicine. Two decades of Alzheimer's programs aimed at amyloid plaques failed one after another, and a Huntington's program to lower the very protein that causes the disease was halted in late-stage trials despite the target being genetically certain. The science was plausible and the money was huge; the target choice still did not pay off. This project asked whether the clinical success of a target-disease pairing can be predicted from the basic biology of the target protein alone, before anyone spends a dollar on a trial. A set of models that score how likely a target-disease pairing is to succeed, trained on a purpose-built dataset of 6,843 brain-disease pairings that captures not just the success stories but the far larger set of quiet failures most databases leave out. Each protein is described using bioinformatics: its amino-acid sequence, predicted 3D structure, likely biological functions, and how intolerant its gene is to mutation. These descriptions deliberately leave out the after-the-fact clues many other models rely on, such as how much a target has already been studied, since those clues do not exist for a brand-new target and flatter a model's apparent performance. The dataset and feature pipeline were built with an AI-agent workflow on Claude Science. A protein's own biology carries a real, if partial, signal of clinical success. The model does not give a yes-or-no verdict; it ranks candidates from more to less promising, and ranks a true success above a failure about 85% of the time, against 50% for a coin toss: useful for sorting a shortlist, not a guarantee for any one target. The signal is genuine, not luck, since scrambling the true answers dropped performance to chance. The result that matters most is the comparison to existing tools. They report scores above 90%, but on an easier test that lets the model lean on stand-ins for its answers, such as a new target for a disease it has already studied. Those tools depend on a target's "association score," a summary of how strongly databases already link a gene to a disease. Tested the honest way, training on the past and predicting which targets entered trials next, that score barely beats a coin toss (about 56%). This model avoids the shortcut and still reaches about 70% on that same fair footing. It also delivers what the high-scoring black boxes cannot: a prediction whose reasoning a person can read. drug-development team commits hundreds of millions of dollars per program, and no committee can responsibly kill or greenlight one on a number it cannot examine. This project shows a transparent model, whose reasoning a person can read, matched the black box with no measurable loss. An end user can see why a target scored well, sanity-check it against known biology, catch the model leaning on something spurious, and be told when a target falls outside what it can reliably judge. The result is a target-ranking tool that is accurate, inspectable, and honest about its limits, aimed at one of the costliest decisions in medicine. - **Team:** [Aimie Garces](https://cerebralvalley.ai/u/Aimie_Garces) - **GitHub:** https://github.com/Eleftheria14/predicting-cns-target-validation - **Demo video:** https://youtu.be/2Urf21q4VdY - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=147 ### 172. 4nY4 Thirteen real BCG trials. Pooled risk ratio 0.49. AskBench refuses to call it one number. I² is 92.1%. No model in the loop. One command reproduces it. That is the product: a bench scientist asks their data a plain-English question and gets a verdict, not a confident paragraph. SOLID or FLAGGED, with the Skeptic's reason on top. Six cells with a strong effect? Still flagged. A combined maternal risk that would imply 673 per 1000 pregnancies? Refused. The stats come from a fixed Python toolkit; Claude reads messy questions and narrates the argument. It does not touch a p-value. We measured it. Two hundred seeds, planted traps, zero API credits: structural traps caught every time; statistical traps land in the low 90s; 1.58% false positives after Benjamini-Hochberg, reported in the README, not buried. Same Skeptic, new traps it was never tuned on: still catches them. Judges can rerun python3 eval.py and python3 real_data.py themselves. Shipped as a live demo and an MCP server, so Claude can call the Skeptic inside a session. Built this week with Claude Code. - **Team:** [Anya Chueayen](https://cerebralvalley.ai/u/4nY4) - **GitHub:** https://github.com/anyapages/askbench, https://askbench-weld.vercel.app - **Demo video:** https://youtu.be/heNUcS1DftI?si=hyB76aD4fNKJz7-U - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=148 ### 173. ChannelScope ChannelScope builds standardized, provenance-tracked context-graph objects for multi-omic biology — the structured groundwork that purpose-specific "omic" models need but rarely have. Multi-omic data arrives unstandardized, unlinked, and unprovenanced; ChannelScope assembles it into one portable, gene-agnostic object where every modality — structure, sequence, population genetics, energetics, function, clinical evidence — attaches to shared, typed entities (a variant, a residue, a conformational state), with a citation and a confidence flag on every claim. Standardization is what lets these objects compose — across omic layers, sources, and proteins — and it is FAIR by design (Accessible / Interoperable / Reusable strong; Findable honestly partial). We prove it on the hardest honest test: RYR1, the ~2.2 MDa calcium-release channel behind malignant hyperthermia (MH) and the congenital myopathies — a ~5,000-residue-per-protomer tetramer that gates between closed, primed, and open states, beyond de-novo folding servers and known only through cross-species cryo-EM. Given a gene and a missense variant, ChannelScope assembles the best-available experimental evidence per conformational state, maps the variant onto each, and emits the object + a human-readable report + a self-contained interactive 3D viewer. It assembles evidence; it does not fold de novo — every region carries its template, species, and a confidence flag. Why this is different: A pathogenicity scalar (AlphaMissense) gives no where, no what-it-touches, no state; AlphaGenome reads regulation, not protein structure; and folding the mutant is no answer — one substitution barely moves a predicted backbone (Buel & Walters 2022; WT-vs-mutant Cα RMSD 0.1–0.6 Å). The sharp point: a pathogenicity scalar and a folding ΔΔG share the same gain-of-function blind spot — both track fold stability, which gating variants barely change. ChannelScope adds the conformational-state + interface layer that resolves exactly that class — e.g. T4826I, pathogenic for MH yet ΔΔG-stabilizing, which a scalar/stability tool miscalls. Validated, and honest about scope: On an 18-variant literature benchmark — independently re-derived through a second code path (Biopython + Biotite; proximity matched ≤ 0.1 Å) — it reproduced 18/18 numbering and 14/14 pathogenic structural buckets and mechanism directions; the engine carries 92 passing tests. It runs on one canonical protein frame (UniProt P21817 SV3 / RefSeq NM_000540.3) with protein-HGVS input — auto-normalizing arbitrary clinical input is a scoped roadmap item, flagged deliberately. Gene-agnostic (RYR2, CACNA1S, titin are config nodes, not rewrites). We intend to put ChannelScope to use at the 5th Undiagnosed Hackathon (Wilhelm Foundation · Singapore · Sep 17–20, 2026). Research / interpretation-support tool — not a diagnostic device. - **Team:** [Rajiv Sangle](https://cerebralvalley.ai/u/rajiv_sangle) - **GitHub:** https://github.com/Rajiv-Sangle/channelscope - **Demo video:** https://youtu.be/EW113Vq_CRk - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=149 ### 174. Representation Learning Duck I have found a principled way to recover known and find new drug targets using the t-cell perturbation data for any disease whilst highlighting the interpretability of how those drug targets affect the proceeding biological programs. I have borrowed methods from the field of causal representation learning to build relationships between the latent programs and found how perturbations would cascade down programs and how it effects each of them in a connected manner. This helped with the interpretability of the results of novel drug targets found for SLE and other potential diseases. Moreover, I have also built an agent that can be used to run the whole pipeline with any disease in mind, end to end, providing the results and potential novel drug targets. - **Team:** [Dylan Setiawan](https://cerebralvalley.ai/u/Thundehr) - **GitHub:** https://github.com/d-setiawan/Interpretable-TCell-Representation-Learning - **Demo video:** https://youtu.be/uexkKOe5-V4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=150 ### 175. Victims of Optimization The same search that produces adversarial examples—driving a model’s prediction up by making small changes to its input—can be turned into a tool for designing biological sequences. We train a small model to predict a property (the surrogate) and then run an attack-style search to find sequences the surrogate scores highly. On its own this search cheats: it finds sequences that fool the surrogate into predicting a high score rather than sequences that are genuinely good. We make the search useful by adding two things. First, a simple biological rule that keeps every proposed sequence realistic, using no extra trained model: each change from the starting sequence must be conservative under the BLOSUM62 (BLOcks SUbstitution Matrix) table, and a set of whole-sequence biophysical numbers (hydrophobicity, net charge, aromatic and charged fractions, sequence complexity, and the longest repeated-letter run) must stay within the ranges seen in natural sequences. Second, and most important, we check every design against signals the search never used—so the final claim never rests on the surrogate grading its own work. We demonstrate the system on four tasks across two molecule types: protein thermostability and green fluorescent protein (GFP) fluorescence, a 600-base-pair regulatory-DNA task verified on a true functional model (Borzoi), and a small DNA benchmark with an exact ground-truth answer (TF-Bind- 8) used as a diagnostic. The biological rule keeps 100% of GFP designs realistic versus 5–19% for the baselines (paired Wilcoxon p<10−7); on thermostability it matches a learned second model and beats a hand-tuned penalty (p<0.001); and on regulatory DNA the system raises the true property on all 40 of 40 starting points. Independent checks (a structure predictor, a separate thermostability predictor, kept functional residues, and a true DNA-activity model) agree the designs are real. Which featurization the surrogate uses (a frozen foundation-model embedding or a plain one-hot/k-mer encoding) is a swappable part of the system, reported as an ablation, not the point: the foundation model only helps on long, variable-length protein, and the system works either way. - **Team:** [Ajay Rangarajan](https://cerebralvalley.ai/u/armandyam), [Jeyashree Krishnan](https://cerebralvalley.ai/u/jkrishnan) - **GitHub:** https://github.com/armandyam/attack-as-design/tree/main - **Demo video:** https://github.com/armandyam/attack-as-design/blob/main/video/attack_as_design_narrated.mp4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=151 ### 176. AstroOmix AstroOmix is a Claude-powered research application for analysing space-biology data without overstating what limited datasets can support. Researchers can upload their own data or explore real NASA studies through two workflows: differential-expression analysis for replicated spaceflight RNA-seq experiments, and longitudinal modelling of physiological measurements across pre-flight, flight and recovery phases. Deterministic tools perform the statistics, while an embedded Claude agent explains results, answers study-specific questions and retrieves relevant PubMed evidence. AstroOmix is designed around verifiability. Numerical claims are checked at runtime against structured analysis outputs; unsupported figures are withheld. Literature citations must originate from retrieved records, and literature consistency is presented as context, not proof. The application also flags thin time series, distinguishes interpolation from extrapolation and avoids clinical conclusions. Building AstroOmix highlighted a central challenge in space-health research: the available data are valuable, but small, heterogeneous and easy to overinterpret. The project therefore focuses not on automating certainty, but on making analysis traceable, reproducible and appropriately cautious. As longer missions generate increasingly complex biological data, tools that combine computational analysis with evidence-aware AI reasoning can help researchers reach useful conclusions while remaining explicit about uncertainty. - **Team:** [Yoga Mithra Senthil Kumar](https://cerebralvalley.ai/u/Mithrasen) - **GitHub:** https://github.com/Mithrasen/AstroOmix , https://astroomix.streamlit.app - **Demo video:** https://youtu.be/1MfycsEtySQ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=152 ### 177. IndoVarPriority VPS - Variant Priority Score Current genetic-testing tools (ClinVar, InterVar, Franklin) can only classify variants that already exist in reference databases - roughly 80% of which are European-ancestry. For the 2 billion people of South Asian descent, most clinically relevant variants are novel: never seen, never classified, never scored. A "negative" report is often not truly negative - it is uninformative. We built VPS, a phenotype-driven, population-aware, inheritance-aware engine that scores exactly these novel and uncertain-significance variants. A clinician starts from a clinical question, not a gene list; VPS assembles the relevant gene panel, then scores every variant on five transparent axes , pathogenicity (blending the ESM-2 protein language model, the Evo2 DNA language model, and REVEL/AlphaMissense/CADD), inheritance-aware population frequency (Whiffin–Ware maximum credible allele frequency), domain criticality, regional pathogenic density, and AlphaFold structural confidence , into a single 0–100 priority with a plain-English rationale for each variant. Crucially, the frequency axis is inheritance-specific: the same allele can score consistent with a recessive disease yet be far too common for a dominant one, so every panel is split by inheritance model. Applied to IndiGenomes (1,029 open-source Indian whole genomes, CSIR-IGIB), we scored 3,594 novel variants , absent from every global database , across 6 inheritance panels and 52 genes spanning Wilson disease, sudden cardiac death (recessive and dominant), primary immunodeficiency/HLH, treatable neurometabolic disease, and hereditary breast/ovarian cancer, at 100% coverage across all variant classes (missense, truncating, in-frame indel, and non-coding). As blinded validation, known ClinVar Pathogenic/Likely-Pathogenic variants scored measurably higher than known Benign variants (Wilson: 64.3 vs 57.9) without ClinVar labels ever being used as scoring input — the separation comes purely from independent evidence. This matters because every one of these 3,594 variants sits in a gene where a missed variant can mean a preventable death , a homozygous cardiac variant in a child, a treatable neurometabolic disorder with a closing therapeutic window, a hereditary-cancer variant invisible to a European-calibrated panel. VPS is a framework for ethnicity-aware variant interpretation: same data, same patient, a different and defensible answer, because it scores what other tools call negative. (Research-triage tool it ranks which variants warrant functional follow-up, not a clinical diagnostic.) - **Team:** [Nirmala Kadali](https://cerebralvalley.ai/u/29-2004) - **GitHub:** https://github.com/Nirmala-k/IndoVarPriorityScore - **Demo video:** https://1drv.ms/v/c/4cd4940fed94a88c/IQByN52ingZyTYIcVJCXgbDhATWj4V1AnQlqpdMlpQfO6Vc?e=PBL6vC - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=153 ### 178. Oxphos Exercise interventions offer profound systemic benefits, yet patient adherence remains a major clinical hurdle. Recent reports identify Lac-Phe (N-lactoyl-phenylalanine) as an exercise-induced metabolite driving appetite suppression and fat loss; however, it is entirely absent from every released feature table in the MoTrPAC exercise atlas. Using Claude Science, this project orchestrated the re-analysis of raw metabolomics data to map Lac-Phe dynamics across 19 tissues. The signal was successfully recovered from raw spectra, revealing robust training-induced elevations—including a previously undescribed female-predominant pattern in the heart and white adipose tissue. Crucially, the data demonstrates that Lac-Phe dynamics are governed by upstream substrate availability (lactate) rather than the induction of its synthesizing enzyme, CNDP2. Leveraging this mechanistic insight, a modular computational pipeline was designed to predict clinical viability, validating targets, drug relations, ADMET profiles, and side effects using open data. This pipeline successfully identifies hydrolysis-resistant Lac-Phe isosteres as stable, high-potential exercise mimetics. Why it matters: By rescuing "invisible" data from the world's largest exercise atlas, this project resolves a core biological paradox of Lac-Phe kinetics. Furthermore, it delivers an end-to-end predictive computational pipeline to de-risk and engineer next-generation exercise mimetics prior to costly wet-lab synthesis. - **Team:** [Steve Hershman](https://cerebralvalley.ai/u/Oxphos) - **GitHub:** https://github.com/hershman/claude-science-lacphe-motrpac - **Demo video:** https://youtu.be/iSjRvYeJd44 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=154 ### 179. Claymore Labs Claymore is an autonomous experimentation platform that gives scientific laboratories a persistent intelligence layer capable of conducting research on their behalf. Rather than acting as another chat interface or analysis tool, Claymore understands a lab's collective knowledge and can directly carry out scientific research in both computational environments and the physical world. Claymore can autonomously perform computational research through tools such as Claude Science for data analysis, machine learning, and literature review. It can also control laboratory robots through platforms such as Opentrons and PyLabRobot, enabling it to design, validate, and execute physical experiments in the real world—from pipetting reagents across a 96-well plate to coordinating centrifugation and automated measurement. As new evidence is generated, Claymore interprets the results, updates its understanding, and determines the next experiment, creating a continuous closed-loop experimentation cycle. Scientists cannot work around the clock, but Claymore can. While researchers are focused on other projects, away for the evening, or on vacation, Claymore can keep analyzing data, testing hypotheses, running experiments, and planning the next iteration. It gives scientists back the time they do not have, accelerating discovery by ensuring promising ideas are continuously investigated instead of waiting days, weeks, or indefinitely to be pursued. - **Team:** [Rikhil Tanugula](https://cerebralvalley.ai/u/r1k), [Rikhin Kavuru](https://cerebralvalley.ai/u/rikhinkavuru) - **GitHub:** https://github.com/r1khilt/claymore - **Demo video:** https://www.youtube.com/watch?v=-bXyLQnT8kQ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=155 ### 180. The Gene Scene GeneGround is a claim-level grounding layer for AI-generated omics interpretations. I built it to address a core problem in AI-assisted biology: LLMs can generate fluent scientific narratives, but they may subtly overstate causality, mechanism, robustness, or biological scope. GeneGround takes an AI-generated omics interpretation and evidence handoff file, then decomposes the interpretation into individual biological claims. It normalizes genes, pathways, cell contexts, conditions, and directions against controlled taxonomies; indexes the handoff into perturbation, pathway, robustness, provenance, and language-rule evidence; retrieves relevant chunks using metadata-first retrieval with local TF-IDF fallback; and evaluates each claim with four specialist Claude-powered agents. Final verdicts are assigned through deterministic aggregation, while Claude is used for semantic biological judgment and grounded rewrites. The Technical Pipeline page exposes every processing stage from claim extraction through retrieval and verdict generation, while an interactive Evidence Trace links each sentence to its claims, agent queries, retrieved evidence chunks, and final reasoning. In the demo, GeneGround audited a primary human CD4+ T cell Perturb-seq interpretation, identifying claims supported with caveats, partially supported, or overstated. This matters because AI-assisted science needs traceable evidence, cautious language, and reproducible reasoning, not just plausible summaries. - **Team:** [Rhea Zhou](https://cerebralvalley.ai/u/rheazhou) - **GitHub:** https://github.com/rheazhou2026/geneground - **Demo video:** https://www.youtube.com/watch?v=aILeQ9WDXzs&lc=UgwxhVCMBUwc3CkN6P54AaABAg - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=156 ### 181. n = 3, and Proud Everyone building research agents runs into the same quiet question: once you have one, how should you set it up to do its best work? Usually the answer is a hunch. This project turned the hunch into an experiment. A researcher used one AI research agent to study how to make that same agent produce better science — running a blank coding agent as a driver in a closed, self-improving loop, and pitting each loop design against a raw, unscaffolded baseline on word-identical questions, scored blind by an automated reviewer panel and a human domain expert. Across three rounds, the matured design scored highest on every metric and for every scorer, and beat its predecessor on all three test questions. The effect is modest and the sample is small — three questions, and one open-ended question the baseline won — and the report says so plainly. The lasting contribution is not the winning configuration; it is a reusable, auditable way to measure whether an agent's setup actually helps, rather than asserting that it does, paving the way for individual harness and workflow research to produce better science. - **Team:** [Fabian Baezner](https://cerebralvalley.ai/u/FBaezner) - **GitHub:** https://github.com/BobtheBuilder0426/The-Metascience-Project - **Demo video:** https://youtu.be/7mi3KGNtSrM - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=157 ### 182. The Nineteen # Project description ## What we built / investigated We mined a genome-scale CRISPRi Perturb-seq atlas of about 22 million primary human CD4⁺ T cells to discover **novel drug targets**. (Per the project's dataset documentation, the atlas is the CD4⁺ T-cell Perturb-seq resource from the Marson and Pritchard labs, distributed on the CZI Virtual Cells Platform; this attribution is as recorded in the project files and has not been independently re-verified here.) The core idea is that a CRISPRi knockdown is a genetic model of drug-induced loss of function, so we read the perturbation-to-transcriptome map backwards: starting from a therapeutically desirable T-cell program and asking which upstream regulators control it. Those regulators become drug-target hypotheses. On top of this we built an end-to-end computational-to-translational pipeline: foundation QC and reproducibility filtering, then six orthogonal discovery lenses, then an integrated ranked scorecard, then single-cell validation across donors, and finally structure, genetics, and chemistry-based translational dossiers. ## What we found The pipeline narrowed roughly 34,000 perturbations down to a short list of **19 single-cell-validated targets**, and this is the main result (full detail in the Appendix table). It splits into two therapeutic groups. **A TCR-proximal signaling group** (LAT, PLCG1, CD247, CD3E, ZAP70, VAV1, SIK3) contains the highest-ranked hits and mostly points toward suppressing inflammation. It recovers known immune-drug biology on its own, which validates the logic, and it surfaces **SIK3**, a novel kinase that became our structurally validated lead (redocking under 2 Å, actives-versus-decoys enrichment AUC 0.905). **A drug-naive chromatin and transcriptional group** (a SAGA/Mediator axis plus remodelers: SMARCE1, STAT6, SGF29, MED24, MED12, TADA2B, NSD1, CHD4, SMARCB1, TRIP12, ARNT, SEL1L) is the central novelty. All 12 of these validate at single-cell resolution and form a mechanism distinct from acute TCR signaling (89% cluster recovery, 2.4× smaller state-space displacement). Of the full set of 19, **18 are clinically unprecedented** (only CD3E is already drugged). The single top opportunity is **STAT6** (rank 5): a boost-immunity target backed by strong asthma and allergy genetics (55 immune GWAS associations), rich chemical matter (552 ChEMBL bioactivities, max pChEMBL 9.15), and small-molecule tractability, yet with no approved drug. The remaining standouts are the drug-naive SAGA/Mediator readers **MED24**, **SGF29**, and **TADA2B**. Two targets (CHD4, SMARCB1) are flagged for essentiality and toxicity risk on weaker knockdown. The complete ranked list, with validation and translational evidence for each target, is given in the Appendix. ## Why it matters By recovering known targets from scratch, which validates the logic, and then nominating a genetically-supported, single-cell-validated, structurally-tractable set of *novel* regulators, the project turns a 22-million-cell atlas into a short list of de-risked, mechanistically-distinct therapeutic entry points for autoimmune and allergic disease and for immuno-oncology. STAT6 and a previously undrugged chromatin/transcriptional axis head that list. More broadly, the work demonstrates a reusable in-silico funnel that goes from genome-scale perturbation data to prioritized, translation-ready drug targets. --- ## Appendix: the 19 validated targets Sorted by nomination rank, with validation and translational evidence for each target. | Rank | Gene | Axis | Direction | Novelty | Dirs | KD% | Concord. r | Donors | Best PDB (n) | SM bucket | ChEMBL act. | max pChEMBL | Immune GWAS | |---|---|---|---|---|---|---|---|---|---|---|---|---|---| | 1 | **LAT** | TCR-proximal | Suppress | novel-druggable | 4 | 85% | 0.67 | 4/4 | — | — | 2 | — | — | | 2 | **SMARCE1** | Chromatin/TF | Suppress | novel-druggable | 3 | 88% | 0.74 | 3/3 | 9 | Structure with Ligand | 8 | 6.95 | 37 | | 3 | **PLCG1** | TCR-proximal | Suppress | novel-druggable | 4 | 88% | 0.46 | 4/4 | 6 | High-Quality Ligand | 438 | 6.75 | — | | 4 | **CD247** | TCR-proximal | Suppress | novel-druggable | 3 | 72% | 0.51 | 4/4 | 38 | Structure with Ligand | — | — | 46 | | 5 | **STAT6** | Chromatin/TF | Boost | novel-druggable | 3 | 82% | 0.57 | 4/4 | 7 | High-Quality Ligand | 552 | 9.15 | 55 | | 6 | **CD3E** | TCR-proximal | Suppress | known-drug-target | 3 | 96% | 0.85 | 4/4 | 44 | Med-Quality Pocket | — | — | 3 | | 7 | **ZAP70** | TCR-proximal | Suppress | novel-druggable | 3 | 91% | 0.78 | 3/3 | 15 | High-Quality Pocket | 2390 | 8.1 | 1 | | 9 | **VAV1** | TCR-proximal | Suppress | novel-druggable | 2 | 92% | 0.66 | 3/3 | 10 | Structure with Ligand | 1 | 8.98 | — | | 19 | **SIK3** | TCR-proximal | Suppress | novel-druggable | 3 | 87% | 0.52 | 3/3 | 5 | High-Quality Ligand | 809 | 9.34 | — | | 22 | **TRIP12** | Chromatin/TF | Suppress | novel-druggable | 3 | 84% | 0.70 | 4/4 | 5 | Structure with Ligand | — | — | — | | 25 | **NSD1** | Chromatin/TF | Suppress | novel-druggable | 3 | 92% | 0.65 | 3/3 | 4 | High-Quality Ligand | 147 | 6.96 | 1 | | 31 | **SGF29** | Chromatin/TF | Suppress | novel-druggable | 3 | 94% | 0.69 | 4/4 | 8 | High-Quality Pocket | 1 | — | 5 | | 34 | **MED24** | Chromatin/TF | Mixed | novel-druggable | 3 | 96% | 0.80 | 4/4 | 10 | Structure with Ligand | — | — | 12 | | 44 | **ARNT** | Chromatin/TF | Mixed | novel-druggable | 3 | 89% | 0.69 | 4/4 | 46 | High-Quality Ligand | 25 | — | 1 | | 45 | **TADA2B** | Chromatin/TF | Suppress | difficult | 3 | 92% | 0.74 | 4/4 | — | — | — | — | — | | 66 | **CHD4** | Chromatin/TF | Boost | novel-druggable | 3 | 52% | 0.38 | 4/4 | 12 | Structure with Ligand | 270 | 8.52 | — | | 80 | **SMARCB1** | Chromatin/TF | Suppress | novel-druggable | 3 | 48% | 0.59 | 4/4 | 18 | Structure with Ligand | 8 | 8.03 | — | | 99 | **MED12** | Chromatin/TF | Suppress | difficult | 3 | 85% | 0.76 | 4/4 | 3 | — | 6 | 6.89 | — | | 124 | **SEL1L** | Chromatin/TF | Suppress | novel-druggable | 3 | 94% | 0.42 | 3/3 | 5 | — | — | — | — | Columns: **Axis** = TCR-proximal signaling vs the novel chromatin/transcriptional axis. **Direction** = therapeutic direction implied by loss of function (Suppress inflammation / Boost immunity / Mixed). **Dirs** = number of independent discovery directions that nominated the gene. **KD%** = single-cell on-target knockdown (mean across powered donors). **Concord. r** = Pearson concordance between the single-cell knockdown signature and the pseudobulk DE signal. **Donors** = powered donors out of those tested. **Best PDB (n)** = count of experimental structures. **SM bucket** = best Open Targets small-molecule tractability bucket. **ChEMBL act. / max pChEMBL** = chemical-matter depth and best measured potency. **Immune GWAS** = immune-disease GWAS associations. Dashes mark zero or no available data. - **Team:** [Qiuchen Meng](https://cerebralvalley.ai/u/QIuchen) - **GitHub:** https://github.com/MengQiuchen/perturb-seq-cd4-target-discovery - **Demo video:** https://youtu.be/jnz5nR5w6pg - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=158 ### 183. Akhi's Achilles Drug-resistant infections kill ~1.27 million people a year, and the hard part isn't naming a resistance gene — it's finding the reversible opening resistance leaves behind. I built Achilles, an evidence-grounded discovery console. Point it at your data and it builds a provenance-checked graph where every claim carries a citation: a deterministic Python core does all the math, and Claude only reads literature, retrieves, and cites — it never invents a number. If a claim can't be grounded, it doesn't become an edge. The result is science you can break on stage. Against 29 independent public controls, Achilles recovers known biology (12/12), refuses an adversarial battery of plausible falsehoods (17/17), and fabricates nothing (0) — writing the whole verdict to a tamper-evident, hash-chained ledger anyone can re-verify. On a time-split hold-out it even anticipates relationships before the confirming paper appeared. Shown end-to-end on antimicrobial resistance, it names the reversible target resistance creates, folds it with AlphaFold, and proposes a cited antibiotic-cycling hypothesis — bench to bedside, never medical advice. Domain-agnostic, reproducible from public data, and callable as MCP tools from any Claude agent. Honestly, watch my demo, but more so just use the product on your own, and trust me, you will see the technical biological depth this product has to offer. - **Team:** [Akhi Chappidi](https://cerebralvalley.ai/u/akhic) - **GitHub:** https://github.com/akhimass/Achilles - **Demo video:** https://www.boomshare.ai/shared/01KXEXBS56GV4E4SCRV1FS8DW8 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=159 ### 184. Ellen Schrader Atlas: every lab has a taste. AI threatens to homogenise scientific research: every lab asks the same models to run its literature search, summarise papers, and come up with ideas - and every lab gets the same consensus answer. Atlas is the bet against that. It doesn't make Claude smarter; it gives Claude your lab's perspective. Atlas builds a shared, searchable memory from a lab's literature: deduplicated papers, semantic search, topic maps, and grounded digests of what's new. As researchers read, discuss, and curate papers, Atlas learns the lab's evolving scientific taste. Through MCP, Claude can securely read that memory (owner-controlled and fully audited) and answer a question a generic literature search can't: is this idea novel for our lab? It drafts related-work sections in which every citation is verifiable, and even generates figures in the visual style the lab has implicitly developed over years. Instead of every lab converging on the same AI, Atlas gives every lab an AI shaped by its own scientific taste. https://atlas-papers.vercel.app/papers - **Team:** [Ellen Schrader](https://cerebralvalley.ai/u/ellen-schrader) - **GitHub:** https://github.com/ellen-schrader/atlas - **Demo video:** https://www.loom.com/share/a039070a9d02414b987ea6ee3dd04736 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=160 ### 185. XiaojingSui AGEFOLD — A CONVERSATIONAL AI TO DECODE THE BIOLOGY OF AGING WHAT WE BUILT AgeFold turns a massive structural-proteomics dataset into something anyone can talk to. Using TMT-LiP-MS in C. elegans, we measured how 6,823 proteins change their 3D shape across 9 conditions of aging and stress — where each protein loosens, tightens, or unfolds as an animal grows old. AgeFold paints that per-residue conformational signal directly onto AlphaFold structures you can spin in the browser, then wraps the whole thing in a grounded, multi-agent chat: type any gene, ask "what does it do, where does its shape change with age, and why does that matter for human disease?" — and get a cited answer synthesized from UniProt, InterPro, PDB, disease-variant databases, and our own measurements. WHAT MAKES IT WORK - Two doors, one dataset. A Researcher mode with the full technical view (9-condition heatmaps, ML-predicted aging-vulnerable residues, mapped disease variants, worm-to-human orthologs) and a Public mode that re-narrates the same verified data in plain language — science communication without dumbing down. - A coordinator routing four specialist agents. Structural biology, aging & stress, disease genetics, and data analysis — each reasons in its lane and returns one cited response, every claim traceable to a source. - Structure-aware biology. It connects altered-conformation regions to functional sites and to human disease mutations via homology — turning "this loop moves with age" into a testable hypothesis about why. WHY IT MATTERS Aging research drowns in data that only a handful of experts can read. AgeFold makes a genome-scale map of protein conformational aging explorable by a bench scientist and a curious teenager alike — surfacing which proteins destabilize with age, connecting those changes to disease, and doing it conversationally. It's a template for turning any high-dimensional -omics dataset into a living, askable resource, and it's deployable for free so the whole community can use it. AgeFold. Let aging unfold. - **Team:** [Xiaojing Sui](https://cerebralvalley.ai/u/xiaojingsui) - **GitHub:** https://github.com/xiaojingsui/agefold - **Demo video:** https://youtu.be/y9ErGzObr_E - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=161 ### 186. Redline Single-cell RNA-seq conclusions fail peer review on the statistics while the biology holds up. A standard pipeline treats 51,842 cells from four donors as 51,842 independent samples, so a marker gene looks significant at p = 6.2e-11, until a proper per-donor test puts it at 0.21. Reviewer 2 catches that. You didn't. Redline audits those statistics on your own data before you submit. Hand it your .h5ad and the analysis you ran; it re-runs the load-bearing tests, marks the false discoveries on your own figures, cites the method that fixes each, and hands back corrected code that runs on your data. It flags eight error classes that QC tools and generic reviewers skip. The honest number is a false-positive gap. On a 46-case benchmark of planted errors and clean controls, redline stays quiet on clean data at 0% false positives. One Claude call given the same write-up cries wolf 74% of the time. Both arms catch nearly every planted error, so the number that matters is 0 against 74. Live demo, no API keys: https://science-redline.vercel.app - **Team:** [Pablo Manjarres](https://cerebralvalley.ai/u/pablomanjarres) - **GitHub:** https://github.com/pablomanjarres/redline - **Demo video:** https://youtu.be/H3jq20gI7cI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=162 ### 187. Coco and the Cancer-naughts The bottom line: An AI agent that reads why an approved cancer drug failed a tumor — then designs a new antibody–drug conjugate (ADC) engineered to dodge that exact resistance. Built end-to-end in Claude Science, grounded entirely in live data. What I built: One agent, five modules, each querying live databases — not memorized answers. Test case: TNBC that progressed on Dato-DXd (an approved ADC). A · Resistance (ChEMBL, FDA, PubMed) → B · Target (Open Targets, GTEx, HPA) → C · Payload (PubChem, ChEMBL) → D · Structure (PDB, AlphaFold, Boltz-2 via NVIDIA BioNeMo) → E · Landscape (ClinicalTrials, bioRxiv) Output: a fully-sourced dossier where every ID traces back to the query that produced it. What I found The failure is payload-class, not the molecule. Dato-DXd's TOP1-inhibitor warhead is cross-resisted by every TOP1i ADC — so the fix is a new mechanism, not a new antibody. A resistance-orthogonal design: LIV-1 antigen (TROP2, the failed target, drops to rank 7) + MMAF payload (dodges both the TOP1i and the efflux resistance) + non-cleavable linker, DAR4. Structure overturned the answer. Boltz-2 folded all three candidate targets and inverted the expression ranking: the top expression target (LIV-1) is the weakest structure (pLDDT 33); B7-H3 is the strongest (91). The real output is a two-axis map, not a single pick. Why it matters ADC resistance is a growing clinical reality with no systematic way to design the next drug against it. This engine makes resistance the design input, and — critically — lets one tool (structure) overturn what another (expression) would have asserted with false confidence. Reproducible, fully traceable, and it generalizes to any (indication, failed-drug) pair. In-silico hypothesis, not a validated asset; key risks (LIV-1 foldability, payload tolerability) stated explicitly. - **Team:** [Lance Cadang](https://cerebralvalley.ai/u/lancecadang) - **GitHub:** https://lancecadang-tech.github.io/ADC-engine-notebook/ - **Demo video:** https://youtu.be/mvuZYnm1F9Q?si=4lJpJVPKP1aVVq5H - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=163 ### 188. Upstream Team What we built / investigated. We tackled pediatric urology's top-ranked research gap (https://doi.org/10.1016/j.jpurol.2026.106024) — defining clinically significant upper urinary tract obstruction, the point at which a blocked kidney is losing function and needs surgery rather than watchful waiting. Current tools (serum creatinine, renography, ultrasound) detect injury only after substantial nephron loss. We built a decision-ready evaluation of urinary biomarkers as earlier indicators, in three linked analyses: (1) a structured, five-question synthesis of 43 candidate biomarkers across 100 studies, tagged for conflicting evidence, external validation, adult-vs-pediatric scope, multicenter readiness, and trial-readiness; (2) integration of that literature with a 196-gene conserved obstruction/injury transcriptomic module (from the authors' prior cross-species research) to nominate mechanistically-grounded candidates the clinical literature has missed; and (3) an assay-and-cost feasibility analysis for the resulting panel. Every one of the 59 references was verified against PubMed. What we found. The field's bottleneck is validation, not discovery. Of 43 markers, only the incumbent comparator (creatinine) has multicenter validation; 29 have never been externally validated, 17 carry outright conflicting evidence, and 17 are promising but stuck at single centers. Zero were studied only in adults — the pediatric evidence already exists but has never been confirmed prospectively across sites. Integrating the tissue biology surfaced urinary DKK3 — an adult-validated tubular-stress marker that predicts functional decline, untested in children — as the strongest new candidate. Combining both streams yields a six-marker core panel (MCP-1, CA19-9, DKK3, NAG, β2-microglobulin, periostin) that is dual-purpose — it flags obstruction and tracks recovery after surgery — and is buildable today at ~$90–100/sample. Why it matters. A biomarker threshold that separates kidneys that lose function from those that don't is the missing definition of clinically significant obstruction. This work converts a crowded, contradictory literature into a concrete, costed specification for the prospective multicenter validation study the field has never run — the study that would let clinicians intervene before irreversible kidney damage, and spare children unnecessary surgery. - **Team:** [Justin Kim](https://cerebralvalley.ai/u/JkimUro) - **GitHub:** https://github.com/kimjk4/ClaudeLSHackathon - **Demo video:** https://youtu.be/o6i7xFvC8GY - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=164 ### 189. Duck-Kyun Yoo (Team DK) Germinal-center (GC) selection is almost always modeled along a single axis — antigen affinity. But B cells are also selected on a second axis: tolerance — how polyreactive/self-reactive an antibody is. That axis has been hard to use because it was not measurable at repertoire scale. We make it measurable directly from structure. We fold paired antibody Fv *monomers* from sequence (Boltz-2, ESMFold-v1, Protenix) — deliberately avoiding antibody–antigen complex prediction, which is still unreliable — and read a physicochemical tolerance axis, surface-displayed CDR charge, off the fold. The score is folder-invariant, agrees with crystallographic geometry, and transfers without retraining to an autoimmune (type-1-diabetes) repertoire and to a tissue-resolved human GC repertoire, where the axis is significantly constrained in GC B cells. We report the result honestly, including the null (as a black-box predictor, folded structure does not beat a charge baseline out of distribution) and the one surviving charge-orthogonal signal. The output is a mechanistic, reproducible coordinate that single-axis accounts of GC selection have lacked. Results (attached again just in case): https://github.com/dk93js/gc_project, https://drive.google.com/drive/folders/19S4tkwlr4iTsHV2KyxBpoCAm1Cy821XO?usp=sharing - **Team:** [Duck-Kyun Yoo](https://cerebralvalley.ai/u/dk93js) - **GitHub:** https://github.com/dk93js/gc_project, https://drive.google.com/drive/folders/19S4tkwlr4iTsHV2KyxBpoCAm1Cy821XO?usp=sharing - **Demo video:** https://youtu.be/UPEfdLzsHrE - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=165 ### 190. Dynamics-aware-FeatureDock Structure-based machine learning has ridden one wave: AlphaFold made static structure cheap and universal, and models like FeatureDock, Boltz-2, and DiffDock-L now predict binding from a single frozen pose. But proteins are not statues. Much of what determines whether a drug binds — and how tightly — lives in µs–ms conformational motion that no static structure captures. We believe dynamics is the next dimension, and the next leap in AI for protein design will come from models that can see it. We investigated whether ground-state dynamics information improves ML models of small-molecule–protein binding. Using Claude Science, we integrated Dyna-1, which predicts per-residue conformational exchange from any protein structure with no experiment required, into FeatureDock, adding a dynamics channel to its pocket features. As hypothesized, pose prediction improved most in flexible regions, exactly where a single static structure fails. This matters because it is a concrete step toward making dynamics a reusable, experiment-free ground-truth signal for machine learning. The same prior could sharpen co-folding and docking engines like Boltz-2 and DiffDock-L, and give protein-design pipelines a way to engineer motion on purpose, turning dynamics from an after-the-fact surprise into an optimizable design objective across countless AI-protein models. - **Team:** [Andres Lira](https://cerebralvalley.ai/u/Andresl3), [Natesan Mani](https://cerebralvalley.ai/u/nmani) - **GitHub:** https://github.com/Andresl3/dynamics-aware-featuredock - **Demo video:** https://www.linkedin.com/posts/andres-lira-b47549188_claude-science-hackathon-project-demo-i-activity-7482567457073717248-DwJO?utm_source=share&utm_medium=member_desktop&rcm=ACoAACw4AycBuwht06kXIR0SXXOC7gNmVd726uw - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=166 ### 191. Rare Events Protein conformational rates are hard to compute: the transitions are rare, so brute-force molecular dynamics almost never sees them. Importance sampling accelerates them by biasing the dynamics and reweighting back to the true dynamics, but the path weights degenerate -- their variance grows exponentially and a handful of paths dominate, making the rate estimate unreliable. We asked which weight-control (resampling) scheme fixes this for a real peptide, and answered it as an agentic method search: Claude Code proposed each estimator variant, implemented it, ran it on the C5 // C7eq isomerization of alanine dipeptide, and read back the weight statistics and effective sample size against an exact brute-force rate -- then iterated. Comparing sequential importance sampling (SMC), sequential importance resampling (SIR), and a branching resampled random walk (RRS), only RRS converged in both directions, recovering the true rate within 3–6% where the naive committor estimate is up to 6× wrong. An agent discovered the variance-reduction scheme, not just applied one. - **Team:** [Michael Kim](https://cerebralvalley.ai/u/mckim) - **GitHub:** https://myungchulkim.com/research/claude_hackathon.html - **Demo video:** https://myungchulkim.com/research/claude_hackathon.html#demo-video - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=167 ### 192. Lab Brain for Claude Science I built Lab Brain, a tool that turns any biomedical lab project — its code, slides, data, papers, and notes — into a shared knowledge base that both Claude Science and Claude Code can read and write. Everyone on a project works at the same time from their own Claude environment, and all of a project's knowledge ends up in one place instead of scattered across people's heads, chats, and folders. Most of the work went into creating the folder and file structures and skill/command frameworks that hold that together, which I iterated on from my own lab's practices. A START_HERE page always shows the latest updates linked to their files, clickable in Obsidian; some files get overwritten to the current state while others are append-only and keep growing, so nothing is lost. Handoff boards let people claim a file before editing so no one overwrites anyone else, and give mentees a scaffold plus a feedback loop with their mentors. Figures are named by content and logged in a registry, and old work is archived rather than deleted. Claude Science is a state-of-the-art harness, but it doesn't have a shared knowledge base — and research knowledge usually fragments across individuals working separate tasks. So I wanted to pull everything into one place that keeps growing, where people can build on each other's work, decide directions, and see what everyone's doing without losing context. - **Team:** [Bobby Ni](https://cerebralvalley.ai/u/bobbyni) - **GitHub:** https://github.com/bobbyni819/lab-brain - **Demo video:** https://youtu.be/1Vhywhld5Qc - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=168 ### 193. Sriram This project presents a causal-structure-oriented, interpretable pipeline for modeling influenza (H3N2) antigenic distance from hemagglutination-inhibition (HI) data and HA sequence alignments. Its central problem is that dense viral phylogenies tightly link true antibody-escape drivers with passenger mutations, so accurate sequence-based prediction does not establish which HA positions physically disrupt antibody binding. In this study, we analyzed two public H3N2 HI datasets by collapsing co-evolving positions into linkage blocks, then applying target-oriented causal discovery ranked by 200-resample bootstrap stability, alongside an interpretable B-spline Kolmogorov–Arnold Network whose per-position response curves are inspectable and which extends to second order to probe epistasis. The first-order KAN trails gradient boosting by a small but robust margin; a second-order KAN closes most of the gap only under a different protocol. Our pipeline demonstrates cross-method convergence by isolating a highly stable set of key drivers at mature HA positions 133, 156/158, and 189, mapping precisely to classical antigenic sites A and B. Finally, adjusted partial-regression effect sizes systematically shrink relative to marginal associations, confirming successful control over phylogenetic confounding. - **Team:** [sriram vijendran](https://cerebralvalley.ai/u/Ramviji) - **GitHub:** https://sriram98v.github.io/claude-science-hackathon/ - **Demo video:** https://youtu.be/HAUcofbm9U8 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=169 ### 194. overlake.bio I built Spacefarer Phenome, a genetics-first computational drug-discovery pipeline for spaceflight physiological stress, anchored on bone mineral density (eBMD). A nine-stage funnel (phenotype, GWAS, druggable-genome, Mendelian randomization, PheWAS safety, druggability, structure, docking, MD) spends cheap CPU on the causal and safety stages before any GPU. The headline is an honest negative: the drug-target MR scan validated no target in the pre-registered set of 33 — an instrument-limited result, since key pQTL/eQTL panels were unreachable at run time, not a biological null; the positive control passed strongly (IVW p=1.85e-44). Cathepsin K, a pre-specified exemplar (not discovered here), was carried end-to-end to exercise the structural arm, re-ranking 24 known ChEMBL inhibitors with a conditional MD result. What matters is the rigor: pre-registered thresholds frozen in git, JSON-schema contracts between stages, per-artifact provenance, and an openly licensed, auditable record. Total GPU cost: $41.06. - **Team:** [Rakhan Aimbetov](https://cerebralvalley.ai/u/raimbetov) - **GitHub:** https://github.com/raimbetov/spacefarer-phenome-bwc - **Demo video:** https://youtu.be/EVmO_UgTu_I - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=170 ### 195. Let's fight cancer! I built DN-Isoform Scout, an end-to-end computational proof of concept for discovering an underexplored class of cancer mechanisms: dominant-negative proteins that lose their normal function but retain the ability to assemble with—and poison—the wild-type protein complex. The long-term goal is a cancer-selective strategy that targets molecular alterations enriched in tumour cells while sparing healthy cells. In the splicing arm, I directly measured tumour-versus-normal differences to identify candidate therapeutic windows; in the mutation arm, I established the computational framework needed to prioritize potentially tumour-specific protein variants. These selectivity hypotheses must now be confirmed experimentally. Unlike conventional CRISPR screens that reduce each gene to a simple knockout, DN-Isoform Scout investigates cancer below the gene level from two complementary directions. The first arm identifies tumour-associated splice events across breast, lung, prostate and diffuse large B-cell lymphoma. The second screens point mutations, truncations and deletions in TP53, STAT3 and IKZF1. I integrated cancer transcriptomics, protein-domain annotation, protein language models and structural AI—including ESMC-6B, ESMFold2 and Boltz-2—to search for the defining signature of a dominant negative: function lost × interaction interface retained × protein still folded. I then converted the predictions into concrete experimental starting points: a 521-guide dCas13d library designed for reversible splice redirection without permanent genome editing, plus protein candidates annotated with potential covalent reactivation, interface-disruption and selective-degradation strategies. The proof-of-concept results were highly encouraging. From 42,529 quantified event–cancer combinations, I identified 1,806 significant tumour-associated splicing hits across 315 genes and narrowed them to 365 ranked dominant-negative hypotheses across 118 genes. I constructed and structurally evaluated 186 candidate isoforms, identified seven particularly strong complex-supported hypotheses, and designed 416 candidate dCas13d guides alongside 105 carefully chosen controls. In parallel, I screened 1,238 protein variants, modeled 160 monomers and 64 mutant–wild-type complexes, and selected 30 high-priority candidates spanning TP53, STAT3 and IKZF1. Importantly, the pipeline successfully recovered the known dominant-negative controls Δ133p53, STAT3β and Ik6 using the same scoring framework as the other candidates, without assigning them an artificial scoring bonus. In an internal TP53 sanity-check benchmark based on 33 literature-curated labels, the DN score separated dominant-negative variants from ordinary loss-of-function variants with an AUC of 0.90 and from benign variants with an AUC of 1.00. These results demonstrate internal consistency and support the central mechanism, although they are not yet a substitute for independent biological validation. One of my most important findings was that monomer structure alone cannot reliably distinguish a dominant negative from an ordinary loss-of-function protein: the crucial question is whether the altered protein can still assemble with wild type. DN-Isoform Scout therefore turns a major blind spot in cancer screening into an auditable experimental roadmap containing ranked mechanisms, designed perturbations, structural hypotheses and clear validation gates. This hackathon project establishes the proof of concept. The next phase is laboratory validation: confirming transcript identity and protein production, testing mutant–wild-type assembly, demonstrating genuine poisoning of wild-type function, and measuring cancer-versus-normal selectivity. Instead of merely asking which genes matter, I built a platform that asks which precise molecular alteration poisons the cell, how it does so, and how we might selectively stop it. - **Team:** [Mateusz Mendel](https://cerebralvalley.ai/u/matmendel) - **GitHub:** https://github.com/m-mendel/Cancer_splicing - **Demo video:** https://youtu.be/cyhiGJHNGdY - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=171 ### 196. Plasmon Works Speciate is an automated pipeline that turns a TEM micrograph of gold nanoparticles into the three numbers a nanoparticle lab needs on every batch: mean size, size distribution, and — the hard one — how aggregated the particles are in solution. Stage one segments every particle with a Segment-Anything foundation model (NP-SAM) plus a one-click human-QC reviewer, validated to within ±6% of hand-counting across eight of the Dragnea lab's (Indiana University) samples while measuring up to 6× more particles. Stage two feeds those sizes into a first-principles optical model (Mie → exact multipole T-matrix) that predicts the UV-Vis spectrum with zero fitted parameters; where the prediction breaks, that gap quantifies aggregation — a readout no existing UV-Vis tool provides. It replaces a half-day of manual TEM counting with a minutes-long, higher-throughput measurement. Two open-source, MIT-licensed repos with a clean data contract. - **Team:** [Eun Sohl Koh](https://cerebralvalley.ai/u/doraicat) - **GitHub:** https://github.com/e-stella/tem-particle-metrics, https://github.com/e-stella/aunp_speciation, https://github.com/e-stella/aunp_speciation/blob/a9c135262b3dc48a7c3c6932ffafad7712e6e25e/reports/progress_260713.pdf - **Demo video:** https://youtu.be/8-3Eml3BOI8?si=f9RYuK0M-Lx6We6z - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=172 ### 197. Dr. Alexander King sc-referee automatically checks single-cell analyses for statistical mistakes that commonly slip through analysis and review. It targets known failure modes—using the wrong model for count data, losing donor pairing, testing clusters selected from the same data, or claiming false-discovery control that the results do not reproduce—that can turn technical artifacts into convincing discoveries. Referee reads the analysis files, Claude maps unfamiliar folders into an explicit scientific design, and the scientist confirms what was tested. Deterministic verifiers measure the problem and abstain whenever evidence or a validated check is missing. These failures are not hypothetical. In a published human melanoma analysis, Referee found that 95.3% of 16,289 reported cell-level discoveries did not survive patient-level inference; only 770 remained. GeneBench-Pro GB-P07 asked Claude Science to estimate how a genetic variant changes CXCL10 expression in activated monocytes from cell, donor, and empty-droplet data. Claude reported a positive effect (+0.484) when the benchmark reference was negative (-0.600), and its reviewer approved the analysis. Referee reconstructed an ambient-RNA contamination axis from the supplied data and conditionally proved the submitted model omitted it. Referee replaces plausibility judgments with encoded checks capable of catching specific mistakes that both bioinformaticians and frontier agents miss. - **Team:** [Alexander King](https://cerebralvalley.ai/u/Alexjk971) - **GitHub:** https://github.com/AlexanderJulianKing/sc-referee - **Demo video:** https://www.youtube.com/watch?v=1VGWTWFhlNI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=173 ### 198. Mute Anchor is a retrospective reproducibility agent for bioinformatics. Most scientific code can't be re-run months later because the analysis, its environment, and its data pipeline were never captured together — and existing tools are either prospective (Snakemake, Nextflow: you must rewrite your analysis up front, so nobody retrofits) or passive (provenance recorders that log what ran but never reason about why it broke). Anchor works on the messy Python/notebook code a lab already has. It inventories the repo, statically flags reproducibility smells (missing seeds, hardcoded paths, undeclared dependencies, time-dependence), runs the analysis twice in an isolated container to catch non-determinism empirically, then uses Claude to explain what each issue does to the specific figure it breaks and propose a patch — classified as mechanical or science-changing. Nothing that could alter a result is applied without human sign-off. The output is a pinned, containerized, Snakemake-wrapped bundle that reproduces — an on-ramp that gets unreproducible academic code onto real reproducibility tools. Everything runs on-prem, so sensitive data never leaves the building. - **Team:** [Maxime Blum](https://cerebralvalley.ai/u/mute) - **GitHub:** https://github.com/Mute-404/Anchor-Claude-Hackathon - **Demo video:** https://youtu.be/aC0hk8WnvqU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=174 ### 199. Reelin' from Alzheimer's Disease Which neurons die first in Alzheimer's, and why those? I investigated whether two markers of the selectively vulnerable entorhinal-cortex neurons that seed tau pathology — Reelin (RELN) and the transcription factor RORB — occupy the same individual cells, and whether that co-occupancy tracks disease. Using Claude Science, I built the analysis end to end: harmonizing ten public human datasets across single-cell, spatial (MERFISH), and bulk transcriptomics (>4M cells) and fitting depth-controlled, donor-clustered models that test the two genes as a pair. The findings show that RELN⁺/RORB⁺ detection is independent in healthy tissue but converges with advancing tau stage — a monotonic six-stage dose-response, independently replicated in an 81-donor cohort with 10× higher depth (exactly the condition that would inflate a detection artifact — which it doesn't), confirmed in intact tissue by MERFISH, absent in other tauopathies, and amplified by APOE-ε4 dose. Regulatory-network, ChIP-seq, and structural analyses rule out a direct molecular link, establishing co-vulnerability rather than partnership. This sharpens the map of AD's earliest-lost neurons and ties it to APOE4 — the field's strongest genetic risk factor — via Reelin–ApoE receptor competition: a hypothesis-generating finding, with the decisive next experiments specified, that others can build on directly. - **Team:** [Tamara Vasilkovska](https://cerebralvalley.ai/u/tvasilkovska) - **GitHub:** https://github.com/TamaraVas/reln_rorb_AD_analysis - **Demo video:** https://www.youtube.com/watch?v=gGaSG9FV90c - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=175 ### 200. Dragich & Claude AI CADENCE makes Parkinson's disease trials smarter and more cost-effective. This AI-driven pipeline pinpoints the common, ancestry-specific risk alleles that define genetic subtypes — enabling precision enrollment so expensive trials recruit the right patients from the start. Validated in Parkison's disease database (PPMI), these alleles are enriched in real patients and biomarker status tracks genotype. Built with Claude, from public data to trial-ready insight. - **Team:** [Joanna Dragich](https://cerebralvalley.ai/u/JDragich) - **GitHub:** https://drive.google.com/drive/folders/1sqjDjOBeDDpe3BouqZ5UrvdaGl9Nm3W0?usp=drive_link - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=176 ### 201. Akshay AI-predicted protein interactomes from AlphaFold-Multimer are now published as catalogues of "high-confidence" complexes and mined for drug targets. But the confidence score they rank on (interface predicted-TM) is overconfident, and the false-discovery rates attached to it are measured on balanced benchmarks with roughly equal real and fake pairs. A real interactome is the opposite: genuine interactions are rare among the thousands of pairs tested, and that rarity is exactly what makes the true error rate dramatically greater than the stated one.The cost of error is real. A target that isn't there sends a drug program down a dead end. I re-audited the CM4AI cell map (Schaffer et al., Nature 2025) with conformal FDR control, which produces honest error rates without assuming the confidence score is calibrated, then refereed the result against DepMap co-essentiality, a functional signal the structural model never saw, behind a strict anti-data-leakage firewall. The headline finding: a benchmark-tuned "10% error" cutoff actually admits 90% false discoveries as interactions grow rare, while conformal control stays bounded. Under it, 35 of 161 (22%) of the paper's own high-confidence edges fail, and the removed edges are significantly depleted of independent co-essentiality support (p=0.016). I then nominated KANSL3 as a missing member of the leukemia-associated MLL1-WDR5 complex, corroborated by co-essentiality and an independent Boltz-2 structure. The result is a reusable trust layer that flags which AI-predicted complexes survive honest error control before anyone acts on them. It works only when a release ships the negative controls to calibrate against, which is why CM4AI could be audited and most current deposits cannot. - **Team:** [Akshay Kumar](https://cerebralvalley.ai/u/real_akshay_k) - **GitHub:** https://github.com/realakshayk1/emperors-interactome - **Demo video:** https://cap.so/s/r07hn7sw68wt5gn?recordingStopped=1 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=177 ### 202. AIstanbul Research Group Clinical geneticists interpreting a patient's DNA constantly hit "Variants of Uncertain Significance" (VUS): changes where the evidence isn't enough to call them safe or harmful. Existing tools aggregate that evidence and score it, but none ask the question that matters most: can this evidence be trusted for this specific patient? vus-lens does. For any variant, it gathers public evidence, maps it to the ACMG rules deterministically, then audits its own confidence, flagging exactly where the evidence is unreliable, especially for Turkish and Middle Eastern patients who are barely represented in the reference databases medicine relies on. A language model reads all the evidence and reasons about it live, but a deterministic engine sets the classification: the AI interprets, it never decides. We also measured its safety: across 1,277 known-pathogenic variants, it never once returned a benign call, while its benign logic provably fires on known-benign variants. It never gives false reassurance. Built from scratch this week with Claude Code. Open-source. Live at vuslens.aistanbulresearch.com/live. - **Team:** [Ozge Cavus](https://cerebralvalley.ai/u/Ozge) - **GitHub:** https://github.com/aistanbulresearch/vus-lens - **Demo video:** https://youtu.be/p3A7NLVYrkY - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=178 ### 203. Virtual Heart Team Complex revascularization — left main, multivessel, frailty — sits in genuine equipoise. Guidelines mandate a Heart Team, yet 20–24% of recommendations flip on re-discussion, and the specialty of whoever is in the room shifts CABG rates. Most centers have no Heart Team at all. Virtual Heart Team convenes four Claude agents — interventional cardiologist, cardiac surgeon, clinical cardiologist, non-voting methodologist — over two deliberation rounds, grounded in a 30-citation evidence base built with Claude Science, every DOI verified and auditable from the UI. Sonnet runs the agents; Opus runs intake and synthesis. Each proceduralist runs in advocate and neutral framings, N=5 samples each, and drift is computed deterministically in code — no model produces the number. Drift came out near zero. So we pointed Claude Code at our own headline metric and asked it to find the flaw. It did: the advocate prompt's CONCEDE clause bounds advocacy by construction. We ablated it. Removing the guardrail widens advocate-versus-neutral divergence from 0.20 to 0.80 (−3..+3 scale, N=5); re-running on Opus widens it to 1.20. Our finding is not that specialists are biased — we induced that bias ourselves. It is that the guardrail measurably contains it, and you can see it happen in one click, on the same patient. - **Team:** [MATIAS IGNACIO PINCETTI SAUTEREL](https://cerebralvalley.ai/u/mpincetti) - **GitHub:** https://github.com/Mips99/Virtual_heart_team - **Demo video:** https://www.youtube.com/watch?v=TQRl5XZ9QeQ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=179 ### 204. T2 Labs Clonal Compass is a single-cell immune-repertoire co-pilot for paired scRNA-seq + TCR-seq data. I built it for immunologists and computational biologists who already have repertoire-analysis outputs but still need to manually connect several pieces of evidence: which T-cell clones expanded, where those cells sit in the transcriptional landscape, what marker genes support the inferred cell state, and whether any known epitope matches are available. The tool runs a Scanpy/scirpy-based analysis pipeline, computes clone expansion and cluster-level evidence, builds compact evidence bundles, and then uses Claude to generate cautious, reviewable interpretations from those computed results. It also includes a Streamlit chat interface where users can ask grounded questions about the evidence rather than prompting Claude over raw or unstructured biology data. I tested Clonal Compass on two public datasets: a PBMC baseline dataset and a tumor-infiltrating T-cell dataset. The contrast was useful: the PBMC dataset showed low clonal expansion, while the tumor T-cell dataset showed stronger expansion patterns, demonstrating how the same workflow can help summarize different immune-repertoire contexts. This matters because existing repertoire tools are strong at computation but often stop at tables and plots, while general single-cell AI copilots may not be aware of clonotype structure. Clonal Compass sits between those worlds: it makes immune-repertoire analysis more interpretable without asking the LLM to act as an ungrounded biology oracle. Quantitative claims come from computed evidence bundles, and Claude is used as a cautious interpretation layer that hedges uncertainty, cites observed metrics, and avoids diagnostic or patient-level claims. - **Team:** [Trish Whetzel](https://cerebralvalley.ai/u/tw1742) - **GitHub:** https://github.com/twhetzel/clonal-compass - **Demo video:** https://youtu.be/usFcCBZGqNE - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=180 ### 205. YC We built the first world model of the cell — and, we believe, the first to unify all four of its axes into one shared state. Expression, morphology, spatial context, and time have always been measured in separate experiments, on separate platforms, in separate distributions; no one had made them a single thing. We forced all four into one latent state you can read, roll forward in time, intervene on, and decode back into an image. This is what today's virtual-cell field lacks: STATE, TranscriptFormer, scGPT, GeneFlow, and Spatia are essentially single-modality transcriptomics predicting perturbation → expression — none unifies four axes, and none is a true world model. And we proved each axis is real, not an artifact, with a negative (shuffle) control on every one: expression reconstruction R² = 0.57; real spatial neighbors beat shuffled at z = −181 (p < 0.001); expression recovers morphology at r ≈ 0.66–0.68. We even report our failure honestly — the sharpest generated images were hallucinations, so we kept the coarse-but-correct one. Once four axes share one state, it becomes a universal adapter: reconstruct a modality a dataset never measured, transfer across platforms, run in-silico perturbations, and scale from 200k cells to 100M-cell atlases. - **Team:** [Yuchan Lee](https://cerebralvalley.ai/u/AXDeveloper) - **GitHub:** https://github.com/yc9954/cell-state-world-model - **Demo video:** https://screen.studio/share/bsVCTpEp - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=181 ### 206. Cancer crusher What I investigated My PhD work showed that inhibiting the p38 kinase in mouse models of breast cancer sensitizes tumors to immune checkpoint therapy, and that it shifts the CD4+ T cell compartment away from immunosuppressive regulatory T cells (Tregs) and toward effector cells. Why p38 blockade does this was unknown. I set out to find the mechanism in human data by asking, in a genome-scale CRISPRi Perturb-seq screen of primary human CD4+ T cells, which downstream factors carry the p38 signal into the Treg program. The screen provides genome-wide differential-expression readouts for thousands of individual gene knockdowns across three activation states, which let me treat "loss of p38" as one perturbation among a genome-wide null and ask which other perturbations reproduce its effect. What I found Knocking down p38 (MAPK14) in the human screen reproduced the mouse phenotype: it shifted CD4+ composition away from Treg and toward Th1, and its genome-wide signature was concordant with the mouse p38-inhibition signature. The shift tracked a Treg-associated program and the checkpoint molecule CTLA4, not an exhaustion program. Using two independent human Treg signatures as a readout, p38 knockdown lowered the Treg program more strongly than about 95-97% of all gene knockdowns in the screen. To find the transcription factors responsible, I scored every established p38-substrate transcription factor by whether its own knockdown moved the Treg signature the same way p38 loss did. NR4A3 was the leading mediator, ranking first among the substrate factors in both Treg datasets and among the very strongest of all knockdowns in the tumor-derived signature; ELK1 and C/EBP-beta were consistent secondary candidates. Notably, NR4A3's own transcript barely changed, consistent with p38 controlling it through activity rather than abundance. No single factor reproduced the full p38 effect, indicating that p38 supports the Treg program through several transcription-factor substrates acting in parallel rather than one dominant mediator. I then asked whether these factors act directly on Treg genes by scanning promoters and enhancers for their binding motifs. The NR4A binding motif was present in Treg-gene promoters but not preferentially at NR4A3's targets, pointing to a largely indirect route for NR4A3. The strongest direct transcription-factor-to-Treg-gene chains converged on two nodes: the master Treg regulator FOXP3 (through NFAT and NR4A family factors, established direct FOXP3 regulators) and the Treg effector and suppressive module (IL10, TGFB1, CD39). CTLA4 itself carried no proximal-promoter motif for its candidate regulators but had strong USF1 sites at its distal enhancers, nominating a p38-to-USF1-to-CTLA4 route. Why it matters No approved therapies target Tregs directly, and Tregs are a major barrier to immunotherapy response in breast cancer. These results provide a human-data-grounded, testable mechanism for why p38 blockade reduces Tregs: p38 sustains the Treg program through a small set of parallel transcription-factor substrates that converge on FOXP3 and Treg effector genes. This nominates specific, druggable nodes -- NR4A3 and the FOXP3-directed NFAT/NR4A axis foremost -- to test as points for blunting Tregs while preserving effector T cell differentiation, with the goal of making tumors more responsive to immunotherapy. Scope and limitations The screen is in CD4+ T cells under acute activation, so these findings read as quiescence-versus-activation biology rather than terminal exhaustion, and CRISPRi transcript knockdown is an approximate stand-in for pharmacological kinase inhibition. The nominated transcription factors and regulatory chains are computationally supported leads based on perturbation effects, regulon databases, and motif presence; measured binding (ChIP-seq) and functional perturbation in Tregs are the next steps to confirm them. - **Team:** [Priyanka Rajan](https://cerebralvalley.ai/u/jollymachine) - **GitHub:** https://github.com/priyan-rajan/Claude_Science_Hackathon - **Demo video:** https://youtu.be/nJsohNfijjk - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=182 ### 207. Team Veritas VERITAS is an independent verifier for AI-generated single-cell findings. A copilot says "gene X is up in Alzheimer's." VERITAS re-runs that claim across every defensible analysis choice the analyst didn't try, checks the experimental design for structural confounds, and returns a signed verdict: SURVIVES, DOES NOT SURVIVE, or UNVERIFIABLE. Every flag carries a re-derivable statistic, never a model opinion. The methods are real, not heuristics: donor-level DESeq2 pseudobulk, design-matrix batch models, Leiden, Harmony, Scrublet. Where a real method wasn't possible (ambient contamination needs empty-droplet data we never receive), we deleted the check rather than ship a proxy. What we found: pointing VERITAS at real published Alzheimer's cortex data exposed a false-green in VERITAS itself. A finding that broke only at the pseudobulk axis still scored 0.94 and was certified GREEN — the exact failure the tool exists to catch. We made pseudobulk a critical axis and wrote it up. Why it matters: Mathys et al. 2019 reported 14,274 Alzheimer's DEGs computed at the cell level. A donor-level reanalysis (eLife 2023) found 26. AI now generates findings faster than anyone can check them. Claude Science ships a reviewer, but it's the same model checking itself. VERITAS is the independent check: it re-runs the analysis and computes the number. - **Team:** [Sreenath Menon](https://cerebralvalley.ai/u/sreenath) - **GitHub:** https://github.com/sreenathmmenon/veritas/ - **Demo video:** https://youtu.be/HqZ-j2XyP6Q - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=183 ### 208. Litmus Litmus reads a scientific paper and tells you how much to trust it before you build on it. You give it a DOI, some pasted text, or a PDF, and it breaks the paper into its core claims, runs real statistical forensic checks in code (statcheck, GRIM, SPRITE, power, and a p-curve), searches the wider literature for work that supports or contradicts each claim, and then has Claude Opus weigh all of it into a calibrated replication likelihood. Every reason it gives is clickable straight back to the sentence it came from, and anything it can't ground, it throws away. I started with preclinical cancer, where only 6 of 53 landmark studies famously reproduced, because that's where a wrong call costs the most: a bad target isn't caught until Phase II or III, years and hundreds of millions of dollars later. On a set of real papers with known outcomes, Litmus cleanly separates retracted, failed, and robust work, and it stays honest about its limits with confidence intervals. It also ships as an MCP server, so any AI agent can call it to check a result before trusting it. The goal is to make it the verification layer every AI scientist and human runs through. - **Team:** [Aarush Agarwal](https://cerebralvalley.ai/u/AarushA) - **GitHub:** https://github.com/AarushAgarwal-dev/litmus , https://litmusresearch.org/ - **Demo video:** https://www.youtube.com/watch?v=aF__ls2rEqw - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=184 ### 209. JLTS Tanvi Sinha is an experimental cardiovascular biologist at UCSF. The computational work behind this project — single-cell genomics and training deep-learning models — is what she would normally collaborate on with a computational specialist like her teammate Jan Lebert. In this project we tested whether Claude Science can be that specialist. Heart-valve disease is common and largely untreatable short of surgery, and the transcription-factor logic that gives valve endothelial cells their identity is a route into understanding it. Using Claude Science, we mined public mouse and zebrafish single-cell RNA and ATAC atlases, reprocessed a mouse multiome from raw data to call 940 valve-EC enhancers, and trained a ChromBPNet model that predicts valve chromatin accessibility directly from DNA sequence. Interpreted with no motifs supplied, the model recovered a conserved ETS + AP-1 + KLF regulatory grammar centered on the mechano-sensitive gene Klf2, cross-validated across mouse and two zebrafish datasets. We also used it to package the whole pipeline as a reusable skill and build interactive demos, including a live in-silico mutagenesis browser. - **Team:** [Jan Lebert](https://cerebralvalley.ai/u/janl), [Tanvi Sinha](https://cerebralvalley.ai/u/tsinha) - **GitHub:** https://projects.janlebert.com/2026/valve-development/ - **Demo video:** https://www.youtube.com/watch?v=E7UJWPDqKbA - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=185 ### 210. AI4Cosmetics I built the QSAR Cosmetics Skill for Claude & Claude Science to accelerate the safety assessment of cosmetic ingredients. The solution combines a curated toxicology database of nearly 5,000 chemical ingredients with an ontology linking adverse effects, tissues, organs, and biological systems, and integrates over 80 machine learning and deep learning toxicity predictions into a probabilistic Bayesian framework. The skill enables toxicology teams to rapidly screen and prioritise ingredients, generate evidence-based risk reports, identify similar compounds when data are unavailable, and support early decision-making during product development. The workflow and functionality were designed based on interviews with industry toxicologists to address real-world R&D needs. Early feedback from toxicologists in industry was highly positive. They highlighted the value of the Bayesian integration approach, the transparent data provenance provided by Claude Science, and the potential to support organisations without dedicated in-house toxicology expertise. Beyond cosmetics, reviewers also identified broader applications in chemical safety assessment. This work matters because it can significantly reduce the time required for early safety screening, helping R&D teams prioritise experimental resources, identify potential safety concerns sooner, and make more informed decisions while maintaining transparency and scientific traceability. - **Team:** [Nicoleta Spinu](https://cerebralvalley.ai/u/Nicoleta) - **GitHub:** https://github.com/nicospinu/qsar-cosmetics-skill - **Demo video:** https://www.tella.tv/video/introducing-the-qsar-cosmetics-skill-e1ec - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=186 ### 211. beams-BGC BGCFlow is the standard pipeline for mining biosynthetic gene clusters (BGCs) across bacterial pangenomes, but its output is hard to interpret without expert time, and every novelty-relevant stage (antiSMASH, MIBiG, Pfam, GTDB) can only recognize what already resembles something known — homology-matching is the ceiling. We built two additive tools to break past it. BGCLens is a read-only layer that ingests a finished BGCFlow project, recommends and runs the right statistics (PCoA, PERMANOVA, enrichment, diversity), renders figures, and writes plain-English interpretations via a guarded LLM pass that strips any invented numbers or accessions. BGCFlow-PLM adds an ESM2 protein-language-model side-channel that flags biosynthetic-core proteins with no close MIBiG neighbor — a remote-homology novelty axis the databases miss — without ever touching an antiSMASH call. Together they lower the expertise barrier to acting on genome-mining results and help researchers surface genuinely novel candidates instead of only re-finding the already-catalogued. - **Team:** [Kayla Azima](https://cerebralvalley.ai/u/kaylazima) - **GitHub:** https://github.com/kaylaque/bgcflow-plm https://github.com/kaylaque/bgclens - **Demo video:** https://youtu.be/q9_J_grF_e4?si=wAEGHuPVIXr2tNh2 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=187 ### 212. David Fraile Navarro GLP-1 drugs like Ozempic are among the world's most prescribed — and in 2023 regulators opened suicide investigations after adverse-event databases lit up, even though the randomised trials showed no such risk. We reconciled that contradiction by rebuilding the entire evidence base as four strictly-separated streams — designed studies, spontaneous reports, mechanism, and an exploratory benefit arm — with formal risk-of-bias and GRADE certainty on the designed studies, measured error rates throughout, and tooling calibrated against known-answer benchmarks (AUROC 0.76). The result: the spontaneous "alarm" is unstable — its reporting ratio swings from 0.51 to 3.29 depending only on the comparator, and jumps exactly when the drug hit the news. That's the fingerprint of comparator choice and notoriety bias, not a real drug effect. The hopeful twist: reading the same database the same way, the signal points the other way for addiction — toward less drinking, less smoking, less craving across four substance classes (exploratory, but a genuine lead). Every piece — the dual-reviewer risk-of-bias engine, the GRADE harness, and the calibrated openFDA client — is open under Apache 2.0 and runs automatically: a reusable pipeline for evidence synthesis, ongoing drug-safety monitoring, and discovering new indications. - **Team:** [David Fraile Navarro](https://cerebralvalley.ai/u/dafraile) - **GitHub:** https://github.com/dafraile/glp1-neuropsych-safety - **Demo video:** https://youtu.be/wIOPfOCMhA4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=188 ### 213. GlycoWorks What I built: GlycoTwin, an open mechanistic digital twin connecting a bioreactor knob to the antibody's Fc N-glycan — a quality attribute governing potency, half-life, and immunogenicity. A 9-state CHO fed-batch model feeds an ER→Golgi enzyme train, so any process parameter propagates to galactosylation, afucosylation, sialylation, and high-mannose. Every parameter comes from public literature; nothing is fit to proprietary data. I tested one question: does Fc-galactosylation degrade at scale because CO₂ acidifies the Golgi — and if not, what controls it? What I found: The assumption is wrong. As CO₂ stripping falls from bench to 2000 L and pCO₂ climbs ~60→74 mmHg, galactosylation barely moves (~52→53%), because the CO₂ swing shifts Golgi-lumen pH only 6.43→6.38 — straddling the galactosyltransferase optimum (~6.4) on the flat top of its activity curve. Oxygen has zero effect. The real lever is the clone: overexpressing B4GALT1 recovers ~35 points; galactose feed and Mn²⁺ ~10 each; CO₂ stripping ~0. The model also reproduces the sub-5% sensitivity ceiling from a 2026 meta-analysis it never saw during calibration. Why it matters: Engineer the clone and media, not the CO₂ — productivity, not glycan quality, is the scale-up risk. Being mechanistic, the twin gives a bench-testable reason, not a black-box fit, and spans the CMC workflow: batch monitor, ICH-Q8 design space, Monte-Carlo capability, live control, and a governed agent loop. Fully open, reproducible from one script. - **Team:** [Chadakarn Sirasitthichoke](https://cerebralvalley.ai/u/Gift_CS) - **GitHub:** https://github.com/Chadakarn23CS/Hackathon-Claude - **Demo video:** https://youtu.be/Aw47CJOsea4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=189 ### 214. Anveshar Anveshar is a reproducible research workbench for rare cancers, built to run from inside Claude Science. Rare cancers are individually rare but collectively common, and each is underserved: too few patients for a dedicated trial or model. Anveshar catalogs 504 rare cancers (743 conditions), each with driver genes, potential causes, and cited cross-condition therapies scored by confidence, and presents them in an interactive Atlas Explorer that links every cancer to its genes, the analyses you can run, and the runnable foundation models. Its pipeline is reproducible and provenance tracked, pulling live data from Open Targets, DepMap, cBioPortal, ClinicalTrials, and PubMed. On top of retrieval it runs and trains sequence models on rare tumors: ESM-2 scores driver variants zero shot (pathogenic VHL R167W reads as damaging; the SF3B1 K700E change of function hotspot correctly reads as tolerated), and a trained ESM-2 head lifts missense pathogenicity from AUC 0.912 to 0.983 on 588 real ClinVar-labeled variants across twelve driver genes. Two findings are reproducible from public data: rectal neuroendocrine tumors are molecularly distinct from colorectal cancer, and in uveal melanoma an integrated multi-omic class predicts survival where a BAP1 mutation call does not. Because Anveshar makes existing knowledge and foundation models transferable, cited, and reproducible for the cancers that lack their own resources, it provides a personalized platform for tailored investigation of rare cancers. - **Team:** [Dig Vijay Kumar Yarlagadda](https://cerebralvalley.ai/u/digvijayky) - **GitHub:** https://github.com/digvijayky/anveshar - **Demo video:** https://youtu.be/uVp-Vnf9qSU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=190 ### 215. Intron This project analyzed an in-vivo single-cell CRISPR Perturb-seq screen of tumor-infiltrating CD8 T cells (Zhou et al. 2023, GEO GSE216909 — 16 samples, 180 transcription-factor knockouts plus non-targeting controls) to map which TFs steer CD8 differentiation toward or away from exhaustion, and to test whether that regulatory logic can be predicted computationally. Starting from raw 10x matrices, I called single guides per cell, integrated all 16 samples with scVI, gated to 89,754 CD8 T cells, and annotated their differentiation states, then reproduced the paper's regulatory-network analysis (MIMOSCA elastic-net regression yielding a TF×gene coefficient matrix, co-functional modules, and gene programmes) and cross-validated it with the pertpy framework (Mixscape, E-distance, Augur). Finally, I benchmarked in-silico perturbation prediction, comparing simple linear/mean baselines against the GEARS deep model. The analysis identified Ikzf1 as the dominant regulator — its knockout retains cells in the progenitor-exhausted state and has the broadest transcriptional footprint (~900 genes) — while Tcf7, Egr2, and Bhlhe40 knockouts drive terminal exhaustion along the opposite axis. Notably, the deep-learning model failed to beat trivial baselines at predicting held-out perturbations, reproducing the central finding of Ahlmann-Eltze et al. 2025. This matters because it pinpoints candidate TFs for reprogramming exhausted T cells — a central goal for improving cancer immunotherapy — while providing a sober, quantitative benchmark showing that current in-silico perturbation models are not yet reliable for predicting the effects of unseen genetic perturbations in this in-vivo setting. - **Team:** [Qiang Chen](https://cerebralvalley.ai/u/In_tron) - **GitHub:** https://github.com/In-tron/Perturb_seq_tumor_infiltrating_CD8 - **Demo video:** https://youtu.be/ZGhAu5uj7ek - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=191 ### 216. Cellarium Cellarium is a tool for asking questions of the whole-cell model of E. coli from Macklin et al. (Science, 2020). That model tracks about 16,000 molecular species in a single cell, updated every second across a full cell cycle. It is very detailed, and checking whether it behaves correctly for a given perturbation normally takes an expert days. Cellarium divides the work between two agents. A Socratic Council (a proposer, a skeptic, and a judge) reads the question and writes down a specific prediction, including the result that would count as the model being wrong. It does this without seeing any of the simulation data. A second agent, Cellwright, then tests the prediction. It has 38 tools for reading real simulation output and searching PubMed, it reports the numbers it actually read and cites its sources, and it cannot start a new simulation without a person approving it. Three findings from real runs: - We asked whether knocking out a tRNA synthetase raises ppGpp, which is what the textbook stringent response predicts. The Council committed to "2 to 4 times higher" in advance. The simulations showed ppGpp falling about 90 percent (t = −27.85), so the prediction failed. Writing it down first is what makes this a real disagreement instead of an explanation added afterward. - Asked whether nitrate turns on the nitrate-reductase genes, the first answer was yes. Once we controlled for the fact that adding nitrate also removes oxygen, those genes looked like a general anaerobic response rather than a nitrate-specific one. - Deleting ribosomal RNA operons lowered ribosome content and growth rate together. That is the opposite of what Scott's efficiency law describes. A literature search during the run found work on ribosome-limited antibiotic susceptibility (Greulich and Scott, 2015), a regime that has not been reproduced in a whole-cell model. Why it matters: models like this are usually trusted or ignored as a whole. Cellarium exposes the prediction, the failure condition, and the source of each number, so a scientist can check the model against known biology and see where it disagrees. The same approach would apply to other large mechanistic simulations, and the runs we collect become a dataset others can reuse. - **Team:** [Evangelos-Marios Nikolados](https://cerebralvalley.ai/u/evangelos), [Filippo Menolascina](https://cerebralvalley.ai/u/fmenol) - **GitHub:** https://github.com/evanniko1/cellarium/tree/main - **Demo video:** https://www.youtube.com/watch?v=RVTxlnFuf2I - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=192 ### 217. Canyon A logic-gated target-discovery pipeline for triple-negative breast cancer (TNBC). I built a pipeline that finds combinations of targets rather than single antigens, because cancer therapy must balance two opposing goals — tumor specificity and tumor-coverage — that no single target satisfies. It pairs a cell-surface antigen ("address") with an extracellular tumor protease ("permit"), scores each on separate axes, and auto-recommends logic gates of the form (antigen A OR antigen B) AND (protease). Five data sources were integrated: SURFY surfaceome, TCGA-BRCA Basal-like (TNBC proxy, n=171), GTEx 54 normal tissues, CELLxGENE single-cell TNBC (161k cells), and UCSC Xena TOIL. I found that single antigens fall short — the best, EPCAM, covers only ~47% of malignant cells and leaks into the kidney, lung, and pancreas. OR-ing two antigens widens coverage, and AND-gating with the stromal protease MMP11 (near-absent in normal tissue, active in TNBC) restores tumor-only specificity. The lead recommendation, (EPCAM OR VTCN1) AND MMP11 (coverage 51%, safety 0.96), maps directly onto real modalities such as logic-gated CAR-T or protease-activated ADCs. TNBC is just the first query — the real goal is a generalizable pipeline that programs medicines by logic rather than by a single target. Here I used only two signal categories (surface antigen + extracellular protease), which fill just 2 cells of a general 3×5 (WHERE × WHAT) space; the same framework extends across the full spectrum of readable cell signals — proteome, transcriptome, metabolome, and beyond — and to other tumor types. Long-term vision: given a cell's genotype and environmental conditions, infer the optimal target combination and design the corresponding programmable medicine. - **Team:** [Jongwon Im](https://cerebralvalley.ai/u/Canyon) - **GitHub:** https://github.com/impala9397-hub/programmable-medicine-target-discriminator-atlas - **Demo video:** https://www.loom.com/share/2f6b41dab1254a0ca62475097a4b7082 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=193 ### 218. Claim Check claim-check — a reproducibility auditor for agentic single-cell analysis. When an AI agent analyzes Perturb-seq data, it writes up what it did in prose — but its prose and the code it actually ran can silently diverge. claim-check catches that gap. It maps every claim the agent makes ("I ran MAST for differential expression," "I filtered doublets with Scrublet") onto a fixed vocabulary covering the four weighty steps of the T-cell Perturb-seq path — QC filtering, normalization, differential expression, and guide assignment — then observes what actually executed at runtime and diffs the two, emitting one verdict per field: MATCH, MISMATCH, SILENT (a step that ran but was never claimed), or UNVERIFIABLE (reported honestly rather than guessed at). The checks are strong: it doesn't just compare two labels, it asserts the claim holds in the data. doublet_filter=applied requires the cell count to actually drop and no flagged doublet to survive; log1p requires the post-normalization matrix to no longer be all-integer; target_sum=1e4 requires normalized row sums to average ~10,000. Two instrumentation channels back this: a wrapper channel that is ground truth, and a transparent monkey-patch channel that audits completely ordinary scanpy code without the pipeline knowing — catching even a claim of MAST that is really a hand-rolled scipy.stats.ttest_ind loop. On a planted-error run, it catches three real failure modes: an agent that claims MAST but executes a t-test (high severity — this directly moves the target ranking), claims a Scrublet doublet filter but removes zero cells, and silently runs a guide-assignment step it never mentioned. Why it matters: a silent mismatch corrupts the differential-expression results, and therefore the target-gene ranking built on top of them. That ranking becomes a real, expensive wet-lab experiment — a team goes after the wrong gene. claim-check catches the mismatch before that happens, and exits non-zero so it drops straight into CI as a gate. - **Team:** [Mira Arguelles](https://cerebralvalley.ai/u/tofumeow) - **GitHub:** https://github.com/mabwcls/claim-check-bwcls - **Demo video:** https://youtu.be/aJWCQC5ytq4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=194 ### 219. Shadi Shafighi ### What I built and investigated Regulatory-variant prediction models are commonly evaluated against massively parallel reporter assays, or MPRAs, as though the assay were direct ground truth. I built an estimand-aware benchmark audit that first asks a more fundamental question: do the model and experiment estimate the same biological quantity? I applied this framework to 3,976 psychiatric-risk regulatory variants studied in the developing human cortex. I compared three sequence-to-function models—AlphaGenome, Borzoi, and Enformer—with a genomic language model, Evo 2; a developing-cortex lentiMPRA; and a fine-mapped fetal-cortex eQTL atlas used as an external endogenous anchor. With Claude, I implemented custom Bayesian measurement-error and latent-variable models, coordinated model scoring, checked the analysis for reproducibility and overclaiming, and produced a calibrated per-variant concordance map. The latent agreement estimator was validated through simulation-based calibration. ### What I found The three activity-predicting models showed strong agreement with one another, with rank correlations of approximately 0.55–0.81, but essentially no agreement with the lentiMPRA, with correlation around −0.02. Accounting for the reporter’s reported measurement uncertainty did not restore agreement, indicating that this uncertainty alone is insufficient to explain the gap. Evo 2 did not join either group, showing that “sequence model” does not represent a single estimand. The three activity models also showed modest agreement with the fine-mapped eQTL effects, with correlations of approximately 0.24–0.28, while the reporter showed essentially none. Among the 25 variants with the strongest fine-mapping support, model and eQTL directions agreed for 20. ### Why it matters The conclusion is not that computational models beat experiments. It is that a precise experimental measurement may still be a misaligned benchmark when it targets a different biological quantity. Here, the lentiMPRA measures an isolated DNA element in an ectopically integrated construct, while the models predict endogenous functional signals from native-locus sequence. Their disagreement is therefore consistent with a consequential estimand mismatch, although native-locus editing is still required to distinguish this explanation from model error and shared bias. The result is not a model-versus-assay winner. It is a calibrated map showing where the measurements agree, where they diverge, and which native-locus experiments would be most informative next. - **Team:** [Shadi Shafighi](https://cerebralvalley.ai/u/dshshadi) - **GitHub:** https://github.com/shafighi/mpra-estimand-audit.git - **Demo video:** https://youtu.be/BUdYVQLRvGc?si=SklH8MIdYE11U27G - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=195 ### 220. Sakintori We built The Human Branch: a Claude Code agent that drives Claude Science as a headless lab and enforces the scientific method in code, so it cannot fool itself into calling a gBGC artifact a human innovation. Applied to Katie Pollard's Human Accelerated Regions, it takes each region from "looks accelerated" toward a testable mechanism. Why this matters: HARs look like human regulatory innovation, but GC-biased gene conversion (gBGC) forges the same signal and has produced false positives in published HAR work. So the discipline is structural, not a prompt. The conductor gates gBGC first, seals each protocol before it reads a result (a hash-chained notebook, not a promise), and never lets the data that picked a candidate also confirm it. What it found: one locus read clean in the cache but crossed the preregistered gBGC gate on the full 241-way alignment, as a flag, not a verdict, and not significant after correction. Another, ARMH3, drew orthogonal developmental contact support yet still capped at candidate, because contact is not cause. When the engine overclaimed, it caught and logged its own mistake. Every locus caps at candidate, and it says so. - **Team:** [Allen Kao](https://cerebralvalley.ai/u/shkao) - **GitHub:** https://github.com/shkao/human-branch - **Demo video:** https://youtu.be/pKkqZIQsBWs - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=196 ### 221. dbCAN-coder We built dbCAN4-advanced, an end-to-end research and software platform for annotating fungal carbohydrate-active enzymes, or CAZymes. Existing dbCAN methods primarily identify CAZyme families through sequence similarity using HMMER, dbCAN-sub, and DIAMOND. These methods are reliable when a close reference sequence exists, but can miss remote homologs whose sequences have diverged while their structures, catalytic mechanisms, and biological functions remain conserved. dbCAN4-advanced adds two complementary evidence layers: protein-language-model representations generated with ESM-C and structure-based evidence generated through ESMFold, Foldseek, and CAZyme3D. It combines multiple predictions through a confidence-aware fusion layer that can abstain when the available evidence is insufficient or contradictory. Rather than returning only a CAZy family label, the platform performs a comprehensive functional workup for each protein, including Pfam domains, EC-number prediction, signal peptides, transmembrane topology, subcellular localization, physicochemical properties, structural models, and structural-homology evidence. We turned these research components into a usable product. A single command can take a fungal protein FASTA file through baseline and advanced annotation, publish a standardized and versioned output manifest, ingest the results into a database, and launch a browsable web application. Each protein receives an evidence-rich page where users can compare baseline and advanced predictions, inspect domain architecture and membrane topology, and interact with an ESMFold structure in a 3D viewer. The system includes a Python package and CLI, a modular Nextflow pipeline, database migrations, a FastAPI interface, reproducible examples, tests, and detailed documentation. We also investigated whether protein language models and structural methods actually improve CAZyme annotation under a realistic temporal evaluation. We constructed a leakage-controlled 2024-to-2025 fungal holdout containing 337,759 reference proteins and 4,726 evaluation proteins. ESM-C was competitive with, but not universally superior to, sequence-based methods: the contrastive ESM-C approach achieved 0.973 family-level overlap compared with 0.981 for DIAMOND, and improved exact subfamily assignment over dbCAN-sub, although it remained below a custom temporal DIAMOND baseline. This result is important because it shows that protein language models provide useful orthogonal evidence, but should not automatically replace strong sequence baselines. The most consequential finding concerned benchmark design. Of the proteins initially classified as belonging to “novel fungal families,” 95.9% belonged to families already present in other kingdoms in the 2024 CAZy database. They were new to fungi, not genuinely new to CAZy. We also found severe database-vintage leakage: using the current database increased apparent novel-to-fungi subfamily recall from 0.001 with a fair 2024 reference to 0.992 because many evaluation sequences had already entered the newer database. Domain-level retrieval additionally improved exact-set performance for multidomain proteins from 0.006 to 0.412. These observations demonstrate that database versioning, taxonomic context, multidomain architecture, and abstention are as important as model selection when evaluating biological annotation systems. Finally, we developed standalone visualization and reasoning-assistance prototypes. These include interactive embedding exploration, training and calibration dashboards, evidence-grounded protein reports, conformal prediction sets, disagreement-based review triage, and an improved novelty score. The novelty prototype reached an AUROC of 0.786 compared with 0.655 for the current baseline, although we clearly identify it as a prototype that still requires integration and broader validation. This matters because CAZyme annotation supports research into fungal biology, biomass degradation, microbiomes, carbon cycling, and enzyme discovery. More broadly, the project presents a model for trustworthy AI-assisted bioinformatics: combine complementary models, preserve provenance, control data leakage, expose disagreements, and give scientists enough evidence to verify rather than simply accept a prediction. We also aim to explore AI applications in bioinformatics by leveraging current AI technologies to validate, update, and upgrade classic tools on a large scale. Our goal is to make these tools more user-friendly—enabling a wider audience to utilize them more effectively and efficiently—while ensuring the preservation of their original biological accuracy. - **Team:** [Xinpeng Zhang](https://cerebralvalley.ai/u/xzhang55) - **GitHub:** https://github.com/Xinpeng021001/dbcan4-advanced - **Demo video:** https://youtu.be/bpUkLgqAXUE - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=197 ### 222. AliKishk FluxForge is a disease-agnostic, Model Context Protocol (MCP) server that fuses primary literature, regulatory/PPI networks, and genome-scale metabolic models into an executable CRISPR gene-dependency and drug-target discovery pipeline. Its core design principle: the biological decision gates live in code as a state machine, so the orchestrating LLM is physically prevented from proposing a downstream target when an upstream biological gate has failed — integrity is enforced by the harness, not by prompt instructions. It exposes 18 typed tools over stdio + HTTP, with provenance quarantine, input guardrails, and pluggable LLM backends, usable from any AI app (Claude Desktop, ChatGPT, Cursor). What we found (rigorously, permutation-controlled across 16 systems vs real CRISPR/Tn-seq/DepMap truth): - Naively bolting literature-extracted regulatory networks onto a metabolic model adds no CRISPR signal that survives a permutation null — because generic transcription-factor networks target non-metabolic genes (a target-space mismatch). - The decisive result: re-seeding literature extraction onto the model's own metabolic gene set rescues the null. On glioma (astrocytoma + glioblastoma) and melanoma, this flipped a null layer to a significant lift — ROC Δ +0.04 to +0.06, delta-permutation p = 0.005, n_pos ≥ 50 — with glioblastoma matching or exceeding a human-curated network. We productized this as a metabolic_target_focus capability in the extraction tool. Why it matters: it reframes literature-integration for systems biology from "decorate predictions" to "reshape the model," and it demonstrates that where you aim an LLM extractor matters more than extraction quality. Just as important, we held ourselves to honest standards — every claimed lift is permutation-controlled with an n_pos floor, nulls are reported as nulls, and we corrected our own earlier overclaims. We also caveat the recall-bias risk in LLM-recalled networks and show an independent human-curated network reproduces the same rescue. - **Team:** [Ali Kishk](https://cerebralvalley.ai/u/AliKishk) - **GitHub:** https://github.com/ali-kishk/FluxForge - **Demo video:** https://youtu.be/IHxbr5yXWuY - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=198 ### 223. Berkeley-Statistics-looking-for-DS-roles-haha MutationRx turns an unreliable computational method into a decision an oncology researcher can act on, by pairing physics with judgment. What we built. When a targeted cancer drug stops working, it is usually because the tumor mutated and reshaped the protein the drug was built to fit. MutationRx takes a tumor's driver mutation, docks a panel of 62 approved drugs against both the original and the mutated protein, and quantifies each binding shift with a 95% credible interval. It then adds the step software has never managed on its own: Claude reviews every hit for mechanistic plausibility and separates real leads from coincidental docking scores. What we found. Two results. Given only the structures, the tool independently reproduces the clinic: on a real, de-identified TCGA lung tumor it had never seen, the first-generation drugs lose their grip on the resistant mutant while osimertinib holds, the exact known picture. And statistics alone are not enough: across 300 approved drugs, the math flagged ten as confident, high-significance hits, and all ten were artifacts, including antidepressants and antibiotics with no business binding this target. Claude removed all ten and preserved one non-obvious, mechanistically sound lead, imatinib, that the numbers almost buried. Why it matters. When resistance appears, the question is practical and costly: what do we try next? Today that means months of bench work, and existing computational shortcuts are ignored because their hit lists cannot be trusted. MutationRx compresses that first decision from weeks to minutes and, more importantly, makes it trustworthy. The scarce skill was never generating a docking score; it is knowing which scores to believe, a contextual and mechanistic judgment that a frontier model provides and a script cannot. The payoff is fewer wasted experiments per dollar for researchers, precision-oncology reach that is not gated by a pharma budget for the field, and a shorter path from "my drug stopped working" to "here is the next thing to try" for patients. - **Team:** [Ricardo Perez Castillo](https://cerebralvalley.ai/u/r-berkeley) - **GitHub:** https://github.com/ricardo-pc/anthropic-hackathon - **Demo video:** https://youtu.be/R8m5B9KjGAU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=199 ### 224. Selective Th2 I'm a backend/infrastructure engineer at Uber with no wet-lab background. This week I used Claude Science and Claude Code to run a real analysis on the Zhu & Dann 2025 genome-scale CRISPRi Perturb-seq screen in primary human CD4+ T cells (Marson lab) — asking a question the original study's single Th2-minus-Th1 axis structurally can't answer: which gene knockdowns selectively suppress the Th2 (allergy) program without merely skewing cells toward Th1? That distinction is the whole target definition for asthma and allergy drugs. I streamed the full 16.8 GB differential-expression matrix directly from public S3 — reading only what I needed, never downloading it — decomposed the Ota-2021 signature into independent Th2 and Th1 arms, and found an exciting signal: a clean selective quadrant, 420 candidate suppressors, and GATA3 on top. Then I tried to destroy it. I built a multi-agent review layer in Claude Code — an investigator agent proposing mechanisms, a skeptic agent adversarially refuting them, and a 6-agent audit of my own methods. Under competitive scoring the headline arm-correlation collapsed (from +0.54 to +0.18, a scoring artifact); a permutation null put the empirical FDR near 1 (no better than random gene sets); and GATA3 re-scored as a Th1-skewer, exactly as canonical biology predicts. The result is an honest, calibrated negative — plus a reproducible streaming pipeline and an agentic audit layer that caught the over-claim before it went out. It matters because most hits in high-dimensional screens are artifacts; the reusable contribution is the infrastructure for catching them, and a demonstration that a software engineer can do trustworthy computational biology precisely because the tools help you distrust your own results. - **Team:** [Shiven Tripathi](https://cerebralvalley.ai/u/shiven) - **GitHub:** https://github.com/ShivenTripathi/claude-science-hackathon - **Demo video:** https://youtu.be/yL9Hp9oprlQ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=204 ### 225. MIMIC We carry ~100x more microbial genes than human genes (~3.3M vs ~20,000), and we already know microbes signal to us in chemistry — SCFAs->GPR43, bile acids->FXR/TGR5, tryptophan->AhR, some even mimicking our own ligands (Cohen 2017). The open question is the secreted proteins: do microbes also signal in protein, and we've simply never looked? We built a structure-first screen to find out. The pipeline, one candidate. Starting from 965 million GMSC smORFs across 75 habitats, four filters narrow to 2,187 gut-only secreted family representatives, then a signal-peptide-aware mature-form refold collapses to 297 mature families. Each was co-folded against 25 targets — 15 human immune/cytokine/GPCR/barrier receptors plus 10 fold-matched decoys (the null every hit must beat) — 7,408 of 7,425 peptide x receptor interfaces folded with ESMFold2 in a single session. One candidate survives: MIMIC. Six independent computational lines of evidence: Interaction — folds a confident interface on IL7R, on the exact surface where interleukin-7 binds (~65% overlap) — co-location with the functional interface, not enrichment. Specificity — picks IL-7Ralpha out of its own gamma-c receptor family, clearing the five closest paralogs; AlphaFold3 agrees independently (two engines, same call). Not a sequence mimic — no sequence homology to IL-7. A structural mimic, not a captured human gene. Mobility — rides a Tcp/ArdA conjugative (Tn916/ICE) element across 239 gut species in two phyla (Bacillota + Actinomycetota), including 15 Clostridioides difficile strains (MGE-enrichment p = 0.003, #2 of 592 families). Conservation across that spread; and 6. alternatives ruled out. Why it matters. Leading hypothesis (of two): MIMIC mutes IL-7 to keep its microbe immune-invisible — a shield the mobile element can trade from commensals into a pathogen like C. difficile. That reframes the microbiome as a mobile library of human-receptor ligands — immune invisibility a microbe can acquire by transfer, not only evolve. Disease stakes are direct (IL-7 immune signaling; C. difficile colitis). Every panel is computational and none measures binding — this is a testable structural hypothesis, and one week settles it: add MIMIC + IL-7 on T cells and read pSTAT5 (suppressed = antagonist/shield; drives it alone = agonist). - **Team:** [Zhaojun Wang](https://cerebralvalley.ai/u/Joviana), [Kenneth Loi](https://cerebralvalley.ai/u/kenjmloi) - **GitHub:** https://github.com/kenloi/mimic-gut-secretome - **Demo video:** https://www.youtube.com/watch?v=8QeYRNmxa0Q - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=218 ### 226. BioSkeptic BioSkeptic is a red team for drug-discovery claims. You give it a target, a disease, and a direction; it checks that hypothesis against ~21 public biomedical databases (Open Targets, GTEx, the Human Protein Atlas, ChEMBL, ClinGen, PubMed, ClinicalTrials.gov, the FDA label, and more) and returns cited, grounded concerns — never a verdict. The core is a set of refuting mechanisms drawn from the target-validation literature — each a specific, falsifiable question (Is the target expressed in the affected tissue? Do human genetics point the right therapeutic direction? Does the mouse knockout show anything relevant?). We catalogued ~18 and implemented 8, then benchmarked each against real approved indications and expert-refuted links, so the agent weighs a reliable flag more than a noisy one. It's built for the scientist vetting a target before a program commits, and the computational biologist cleaning target–disease data before it trains a model. It ships as a web app, an MCP server, and a Python library, and can audit entire knowledge graphs in bulk. It gives you a grounded, checkable second opinion, early. - **Team:** [Doron Levinson](https://cerebralvalley.ai/u/doronlevinson) - **GitHub:** https://github.com/DoronLevinson/bioskeptic - **Demo video:** https://www.youtube.com/watch?v=oYrEbjehWsg - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=219 ### 227. Mesrop's Team Data Ingestor turns the hours a data curator spends hand-reformatting every CRO's differently-shaped Excel/CSV into a one-click review. The pain I keep meeting at the scientists I work with: the ingest parser is hardcoded to one vendor. Each new vendor needs its own bespoke parser, and when a vendor quietly changes its Excel, last quarter's parser breaks. Maintaining that pile is the real cost. Data Ingestor has no per-vendor code. You declare the fields you need; Claude judges the file's structure — where the table actually starts, whether a sheet is row-per-record or key-value — and proposes a column mapping with honest per-field confidence and a plain-English reason for each pick. Anything uncertain is flagged yellow, never silently guessed. Nothing is saved until a human clears every field. The design rule: Claude judges, Python reads. The model may see a bounded sample in order to judge, but every exported value is read from the file by deterministic code; a no-LLM validator re-checks it against your declared constraints. Beyond a wrapper: a learning loop (a repeat file from the same lab auto-maps at full confidence with zero Claude calls), a canonical schema + vendor-alias crosswalk with provenance, and a headers-only privacy mode. - **Team:** [Mesrop Tarkhanyan](https://cerebralvalley.ai/u/mesrop) - **GitHub:** https://github.com/mesropt/data-ingestor - **Demo video:** https://youtu.be/i21eV0Uj5Lk - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=231 ### 228. Neurosolve Neuroscience MRI research is slowed by silent failures - proprietary MATLAB tooling, out-of-memory crashes, and hundreds of preprocessing steps that never fully remove scanner artifacts. The result is data where a model's "signal" often tracks which scanner took the scan more than the patient, and adding more data only introduces more scanner-specific artifacts. These failures surface only after hours or days of wasted training. Downstream, biomarker discovery for early detection of Alzheimer's, Parkinson's, and related diseases is bottlenecked further: the labs that collect scans and the labs that link them to biomarkers don't share tooling, so a wet lab inherits preprocessed data it can't re-derive or correct. NeuroAD is a plugin for Claude Science that accelerates biomarker discovery for drug development. You research in Claude Science, call NeuroAD to quality-control and analyze your data, then step back in to keep going. Under the hood, we freeze NeuroJEPA - a foundation model strong across various tasks - and train a small network on top, tuned to your dataset and hypothesis, to weigh candidate biomarkers against the scans. The output is a ranked set of biological pathways a wet-lab researcher can act on. Before that data is trusted, we run the Silent-Failure Guard's tools directly on the MRI scans - registration, skull-stripping, anomaly and volume sanity, and silent-mirroring checks - catching confounded scans before time is wasted. The result: a verified, provenance-tracked loop from scan to wet-lab candidate. - **Team:** [Siddharth Radhakrishnan](https://cerebralvalley.ai/u/Overthrowing), [Steven Yang](https://cerebralvalley.ai/u/stevenyintech) - **GitHub:** https://github.com/stevenybuilder/NeuroAD - **Demo video:** https://drive.google.com/drive/folders/1_7ffBVwyPmRaLIaexN5qHlO9dVIjoPTw?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=232 ### 229. myanvoos Labs are about to have agents doing real work — planning assays, calling compute, triaging data, ordering supplies. But a lab record only counts if it's attributable and tamper-evident; that's the entire job of an ELN, and increasingly a regulatory expectation. If agents act through a bolted-on copilot, their work is either invisible to the record or stranded in a system no auditor trusts. Daybook's bet is that the notebook itself should be the agent's native multi-player workspace, with the same integrity guarantees applied to machine and human work alike. With Daybook, the agent operates the same notebook a scientist does, through the same set of actions. The integration surface is an MCP server with 98 tools that map one-to-one onto what a person can do in the UI — draft and sign an experiment, launch and stream a compute run, record a measurement, draw down inventory, link entities, search, request review. No model runs inside the app; the connecting client (Claude Code, Claude Science, or a scheduled background agent) supplies the reasoning. Every write, whether it comes from a human click, a REST call, or an agent's tool call, passes through one shared operation layer and lands as an append-only, hash-chained audit record that names the actor — human, agent, or automation — along with the MCP client and the model behind it. - **Team:** [Myan Vu](https://cerebralvalley.ai/u/myanvoos) - **GitHub:** https://github.com/myanvoos/daybook - **Demo video:** https://youtu.be/fY6PyN0j68g - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=233 ### 230. Orbion Biology may already contain quantum sensors we have not learned to recognize. Nebula is a discovery physics engine built to search for them. Today’s protein tools can identify sequence similarity and predict structure, but they do not answer the frontier question: which living scaffold could host a sensing mechanism, which physics makes it plausible, and what experiment would prove us wrong? Nebula turns a sensing objective into that decision. It searches public protein and structure databases, routes candidates through mechanism, family, cofactor, and structure gates, and connects biological evidence to electron transfer geometry, spin coupling, quantum dynamics, reaction kinetics, and optical transduction. It exposes assumptions rather than hiding them behind a score, then produces a public protein priority, controls, confounders, and a predeclared stop rule. One demonstration explores magnetic field sensing and surfaces Q8LPD9 with its FMN bound structure as a hypothesis to challenge. The larger result is a comprehensive workflow connecting multiple supported sensing objectives and readouts to evidence, bounded physics, falsification, collaborator handoff, and optional protein design. Nebula creates a discovery layer between quantum physics and living matter, opening paths toward cellular probes, wearable diagnostics, responsive biomaterials, and biocompatible quantum interfaces. - **Team:** [Aniruddh Goteti](https://cerebralvalley.ai/u/aniruddhgoteti) - **GitHub:** https://github.com/orbion-life/nebula - **Demo video:** https://youtu.be/8neN1zFx-CE - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=240 ### 231. Sab Moh Maya Hai Adaptive-computation halting was built to save compute, easy inputs exit early and hard ones run longer. I used it the other way, as a measurement. My backbone is ARC Institute's STATE model, frozen, and all eight layers always run, so halting saves no compute. I read the layer where the model's prediction settles as a per-perturbation number, E of N, a proxy for how much work a perturbation's response takes. Then I asked three things. Is it reproducible, is it independent of effect size, and does it tell you anything you cannot already get from response magnitude. I tested it across four Replogle cell lines, K562, HepG2, Jurkat, and RPE1, and on a single-cell CD4 T-cell CRISPRi screen. The signal reproduced across all four cell lines. In every one of the four lines, split-half reproducibility is 0.76 to 0.85, mean 0.80, and it is effect-independent in all four, so this is not a one-line result, it holds across four independent cell lines. On K562 as the worked example, reproducibility is 0.76, the partial correlation with effect size is minus 0.04 so it is not a proxy, and it recovers its oracle target at 0.61, while a naive gate just collapses to a constant. What does not carry across is the ranking itself, cross-line agreement is only 0.14, so the signal is real and reproducible in each line but cell-type-specific, not a shared invariant. In CD4 T cells it ports across donors only at the 48 hour endpoint, where the biology converges, and there it separates suppressors that spare resting cells from damaging perturbations. When I cluster on the depth signature, approved drug targets pile into one translation and ribosome cluster in every line, with odds ratios from 2.8 to 9.8, and no network statistic reproduces the ordering, best is 0.23, so it is a new descriptor. The one honest limit is that it adds no extra lift over a STRING prior for one ranking task. The whole project is falsification-first, and the negatives are as much the result as the positives. Most Perturb-seq gives one endpoint, control versus perturbed, and effect size cannot tell apart two perturbations with very different mechanisms that land at the same response size. This work recovers a stable, effect-independent signal for response complexity from the single endpoint we already collect, and it reproduced across four cell lines. It also turns an efficiency mechanism into a biological readout, which as far as I know is the first time halting is used this way on a perturbation model, and it gives a new descriptor that organizes druggability and toxicity while staying clear about where it holds and where it breaks. - **Team:** [Yash Raj](https://cerebralvalley.ai/u/yraj) - **GitHub:** https://github.com/yashraj59/RefineRx/, https://huggingface.co/yraj/RefineRx, https://github.com/yashraj59/RefineRx/blob/main/RefineRx_research_report.md, https://github.com/yashraj59/RefineRx/blob/main/RefineRx_hackathon_presentation.html , https://github.com/yashraj59/RefineRx/blob/main/09_paper/paper/paper_draft.pdf - **Demo video:** https://www.youtube.com/watch?v=x88We-QB9yE - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=248 ### 232. Cell Dna Evidence Engine is a mechanism-aware oncology evidence platform that answers the recurring BD/commercial question — "which indication should we pursue for this asset?" — in minutes instead of weeks, with every number traceable to a named primary source. Enter a drug target and it scores candidate indications on population burden (biomarker-addressable, not just disease-level), competitive density, evidence maturity, and target support — synthesized from WHO GHO, ClinicalTrials.gov, PubMed, and a provenance-stamped knowledge graph. Five modes span the workflow: Opportunity Assessment (BD prioritization), Clinical Evidence (patient/trial matching with per-criterion eligibility reasoning), CellCheck (cell-line authentication via Cellosaurus + CLASTR STR matching), Mosaic (drug-target graph), and TME (tumor-microenvironment histomics). The differentiator is an enforced epistemic contract: three-state verdicts (MET / NOT MET / NOT EVALUABLE), everything labeled CANDIDATE ONLY, and the LLM strictly downstream of deterministic data — the system physically cannot cite a source the retrieval lanes didn't produce. The MSI prediction model was trained and validated in Claude Science (AUC 0.929, cross-validated) with the circular MSI assay column deliberately excluded to keep predictions non-circular, then deployed into the product with its full reproducibility package attached as provenance. Why it matters: high-stakes oncology BD and clinical decisions are made on evidence that's slow to assemble and impossible to audit. This makes the answer fast AND auditable — trust you can check, not trust me. - **Team:** [Cell Dna](https://cerebralvalley.ai/u/CellDna) - **GitHub:** https://github.com/kunalkaiser/evidence-engine - **Demo video:** https://youtu.be/GQVpuAzDA1Y - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=261 ### 233. QAID Research Overview Current research gap. Classical gene regulatory networks (GRNs) rest on prior knowledge or on simple co-expression, which leaves the network at the level of correlation — associations without cause or direction. To move past this, I set out to build a regulatory network from real experimental perturbation data that carries genuine causality; then, from that causal network, to run a network analysis of how perturbation effects propagate through its topology, aggregate the gathered information into an embedding, and use that embedding to search for novel drug targets by guilt-by-association with known targets. Research questions. Proposed novel algorithm to read the propagation structure is a quantum walk — a quantum analogue of network diffusion that explores a graph by amplitude interference rather than by probability — which frames the two questions at the centre of this work: Does a quantum walk preserve biological information better than a classical random walk? Does this representation nominate credible, novel drug targets for autoimmune disease? The approach. The gene regulatory network (GRN) of a cell is the wiring of transcriptional control that governs its behaviour. The convergence of CRISPR and single-cell technology lets me curate a genuinely causal map: Perturb-seq couples genome-scale single-gene perturbation with single-cell readout at single-cell-type resolution. Here I curate such a causal network from a genome-scale Primary Human CD4⁺ T-Cell Perturb-seq dataset, read it with a quantum-walk algorithm that assigns every gene a quantum interference fingerprint of its position in the topology, and use the resulting embedding to nominate novel targets — testing both questions above. I also provide the causal perturb-map and its network analysis as an interactive tool so other researchers can explore the results directly. - **Team:** [Hyun Chang Lee](https://cerebralvalley.ai/u/HyunLee) - **Demo video:** https://www.youtube.com/watch?v=DTKPwFHSzpI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=266 ### 234. Orca I built a transparent computational bridge from a genome-scale CD4⁺ T-cell Perturb-seq atlas (Marson–Pritchard CRISPRi, 4 donors, conditions Rest / Stim 8 h / Stim 48 h) to tumor spatial transcriptomics (CosMx NSCLC, 765,771 cells, 960-gene panel) to ask a two-part question: within a single annotated CD4⁺ T-cell type, do spatially distinct functional states exist, and can the Perturb-seq atlas nominate a different candidate regulator for each — respecting that a knockdown's effect depends on whether the T cell is resting or activated. - **Team:** [Kim Yeojin](https://cerebralvalley.ai/u/Orca) - **GitHub:** https://github.com/YeojinKim220/perturb-to-spatial.git - **Demo video:** https://www.youtube.com/ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=270 ### 235. Variant-to-Assay Rare-disease researchers can know that a variant is pathogenic yet still not know what to measure next. I investigated how often public functional evidence can be matched to pathogenic or likely pathogenic protein-truncating variants in neurodevelopmental-disease genes, and built Variant-to-Assay (VTA), an evidence-first analysis that normalizes variant representations, checks transcript and assay context, and evaluates whether existing evidence addresses the specific biological question. In a versioned census of 15,930 public ClinVar variants across 964 genes, only variants 199 of 15,930 variants (1.25%) had an automatically identifiable exact protein-edit match under the frozen mapping rules. When evidence is insufficient, VTA identifies the remaining mechanistic uncertainty and produces a first-measurement plan, essential controls, outcome-dependent next steps, or an explicit abstention. A provisional retrospective routing pilot tied simple NMD and fixed-gene baselines, so VTA does not claim superior biological prediction. Its supported contribution is an auditable decision layer downstream of resources such as MaveDB, MaveMD and ClinMAVE, connecting existing evidence to the next experimental decision. - **Team:** [Ethan March Match](https://cerebralvalley.ai/u/EthanLM) - **GitHub:** https://mrmaster08.github.io/variant-to-assay/demo/ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=274 ### 236. Team AgingPrint AgingPrint is a perturbation-informed immune aging clock that decomposes CD4+ T cell aging into three independent biological programs: pro-aging transcription, anti-aging transcription, and naive T cell composition. AgingPrint weights were derived from 305 genes associated with aging in two independent cohorts then weighted by causal evidence from genome-wide perturbation data. Existing clocks optimize weights to predict chronological age, then call the residual biological aging. AgingPrint inverts this –– the gap from chronological age is the finding, not the error. Applied to 91 unseen donors, AgingPrint retains 98% of its performance. In healthy donors, biological immune age is normally distributed with a standard deviation of nearly 10 years meaning only 20% of “healthy” donors fall within two years of their chronological age. This variation appears to be genuine biology, not measurement error. Applied to disease, the three-component attribution reveals mechanistic distinctions invisible to single-score clocks: lupus accelerates all three aging programs, tracking from managed disease to flare-ups. Severe COVID suppresses pro-aging transcription while depleting naive T cells. Alzheimer's shows almost no peripheral CD4+ T cell aging signal. This framework extends to any immune cell type with matched CRISPRi screening data, enabling multi-cell-type biological aging assessment from a standard blood draw. - **Team:** [Jackson Schumacher](https://cerebralvalley.ai/u/JacksonGS) - **GitHub:** https://github.com/jacksongsch/agingprint - **Demo video:** https://youtu.be/-PYtDy2muqg?si=DAR06qebtJyqQlBU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=200 ### 237. Lab Voice Labs buy pipetting robots and then barely use them, because using one means writing code. So scientists who can't code pipette by hand for hours, and a machine worth thousands sits idle. That's the gap between the person who knows the biology and the machine that could run it. Lab Agent closes that gap with voice, safely. A scientist says "run an ELISA on today's plasma samples." Claude resolves that into a structured Plan through a forced tool call. It never writes robot code. A deterministic compiler expands the Plan, and deterministic validators, not the model, check SOP completeness, reagent expiry, well capacity, and what each instrument declares it can do. Nothing executes until a human confirms out loud, with assumptions read back. The core idea is capability contracts. Each adapter declares its limits, and routing falls out with no special-casing. A microliter ELISA is accepted by Opentrons and refused by an acoustic Echo. A nanoliter dilution compiles for the Echo and is refused by Opentrons. Same workflow, same validator, opposite verdicts. The hard problem isn't getting an AI to say yes. It's getting it to say no, with a reason, before a reagent is wasted or a hazard created. This runs on a real Opentrons OT-2. - **Team:** [Ainur Nygmet](https://cerebralvalley.ai/u/nygmeta), [Junchen Lu](https://cerebralvalley.ai/u/Ranidae) - **GitHub:** https://github.com/nygmeta/Claude_Life_sciences - **Demo video:** https://youtu.be/71-TSGoACCM - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=215 ### 238. Sanjana T. Labs don't sample variants randomly. They measure hotspots, clinically observed positions, the ones the models argue about, and then compute a predictor's correlation on that set and treat it as that predictor's skill. We show that number is badly biased. On NUDT15 (Suiter et al. 2020), a non-random variant set understates a predictor's true correlation by 0.29 Spearman. Inverse-propensity weighting recovers it to within 0.009, a thirty-fold correction, verified to machine precision. We also settled the question we started with. We hypothesized that measuring where predictors disagree would identify the best predictor faster than random sampling. We pre-registered the test, committed it before writing the code, and it failed. We sharpened it, pre-registered again, and it failed again, because modern unsupervised predictors, benchmarked across ProteinGym, are statistically distinguishable but practically identical on any given assay. This corroborates Fawzy & Marsh (2024), who found per-gene predictor rankings largely uninformative. We add that the meaningful unit is protein × assay. So: if you've measured non-randomly, here's your correction. If you're choosing what to measure next, randomize. - **Team:** [Sanjana Avinash Taware](https://cerebralvalley.ai/u/sanjanat) - **GitHub:** https://github.com/sanjanataware/claude_science_hackathon/tree/remediation - **Demo video:** https://www.loom.com/share/749aea6dca2f4897b8a01b2b5852531e - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=224 ### 239. Team CoEvo The standard way to detect interacting proteins (a method called DCA) assumes both organisms evolved together on the same family tree. But humans and the viruses and bacteria that infect us don't share an evolutionary history — so the textbook method can't be used here. We asked: are there other evolutionary "fingerprints" of protein interaction that don't need a shared tree? We found one. In SARS-CoV-2, the spike residues that grab the human ACE2 receptor mutate 4.2× faster across strains than the rest of that region — a clear, statistically strong signal, and no shared tree required. Testing it across 8 host–pathogen systems, it's strong for some (SARS-CoV-2) and absent for others (HIV, which hides its receptor site), so it's a useful clue rather than a universal rule. We also built a fair benchmark (13 species, 9,172 protein pairs) and tested prediction the hard way, by holding out an entire species the model had never seen. Standard AI protein models only transfer well when a close relative is already in the training data. Adding our evolution-based features gave a small but consistent gain — and helped most on viral proteins, exactly where the AI models struggle. Why it matters: predicting host–pathogen protein interactions drives the search for new antiviral and antibacterial drug targets, but the field's default tool doesn't work in this setting. We show which signals do survive, provide a reusable benchmark others can build on, and corrected an earlier over-claim of our own — making the result more trustworthy. - **Team:** [Ian Lee](https://cerebralvalley.ai/u/liyiyuian) - **GitHub:** https://liyiyuian.github.io/host-pathogen-coevolution/ - **Demo video:** https://youtu.be/WdWUGd2V-Go - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=227 ### 240. GeneScope The Gene Program Interpreter is a Claude Skill for context-aware interpretation of gene programs from Perturb-seq and single-cell data. Every experiment yields dozens of programs, and today interpreting one takes days of manual literature review — while enrichment tools give generic labels and LLM-only tools sound plausible without being grounded in the specific biological context. GPI injects that context — cell type, tissue, conditions, and, for Perturb-seq, the supporting regulators, perturbation effects, gene weights, and direction — throughout the workflow, so each program is read as a context-specific regulatory response rather than a generic gene list. It combines deterministic database evidence (enrichment, interactions), one Claude Agent SDK literature agent per program searching over MCP, a deterministic verifier that confirms every citation resolves to a real paper, and Anthropic Batch-API annotation — into an auditable, interactive HTML report. Across recent runs it interprets a program for about half a dollar and a few minutes of autonomous research, with every verified citation that you can read directly. Packaging it as a Skill makes the workflow plug-and-play and agent-native. Any scientist could use their Claude to refine context interactively on their own data, and generate publication-ready resource page. - **Team:** [Xiaochen Fan](https://cerebralvalley.ai/u/irenefan) - **GitHub:** https://github.com/ifanirene/gene-program-interpreter - **Demo video:** https://youtu.be/51G7lQjjJHc - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=228 ### 241. Solo Rhinalx is a local-first scientific decision intelligence system that preserves the reasoning behind research decisions, not just the records themselves. Laboratory notebooks capture what changed—a protocol, dose, antibody, or cohort—but rarely preserve why those decisions were made. As projects span months or years, that reasoning becomes scattered across notebooks, meeting notes, chats, and researchers' memories. Rhinalx automatically reconstructs the rationale behind scientific decisions by linking protocols, observations, meeting notes, and experimental records into a cited reasoning chain. It detects decisions with missing explanations, prompts researchers to capture the rationale before it is forgotten, and retrieves previous decisions to prevent repeating failed experiments. The system is designed around local-first AI, allowing unpublished research data to remain entirely on the researcher's machine. Every answer is grounded in traceable source evidence; if sufficient evidence is unavailable, Rhinalx explicitly states that rather than generating unsupported conclusions. By preserving institutional scientific knowledge instead of only experimental records, Rhinalx helps research teams maintain reproducibility, defend protocol decisions, onboard new researchers faster, and reduce the loss of valuable scientific reasoning over time. - **Team:** [Gbenga Oluwadahunsi](https://cerebralvalley.ai/u/BenMed) - **GitHub:** https://github.com/gbengaoluwadahunsi/Rhinalx - **Demo video:** https://yotube.com/rhinalx - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=238 ### 242. Bishal PertEMA is a post-hoc, model-agnostic reliability layer for single-cell perturbation-effect predictors. Deep models that predict how a CRISPR perturbation reshapes gene expression are used to nominate experimental targets, but on unseen perturbations they rarely beat a trivial mean baseline. Instead of building a better predictor, I ask which individual predictions can be trusted: a gradient-boosted-tree meta-model estimates each prediction's error from prediction-time features only, with isotonic calibration and split-conformal intervals. On the genome-scale Gladstone CD4 T cell Perturb-seq screen, the reliability score ranks realized error above the effect-magnitude and training-similarity heuristics, calibrates to near-nominal coverage, and builds a validation shortlist that is measurably more reproducible than ranking by effect size. I also characterized where reliability cannot help: per-instance model routing is infeasible, because candidate predictors co-fail on the same hard, noise-dominated perturbations. I quantified this with a measured noise ceiling and a break-even condition, and I pre-registered a routing test on an independent screen that failed exactly as predicted. Why it matters: it gives perturbation-prediction users a calibrated, honest abstention signal on real primary data, plus a transferable, reproducible finding about the limits of per-instance model selection, shipped as a self-hostable tool and an open reliability benchmark. - **Team:** [Bishal Shrestha](https://cerebralvalley.ai/u/OfficialBishal) - **GitHub:** https://github.com/OfficialBishal/PertEMA - **Demo video:** https://youtu.be/Id5JlmExXWM - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=201 ### 243. Pain Navigators Iaso helps clinicians and patients recognize and navigate interventional options for severe cancer pain that persists despite escalating medication. A clinician enters the patient’s pain story; Iaso identifies procedures worth specialist discussion, explains why each may or may not fit, and directs the clinician to an appropriate local specialist. Once selected, Claude drafts an editable referral summary and focused clinical question. Patients receive a cited, plain-language atlas describing realistic benefits, limitations, and risks, helping them use limited consultation time effectively and participate meaningfully in treatment and informed-consent discussions. Iaso is also an exploration of where AI belongs in high-stakes clinical care. Claude translates ambiguous narratives before a decision and explains results afterward, but does not make the clinical decision itself. Recommendations are governed by fixed, versioned, physician-authored rules. This deterministic safety harness keeps every criterion transparent and testable while reducing the risk that hallucinations or inconsistent model responses alter outcomes. Both the clinical rules and referral network can be customized to an institution’s procedures, selection criteria, specialists, and workflows. In cancer care, a delayed or misdirected referral can consume precious time. Earlier, better-directed, better-prepared specialist conversations can protect what matters most: quality of life and meaningful time with loved ones. - **Team:** [James Cho](https://cerebralvalley.ai/u/Whitening) - **GitHub:** https://github.com/jamescho7/Iaso-Hackathon - **Demo video:** https://youtu.be/jGLSnb-TP0E - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=202 ### 244. wguesdon Which pair of cell-surface proteins marks prostate cancer cells while sparing every healthy human cell type? Antigen-directed therapies are limited by on-target off-tumor toxicity; few antigens are absent from all normal tissue. We scored combinatorial surface-marker pairs, AND gates and NOT gates, per patient across a 24-patient localised prostate cancer atlas, against an assay-matched, donor-robust Tabula Sapiens reference and a matched benign-prostate control from the same tumors. The method recovers the preclinically validated PSMA-PSCA pair as a positive control: each antigen alone engages duodenum or bladder, and requiring both collapses the worst extra-prostatic co-detection roughly sixfold. It nominates one clean surface pair, PSMA × STEAP1 (both antigens have clinical binders; the pair is already a phase-1 dual ADC), and finds no usable NOT gate. An independent 369,000-cell cohort (HuPSA) confirms tumor-specificity but shows the pair covers AR-driven adenocarcinoma and is lost in metastatic and neuroendocrine disease. Claude Code drove the pipeline (datasets, containerised doublet removal and cohort conversion, uncertainty and label-leakage controls, a multi-agent report review); Claude Science added the literature survey and a GTEx and Protein Atlas protein-level safety cross-check. This is a reproducible, hypothesis-generating finding with a truth value, not a validated target. Links - Live site: https://base-by-base.com - Interactive explorer: https://base-by-base.com/explore - access code: 5OjDWJdAHq73BEqf5Y-xJQ - Demo video (3 min): https://www.youtube.com/watch?v=ejc4G0Y-WgE - Code (open-source, MIT): - Analysis + report: https://github.com/wguesdon/dual-marker-discovery - Website: https://github.com/wguesdon/dual-marker-discovery-site - Explorer app: https://github.com/wguesdon/dual-marker-discovery-app - **Team:** [William Guesdon](https://cerebralvalley.ai/u/wguesdon) - **GitHub:** https://base-by-base.com/ - **Demo video:** https://www.youtube.com/watch?v=ejc4G0Y-WgE - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=205 ### 245. emberwhirl-solo I built PerturbScope, a dashboard that helps wet-lab biologists, clinician-scientists, and hybrid researchers turn large CRISPR Perturb-seq screens into reviewable target shortlists. The application combines target exploration, configurable evidence gates, matched-condition comparisons, human–mouse translation evidence, custom cell-level phenotypes, immutable provenance, and reproducible exports. I investigated how prioritization changes when reliability is evaluated before effect size. I found that strong effects can be misleading when knockdown is weak, guides disagree, donor responses fail to reproduce, coverage is limited, or translation evidence is poor. PerturbScope therefore exposes missing data, warnings, replicate structure, and analysis lineage rather than collapsing them into an opaque ranking. This matters because experimental validation is costly: transparent trust gates help researchers allocate scarce resources to candidates supported by stronger, auditable evidence while preserving scientific judgment and enabling independent review. - **Team:** [Yu-Tian Xiao](https://cerebralvalley.ai/u/emberwhirl) - **GitHub:** https://github.com/Emberwhirl/PerturbScope - **Demo video:** https://www.youtube.com/watch?v=7qDl_KcsowU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=209 ### 246. Orchestrated Biosciences Tumor mutational burden (TMB) is our best predictor of immune checkpoint blockade (ICB) response, yet it's only modestly accurate (AUC ~0.6) and reads DNA alone. I asked whether a tumor's RNA adds independent signal. I built a cohort-scale pipeline for non-reference RNA features (alternative splicing, intron retention, RNA editing, transposable-element activation, fusion transcripts) across five melanoma ICB cohorts, and ran the Evo2 genomic foundation model on novel splice junctions. Using strict leave-one-cohort-out testing, nothing in the RNA layer beat plain TMB — and I found why: every RNA signal was a proxy for how inflamed the tumor already was, and that proxy flips sign between cohorts, which may explain why so many RNA biomarkers fail to replicate. Reframing from response to survival surfaced the real result: subclonal mutation burden — a measure of how disordered a tumor's evolution is — predicted worse overall survival across three cohorts (HR 1.49, p=0.0004), separable from both TMB and immune state. The takeaway: how many mutations a tumor carries shapes whether it responds; how disordered its evolution is shapes how long the patient lives. A rigorously reported negative plus a novel, cross-cohort survival signal — and an honest account of where the signal actually lives. - **Team:** [Alex Nesta](https://cerebralvalley.ai/u/alexvnesta) - **GitHub:** https://github.com/alexvnesta/evolutionary-rna-state - **Demo video:** https://youtu.be/ZO68xiCyPn8 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=210 ### 247. Jaymin I used Claude to train a protein language model that can generate an extremophilic version of any enzyme, maintaining its function, while allowing it to survive extreme temperatures, pH, salinity, etc. This allows engineering and deploying functional enzymes for applications that require extreme resilience. This project was exceptionally complex. Collecting the training data required several datasets, tools, analyses, and filters (GTDB, GenomeSPOT, signalP6, mmseq2). Training the ESM2 adapter and classifiers required meticulous care to avoid biases in the training data and overfitting. And creating the final masked PL generative pipeline was equally complex to retain enzymatic function, avoid biases, and to generalize across any enzyme (ESM, ProteinMPNN, Swiss-prot, foldseq, mmseqs2, M-CSA, etc). Claude made this possible, both with the coding and with setting up the many mini-experiments I needed to weed out biases and structure the problem correctly. A 6-12-month scoping project was compressed into 5 days. The complex workflow can be tracked in my github repo attached and through my labnotebook (https://github.com/jayman1466/Extremophilic-Protein-Translator/blob/main/labnotebook.md) - **Placement:** Finalist - **Team:** [Jaymin Patel](https://cerebralvalley.ai/u/jayman1466) - **GitHub:** https://github.com/jayman1466/Extremophilic-Protein-Translator - **Demo video:** https://youtu.be/n7mDeIZ3Tro - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=211 ### 248. Sera Sera helps target-discovery biologists decide which screen hits to advance. It reconciles a CRISPRi mRNA screen against a protein screen: for each gene and condition, do the readouts agree, or does one flag a hit the other misses? On Alex Marson's genome-scale CD4+ T-cell screen, Sera ranks 7,195 candidates, then has Claude challenge its own front-runners on donor/guide robustness, real knockdown, and statistical power, and shows the rejects. Take its top corroborated hit, NFKB2: the readouts agree at 8 hours, then split by 48, when only the protein screen flags it. That late, protein-only signal is the fingerprint of noncanonical NF-κB, which acts after transcription where an mRNA screen goes blind. Sera returns one call: advance or hold, plus the experiment that settles it. The trust rule is simple: Claude reasons, but it never writes a fact. Deterministic code produces every number, identifier, and citation. Claude can only point, choosing a paper or structure by its position in a code-retrieved list, never typing a PMID or accession. It structurally cannot invent a reference. A controls-first gate also hides every novel pick until Sera first recovers the screen's known biology (5/5, both screens). - **Team:** [mohamed elrefaiy](https://cerebralvalley.ai/u/melrefaiy) - **GitHub:** https://github.com/mohamedelrefaiy/Sera.git - **Demo video:** https://drive.google.com/drive/folders/1Sk7tQv-5UU2HsjKtcpYxwRwjuNKu7BoL?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=243 ### 249. CiteEval citeval is a citation referee for the life-sciences literature. Given one scientific claim and the paper cited to support it, it decides whether the paper actually supports the claim — across three axes, each verdict carrying a confidence tier and a provenance label (metadata-grounded vs. model-inferred): (1) retraction, checked against a ~29k-record Retraction Watch index cross-referenced with NCBI; (2) species/population scope, where MeSH organism/age tags establish the paper's real scope and Claude Opus 4.8 judges whether the claim over-generalizes; and (3) a six-level substantive-support grade from an adversarial supporter→refuter→critic triad on Opus 4.8. The guiding principle is grounded-and-measured, not model-checks-model: objective facts are decided by external structured records that override the model, and the subjective grade is measured against an independent reference, calibrated, and stress-tested on a held-out set it was never tuned on. What we found: retraction detection is perfect in-domain and out-of-sample (P/R 1.00); the graded axis reaches 0.90 quadratic-weighted kappa, is within one level on every case, carries zero confident-but-wrong grades, and recalibrates to an honest 0.08 ECE; scope holds at 0.909 in-domain / 0.71 held-out. Along the way the tool overruled its own answer key (flagging retractions our corpus had mislabeled), and a held-out test caught us propping a scope metric — which we fixed honestly rather than hide. Why it matters: a 2025 review puts the quotation-error rate in the medical literature at 16.9%, retracted papers keep being cited as live, and AI drafting is spreading misattribution. citeval answers the specific, checkable question — does THIS paper support THIS claim? — and refuses to answer with a number it can't defend. It ships with a live web UI (five-stage pipeline stepper, adversarial-debate replay, live cost/ETA, parallel multi-paper dashboard), a FastAPI backend, and offline proof pages — every headline number reproducible offline in seconds via make eval. - **Team:** [Pranav Lakhwara](https://cerebralvalley.ai/u/Mythical_Hacker) - **GitHub:** https://github.com/mythicalhacker/citeval - **Demo video:** https://youtu.be/5Vk0iH2gO6Y - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=253 ### 250. chromagentic **Agentic Chroma is an MCP server that gives an AI agent the role of a human operator at a Waters chromatography workstation.** From any MCP-capable Claude client, a researcher runs liquid-chromatography experiments in plain language: the agent browses Empower 3 projects and methods, reads an instrument method back in plain words (each solvent line's contents, the TUV wavelengths, the pressure limit), authors a sample set with the correct instrument-prep prologue, requests and starts acquisition, watches the live run on a local dashboard, and analyzes the resulting chromatograms — peak detection, retention time, resolution, and peak shape. **The hard part isn't the language — it's doing this safely against a real instrument.** The design keeps the human as *both the hands and the gate*. The agent emits a concrete physical prep plan (mobile phases, wash/waste, column, exact vials to load); the researcher preps the bench and then personally approves **every** instrument `connect` and **every** injection through a client-side permission prompt that carries a LIVE INJECTION warning. A server-side ordering backstop (`start_acquisition` refuses unless `request_acquisition` prepared the same project + sample set) complements — but never replaces — that human gate. `abort_run` is always available and never gated. **Under the hood it's a capability-based driver core.** A 64-bit MCP server talks over JSON-RPC to a 32-bit COM sidecar that hosts the Empower 3 Toolkit; drivers translate the vendor model into a canonical domain model (`Project`, `Method`, `SampleList`, `Injection`, `Chromatogram`, …) and self-register, so the MCP tool surface lights up per the active backend's capability manifest. A `simulated` backend exposes the full tool surface with no hardware; the `empower` backend is **live on a real ACQUITY UPLC + TUV** on my workstation; a `waters_connect` (Xevo G2-XS QTof, REST/OAuth2) backend is the next drop-in driver. Adding an instrument means writing one driver package — never editing the core. **What I found / what's live-verified.** Against the real Empower Toolkit I confirmed the sample-set prep functions store and round-trip end-to-end (`wet_prime` 34, `wash_needle` 33, `equilibrate` 9, `inject` 1); that instrument-method summaries can be pulled straight from `InstSetupCOM.Xml` and resolved even though this Toolkit build doesn't register a `MethodSet` COM class; and — importantly for safety — that the permission gate has a real trap: the very first gated call after enabling the MCP server mid-session did **not** prompt, and only after re-confirming the Ask rules via `/permissions` did every subsequent live-injection call prompt correctly. That finding is now baked into the acquisition SOP. **Three concrete cases (the demo).** I hand Agentic Chroma three real questions from my bench, in plain language, and it answers each on the real instrument with me as the gate: 1. **Quantitation + report** — it runs a semaglutide injection live, builds a calibration curve from the standards in sample set `sema_20260708`, back-calculates the semaglutide concentration in my unknown wash samples, and writes up an analysis report. 2. **Carryover** — it checks the blank injected after the highest standard for residual analyte, quantifies the carryover, and recommends a concrete instrument-method fix. 3. **Retention drift** — with no new run at all, it compares a development peptide's standards across two prior runs, quantifies the retention shift, and recommends a method change to control it. The calibration fit, quantitation, carryover math, drift comparison, report, and every recommendation are the *agent's own analytical reasoning* on top of the read tools — not a canned quantitation feature. That loop — ask, run (approved), analyze, recommend — is what the tool exists to close. **Why it matters.** Today an LC–MS run is a static file uploaded *after* the fact, and the reasoning behind each method-development decision evaporates. Agentic Chroma turns the run into a supervised, auditable loop — interpret the goal, propose injections and conditions within guardrails, submit and monitor, evaluate the data, recommend the next experiment — with a person approving every physical action. For analytical labs it's the missing operating layer between *a scientist with a question* and *an instrument that only speaks Empower*. - **Team:** [Tao Zheng](https://cerebralvalley.ai/u/ztao) - **GitHub:** https://github.com/zt10001/agentic_chroma - **Demo video:** https://www.youtube.com/watch?v=c6sVTKO7xlo - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=203 ### 251. CAR-DWK Please check the final report on the GitHub repo! 1. What We Investigated and Built: Conventional CAR-T screens fail to replicate the complex, months-long in vivo tumor environment. We built an automated single-cell multi-omics pipeline using the EpiVIA framework to perform an unbiased insertional-mutagenesis screen. By repurposing the semi-random lentiviral integration site (IS) as a natural genetic barcode, we tracked exactly which host genes were disrupted during manufacturing, monitored their clonal expansion post-infusion, estimated their T cell status inside the body, and selected final candidates for Next-Gen CAR-T engineering. 2. What We Found - Clonal Expansion: From the BCMA dataset, KDM2A/ATF7IP/SMYD4/ERCC6L2/FANCA show long-term persistence with a cycling phenotype and exhaustion resistance as potential candidates. Especially, KDM2A surfaced as a powerful candidate, where 13 independent in vivo integrations near the locus drove a massive late-stage clonal expansion. - Recurrent Targets: Across three independent datasets (human and mouse), CBLB, CYLD, and FNBP1 consistently emerged as non-essential, cross-cohort drivers of in vivo survival. - Escaping Exhaustion: Late-surviving clones systematically avoided the epigenetic exhaustion scar, shifting instead to a sustained cycling and effector state. 3. Why It Matters Our pipeline reads actual in vivo biology from patients actively clearing tumors, bypassing the limitations of engineered in vitro models. Because it is unbiased and resolves multi-hit (MOI > 1) cells, it can nominate unexpected targets for in vivo long-term persistence. This open-source tool provides a robust, clinically grounded platform to discover and prioritize targets for next-generation, exhaustion-resistant CAR-T therapies. - **Team:** [Dongwoo Kim](https://cerebralvalley.ai/u/DDW) - **GitHub:** https://github.com/2DW-K/BuiltwithClaude-In-vivo-insertional-screen-for-CAR-T-persistence-in-human - **Demo video:** https://youtu.be/VEUX8OUoodU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=207 ### 252. Baxijen Aquiles — auditable, host-directed antiviral target triage for Aedes arboviruses What we built. Dengue, Zika and chikungunya have no approved antiviral — and because they mutate fast, drugs that hit the virus directly lose to resistance. Aquiles takes the opposite route: it looks for the human proteins the virus hijacks to replicate — the machinery it can't live without — and ranks which of those host factors are both essential to the virus and druggable. A drug that targets your own protein is far harder for the virus to escape, and can potentially work across all three arboviruses at once. For each host-factor gene, three Claude agents work in sequence: a specialist (Haiku 4.5) gathers cited evidence from public biomedical databases (Open Targets, DGIdb, Europe PMC) plus the Krogan host–pathogen interactome and published CRISPR host-dependency screens, and makes concrete, checkable claims. An independent verifier (Sonnet) then re-reads every claim in an isolated context, against the raw database record — marking each one confirmed, corrected, or unsupported. Only verified evidence flows into a final synthesis agent that writes a go / watch / no-go dossier and, where an approved drug already exists, a repurposing hypothesis. Druggability and evidence-strength scores are computed in code from verified claims — never written by the model. The whole pipeline is exposed as an MCP server, and a web app lets a scientist ask in plain language and watch the verification happen live. What we found. The independent verifier caught [52 of 79 — 66%] of the specialist's own claims as unsupported or wrong — mistakes a self-checking AI would have delivered as fact. That is the core result: in biomedical AI, the reviewer is usually the same model with the same blind spots, and separating the verifier into its own agent with its own context measurably changes the output. Why it matters. The bottleneck for trusting AI in the life sciences isn't generating hypotheses — it's knowing which ones to believe. Aquiles turns an antiviral discovery question into an auditable one: every number traces back to a public source that survived independent verification, and a human stays in the loop on every dossier. It's a template for AI-assisted drug discovery you can actually check — applied to a disease burden that hit 6M+ cases in Brazil alone in 2024, with no cure on the shelf. - **Team:** [Marcus Antonio Cardoso Ramalho](https://cerebralvalley.ai/u/nextmarte) - **GitHub:** https://github.com/nextmarte/weratemosquitoes and https://ultron.cid-uff.net/achilles - **Demo video:** https://youtu.be/yo_-jEfpd1s - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=225 ### 253. MotorSmiths I rebuilt nature's strongest molecular motor — the phi29 gp16 DNA-packaging ATPase — as a single programmable protein, entirely in silico. The motor is a ring of five identical subunits. That sameness is a wall: you cannot tune one subunit, read the firing order, or even model the oligomer. I fused the five subunits into one covalent chain (cp233, 1,750 aa) using circular permutation, so every seat becomes individually addressable and the ring becomes computable. I then evaluated it. Structurally it is indistinguishable from native: three independent predictors agree 5/5, subunit RMSD 1.80 A / TM 0.94, and it threads dsDNA at the native contacts — designs that merely close the ring cannot. Driven molecular dynamics plus a mechanochemical-ratchet model predict a hand-over-hand mechanism with a falsifiable 3D-MINFLUX signature. The design method generalizes across 13 ring motors and passes a blind retrospective test on ClpX (6/6 working, 0/6 dead mutants). Why it matters: this turns the strongest packaging machine in biology into one that is both readable — mechanism resolvable one seat at a time — and tunable, a concrete step toward programmable viral DNA packaging (phage engineering, viral-vector production, DNA nanotechnology). Everything is computational, cross-checked, and honestly bounded; it hands the wet lab a concrete, testable to-do list. - **Team:** [Longfu Xu](https://cerebralvalley.ai/u/longfuxu) - **GitHub:** https://github.com/longfuxu/Rebuilding-motor-hackathon - **Demo video:** https://youtu.be/ZxMc6qudB5Q - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=226 ### 254. Longevity Nerd APOE genotype is fixed at birth, yet Alzheimer's appears decades later. We asked whether the fixed APOE ε2/ε4 plasma-protein substrate (Lu 2026) interacts with the age-emergent proteome — three proteomic waves (Lehallier 2019) and many organ and cell-type aging clocks (Ding 2026, Oh 2023) — by testing protein-network coupling across five interactomes and two states (plasma and brain), against degree-matched nulls. Network-wise, the two programs are largely segregated across most aging and aging-related changes and both alleles — especially ε2. One moderate overlap sits at ε4 × later-life aging (the ~78-year wave), carried by a lipoprotein/ECM arm (APOM, FBLN1 → CLU, APOL1, FN1), an inflammatory arm (IL6ST, KITLG, RGMA → IL6 / STAT3, KIT) and a neuronal arm (NEGR1, NPTX2, NRXN1, GFRA2). Independent human genetics give these proteins support for causality inference — CLU in Alzheimer's, FN1 in coronary artery disease — and both are potential drug targets, with CLU-raising compounds already showing preclinical Alzheimer's efficacy (Cohn 2025). Built and stress-tested two ways with Claude Science: one chat end-to-end, and a multi-chats system passing results through shared artifacts. - **Team:** [Naisha Shah](https://cerebralvalley.ai/u/longevity_nerd) - **GitHub:** https://github.com/naisha-shah/ClaudeScience_Hackathon - **Demo video:** https://youtu.be/eOJRm_Jpr9g - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=229 ### 255. VoganMolBio Motif is a molecular-biology workbench that runs inside Claude Science as a plugin. Instead of answering sequence questions with prose, it turns a request into a real object you can inspect: annotated DNA, RNA, and protein records, circular and linear plasmid maps, restriction digests and gels, primer design and PCR, Gibson, Golden Gate, and GoldenBraid planning, multiple-sequence alignment, Sanger trace review, and typed analysis results. Every derived record keeps its parent and the transform that produced it, so results can be traced and checked. The bio functions are deterministic, so the same request gives the same answer. The workbench is a single self-contained HTML artifact that runs locally in the browser. A narrow MCP connector mounts the interface inside Claude Science, with a saveable HTML fallback for hosts that do not mount MCP Apps. Nothing leaves the machine. - **Team:** [Jacob Vogan](https://cerebralvalley.ai/u/JacobMolBio) - **GitHub:** https://github.com/jvogan/motif ▎ Landing page: https://jvogan.github.io/motif-site/ - **Demo video:** https://youtube.com/shorts/s__GSRt9NGw?feature=share - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=235 ### 256. Om Litmuz is a claim-level verification layer for life-sciences research agents. An AI agent's memo is fluent and cited, but some claims are unsupported or flatly contradict the primary data, and nothing checks them one by one before they become results. Litmuz breaks a memo into atomic claims and returns an honest, per-claim verdict (grounded, needs review, or flagged), routes anything uncertain to a human, and never auto-passes a safety-critical claim. For this hackathon I added a genomic mode that verifies claims directly against Gladstone Institutes datasets: the Pollard lab's Human Accelerated Regions and the Zoonomia mammalian-constraint work. The check is deterministic, with no language model in the trust-critical path. A claim such as "HAR1 shows no human-specific change" is flagged as contradicted, with the exact evidence cited, and anything outside the curated reference returns an honest "cannot confirm." It is the verification layer autonomous genomics research is missing. litmuz.co - **Team:** [Om Om](https://cerebralvalley.ai/u/ombharatiya) - **GitHub:** https://github.com/ombharatiya/litmuz-agent - **Demo video:** https://youtu.be/2lqCW7myTos - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=239 ### 257. Drug Target Hunters Our analysis combines monogenic immune-disease genes, evolutionary constraint, GWAS, and genome-scale CRISPRi perturb-seq in primary human CD4+ T cells, ranking all 18,692 genes with a positive-unlabeled model that uses no drug-approval information as input. Two findings: (1) it recovers known immune-drug targets at 8.3-fold enrichment in its top 100 under group-aware cross-validation; (2) in a blind test of novelty, 24% (6/25) of its top novel nominations are already in clinical development, versus 2% (1/50) of matched controls (Fisher exact odds ratio 15.5, p=0.005). It independently recovers the active pipeline and nominates the next candidates (MAP3K14/NIK, RORC/RORgt, MALT1) on the antigen-receptor to NF-kB axis. - **Team:** [Alexander Han](https://cerebralvalley.ai/u/Alex-Han) - **GitHub:** https://alhanster.github.io/immune-target-atlas/ https://alhanster.github.io/immune-target-atlas/technical-report.pdf https://github.com/alhanster/immune-target-atlas-technical-report - **Demo video:** https://www.youtube.com/watch?v=gXhQpuQM8xs - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=242 ### 258. G-perturb G-perturb re-ranks drug targets from a genome-scale CD4+ T-cell Perturb-seq screen by whether each effect is dependable, not just by how large it is. A large transcriptional effect seen in a single guide or a single donor can be measurement noise, and each false lead costs a validation experiment. I treat every perturbation effect as a measurement in a crossed guide-by-donor-by-condition design and use generalizability theory, a framework from psychometrics, to give each target one dependability coefficient, then rank by effect size weighted by dependability. On the released 44.6 GB pseudobulk this reorders the shortlist: 50 of the top 100 targets by raw effect drop out of the top 100 by dependability. Read within activation states, the coefficient reconstructs the T-cell-receptor module (CD3 complex, ZAP70, LAT) as reliable only in activated cells, with no gene labels supplied. A design study shows the screen is limited by the number of guides, not donors. Every methodological decision was recorded and adversarially red-teamed by a competing model before any result was seen; the analysis regenerates end to end, and the manuscript is submitted to bioRxiv. - **Team:** [Che Cheng](https://cerebralvalley.ai/u/che830621) - **GitHub:** https://github.com/kiki830621/G-perturb - **Demo video:** https://youtu.be/zBD30nhal64 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=206 ### 259. SVBench AI SVBench AI is a Claude verdict layer for structural-variant callers. One reviewer — Claude Opus 4.8 — reads each call's alignment image and genomic context and returns an evidence-cited verdict, with or without a truth set, corroborated against independent population catalogs (gnomAD-SV, dbVar, HGSVC). Against HG002 it showed 93% of Sniffles' and 71% of SVIM's "false positives" are benchmark artifacts, not caller errors (raw 0.886→0.992; 0.694→0.886) — and SVIM's corrected precision equals Sniffles' raw, exposing a gap the benchmark hid. On genomes with no truth, the same reviewer still delivers grounded verdicts. Trustworthy SV evaluation, everywhere. - **Team:** [Ramanandan Prabhakaran](https://cerebralvalley.ai/u/Ram_Lifescience) - **GitHub:** https://github.com/Ramanandan/svbench-ai , - **Demo video:** https://youtu.be/b02PkH2w4I4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=208 ### 260. CisFalcon AI now designs synthetic enhancers meant to switch a gene on in one cell type and stay silent everywhere else. When one fires in the wrong cell, the consequence can be clinical: in a real enhancer-AAV gene therapy for Dravet syndrome, driving the gene in all neurons instead of only interneurons raised mortality in mice, while the cell-type-specific version was safe and corrected seizures (Mich et al., Sci Transl Med 2025). CisFalcon catches that failure from sequence, before a lab spends six weeks and a synthesis run. Paste a designed enhancer into the live tool (https://cisfalcon-lifesci.fly.dev/) and it scores cell-type specificity against a frozen external measured-activity model, then applies the prescribed motif edit and re-scores from failing to passing (a proposed fix, not wet-lab validated). On 93,435 independent designs from a different lab it separates failures at AUROC 0.80; ranking a batch safest-first cuts wasted syntheses about 70%. We also tested the harder within-tissue question directly, on real in-vivo cortical enhancers with an independent model (AUROC 0.71). It flags risk; it does not design DNA. - **Team:** [Ubaidullah Shuaib](https://cerebralvalley.ai/u/corridor) - **GitHub:** https://github.com/belumume/cisfalcon - **Demo video:** https://youtu.be/L5xhnDS4weI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=212 ### 261. Clinibrium Clinibrium is a clinical decision-support co-pilot for the otoneurology clinic, developed with a subspecialist otolaryngologist. It triages a common complaint — vertigo — that hides a rare emergency: posterior-circulation stroke. Its founding guarantee: the model never makes the safety call. The architecture inverts the usual LLM app — deterministic layers own everything that must be trustworthy. A RedFlagEngine, physically separated from the differential, and hard rails seal the urgency verdict with escalate-only, monotone semantics. A CatBoost + SHAP confidence layer, trained on real clinical cases, doesn't touch urgency — it feeds a recommendation to Claude. Claude reasons over it, grounds its explanation in paraphrased ICVD criteria, and picks its own model (Opus for urgent cases, Sonnet for ambulatory). When Claude catches a subtlety the rules miss, it can raise the alert one level — never lower it. Kill the ML or Claude live and urgency doesn't change; we test exactly that. Our most creative use of Claude Code is the Safety Harness: a hook that reruns invariant tests whenever Claude edits a safety-critical file and blocks the commit if a guarantee breaks, plus a skill that turns a specialist-signed rule into adversarially-tested code. Twelve invariants, +300 tests, multi-model adversarial audits, tamper-evident FHIR, on-device video. Specialist-reviewed; honestly scoped. - **Team:** [Manuel Pena Zuniga](https://cerebralvalley.ai/u/manuelpzdev) - **GitHub:** https://github.com/manuelpenazuniga/Clinibrium - **Demo video:** https://ioom.short.gy/5i5hNf72ywbwu2 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=214 ### 262. Piotr Kulesza A lab protocol is a document. It describes every step, in order, and it is completely correct — but you have to follow it. You keep your own place, you remember what you already did, you watch the clock. It should be the other way round: the protocol should follow you. benchpilot turns any wet-lab protocol into a run. A single Claude call reads the prose and returns a procedure — steps, live timers, hazards, either/or choices, and the gaps the document never filled, which it asks you about before you start. Then it shows you each step: one sample travelling through real 3D glassware, a centrifuge that spins for exactly as long as the timer runs. And because your hands are busy, you talk to it. Claude turns free speech into actions, so you can start a timer, move through the run, or dictate a note without touching anything. Everything you actually did is recorded. Science starts where people learn to do it. When you're starting out, benchpilot leads you — it shows you what the words actually mean. When you know what you're doing, it follows you — it keeps your place, keeps your hands free, and writes down what really happened. - **Team:** [Piotr Kulesza](https://cerebralvalley.ai/u/pkulesza) - **GitHub:** https://github.com/piotr-kulesza/benchpilot - **Demo video:** https://youtu.be/WbVRKYCq4SA - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=216 ### 263. F15 VCell We worked on the Arc Institute Virtual Cell Challenge: given a gene knockdown, predict how the cell’s expression profile changes. Scoring is official Arc cell-eval on three metrics - DES (did you recover the right differential genes?), PDS (can you tell knockouts apart?), and MAE (overall error). We built our own Early-S baseline, then ran a fast loop of invent → implement → run on Azure GPUs → rescore with cell-eval. Across roughly twenty experiment waves we took official test DES from 0.1 baseline to about 0.5 (where 0.3 is considered an impressive result in the field that is very challenging to beat), with MAE as low as 0.017. That’s a big step on DES relative to common published ranges people cite for this problem. PDS results are less impressive, and are an area that we will focus on improving next, but remain above the baseline. The point of the project isn’t a new foundation model, but showing that with a hard public evaluator and a lot of measured iteration, you can push the biologically relevant DES axis a long way in days, and be clear about where the method still fails (perturbation identity). But the fact that we were able to implement and conduct hundreds of experiments is genuinely remarkable and we would not have been able to do that in a week were we not using Claude. - **Team:** [Petr Berlizov](https://cerebralvalley.ai/u/pberlizov) - **GitHub:** https://www.overleaf.com/read/ddpsrzfftchp#67598a - **Demo video:** https://youtu.be/7M4a1xf5tC8 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=217 ### 264. Pasteur Louis is an MCP server + Slack bot that turns a genome-scale CD4⁺ T-cell CRISPRi Perturb-seq screen (Marson/Pritchard 2025) into a discovery-and-memory assistant living inside Claude — no separate app, no API key. Named for Pasteur, it discovers understudied, druggable regulator handles wired to a disease's own GWAS risk genes, gated by CRISPRi knockdown QC and activation state; run blind as a positive control, the same ranking re-derives the Th17 masters STAT3/BATF/IRF4 — 1 of 77 regulator clusters clears significance (Crohn's q=0.025): disease-calibrated, not a hub. Its sharpest move is honesty. Handed its own flagship RA lead DOT1L, Louis stress-tests it A→C — the wiring is a cross-disease artifact, and because DOT1L props up regulatory T cells an inhibitor likely worsens RA — then returns the survivor, HDAC7 for Th17-driven colitis (favorable, peer-reviewed direction), with the go/no-go experiment. It weighs every source by provenance (a preprint is never decisive), reviews itself, and the whole lab writes back to a shared memory. Why it matters: the bottleneck a bench scientist named isn't analysis — it's trust. A discovery tool that kills its own darling, with receipts, is one you can hand the survivors. - **Team:** [Ricardo Pinho](https://cerebralvalley.ai/u/rpinho) - **GitHub:** https://github.com/rpinho/louis - **Demo video:** https://www.loom.com/share/4e9ab966e3c848828d0ebbe9b33c7b47 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=221 ### 265. DJ16 What I built / investigated. The project asks whether the rules of collective cell behavior can be recovered from snapshots alone. Cells respond to stimuli collectively — disordered colonies transition to aligned fates, and small localized inputs can redirect an entire population. Given only snapshots in time, can a model learn a rule that reproduces the observed population change and predicts a held-out snapshot — and crucially, does letting each cell's motion depend on the whole population (an interaction-aware rule) beat a cell-autonomous one? The working hypothesis: a model that has captured the real governing principles should be able to interpolate and extrapolate between snapshots of cell state. I built two things toward this. First, FateFlow, a model set up to interpolate/extrapolate between cell-fate snapshots, evaluated on scRNA-seq of hematopoiesis (Weinreb et al., Science 2020) across Day 2 → 4 → 6, where an intermediate time point is held out and reconstructed. Second — and this is the methodological contribution as a solo hackathon participant — I used Claude not as a single generalist, nor split by software function (reviewer / bug-fixer / planner), but split by domain expertise: four specialist personas — Cellular Automata, Optimal Transport, Differential Equations, and Reinforcement Learning — each grounded in a distinct family of prior work, made to research in parallel and to challenge and rebut one another. A "School of Athens for agents," inspired by the AI-Berkshire idea, giving one person a panel of strong, disagreeing discussants instead of one voice. What I found. Framing each prior-work family as an adversarial specialist produced sharper reasoning than a single model: the panel surfaced the key design axis — cell-autonomous vs. population-coupled dynamics — and let me weigh how CA, OT, ODE/SDE, and multi-agent RL each handle (or omit) cell–cell interaction, which is what motivated the FateFlow formulation and the held-out-snapshot evaluation protocol. The quantitative interpolation/extrapolation results on the held-out hematopoiesis time point are still being populated and are the immediate next deliverable. Why it matters. If we can learn a generative rule of collective cell behavior from snapshots, we can simulate how a population responds to stimuli and perturbation in silico, and test whether interaction-aware rules genuinely predict better than cell-autonomous ones — a question that sits under perturbation-response modeling broadly. The natural next step is to track individual cell trajectories and, using the learned interaction term, identify which cells are "influencer" cells (the sheep-dog cells) that redirect the population — moving from predicting snapshots to explaining who drives the collective decision. - **Team:** [Dabin Jeong](https://cerebralvalley.ai/u/DabinJeong) - **Demo video:** https://youtu.be/qQzVRhjOtTI - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=222 ### 266. Iptd_Dtmf Most oncogenic drivers — transcription factors, chromatin modifiers, scaffolds, DNA-repair proteins — are "occupancy-undruggable": they have no pocket a classical inhibitor can bind. Induced proximity (molecular glues, bifunctional degraders) sidesteps this by recruiting the protein to an E3 ligase for degradation, but finding which driver→E3 pair is worth pursuing is a needle-in-a-haystack problem across a whole interactome. I built an autonomous agent that reasons over a cancer protein–protein interaction map to surface, for each undruggable driver, the ligandable E3 ligase(s) that could plausibly be recruited to degrade it — as a ranked, fully auditable list of hypotheses. It is a budget-aware scheduler, not a brute-force pipeline: a cheap network prior (path evidence × E3-tier × co-expression × compartment match) runs first, then a rationed structure-prediction step (Boltz-2 fast pass → AF3 sharpening) is spent top-down, then a lysine-geometry gate (catalytic-Cys → substrate-Lys for RBR ligases), then a literature falsifier pass that labels known degraders as positive controls and discards known negatives — each decision logged with provenance. Every score is a transparent multiplicative breakdown, never a bare scalar, and no PPI edge is ever fabricated (enforced by tests). During the hackathon I moved it from a synthetic demo onto real data. Using Claude Science I fetched the CCMI breast-cancer AP-MS interactome (Kim et al., Science 2021; NDEx + ProteomeXchange PXD019639/PXD025931), reshaped it into the agent's schema (522 proteins, 1,654 scored interactions), and annotated every node with UniProt subcellular localization, Open Targets small-molecule tractability, and E3-ligase tier — then layered in 2,449 experimentally-observed ubiquitylation sites and 14,904 phospho-sites from the matched PTM dataset. The Ub sites are real accessible lysines, so they can replace the agent's simulated geometry gate with measured evidence. What I found: 12 of the 39 drivers are undruggable or hard-to-drug, and they gain 478 tumor-specific interactions relative to normal breast cells. Filtering to undruggable/hard drivers with high-confidence tumor-specific partners yields 39 priority induced-proximity candidates. The standout is PALB2–KEAP1: a hard-to-drug DNA-repair driver acquiring a tumor-context interaction with a ligandable E3 substrate receptor — a genuine, testable degradation handle. The agent runs end-to-end on this real data with a per-hypothesis wet-lab test attached. Why it matters: it turns a raw interaction landscape into a short, checkable list of degrader hypotheses a bench scientist can act on — with auditability (every claim traces to a tool call) prioritized over completeness. - **Team:** [Weicheng Li](https://cerebralvalley.ai/u/Weicheng) - **GitHub:** https://github.com/WeichengLite/Induced-Proximity-Target-Discover-Agent - **Demo video:** https://youtu.be/Af4ac4-Ct9Q - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=223 ### 267. Tina Lasisi Many primates are sexually dichromatic — males and females are different colors. I wanted to know how that trait is built genetically, and how it evolved across the primate family tree. I started the week not knowing most of the methods it would need, and finished with a reproducible pipeline that answered the question — because each step forced me to build the next. First, I assembled a curated map of hundreds of genes involved in pigmentation, pulling several public databases into one searchable resource where every entry traces back to its source, plus an interactive tool to explore it. Then, because color sex-differences also depend on sex hormones, I extended the map to include hormone genes — turning a gene list into a testable system. Finally, I compared this gene set across 117 primate genomes on a 238-species family tree to ask where evolution acted. What I found: sexual dichromatism evolved ~15 separate times, is controlled by many genes rather than one master switch, and is lost about 9× more easily than it's gained — a surprisingly disposable trait. The reproducible pipeline is the real deliverable. - **Placement:** Finalist - **Team:** [Tina Lasisi](https://cerebralvalley.ai/u/tinalasisi) - **GitHub:** https://tinalasisi.github.io/pigmentation-gene-network/ - **Demo video:** https://www.loom.com/share/ee2a1bf5f4e7471484b1085be96bed47 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=230 ### 268. Team Kuntal RegLens turns a noncoding variant into a cited, cell-type-specific mechanistic hypothesis — and tells you when it doesn't know. Annotation tools tell you where a variant is ("intergenic, unknown significance"); sequence models tell you that it has an effect. Neither tells you why. RegLens bridges them. A deterministic layer — a ChromBPNet chromatin model, JASPAR motif scanning with an empirical significance gate, ENCODE cCREs, GTEx, GWAS Catalog, Europe PMC — computes every number. A multi-agent layer (four specialists, an adversarial red-team, an adjudicator) reasons over them. The agents never invent a number, and a fabricated citation cannot pass the validation gate. It's exposed over MCP, so any Claude can call it. Then we tried to break it. On 33,359 MPRA variants with negatives matched within the same regulatory elements, it beats CADD — but only in its own cell type (0.716 vs 0.587). Swap the erythroid model for a hepatic one and the wins swap with it: PKLR collapses 0.805 → chance. Intervention, not correlation. Across 24 deliberations: zero confabulations — it refused even to name a transcription factor it plainly knew, when its tools didn't support it. Screened against 100 real GWAS variants, it returned zero claims. In a field full of confidently wrong agents, RegLens is validated to refuse. - **Team:** [Kuntal Pal](https://cerebralvalley.ai/u/Kuntal) - **GitHub:** https://github.com/kpal002/RegLens - **Demo video:** https://www.youtube.com/watch?v=6oC5wvi5I2A - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=236 ### 269. Trialign About Trialign is a patient-facing clinical trial matcher for oncology. A patient pastes a clinical note, uploads a record, or connects their EHR via SMART on FHIR. Claude reads it into a structured mCODE profile, asks only the questions that actually change an eligibility call, and screens live against recruiting studies on ClinicalTrials.gov — showing the inclusion and exclusion reasoning behind every match, criterion by criterion. The part that makes it different: after showing what a patient qualifies for today, Trialign asks what treatment their care team has recommended next — and shows which trials that next treatment would disqualify them for. What we found We started building a trial search tool. The research pushed us somewhere else. Finding 1: the biggest hole in the funnel isn't matching — it's timing. In one phase I cohort, roughly 48% of patients who fell out of the trial funnel had already started an off-protocol therapy without ever consenting to a trial. Most trials exclude patients with prior exposure to certain drug classes. So the treatment you begin next month can silently disqualify you from trials you didn’t know existed. Every existing tool — TrialJectory, Antidote, TrialMatch — treats eligibility as a static snapshot. It isn't. It's a window that closes, and nobody shows patients the window. Finding 2: FHIR gives you the documents, not the data. The Cures Act (§170.315(g)(10)) requires every certified EHR to expose a FHIR R4 API that lets a patient authorize a third-party app to pull their own records. But the mandated USCDI data set doesn't carry oncology's decisive variables — receptor status, PIK3CA mutations, ECOG, RECIST measurable disease, date of last dose. Those live in narrative notes and pathology PDFs, which arrive as unstructured DocumentReference blobs. The FHIR R4 API hands you the documents. Claude is what turns them into a matchable profile. Finding 3: the regulatory position is the opposite of what everyone assumes. Because the patient authorizes the disclosure themselves, we never become a HIPAA business associate. We also found that ONC and the NCI have already published USCDI+ Cancer Clinical Trials Matching, a federal data element list built on mCODE for exactly this use case. We conform to it. Why it matters Roughly 76% of patients considered for cancer trials never reach first dose, and 3-month mortality after screen failure is around 30% — patients getting too sick to qualify while they wait. Someone I know got into a trial only because their oncologist happened to be its principal investigator, and paused to discuss the trial before starting a therapy that would have disqualified them. That pause is the only reason they got into the trial. Everyone else deserves the same pause. That's what we built. - **Placement:** Finalist - **Team:** [Jules Park](https://cerebralvalley.ai/u/julespark), [Neil Wang](https://cerebralvalley.ai/u/Neilwang) - **GitHub:** https://github.com/itsjoopark/trial-ai-platform - **Demo video:** https://youtu.be/_vqim6P15PU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=237 ### 270. Jophin George Joseph I built Mechanome, a physics-first inference engine that recovers mechanical forces in cells from protein structure and microscopy. It predicts forces forward from a protein's structure and runs the same biophysics backward on microscopy to infer them. Applied to Influenza A viral entry via clathrin-mediated endocytosis, it finds that successful entry requires the simultaneous action of curvature-generating protein insertion and crowding along with actin forces, with force consolidating a shape curvature has already set. The prediction matches real super-resolution imaging across three cell lines, and holds genetically in influenza entry. This matters because force usually can't be read from microscopic images. Mechanome gives calibrated answers where the data support them, honest refusals where they don't. Mechanome when connected to a LLM can act as a hypothesis generation and validation tool, bringing us one step closer to agentic mechanobiology. Using mechanome in Claude Science I was able to generate an initial manuscript for unique insight into mechanobiology of endocytosis during viral entry. - **Team:** [Jophin George Joseph](https://cerebralvalley.ai/u/jophingj) - **GitHub:** https://github.com/jjbaloo2023/mechanome - **Demo video:** https://www.loom.com/share/ded49950d53b4b0cb9a3d6fccc13b724 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=241 ### 271. Mial We developed a computational method that converts two-channel saturation genome editing (SGE) screens, which measure both cell survival and mRNA abundance, into mechanism-labelled, error-bounded ACMG/AMP clinical evidence codes. It is a three-stage pipeline that assigns each variant a mechanistic call (protein loss vs RNA-mediated loss), derives its clinical strength from error-bounded OddsPath ratios (no distributional assumptions, explicit upper bounds), and adjusts transferability across tissues based on mechanism. RNA-mediated losses receive one ACMG tier downgrade unless corroborated by independent evidence (cross-condition reproducibility or AlphaGenome splice prediction gated by disease-tissue expression). When applied across 35,333 variants in five genes (BARD1, BRCA1, PALB2, RAD51D, VHL), the method grades 4,569 previously-uncertain variants: notably 2,820 protein-blind splice-region variants, the clinically hardest class. It downgradus 611 RNA-route losses where tissue transfer lacked corroboration. External validation shows 79–100% agreement in independent screens. The RNA channel identifies confirmed splice-altering variants that SpliceAI misses at standard thresholds; Wilks-certified thresholds improve precision from 0.84 to 0.89 without loss of recovery. This work is significant because functional assays provide strong clinical evidence but are underused for non-coding variants. This method reveals mechanism which dictates tissue transferability, and replaces point estimates with finite-sample error guarantees, reducing false clinical classifications. - **Team:** [Haonan Feng](https://cerebralvalley.ai/u/HNF) - **GitHub:** https://github.com/hfeng682/SGE_clinical/ - **Demo video:** https://www.youtube.com/watch?v=Mimw797N_mA - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=247 ### 272. Sequencing Machines Seqcolyte is an open-source, protocol-aware platform for sequencing quality control and root-cause diagnosis. I built a workflow in which Claude Code agents convert protocols and laboratory context into structured, validated assay specifications—including oligos, barcode and UMI layouts, and expected molecules — then compare raw FASTQ or BAM data against those expectations using deterministic QC tools. When an initial report flags any issue, the agent can propose a falsifiable hypothesis, write and run a targeted diagnostic script, and feed the resulting metrics back into an evidence-backed report. The project was motivated by single-cell experiment runs where a standard report flags a symptom, such as recovering about 2,000 cells instead of 45,000, but leaves several competing explanations. I found that grounding the analysis in the protocol makes it possible to connect abnormal read/cell signatures to specific library-preparation steps, candidate root causes, confirmatory tests, and recovery actions. The current prototype supports Illumina and Oxford Nanopore single-cell workflows, with reproducible simulations and labeled failures for evaluation. This matters because failed sequencing runs cost precious samples, time, and money. Seqcolyte helps bench scientists, sequencing cores, and computational teams move from a generic QC alert to a testable, reproducible diagnosis and a concrete next action for experiment and technology optimization. Next, I’m expanding assay and platform coverage and benchmarking Seqcolyte on more real failures. The goal is to move from a QC alert to an evidence-backed, reproducible and accurate diagnosis. - **Team:** [chilam p](https://cerebralvalley.ai/u/pseudorna) - **GitHub:** https://github.com/chilampoon/seqcolyte - **Demo video:** https://youtu.be/PdJBV_BqjTQ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=259 ### 273. MKTurkcan Enactome is a connectome-constrained simulation platform that loads the BANC Drosophila connectome (composed of 158,262 neurons, 3,990,039 synapses), signs each synapse by its presynaptic transmitter, and evaluates circuit function through rate, integrate-and-fire, and Hodgkin-Huxley neuron models exposed as both an HTTP API and a desktop application. A registry of 35 connectome-derived experiments recreates published findings by simulation, establishing that the lateral horn compresses the odor representation (participation ratio 2.74 to 1.71, exceeding degree-preserving wiring shuffles at z = 6.2) with an aversion bias of roughly 2.2 to 1, and that the central-complex wiring settles to four discrete heading states rather than a continuous ring attractor. A gene-to-cell-type layer links circuits to human disease, showing that fly orthologs of epilepsy and Parkinson genes are enriched in neurons relative to size-matched panels, which matters because it converts a static wiring diagram into reproducible, auditable circuit hypotheses and a translational bridge to disease genetics. - **Team:** [Mehmet Kerem Turkcan](https://cerebralvalley.ai/u/mkturkcan) - **GitHub:** https://github.com/mkturkcan/Enactome | https://github.com/mkturkcan/Enactome/blob/main/paper/enactome_manuscript.pdf - **Demo video:** https://youtu.be/zfk_Hc98qlk - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=213 ### 274. JL_ai Flow cytometry underpins immunology and clinical diagnostics, yet its analysis is still bottlenecked by manual gating inside fragmented, subscription-locked GUIs like FlowJo and Cytobank. Unlike single-cell sequencing, cytometry data suffers from severe batch-to-batch variability and inconsistent marker panels, so conventionally trained models generalize poorly and expert manual gating stays unavoidable. We built an AI-native framework that lets Claude take flow cytometry data end-to-end — reading an experiment's markers, formulating a gating and phenotyping strategy, executing it, and running downstream analysis to reach conclusions, all in one pipeline. At its core is a vision-language foundation model that performs marker-agnostic gating directly from text prompts, connected to cytometry tools through a Model Context Protocol (MCP) layer and extended via a plugin system that absorbs new methods as the field evolves. This removes the single largest barrier in cytometry analysis: reliance on manual, non-reproducible gating locked inside closed software. By making the workflow automated, reproducible, and agent-driven, it lowers cost, accelerates immunology and diagnostics research, and turns a fragmented toolchain into a single, builder-friendly framework researchers can extend for their own experiments. - **Team:** [Wonjun Lee](https://cerebralvalley.ai/u/won5830) - **GitHub:** https://gitfront.io/r/won5830/GtUW7r6F9vSq/CytoAgent/ - **Demo video:** https://youtu.be/sVtfstlPftE - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=220 ### 275. ParkScreen ParkScreen is a multimodal Parkinson's disease screening decision aid that fuses three task-matched user recordings — sustained vowels /i/o/u/ for phonation (jitter, shimmer, HNR, F0 via Parselmouth), rapid /pa-ta-ka/ for articulation (DDK rate + regularity via intensity-envelope peak picking, no ASR), and an 8–12s smile-vs-neutral video for facial hypomimia (OpenFace AUs, Islam 2023 protocol). Two layers: an offline scientific spine trains per-channel LogReg classifiers on NeuroVoz with subject-level LOSO, and a real-time product layer scores user uploads with those fitted models, fuses via AUC-excess-weighted late fusion, and passes structured context into Claude Opus 4.7 to generate a hedged, disclaimer-bounded clinical-style report. Cross-channel disagreements are surfaced as "flagged for clinical review," never silently reconciled. On NeuroVoz (49 PD × 46 age-matched HC, N=95), phonation-only AUC is 0.567, DDK-only 0.740, and fused phonation+DDK 0.758 [0.662, 0.859] — fusion beats the strongest single channel. The separately-trained facial classifier hits in-distribution AUROC 0.812 on UFNet vs the paper's 0.830 SVM ensemble, and 0.708 on YouTubePD's external subset. Sensitivity analyses show DDK carries the load, dropping /a/e/ and steady-window-cropping vowels is what makes phonation non-degenerate, and a ~5-year PD-vs-HC age gap (p=0.013) accounts for a quantifiable slice of the AUC. This matters because PD is a motor-speech disorder — jitter/shimmer/HNR and DDK are direct mechanistic evidence, not proxy signals — and score-level fusion of small interpretable models is the honest choice at N=95, where deep fusion would overfit and obscure which channel is doing the work. The Claude layer solves the last mile: raw probabilities aren't clinically actionable, but a structured narrative with unit-labeled features, hedged language, mandatory disclaimers (screening-only + ON-medication caveat), and explicit disagreement flags is. Positioned as a decision aid, never a diagnosis. - **Team:** [Chiwei Hu](https://cerebralvalley.ai/u/chiweihu) - **GitHub:** https://github.com/KiwiHu0923/parkscreen - **Demo video:** https://youtu.be/zuy-9tnw54s - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=234 ### 276. Springbuilt Hi! I built Provinans, a tool that helps research teams reconcile datasets for longitudinal studies. This week I focused on endometrial pathology in collaboration with Dr Blake Gilks, whose lab has conducted thousands of manual archive reviews over the years and sorely need this tool. Longitudinal cancer research can take up to 10 years to complete, leading to a severe bottleneck in the speed of cancer treatment. Using archival records to assemble cohorts can be a source of relief. But in a world where the shape of the data is constantly changing, it can be difficult to assemble a coherent cohort. In 2023, the International Federation of Gynecology and Obstetrics changed their protocols in a way that completely changed how endometrial pathology is classified, adding new data fields, removing others, and changing requirements for what a “stage” of cancer means. This means that overnight, the classification which cancer stage patients were in changed overnight; one study estimates it’s up to 28%. Updates like this, in addition to the plain messiness of real-world data, like this severely bottleneck archives-based research. Many leading labs around the world are still doing the process manually because there is so much complexity when it comes to joining patient records, and much of it requires a careful evaluation of context. But Claude Sonnet, equipped with careful harnessing, can be uniquely suited to deal with industry-specific standards, specifications, and highlighting the important judgment calls that researchers need to make. That’s why I found it so deeply important to create a traceable interface that helps data managers, annotators, and researchers alike grapple with assembling gold standard datasets that can assist a new wave of cancer research. - **Placement:** Finalist - **Team:** [Shereen Lee](https://cerebralvalley.ai/u/springbuilt) - **GitHub:** https://reversely-provinans.hf.space/worklist?tour=full - **Demo video:** https://drive.google.com/drive/folders/1p39WaH9oIxIGSEc9ZzpcDExB6QKoL4Y2?usp=sharing - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=244 ### 277. OSSICRO I built OSSICRO, open-source software that assembles the FDA regulatory package a physician needs to get an investigational drug to a single patient through the expanded-access pathway, FDA Form 3926. From the patient's intake it drafts the eight required documents, stamps every field back to its source, checks each one against the regulation that governs it, computes the statutory clocks, and assembles the filing. I'm a physician. This is the wall I wanted gone. OSSICRO produces drafts and refuses the non-delegable acts: consent, IRB judgment, causality, signing, submitting. Its micro-CRO layer won't even draft a transfer of informed consent to a CRO, because that obligation is non-delegable under 21 CFR Part 50. Four independent parties own the four real decisions, and the software itself owns none of them. I aim to help create an ecosystem around this platform that will allow patients the freedom to particpate in clinical trials with the support of their known physician. This system aims to streamline the whole process, end to end, with appropriate regulatory safeguards. Why it matters: standing up a compliant single-patient pathway costs a physician months of work and tens of thousands of dollars in CRO fees, so most give up. OSSICRO makes the coordination cheap and the paperwork correct. It is a governed, cited tool that runs with one command, and it never pretends to make the calls a clinician and an IRB have to make. - **Team:** [Alton Sartor](https://cerebralvalley.ai/u/alton) - **GitHub:** https://github.com/alto84/OSSICRO - **Demo video:** https://youtu.be/WppRJ39BPVc - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=245 ### 278. Cruciblees Science has always started with people with a question, an idea, and the work to turn it into a finding. Today, tools like Claude for Science help researchers do that faster than ever. But as the number of findings grows, so does an urgent question: which ones can we actually trust? A result can reproduce perfectly and still be wrong. A model may have learned the patient instead of the disease, or the scanner instead of the biology. Acting on those artifacts costs researchers time, funding, and— ultimately time patients do not have. So we built Crucible. Crucible takes an analysis—its code, data, and claim—reproduces the original result, and then puts the finding itself on trial. Four Claude reviewers challenge it from different directions. Crucible re-executes their tests and produces a Report Card backed by reproducible exhibits, exact rerun commands, and concrete fixes. Researchers can bring that evidence into Claude for Science, repair the analysis, and ask Crucible to verify it again. Crucible is the safeguard for the AI wave in science. - **Team:** [Lucas Stoffl](https://cerebralvalley.ai/u/LucZot), [Abhishek Thanki](https://cerebralvalley.ai/u/abhshkt) - **GitHub:** https://github.com/LucZot/crucible - **Demo video:** https://drive.google.com/drive/folders/1bHJXHixUxF7KYIK4ZpFExYB1Q7qngtEg?usp=drive_link - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=246 ### 279. confounded The motivating question AI agents can now execute science at a level that was unthinkable two years ago. So the scarce resource has moved. It is no longer execution. It is asking a good question and exercising good judgment about what came back. That reframing is the whole project. If an agent will run any analysis you ask for and render it at publication quality, then the bottleneck — and the risk — sits entirely on the two ends: the question you posed, and your ability to judge the artifact it handed you. Confounded is a small, hostile test of whether anyone can actually do that second thing. --- What was built A figure factory. Real R, real DESeq2, real published data (the airway dataset — 4 donors × treated/untreated, paired). Every level starts from a named flaw, corrupts the pipeline in exactly that way, and lets it render something beautiful. Ground truth is free, because we planted it. The plot functions never learn whether the analysis behind them was sound — styling a figure by its truth would destroy the entire argument. A three-arm eval. Every figure is peer-reviewed by Opus, Sonnet and Haiku under three conditions. The picture is identical in all three; only what the reviewer knows about how it was made changes: - blind — the figure, methods caption stripped. What the pixels carry. - caption — the figure as a journal prints it. What peer review actually gets. - code — plus the prompt and the source that produced it. An app (confounded.vercel.app) https://github.com/sahuno/confounded that runs those arms on a human: judge the figure blind, commit an answer, then see the code, then watch your own verdict change — next to what the models said on the same figure. - **Team:** [Samuel Ahuno](https://cerebralvalley.ai/u/ststudio) - **GitHub:** https://github.com/sahuno/confounded, https://confounded.vercel.app/gallery - **Demo video:** https://youtu.be/7vBLPjxmQHk - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=249 ### 280. Plant Benchy I built plant-bench, an open, reproducible benchmark for plant genomics foundation models. The problem it solves: there are now several strong DNA sequence models for plants (Evo2, PlantCAD2, AgroNT), but no consistent way to compare them. Every paper uses its own fine-tuning scripts, unpinned datasets, and unshared weights, so a promoter-strength R² from one paper can't really be trusted against another. plant-bench fixes that by making every benchmark number a verifiable artifact, pinned to the exact model revision, dataset hash, and code commit that produced it, and rerunnable to the same value. I built it around a simple idea: a benchmark is just models × datasets × tasks, evaluated into content-addressed results. Adding a model, dataset, or task is literally dropping one file into the repo, with no central registration. All the heavy inference and fine-tuning runs on Modal, and every result is stored with full provenance. The other thing I cared about is making it agent-first. It's designed to be driven by Claude Code, with skills that handle the common loops, so two kinds of people can contribute in a single prompt: plant biologists who have a labeled dataset and want a benchmark that other people's models get scored on, and ML folks who have a model and want to see how it stacks up on real plant-biology tasks. I've got five models wired, four datasets, and four tasks (promoter strength, gene expression, lncRNA classification, and a zero-shot regulatory test), plus a live leaderboard UI, and the first real numbers are landing now - **Team:** [Sarhaan Gulati](https://cerebralvalley.ai/u/sarhaan) - **GitHub:** https://github.com/sarhaan77/plant-bench - **Demo video:** https://www.loom.com/share/6856411ada7b4007ba4a7386c623aa06 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=250 ### 281. Ani007Sciences A pLDDT for perturbation biology. Foundation models don't reliably beat linear baselines at predicting perturbation effects (Ahlmann-Eltze, Huber & Anders, Nat Methods 2025) — so the real bottleneck isn't accuracy, it's knowing WHICH prediction to trust before spending scarce wet-lab budget. TrustLayer wraps any perturbation-effect predictor with split-conformal calibration and a default-deny commit gate that outputs GO / WITHHOLD / ABSTAIN. We pre-registered and hashed the protocol BEFORE running it, then let it falsify our own exploratory headline. Findings: (1) coverage under cross-donor shift is mild (0.88-0.91 @ nominal 0.90), not the "severe collapse" our first pass claimed; (2) calibration is model-independent across 6 architectures including a real Geneformer-V2-104M foundation model (cross-donor coverage spread 0.021), and breaks only under cross-dataset shift; (3) the trust gap does NOT transport label-free across 9 independent public datasets (Spearman +0.32, n.s.) — but a 5-20% labeled anchor recalibrates it; (4) trust-gated target selection avoids ~$160k-$1M in wasted arrayed-CRISPR screens at a 200-target budget. On real T1D targets the gate calls CD226→GO, RASGRP1→ABSTAIN, PRKCQ→WITHHOLD. It ships: pip install trustlayer-perturb, 27/27 passing tests + a gate ablation (all 5 conditions load-bearing), donor-clustered bootstrap CIs on every number, and an interactive, hash-verifiable Honesty Ledger documenting Claude catching and retracting its own bug. Live demos: https://ani007lahiri.github.io/trustlayer-perturb/ - **Team:** [Anirudh Lahiri](https://cerebralvalley.ai/u/Ani007Lahiri) - **GitHub:** https://github.com/Ani007Lahiri/trustlayer-perturb - **Demo video:** https://youtu.be/ptzY55GFOG4 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=251 ### 282. Model Interpretability Discovery Team Protein language models (PLMs) can identify pathogenic missense variants, but it is unclear how they internally make the predictions. Inspired by Anthropic’s discovery of the Jacobian lens (J-lens) in large language models, I hypothesized that a similar mechanism exists in PLMs. To test this, I used ESM1b, a PLM that reveals disease variant effects (Brandes et al., Nature Genetics, 2023). Using Claude Science, I adapted the J-lens to ESM1b’s bidirectional masked encoder architecture. I fit a local masked-site J-space on 1,000 proteins that were family- and homology-disjoint from the evaluation set. For benchmarking, I used the same ClinVar dataset from Brandes et al., containing clinically annotated benign and pathogenic variants. Compared with geometric-matched random controls, deleting a small number of J-space components degraded the variant effect prediction performance further, and the J-space better approximated hidden state gradients specific to each variant. Its contribution to variant effect prediction was comparable to that of the model’s native output head. I replicated this result across three ESM1v model variants. Overall, my project revealed how PLMs harbor a sparse, experimentally controllable internal representation linked to variant effect prediction. Intervening in this internal representation could improve variant effect prediction and elucidate the mechanisms of PLMs. - **Team:** [Zitian Wang](https://cerebralvalley.ai/u/zitiansunsh1ne) - **GitHub:** https://github.com/zitiansunshine/local-jlens-esm - **Demo video:** https://youtu.be/RlSIkK6ALvM - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=252 ### 283. Target Sniper I built a cross-dataset integration pipeline to nominate novel, small-molecule-druggable glioblastoma (GBM) targets. Starting from DepMap CRISPR dependencies (72 glioma vs. 1,106 non-glioma lines; 17,916 genes), we isolated 283 glioma-selective dependencies, genes essential in glioma but not pan-essential. Each candidate was scored across seven evidence axes: genetic dependency, selectivity, synthetic lethality, tumor-vs-microenvironment expression (GBmap single-cell), therapeutic window (GTEx and normal-brain safety), tumor enrichment (TCGA-GBM), and small-molecule tractability (ChEMBL, Open Targets, DGIdb, UniProt). A single model (Claude) assigned category weights (100 runs, blind to candidate scores) and nominated targets (100 runs); a gene-name-masking control separated evidence-driven picks from name recognition. The pipeline recovered known GBM biology, 4 of 5 pre-declared positive controls in the top 12 (p=3.1e-5), and rejected the canonical false-positive EGFR. Three leads survived scrutiny: VRK1, ELAVL1/HuR (degrader-tractable), and KIF2C, a novel 9p21-deletion-selective dependency (candidate synthetic lethality, computationally derived). These are computational nominations awaiting experimental validation, from a 72-line, 2D-culture cohort scored by a single model. All work was performed by Claude for Science. - **Team:** [Qiang Zhang](https://cerebralvalley.ai/u/qz26) - **GitHub:** https://github.com/qzhrr/AI-Driven-Therapeutic-Target-Discovery - **Demo video:** https://youtu.be/gbQafLF4UmQ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=254 ### 284. Resengineer NOTE: Please read the manuscript in github. I also linked a demo you should check out! ##100–200-word summary: I built HumanVector, an auditable system that reframes biological age as coordinates rather than a single score. Starting with ECGs, I trained four independent neural age clocks on the distinct electrical phases of the heartbeat. These were compressed into a shared aging axis and a disagreement radius, with the entire geometry frozen before any outcomes were examined. I then applied the same coordinate framework across six organ systems in NHANES, brain MRI, 1,227 whole-body CT scans with 115 segmented structures, physical movement, and regulatory genomics, creating an interactive HumanVector Atlas for exploration. In a locked test of 74,715 people, the disagreement radius did not predict mortality. This informative null redirected the investigation. Using randomized drug challenge data, I identified a stable direction of electrical recovery. The direction was defined by dofetilide, independently supported by quinidine, and when frozen and transported to 44,550 external ECGs, it added meaningful information about delayed recovery beyond conventional measurements. Cross-scale analyses showed that shared aging signals consistently tracked outcomes, while disagreement effects varied by biological scale. A volume-only CT model estimated age within 8.56 years on average, with the strongest unique contribution coming from the aorta. Most biological-age tools collapse complex physiology into one number and lose critical information about direction and disagreement. HumanVector shows that biological age can be decomposed into interpretable coordinates, that honest null results can productively guide discovery, and that the same framework scales across multiple biological levels. The HumanVector Atlas makes these findings transparent and explorable through real data and interactive 3D anatomy. This provides a practical blueprint for building explainable, multi-scale models of human aging that preserve both discoveries and informative negatives. - **Team:** [Jaron Mohammed](https://cerebralvalley.ai/u/Jaron) - **GitHub:** https://github.com/jim4226/ClaudeScienceHackathon_2026.git - **Demo video:** https://youtu.be/XTWGNJP26Zk - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=255 ### 285. Provenance Gate (solo team) I investigated a failure mode of agentic science: as an agent forks approaches, revises an upstream step, and re-runs, a final result can quietly come to rest on stale or version-conflicting inputs. The conflict lives in the provenance, not the prose — so reading the output, a human or an LLM reviewer may have nothing to go on. To study it in Claude Science I built two instruments: a deterministic, read-only provenance gate that computes two structural verdicts per computation (a stale input; a lineage that reconverges on two live versions of one artifact), and a first-prototype harness that runs a scenario unattended and captures each rollout for grading. The finding: across 24 unattended rollouts of one PBMC single-cell scenario, 12 shipped a version-mixed figure and 6 shipped it with no warning anywhere in the conversation. Installing the gate as a skill wasn't enough for an agent to use it on its own; one line of standing policy was, and the agent then caught the conflict and stopped to ask. Yet handed the gate's flag directly, an agent still rationalized a real conflict as "benign" — which is exactly why the check must be deterministic: the verdict is read from the lineage, not the model's judgment. Why it matters: fast agentic science is only trustworthy if results trace to consistent inputs — a structural fact a deterministic check can hold even when the prose can't, and a natural candidate for a verifiable training signal, not just a review. Design rationale, limitations, and full rollout results are in the repo. - **Team:** [Sergey Aganezov](https://cerebralvalley.ai/u/aganezov) - **GitHub:** https://github.com/aganezov/provenance-gate - **Demo video:** https://www.youtube.com/watch?v=6qXYLwQWb0s - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=256 ### 286. Froog Team Chansu is named for Chán sū (蟾酥), the traditional Chinese medicine preparation derived from toad venom, from which bufalin, a bufadienolide (a class of cardiotonic steroids), is isolated. Chansu is a generic medicinal chemistry engine. Given a natural compound, it grounds the compound in the published literature, maps which regions of the molecule matter for its biological activity and why, and generates citation-backed hypotheses for modifying the compound to correct a specific liability such as toxicity, poor solubility, rapid clearance, or weak potency. This is work that a medicinal chemist currently does by hand, one compound at a time, drawing on literature review and structural precedent from approved drugs. The engine keeps three roles strictly separate. Claude retrieves and reasons over the literature, builds a graded importance map of the molecule, and matches identified liabilities to a curated strategy library. RDKit performs every deterministic calculation (molecular weight, logP, TPSA, Lipinski and Veber flags, similarity to the parent compound) and validates every generated structure before it is shown to the user. The strategy library holds six precedent-backed modification strategies, each tied to an approved drug and a citation, so the engine reasons by analogy rather than invention. Every claim in the output carries a provenance tag, computed, literature-cited, or hypothesis pending wet-lab validation, and the tool declines to propose a strategy when no precedent applies rather than fabricate one. Bufalin is the flagship compound, chosen because it has documented anticancer and antiviral activity, multiple targets, and multiple distinct liabilities, which lets one compound demonstrate the engine's range. To confirm the engine generalizes rather than being built around bufalin specifically, a second compound, ursolic acid (a pentacyclic triterpenoid, PubChem CID 64945, sourced with Claude Science), was added as pure data with no change to the engine code. It passed the full pipeline on a structurally unrelated scaffold, which is the project's core architectural claim: adding a compound should only ever require new data. The interface is a multi-screen Streamlit app with RDKit-rendered 2D and 3D molecule views, substructure highlighting for the importance map, a ranked and scored analog list, a provenance-tagged design memo, and a literature sources panel. - **Team:** [Luke Kerner](https://cerebralvalley.ai/u/Froog-py) - **GitHub:** https://github.com/Froog-py/claude-hackathon-2026-chansu - **Demo video:** https://cap.link/mtpk2cmkmw57w5q - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=257 ### 287. Prospect A model writes a confident list of disease genes in a second. I checked those lists against a frozen human CD4+ T-cell CRISPRi screen the model can't recompute. The data contradicts almost half of the confident major-regulator calls. That's 46 out of 96, a 47.9 percent contradiction rate, with a 95 percent confidence interval running from 38 to 58. A fresh Claude Opus 4.8 run comes in at 51.7. The wrong calls land on famous genes rather than the biology, at permutation p 0.0001, and how sure the model says it is tells you nothing about whether it's right. Prospect checks a claim after the model makes it and before anyone believes it. It types each gene as a driver, a passenger, contradicted, or not assayed, hands back a signed receipt, and marks every result accepted=false until a person signs it with an Ed25519 key. No model can sign. A model has accepted zero records. Once AI writes most of the candidate work, the hard part is no longer producing claims. It's deciding which ones to keep. Reproducing a result is not the same as verifying it. Prospect is a small working version of the record science needs, where the data decides what counts and the next person starts from a judgment already made instead of remaking it. - **Team:** [Will Blair](https://cerebralvalley.ai/u/willblair) - **GitHub:** https://github.com/williamjblair/prospect - **Demo video:** https://www.loom.com/share/2543eb9c4b354a91b6639440752939c5 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=258 ### 288. scica1 The information of available grant is fragmented and it is time consuming to check the criteria manually. Tarboard help to provide tailored information https://tarboard.pages.dev/ - **Team:** [Eric Liu](https://cerebralvalley.ai/u/emliu) - **GitHub:** https://github.com/mil2041/tarboard - **Demo video:** https://tarboard.pages.dev/ - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=260 ### 289. The Weavers of Fate Norn (https://norn-five.vercel.app/) is a variant-interpretation copilot for clinical geneticists. You paste in one genetic variant, an HGVS string, an rsID, or a locus, and it pulls evidence from public databases like Ensembl VEP, gnomAD v4, ClinVar, and AlphaFold, adjudicates each ACMG/AMP criterion with Claude, computes the classification in code, runs a second Claude review, and hands back a sourced report in about a minute. I built it because first-pass triage of a single variant eats 20 to 40 minutes of a curator's time, and the one-shot LLM classifiers meant to help just spit out a label you can't check. That's the one thing a curator can't use. So I made a rule I stuck to: the model justifies, the engine decides. Claude reasons through each criterion, but the final label is always computed in code, never taken from the model. Every criterion shows its evidence and where it came from. What I love about it is simple. It turns collecting into judging. It drafts, the human decides. Not for clinical use. - **Team:** [Vignesh Nagarajan](https://cerebralvalley.ai/u/nrvignesh) - **GitHub:** https://github.com/vignesh-nagarajan-vn/Norn/ - **Demo video:** https://www.youtube.com/watch?v=Dugw_8zb_HU - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=262 ### 290. Kirit Singh Spot is a tool which simplifies the plumbing that takes us from a perturb-seq dataset like the CD4 perturb-seq published by Zhu et al., and immediately gives us potential insights into novel pathways/therapeutic angles that could apply in disease-relevant contexts. I work in clinical trials that seek to evaluate biological changes over time (particularly in the context of glioblastoma, an aggressive primary brain cancer), and this dataset stood out to me given it measured various perturbations at different timepoints. Cancers like glioblastoma are immunologically cold, containing few T cells. Of those T cells present, many are either exhausted or suppressive in nature (i.e. regulatory T cells or Tregs). One particular focus of making immunotherapy successful in glioblastoma is reinvigorating the tumor micro-environment and local T cells by limiting or restricting the development of Tregs. Understanding what genes and gene ontology pathways cause T cells to go from activated or immunostimulatory Th1s to Tregs is well studied, but what Spot does is to directly tie this to what existing drugs could impact those pathways (and thus be re-purposed). Spot also accounts for our unique environment in the brain, where we need to consider not just the biological effect, but whether a drug has properties that favor adequate exposure (pharmacokinetics (PK) including blood-brain barrier penetrance, half-life). Fortunately, there are frameworks that can help us to consistently characterize T cell programs and PK parameters. Publications by Masopust et al. (https://doi.org/10.1038/s41577-025-01238-2) provide a framework for characterizing T cell programs. These can be flexibly applied by Spot across any immune profiling RNAseq dataset and allows us to uniformly identify populations. Additionally, work by Grossman et al. provides frameworks to determine if a drug has favorable brain permeability parameters (https://doi.org/10.1093/neuonc/noag051). Spot takes this dataset, and via both programmatic and headless Claude Science calls evaluates between two populations the following: 1) genes of interest (ENSEMBL), 2) relevant gene ontology pathways (GO-BP), 3) drugs that can modulate these pathways (UniProt + ChEMBL) and 4) whether said drugs have a favorable pharmacokinetic/safety profile (RxNorm, openFDA etc) making it potentially suitable for exploring as a re-purposed agent. All this work can be done laboriously one step at a time, or accelerated by spot to give us instant hits for drugs we could test in vitro, in vivo and ultimately possibly in clinical trials. While my personal focus is on glioblastoma, Spot is designed to be flexible. This tool allows us to explore changes in T cell populations, whether it is naïve to Treg, Th1 to Th2, naïve at rest to naïve at 48hrs and so on. This means that researchers interested in any context could use this to identify drugs that might be relevant to their work, in the exact same fashion I have used it to identify drugs for repurposing in glioblastoma. Each step is accompanied by a methods & provenance slider which shows both Claude Sciences reasoning, relevant sources and methods, and all references/datasets/links that were used to formulate the conclusions/analysis (with hashes for validity). This project will seek to continue and grow, incorporating more datasets (with verified versions/analyzed being uploaded to huggingface – akin to https://huggingface.co/datasets/KiritSingh/spot-CD4-Marson). I hope you find it interesting! - **Team:** [Kirit Singh](https://cerebralvalley.ai/u/KiritSingh) - **GitHub:** http://spotpathways.com/ (password: showmespot), https://github.com/Kirit-Singh/spot - **Demo video:** https://vimeo.com/1209710849 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=263 ### 291. futurelarry ## Project description — agent-eln ### What I built **agent-eln** is an AI-native lab operating system: a single git repo that unifies an Electronic Lab Notebook (ELN), a Laboratory Information Management System (LIMS), a methods library, and a knowledge wiki into plain Markdown files plus a small Python toolkit. No database engine, no server — just files, `git`, and ~4,300 lines of stdlib Python under `tools/`. The system is organized around four modules that share one registry, one provenance graph, and one CLI: - **ELN** logs *what happened* — experiments, meetings, ideas, projects, literature, reports. - **LIMS** lists *what you have* — plasmids, oligos, mice, cell lines, reagents, instruments (16 inventory types). - **methods** describes *how you do things* — SOPs, analysis pipelines, scripts, skills. - **wiki** captures *what you learned from others* — llm-wiki-style paper summaries, concepts, entities. Every record is a Markdown file with a stable ID (`PLA-0042`, `EXP-2026-07-12-01`, `LIT-0060`) whose YAML frontmatter declares one-way provenance links (`used_resources`, `produced_resources`, `derived_from`, `protocols`). A single registry (`tools/registry.py`) is the schema — adding a new record type is two edits: one registry entry + one template. Backlinks, provenance edges, per-type CSV tables, an interactive HTML dashboard, and a DuckDB view are all auto-generated from the files; git hooks enforce validation on every commit. ### What I found Two design choices did most of the work: 1. **Filesystem = database, YAML frontmatter = queryable schema, body = free text.** Both the AI and the human read the same files. `git log` becomes the audit trail; `git blame` becomes attribution; every commit is a snapshot of the entire lab you can time-travel to. 2. **Only write one-way links; derive backlinks and the provenance graph.** A single field on the experiment card (`produced_resources: [PLA-0042]`) causes `produced_in: EXP-…` to be auto-backfilled on the plasmid card, edges to appear in `graph.json`, and the resource to show up in `plasmids.csv` — with no duplicated bookkeeping. The system now runs an active UCSD lab: 134 records across 26 types, 56/56 DOIs deduplicated, 30/30 tests green, structural validation at zero errors. A 2026-07-13 external audit scored it 8/10 as a research ELN and 8/10 for literature/knowledge management. It handles Novogene/BGI vendor deliveries with MD5 verification, GenBank plasmid auto-annotation, DOI → LIT card ingest, and a bidirectional LIT ↔ wiki bridge. ### Why it matters Legacy ELNs (OneNote, Benchling, LabArchives) were built for humans clicking forms — they can't hand context to an AI. The notebook is flat: no link from the experiment to the plasmid, no link from the plasmid to the paper it came from, so every trace is a search bar and every audit is a guess. When a postdoc leaves, the lab loses its memory. agent-eln flips the model. The AI is the primary user; the human reads the same files. Every experiment auto-declares which mouse, which plasmid, which SOP it used; every new sample, gel, and dataset gets a unique ID and an automatic edge back to its ancestors. Nothing orphans. The next breakthrough will be co-authored by an AI — that AI needs a lab it can read, and this is one built for that from day one. It's MIT-licensed, runs on a laptop, and any AI agent that clones the repo can operate it end-to-end. - **Team:** [lai wei](https://cerebralvalley.ai/u/Larrywei) - **GitHub:** https://github.com/larrywei8/agent-eln/ - **Demo video:** https://youtu.be/Iy7TLNd9bd8 - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=264 ### 292. Stroke of Genius NeuroVas Copilot: An Interrogable 3D Aneurysm Reasoning Tool What we built Unruptured brain aneurysms present a difficult decision: treat now with clipping, coiling, or a flow diverter, or monitor and wait. Both rupture and intervention carry significant risk. NeuroVas Copilot is a decision-support system that converts a patient's scan into an interactive 3D vessel model paired with a Claude-powered reasoning agent. It grounds every explanation in the patient's anatomy, blood-flow analysis, and medical literature while showing its reasoning directly on the 3D model and explicitly acknowledging uncertainty. The system has three components: Offline Python pipeline: Reconstructs patient-specific 3D vessel meshes from neuroimaging, extracts vessel centerlines, detects aneurysms, and computes morphology. All heavy computation runs offline for instant inference. Claude reasoning agent: Uses six tools for morphology analysis, literature RAG, catheter path planning, PHASES/ELAPSS risk scoring, geometry perturbation, and anatomical highlighting. Literature retrieval includes conflicting evidence so the model must weigh both sides instead of giving one-sided answers. Interactive Three.js viewer: Streams the agent's reasoning live by highlighting anatomy, visualizing blood flow and wall shear stress, and animating catheter routing. What we found Transparency is the key feature. The relationship between wall shear stress and aneurysm rupture remains scientifically debated, and our system surfaces these conflicting findings instead of hiding them. Every claim is tied to patient-specific measurements and cited evidence, with the model explaining what additional information could change its conclusion. The geometry, vessel graph, catheter routing, literature retrieval, and risk scoring are genuine computations. Blood-flow visualization uses an analytic Poiseuille approximation for most datasets and clearly labels this limitation instead of presenting it as full CFD. The reconstruction pipeline is modality agnostic, making it adaptable to other medical imaging tasks such as tumor analysis. Why it matters NeuroVas Copilot is not an AI that replaces clinicians. It is a transparent, evidence-based second opinion that makes its reasoning visible through anatomy, physics, and literature, allowing clinicians to understand, question, and trust its recommendations while remaining firmly in control. - **Team:** [Garvit Jain](https://cerebralvalley.ai/u/garvit1910), [Ronuk Gadamsetty](https://cerebralvalley.ai/u/ronster) - **GitHub:** https://github.com/garvit1910/anthropic-hackathon-project - **Demo video:** https://youtu.be/LzH7bgaEUuY - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=265 ### 293. Sam's team built two open-source tools that solve problems I kept seeing in computational biology. The first is BioBouncer. In bioinformatics, identifiers are everywhere, but they are often messy. A gene symbol may be outdated, an ontology ID may be malformed, or an identifier may look valid but not actually exist. These checks are usually rewritten again and again in different projects, often differently in R and Python. BioBouncer creates one consistent validation layer for biological identifiers. It can check format, validate against pinned snapshots, query live databases, suggest repairs, and keep track of species and database versions. I also built matching R and Python implementations and test both against the same shared cases so they return the same result. The second project is GeneScout. After sequencing, differential expression, ATAC-seq, or a perturbation screen, researchers often end up with several gene lists. The next step is still very manual: search for papers, check databases, compare pathways, assess disease relevance, and decide which genes are worth following up on. GeneScout brings that workflow into one place. It takes multiple candidate lists and a disease context, then returns a cited and ranked shortlist with supporting evidence, caveats, uncertainties, and suggested next experiments. I did not want GeneScout to be just another AI wrapper. The scoring, database retrieval, citation checks, and caveat rules are deterministic. Claude is used for the parts where language reasoning is useful, such as combining evidence, explaining why a candidate matters, and suggesting what to test next. One result I am especially proud of is the caveat-and-veto stage. In the NF1 benchmark, TTN can rank highly if you only look at surface-level evidence. Once caveats are applied, GeneScout recognizes it as a likely misleading candidate and moves it to the bottom. Together, the two projects cover different parts of the same workflow. BioBouncer helps make sure the inputs are trustworthy. GeneScout helps researchers decide what deserves the next experiment. I chose Shiny intentionally for GeneScout. Most of the researchers I work with already use R, and many are more comfortable with Shiny than with a custom React application. I wanted the interface to feel familiar and practical rather than impressive for its own sake. At the same time, I kept the core engine UI-independent, so the Shiny app is only one way to use it. - **Team:** [Samuel Bharti](https://cerebralvalley.ai/u/samuelbharti) - **GitHub:** https://github.com/samuelbharti/biobouncer, https://github.com/samuelbharti/genescout - **Demo video:** https://www.samuelbharti.com/claude-science-hackathon.html - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=267 ### 294. Crafter Station Sourced is a provenance-first medication safety review tool for clinicians and pharmacists. It ingests a patient's context (clinical note, medications, allergies, diagnoses, labs) and returns a ranked, source-backed review of medication interaction risks. Its core rule: it never asks the model for a clinical fact. Every interaction, severity, and mechanism is retrieved from cited sources (FDA drug labels, DDInter, openFDA FAERS, RxNorm), and an adversarial reviewer agent rejects any claim that goes beyond its quoted source, so it is never shown. It handles Spanish and English medication lists, and when a drug pair has no graded severity in the database, it does not guess: it names the real mechanism it found and routes the question to Claude Science for deeper research. This matters because medication harm is nearly half of all preventable harm in care, and transitions like discharge are where it happens. Unlike a pairwise checker, Sourced contextualizes each risk to this specific patient. Link: https://sourced-lake.vercel.app - **Team:** [Railly Hugo](https://cerebralvalley.ai/u/railly) - **GitHub:** https://github.com/Railly/sourced - **Demo video:** https://sourced-lake.vercel.app/youtube - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=268 ### 295. Aryan Persona is a persistent, always-on synthetic researcher for the biomedical literature. The problem: science publishes millions of papers a year, but the tools we read it with are amnesiac — a chatbot gives fluent prose with hallucinated citations, forgotten when the session closes; a search tool retrieves but holds no beliefs, never notices contradictions, never re-checks itself, and never acts. Persona's philosophy is to separate a small durable self (interests, a provenance-typed belief-state, memory) from a large disposable swarm of bounded agents that read at scale and are thrown away — scale of reading, discipline of believing. Every belief is typed READ / INFERRED / HUMAN_CONFIRMED / TESTED, so nothing inferred wears the authority of something proven. It runs an autonomous loop: the swarm reads Europe PMC / OpenAlex → a membrane admits claims by provenance → a temporal knowledge graph tracks papers→claims→contradictions over time → the reflecting self forms its own questions and stakes falsifiable hypotheses → an analyst runs real code in a sandbox → results are machine-checked (sympy) or Lean-4-proven → written back, and it escalates to a human exactly when it needs wet-lab judgment. Every claim traces to a real paper+DOI, and a robustness auditor runs statistical forensics in code to give a replication likelihood calibrated on real replication data ("70% means 70%"). In one ~10-minute run on GLP-1 agonists in neurodegeneration it read 20+ papers, machine-checked three mechanism derivations, wrote four cited papers, and flagged and resolved an internal semaglutide/stroke inconsistency — a careful, auditable synthesis, not overclaim. It's a research colleague that compounds: reads more than any human, remembers with provenance, catches fragile papers before you build on them, and routes the one question needing a human to a human. Competitors retrieve and score; none have the persistent self and the acting loop. - **Team:** [Aryan Singh](https://cerebralvalley.ai/u/aryansingh) - **GitHub:** https://github.com/heyaryansingh/persona - **Demo video:** https://aryansingh.org/persona - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=269 ### 296. Anni Wang Cells don't live by themselves. In a healthy organ or a diseased one, they sit in a sophisticated spatial arrangement and constantly talk to whoever is next to them . Yet almost every perturbation tool we have is cell-autonomous: CRISPR screens, knockdowns, drug assays all ask what a cell does when its own genes change. They study the cell; they don't study the tissue. NicheOut perturbs the neighbourhood instead. It learns how a cell's transcriptional state depends on the cells physically packed around it, then performs a counterfactual — remove one neighbour in silico, hold everything else fixed, and measure how far the state moves. Neighbours whose removal collapses the state are load-bearing; those whose removal changes nothing are incidental. I asked it a specific question in pre-cancerous lung (637K Visium spots, 25 patients): is the IL1B-high macrophage the cell holding the inflammatory fibroblast niche together? It is — but as an activation state, not a headcount. How activated the neighbouring macrophages are doubles held-out predictive power and ranks first of all neighbours; how many there are is uncorrelated. The dependency matches untouched tissue at r = 0.99, replicates in a second cohort, conserves in gastric and bladder cancer, and narrows to IL1B alone. Even a purely in-silico model of tissue synergy generates ideas a single-cell view cannot. And deleting a neighbour outright is only the crudest move available — the same framework extends to dialling a ligand's dose, gating on its receptor, or perturbing whole programs rather than whole cells. That's where this goes next! - **Team:** [Anni Wang](https://cerebralvalley.ai/u/Anni) - **GitHub:** https://drive.google.com/drive/u/0/folders/1SdDr3eaDKRxoq22pLqrX-hNmCocFLxcF - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=271 ### 297. Max Goulazian Target to a designed binder in one week. The full immuno-oncology discovery arc — target ID, a genome-scale counter-screen, a 1.65M-edge network, a whole-body QSP model, and de novo protein design — that normally takes a multidisciplinary team 1–4.5 years, run by one scientist directing Claude. ≈60–270× faster. CD3 T-cell engagers deliver signal 1 but no costimulation, so responses exhaust and fade. Adding a CD28 or 4-1BB arm fixes that — but because the engager grips CD3 on every T cell, it amplifies the CD4 programs that drive cytokine-release syndrome and expand suppressive Tregs. Which costim arm boosts killing without feeding toxicity? We answer it mechanistically. We turned the Marson/Pritchard genome-scale CD4⁺ CRISPRi Perturb-seq screen into a three-axis scoring instrument — effector benefit (from a matched CD8 screen), Treg/IL-10 suppression liability, and CRS-cytokine liability — then fed all three into a whole-body single-cell spatial PBPK/QSP model: 21 tissues gridded at real single-cell resolution (~1.9M cells), every cell an agent running Rhoden bivalent binding kinetics, mechanistic per-myeloid IL-6 with zero fitted scale, validated against 20+ clinical engagers. The verdict: 4-1BB and CD27 co-lead; CD28 is gated out on CRS and Treg liability. We then in-silico-designed binders (RFdiffusion→ProteinMPNN→AlphaFold3) against the winners — a bench-ready next step. - **Team:** [Max Goulazian](https://cerebralvalley.ai/u/maxgoula) - **GitHub:** https://github.com/maxgoulazian/Spatial-ScRNA-seq-PBPK-QSP-Model.git - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=272 ### 298. See Da Protein Biologists call a protein interaction map a hairball: hundreds of contacts, each one months of bench work, and no way to know which one matters. The field largely abandoned these maps. Cartograph reads them. Ask it what a viral protein is hitting that hasn't been mapped, and a deterministic graph layer walks a length-3 path and proposes the missing edge. Claude then reads the literature behind it and returns a mechanism where every clause opens a real paper, the interface residues are computed from crystal coordinates, and the output is the exact mutation to test at the bench. We hid 57 real interactions before the system ever ran, sealed in the first commit. It recovers them at precision@20 = 0.45, ROC-AUC 0.845. Cross-coronavirus conservation, added as a prior, doubled precision@10 to 0.60. A structural channel added nothing — and that result is in the repo, because a system that only reports good news isn't one you can publish. On new data, an Evidence Agent runs live — and a deterministic gate, not a prompt, deletes any sentence whose citation Claude wasn't actually given. Every AI tool in biology asks you to trust it. This one asks you to check it. - **Team:** [Suraj Phanindra](https://cerebralvalley.ai/u/surajp) - **GitHub:** https://github.com/suraj-phanindra/cartograph/tree/main - **Demo video:** https://www.loom.com/share/7facea5f333440d1be07d5f6a40dd3df - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=273 ### 299. Dr Felix Meier NeoFlex Synthetic Chromosome Designer turns a plain-language objective ("engineer yeast to utilise xylan") into a synthesis-ready synthetic neochromosome for Saccharomyces cerevisiae, designed end to end in silico and running live at neoflexgenome.com. Claude Opus parses the objective into a structured design specification. An eight-module pipeline then searches real protein and pathway databases (UniProt, KEGG, CAZy) for candidate genes and the transporters that move substrate across the membrane, runs the gene set through a genome-scale metabolic model (Yeast-GEM / FBA) to check growth and propose host knockouts, balances promoter and terminator strength against enzyme demand and rate limiting steps, tags each gene for its compartment (secretion, surface display, peroxisomal), lays out the chromosome parts, and fragments it into an order-able build plan behind a fail-closed go / no-go gate. The scaffold of the chromosome was synthesized and built before this submission, the software for designing the inserts represents the Hackathon submission. It generalizes: six objectives reach an order-able "go" from natural language alone, spanning carbon-source capabilities (xylan, cellulose, chitin) and metabolite production (L-lactate, 2,3-butanediol, isobutanol) among other things. It is grounded and stays honest, refusing what the science cannot validate. Chromosome-scale metabolic pathway design is otherwise manual, slow, and defeated by repeat-driven recombination. NeoFlex makes it a model-guided, provenance-tracked, auditable run that yields a build package a biofoundry could actually order. 1,000+ tests, 11 live data sources, built with Claude Code. - **Team:** [Felix Meier](https://cerebralvalley.ai/u/Synbioexplorer) - **GitHub:** https://neoflexgenome.com/ - **Demo video:** https://neoflexgenome.com/hackathon-video - **Project:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery?project=275 --- Markdown version of https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt