GeneSleuth
Gemini 4 NYC HackathonGeneSleuth helps a clinical geneticist find the variant behind an unsolved rare epilepsy case. Sequencing returns thousands of rare variants, and the hard ones sit outside protein-coding genes, where a variant can change splicing in one tissue (for example brain) and look silent in blood. A published DNM1 case motivated this: a deep intronic variant altered brain-specific transcripts while blood RNA tests stayed ambiguous. The input is a referral letter (PDF, including noisy scans) plus candidate variants. The system reads the letter, narrows 282 rare suspects in a 58-gene epilepsy panel, investigates with live AlphaGenome 2 calls in the tissues that matter, and returns a ranked shortlist. Every claim cites a tool call, and a checker rejects claims the evidence does not support. A reviewer can accept or reject candidates, reveal the hidden answer and export a Markdown evidence report. Everything was built on Nov 21 at the event: the data pipeline (1000 Genomes NA12878 background, gnomAD v4.1 frequency filter, real ClinVar variants planted as spike-ins), 126 synthetic referral letters (about a third rendered as scans), the agent, the claim checker, the benchmark and the web app. We tested on 20 held-out cases against four baselines and report the result as measured. The full agent did not beat the simple baselines: hit@1 was 50% against 85% for Gemini 4 alone, and 9 of 20 runs hit the demo time limit (the 11 that finished all placed the hidden variant in the top 10). What GeneSleuth adds is an auditable evidence trail: tissue-specific AlphaGenome tracks, cited claims, and 82 of 316 claims rejected by the checker. Limitations are in the README.





