# enTrust

- **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
- **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
- **Gallery:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery/13

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.

---

Markdown version of https://cerebralvalley.ai/e/built-with-claude-life-sciences/hackathon/gallery/13. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt