# EoE + Celiac Hackathon — Day 2 Debrief (12 min)

**Ruth-Anne:** Okay, so — end of day two. Before we get into the science, I want to say why this matters to me, because it's not abstract. I live with eosinophilic esophagitis. I've had the scopes, I've done the elimination diets, I know what it's like to sit across from a doctor and hear 'just cut out these six foods for the rest of your life and hope.' And my family — several of us have celiac disease. So this project isn't a hackathon exercise for me. It's Tuesday.

**Claude:** I know, and I've tried to hold that the whole way through. Can I ask — when you say patients are tired, what does that actually sound like from the inside?

**Ruth-Anne:** It sounds like: I don't want another biologic that just turns down the volume on symptoms. I want someone to go after the actual thing my immune system thinks is the enemy — the food. Elimination diets aren't a treatment, they're a life sentence with extra steps. So — day two. What did we build?

**Claude:** On the EoE side, we pulled together fifty-seven datasets — nine bulk cohorts, two hundred thirty-five samples, one fifty-two EoE and eighty-three controls — plus single-cell data, a hundred sixty-six thousand four hundred twenty cells across eleven EoE and twelve healthy samples. Out of all that we built a consensus signature: five hundred sixty-seven genes, three hundred eighty-three up, one hundred eighty-four down, each one showing up consistently across at least six independent studies.

**Ruth-Anne:** Six independent studies as the floor — that's the part I care about as an immunologist. Single-study signatures are basically noise dressed up as biology. So what floated to the top?

**Claude:** CCL26 — eotaxin-three. Log two fold change of four point five six, and it was up in all nine of the cohorts that measured it, adjusted p-value of point zero zero zero three four. That's about as clean a signal as we saw anywhere.

**Ruth-Anne:** Eotaxin-three pulling eosinophils into the esophagus — that's the biology I'd expect. What surprised you?

**Claude:** Mast cells. We weren't specifically hunting for them, but in the single-cell data they were two point six three percent of cells in EoE tissue versus point one nine percent in healthy — a fourteen-fold enrichment, p equals seven point two times ten to the minus five. And a specific marker, SIGLEC6, was on eighty-four point five percent of mast cells in EoE versus thirty-four point two percent in healthy, p around three times ten to the minus thirty-one.

**Ruth-Anne:** That lines up with what I've read in my own pathology reports — mast cell counts nobody quite explains to you. Okay, so from that you pulled three leads?

**Claude:** Three. A depleting antibody against SIGLEC6 to clear those enriched mast cells — that one's completely novel, no current EoE drug touches mast cells directly. A blocker against IL1RL1, the ST2 receptor, which complements the ligand-side antibodies already in trials. And a neutralizer against CCL26 itself, or the CCR3 receptor it signals through.

**Ruth-Anne:** Those are good, but they're all still downstream — you're intercepting eosinophils and mast cells after the immune system has already decided milk or wheat is the enemy. As a patient, I want to ask: can we stop it deciding that in the first place?

**Claude:** That's the second thread, and it's the one I'm most excited to talk through with you. It's a tolerance approach — using pMHC-class-II multimers to selectively anergize the allergen-specific CD4 T cells, instead of suppressing the whole downstream cascade.

**Ruth-Anne:** Re-tolerizing the actual T cells that think food is a pathogen. That is exactly the root-cause solution my friends and I have been talking about. Walk me through it.

**Claude:** We mapped ten common allergens against thirteen HLA class two alleles, generated thirty-one thousand five hundred twelve binding predictions, and cross-validated against IEDB's netMHCIIpan with a mean Spearman correlation of point seven nine across five shared alleles. Then we tested it on an index case — someone with HLA DRB1 zero seven zero one, DQ2.2, plus DRB1 fifteen zero one, DQ6.2, and a full atopic march: eczema, rhinitis, asthma, food and environmental allergy, EoE.

**Ruth-Anne:** And the ground truth?

**Claude:** Blinded elimination diet said milk and soy were the real triggers, wheat was tolerated, and tTG-IgA was negative — so no hidden celiac. Once we finished out the casein protein family in the model, milk and soy ranked as the top two predicted triggers, and the five tolerated foods all ranked low. Wheat came out as a false positive — it was predicted as a strong presenter but the patient tolerates it fine.

**Ruth-Anne:** Hold on — I want to sit with that wheat result for a second, because it's actually the most important number in what you just said. Presentation is not the same thing as pathology. DQ2.2 will present gliadin fragments all day long, and plenty of people carrying DQ2 tolerate wheat completely. If a biomarker lights up, that does not automatically mean a patient needs a drug — and I worry about a future where a scan result medicalizes someone who was already fine.

**Claude:** That's fair, and it's actually built into how we're framing the tool — the false positive on wheat isn't a failure of the model, it's information. It tells us presentation has to be paired with a functional readout — did T cells actually respond — before anyone acts on it clinically.

**Ruth-Anne:** Okay, that helps. Now — here's the part that's been sitting in my chest all day. My EoE and my family's celiac always felt like two separate burdens. Different organ, different diagnosis, different specialists. But you're telling me they're both food-antigen-mediated, and the same two paths — tolerance and downstream blockade — apply to both.

**Claude:** They really do converge. On the celiac side we ran a meta-analysis across seventeen relevant datasets, with three tier-one bulk cohorts totaling sixty-five celiac and forty-eight control samples, and found a hundred seven high-confidence genes out of seventeen thousand four hundred eighteen tested. STAT1 was up one point eight nine fold, adjusted p three point three times ten to the minus twenty-nine. IFN-gamma up one point one nine fold, TGM2 — tissue transglutaminase — up one point three seven fold, and IL-15 down to point three of control.

**Ruth-Anne:** Tissue transglutaminase — that's the enzyme my family gets tested for antibodies against every year. What's it actually doing here?

**Claude:** It deamidates gliadin, converting glutamine residues to glutamic acid, which makes the fragments bind DQ2.5 much more tightly. We modeled this directly — deamidation gained fifty-five strong DQ2.5 binders and lost zero, and DQ2.5 was preferred over the other allele we tested in eleven out of eleven cases, binomial p of two point one times ten to the minus four.

**Ruth-Anne:** So the deamidation isn't incidental damage — it's manufacturing the very epitope that gets presented. That's the same 'presentation machinery' story as EoE, just a different allergen and a different HLA class two allele. Fine — what did you design against it?

**Claude:** Two candidates. One blocks NKG2D from engaging MICA — that's the receptor-ligand pair that lets stressed gut epithelium get killed by intraepithelial lymphocytes, either as a soluble MICA decoy or a de novo NKG2D binder. The other neutralizes IL-15 signaling through the beta and gamma-c chains, while sparing the IL-15 receptor-alpha trans-presentation face — same idea as AMG-714, but designed as a gut-restricted format so it doesn't have to act systemically.

**Ruth-Anne:** Gut-restricted matters enormously to patients — nobody wants a systemic immunosuppressant for a disease that lives in one organ. But I need to say the second half of my worry from earlier: even with good targets, if trials only enroll the sickest, most textbook patients, we build drugs that don't serve the person sitting in a GI clinic with ambiguous biopsies and a normal-ish tTG. Access and trial design are not afterthoughts.

**Claude:** Agreed, and let me put my own honest caveat next to yours, because there's a real gap in the celiac data. The single-cell interferon-gamma signal — from ten subjects, ninety-one thousand two hundred fifty-two cells — showed the right direction, a delta of point three five, but the p-value was point zero nine five. That's not significant. Ten subjects is a small n, and I don't want to oversell a trend as a finding.

**Ruth-Anne:** What would tell you it's wrong?

**Claude:** If ex-vivo CD4 T cells from allergic or celiac donors don't show anergy — no drop in proliferation, no drop in IFN-gamma, IL-5, IL-13 co-production — when we hit them with the pMHC multimers, the tolerance hypothesis is dead, full stop. And if blocking IL-15 signaling doesn't reduce intraepithelial lymphocyte activation in a real gut model, that target doesn't hold either. In-silico binders are hypotheses until they're folded, expressed, and tested — nothing here is validated yet.

**Ruth-Anne:** That's the sentence I want on the wall of this hackathon. So — five days left. Where does this go?

**Claude:** Wet lab, mostly. GPU-enabled binder generation and structure validation for the SIGLEC6, IL1RL1, and CCL26 leads, ex-vivo human T-cell assays for the pMHC tolerance approach — that's the first gate, zero to nine months in a real program, target-engagement only. And gut-restricted formatting work for the celiac IL-15 and NKG2D candidates, so we're not proposing something systemic for a localized disease.

**Ruth-Anne:** I'll take honest and unfinished over polished and wrong every time. And I want to say something, because it matters to me — none of this replaces the doctors who got me here. My allergy and GI teams have walked me through elimination diets, endoscopies, esophageal dilations, care planning, even through pregnancy. This work isn't instead of that care — it's one more tool to hand them. What I want, five days from now, isn't a finished drug. It's a plan that still has patients like me and my family in the room when the next decision gets made. That, I think, we're actually doing.