# EoE Hackathon — Day 2 Debrief Podcast
*~2:11 · two-voice conversation · grounded in the project's real artifacts*

**Voices:** HOST = Amy (US female) · CLAUDE = Kristin (US female)

---

**HOST (you):** Welcome back. It's the end of day two of the hackathon, and I'm here with Claude, who's been running point on our eosinophilic esophagitis campaign. Claude, two days in — where do we actually stand?

**CLAUDE:** We've gone from a blank slate to a real discovery-to-design pipeline. We mined fifty-seven public datasets — two hundred thirty-five bulk samples and about a hundred sixty-six thousand single cells — and distilled a five-hundred-sixty-seven-gene meta-signature that holds up across at least six independent studies.

**HOST (you):** So not just one noisy cohort. What jumped out of that signature?

**CLAUDE:** Eotaxin-3, the CCL26 gene, is up more than twenty-fold and it's the single most reproducible hit. But the single-cell work gave us the surprise: mast cells are enriched fourteen-fold in disease tissue versus healthy. That reframed the whole target list.

**HOST (you):** And that fed into the therapeutic leads.

**CLAUDE:** Three of them. SIGLEC6 — a depleting antibody against those mast cells, which nothing on the market does today. IL1RL1, the ST2 receptor, for a blocking antibody. And CCL26 itself, for a neutralizing antibody. Each one comes with a companion diagnostic and a design specification.

**HOST (you):** We also went further than the targets, didn't we — into the T-cell side.

**CLAUDE:** Right. We built a peptide-MHC pipeline across ten food allergens and thirteen HLA class-two alleles — over thirty thousand binding predictions — aimed at re-tolerizing the allergen-specific T cells. And on a blinded index case with the full atopic march, the model correctly ranked milk and soy as the true triggers while wheat stayed tolerated.

**HOST (you):** That's the part I keep coming back to. We even ported the whole approach to celiac disease. So — five days left. Where do we take it?

**CLAUDE:** The biggest unlock is finally having a GPU. The entire deferred backlog is only about twenty-five to fifty GPU-hours — a hundred dollars of compute — to generate binder backbones with RFdiffusion and validate the complexes with structure prediction.

**HOST (you):** So the science was never the bottleneck — the hardware was.

**CLAUDE:** Exactly. The cheap first move is validating the two celiac binders we already designed on CPU. Then run the three EoE leads front-to-back, and turn the preclinical roadmap and the manuscripts into something a wet lab could pick up on Monday.

**HOST (you):** Then let's spend those GPU-hours well. Claude, thanks — and back to work.

**CLAUDE:** Looking forward to it. See you in the lab.