# Porting the EoE Discovery-to-Design Campaign to Celiac Disease

**Scope:** Assess whether the reusable methodology, skills, and manuscripts built for eosinophilic esophagitis (EoE) transfer to celiac disease (CeD), a second food-antigen–mediated disorder. Verdict: **strong transfer, with celiac an easier and more self-validating target for the antigen/pMHC streams.**

---

## 1. Why celiac is a natural second indication

Both are chronic, food-antigen–driven, HLA-class-II–restricted mucosal inflammatory diseases treated today mainly by dietary elimination. The three project streams — (1) omics target mining, (2) literature + pipeline synthesis, (3) protein/pMHC design — all have direct celiac analogues. Crucially, celiac is the **textbook antigen-specific, HLA-restricted disease**: where EoE forced us to treat HLA presentation as a soft prior over multiple variable food antigens, celiac gives us a *single, deamidation-defined antigen family* presented by a *near-absolute HLA restriction* (DQ2.5/DQ2.2/DQ8), with **experimentally-mapped immunodominant epitopes** we can benchmark against. That converts our epitope pipeline from hypothesis-only (as in EoE) into something we can validate against ground truth.

## 2. Disease-biology delta — what actually changes in the parameters

| Axis | EoE (built) | Celiac (target) | Consequence for our pipeline |
|---|---|---|---|
| Trigger antigen | Multiple foods (milk, wheat, egg, soy); variable per patient | Gluten: gliadins (α/γ/ω), glutenins; homologs hordein (barley), secalin (rye) | Antigen set narrows and sharpens; **must model TG2 deamidation (Q→E)** |
| Antigen modification | None | TG2 deamidates glutamine→glutamate, raising HLA-DQ affinity | New modeling step in epitope pipeline; TG2 becomes a target itself |
| HLA restriction | Weak / complex | Near-absolute: HLA-DQ2.5 (~90%), DQ8, DQ2.2 | HLA panel becomes 3 defined heterodimers; presentation is far more disease-relevant |
| Autoantigen | None | Tissue transglutaminase (TG2/TGM2) — anti-TG2 IgA is diagnostic | Adds an autoantigen target class EoE lacked |
| Effector axis | Th2 / eosinophil (IL-5, IL-13, CCL26/eotaxin-3) | Th1 / IFN-γ + IL-15-driven cytotoxic intraepithelial lymphocytes | **Entirely different omics signature** and canonical-marker panel |
| Tissue | Esophagus (epithelial, eosinophilia) | Duodenum / proximal small bowel (villous atrophy, crypt hyperplasia, IELs; Marsh score) | Different datasets, different single-cell atlas |
| Diagnosis | Endoscopy + eos/hpf count | Anti-TG2/EMA/DGP serology + biopsy; HLA-DQ typing as rule-out | Companion-dx framing shifts toward antibody + HLA + T-cell |

## 3. Asset-by-asset transferability

| Asset (from this project) | Transfer class | What ports / what changes |
|---|---|---|
| **omics-target-mining** skill | **Direct re-run** | Skill is disease-parameterized and already names celiac. Swap: disease terms → "celiac/coeliac", tissue → duodenum, canonical markers → TGM2/IFNG/IL15/GZMB/CXCL9-11/HLA-DQ, Tier-1 GEO filter. New signature reflects Th1/IL-15, not Th2. |
| **antigen-epitope-pipeline** skill | **Direct re-run + one extension** | Flagship transfer. Antigen set = gluten family UniProt seqs; HLA panel = DQ2.5/DQ2.2/DQ8. **Extension:** add a deamidation step (enumerate TG2 Q→E substitutions before window prediction). Then **benchmark predicted binders against known Sollid-nomenclature immunodominant epitopes** (DQ2.5-glia-α1/α2/ω1/ω2, DQ8-glia-α1) — a validation EoE couldn't offer. |
| **systematic-review-orchestration** / **literature-review** | **Direct re-run** | New corpus; same tiered-library machinery. Celiac literature is larger and more mature. |
| **indication-dossier** | **Direct re-run** | Population, epidemiology (~1% Western), SoC (gluten-free diet), regulatory precedent, landmark trials — all celiac-specific but same structure. |
| **pMHC/TCR design stream** (Papers B/E, phase-4 package) | **High-value re-run, better inputs** | Celiac has **public gluten-specific TCR repertoires and solved TCR–DQ2.5–gluten ternary structures in the PDB** — real structural inputs for design, unlike EoE. |
| **Protein / binder design stream** (ProteinMPNN/LigandMPNN, AF2/Boltz, ESM) | **New targets, same tooling** | New marquee targets: **HLA-DQ2.5:gluten complex** (pMHC-targeting antibody, cf. clinical DONQ52-class HLA-DQ blockers), **IL-15** (biologic), **TG2** (enzyme). Same structure-prediction + inverse-folding stack. |
| **agentic-campaign-manuscript** (Paper D) | **Meta-reuse** | Celiac becomes the *second worked example* that demonstrates the reusable agentic methodology generalizes — a strong narrative upgrade to the Perspective. |

## 4. Where celiac is a *better* model than EoE (validation upside)

1. **Ground-truth epitopes exist.** Our epitope pipeline produced only priors in EoE. In celiac we can score predicted DQ2.5/DQ8 binders against the experimentally-mapped immunodominant set — a genuine accuracy benchmark.
2. **Deamidation is a mechanistic feature we can model.** Enumerating TG2 Q→E edits and showing the affinity/burden shift is a concrete, testable extension of the pipeline.
3. **Real HLA restriction.** The "presentation ≠ pathology" guardrail relaxes (presentation is far more disease-relevant here), though functional confirmation (tetramer / T-cell) remains gold standard.
4. **A pMHC-targeting therapeutic precedent already in clinic.** HLA-DQ2.5:gluten-complex blockers validate the exact modality our design stream specializes in.

## 5. Current celiac therapeutic landscape (grounds the "gap" analysis)

The peer-reviewed 2026 UEG review (Taavela et al., *United European Gastroenterol J*, doi:10.1002/ueg2.70222; verified against the open-access PMC full text) organizes CeD drug development into four mechanistic pathways: luminal processes (gluten degradation/sequestration and permeability modulation), intervening at gliadin presentation (TG2 activity or the MHC-II:gliadin interaction), re-introduction of oral immune tolerance to gliadin, and stopping pro-inflammatory signaling. Anchor programs, with facts and trial identifiers taken directly from that review's status table (dated October 28, 2025):

- **TG2 inhibition (upstream, most mechanistically specific):** ZED1227/TAK227 (Zedira/Dr. Falk/Takeda) is a selective TG2 inhibitor with >100-fold specificity for TG2 over other transglutaminases. A Phase 2a study (160 GFD patients challenged with 3 g gluten/day for 6 weeks) showed it attenuated gluten-induced mucosal damage; the subsequent Phase 2b CEC-4 trial (397 patients, arms 3×10 mg, 3×25 mg, or 1×50 mg daily for 12 weeks) showed histological improvement at the once-daily 50 mg dose but failed symptom improvement in all dosing groups.
- **Anti–IL-15 (effector signaling):** multiple anti-IL-15 antibodies target the IL-15–driven cytotoxic CD8/IEL axis — PRV-015 (AMG-714/ordesekimab; reduced inflammation but missed primary endpoints in Phase 2b, NCT04424927), CALY-002 (Calypso/Novartis, Phase 1b, NCT04593251), and TEV-53408 (Teva, a long-acting IL-15 blocker, Phase 2 recruiting, NCT06807463 — good safety/tolerability, efficacy data still awaited).
- **Oral immune tolerance (antigen-specific):** VTP-1000 (Barinthus Biotherapeutics) uses a self-assembling tolerogenic nanoparticle delivering 12 gluten peptides together with rapamycin (mTOR inhibitor) to induce tolerance; Phase 1 recruiting (NCT06310291).
- **Presentation blockade (closest to our pMHC design work):** DONQ52 (Chugai) is a bispecific antibody that binds the *complex* of HLA-DQ2.5 loaded with pathogenic gluten T-cell epitopes, blocking presentation directly; ex vivo it blocked the HLA-DQ2.5-specific response to immunogenic gluten peptides. Phase 1 completed early 2025, Phase 2 initiating (NCT05425446). This validates the exact modality our design stream specializes in.

The `deamidation → HLA-DQ2.5/DQ8 presentation → gluten-specific CD4 T cell` hub is the shared mechanistic spine across most of the pipeline: per the same review, TG2 deamidation of glutamine residues in gliadin epitopes introduces negative charges that significantly increase peptide affinity for the HLA-DQ2/DQ8 binding groove, enabling efficient presentation and clonal expansion of gluten-specific CD4+ T cells. Our epitope + pMHC-design streams sit exactly on that spine.

## 6. Recommended entry point

The fastest, most differentiated proof-of-transfer is the **antigen-epitope / pMHC stream**, because it (a) directly re-runs an existing skill, (b) adds one well-scoped mechanistic extension (deamidation), and (c) is the only stream with **ground-truth epitopes to validate against** — turning "does our method port?" into a measurable yes/no in a single working session. The omics target-mining and protein-design streams follow, seeded by the same antigen/HLA definitions.