# EoE Phase-3 Report: Computational Protein-Therapeutic Design

## Executive summary

Phase-3 converted the Phase-2 validated targets into structure-grounded design specifications.
Three protein-tractable leads — **SIGLEC6** (mast depletion), **IL1RL1/ST2** (alarmin-receptor
blockade), and **CCL26** (eosinophil-recruitment block) — were carried through interface mapping,
competitor differentiation, ESM-2 epitope analysis, binder design briefs, an epithelial
antigen-presentation (pMHC) analysis, developability screening, and companion-diagnostic
specification. All three have favorable, well-conserved, developability-clean epitopes and paired
patient-selection assays. Two GPU-dependent steps (de novo backbone generation, designed-complex
refolding) are delivered as executable protocol specifications because no GPU host was available.

## 1. Design roster (locked)

| Lead | Modality | Epitope | Anchor structure | Companion Dx |
|------|----------|---------|------------------|--------------|
| SIGLEC6 | depleting mAb / ADC | Ig-V (28-123) | AlphaFold (no exp. human) | SIGLEC6+ mast fraction |
| IL1RL1/ST2 | blocking mAb / decoy | Ig1-Ig2-Ig3 (22-313) | 4KC3 (experimental) | soluble ST2 |
| CCL26 | neutralizing mAb / CCR3 antagonist | mature chemokine (24-94) | AlphaFold + eotaxin homologs | eotaxin-3 |

ALOX15 routed to a small-molecule program (intracellular, not antibody-tractable). MSLN parked
pending protein-level surface-expression validation.

## 2. Binding interfaces (Step 2)

- **IL1RL1/ST2** — mapped from the experimental IL-33/ST2 complex (4KC3): a large **1,781 Ų**
  buried interface spanning all three Ig domains, hotspots Tyr119, Phe245, Leu308, Lys22, Arg38,
  Tyr132. Ideal blocking-antibody epitope.
- **SIGLEC6** — Ig-V domain, 54 surface-exposed residues; sialic-acid-binding Arg cluster
  (R122/R100/R109-114) is the functional site, acidic face (E28/E37) an alternative
  non-competitive epitope.
- **CCL26** — CCR3-engagement surface (N-terminus, N-loop, 40s loop), disulfide-constrained; small,
  supporting a parallel CCR3-antagonist route.

## 3. Differentiation from existing/pipeline therapeutics (Step 3)

Structural benchmarking against the tozorakimab/IL-33 complex (9WWH; 4KC3-derived comparison):
clinical anti-IL-33 antibodies target the **ligand** (tozorakimab epitope shares only 9/25 residues
with IL-33's ST2 site). Our IL1RL1 mAb targets the **receptor** ligand-binding face — isoform-
independent, complementary. SIGLEC6 mast depletion has no EoE clinical precedent. All three act at
narrower, more targeted nodes than dupilumab's broad IL-4Rα blockade.

## 4. ESM-2 epitope conservation & developability (Steps 4, 8)

ESM-2 650M per-residue analysis: all epitopes well-conserved (mean entropy 0.6-0.8); escape-risk
positions largely avoid the interface for SIGLEC6 (0/4), less so for CCL26 (its sole escape-risk
residue H39 sits in the docking surface — flagged, mitigations in the brief) and IL1RL1 (6/18
interface-overlapping — steer CDRs to conserved aromatic hotspots). Epitope clean-fraction (paratope-
available surface): SIGLEC6 0.93, CCL26 0.93, IL1RL1 0.84 — all favorable.

## 5. Epithelial antigen presentation / pMHC (Step 7)

EoE is a food-antigen-driven CD4/Th2 disease, so the disease-central antigen-presentation arm is
**MHC class II** (CD4), not MHC-I (CD8). Both arms were analyzed:
- **MHC-I / CD8 (disease-associated CHANGE):** EoE epithelium acquires HLA class-I presentation
  (suprabasal Δ+0.82, basal Δ+0.55), IFN-γ-driven (r=0.94, JAK/STAT). Points to a cytotoxic/IFN-γ
  CD8 component in injured epithelium — mechanistically interesting, exploratory.
- **MHC-II / CD4 (disease-CENTRAL):** carried by professional APCs (myeloid MHC-II score ~3.1,
  constitutive; epithelial MHC-II inducible per literature — Mulder 2011, doi:10.1016/j.ajpath.2010.10.027).
  This is the causal, food-antigen-priming arm behind the Th2 response.

Food-antigen presentation load (predicted, milk β-lactoglobulin / egg ovomucoid): class-II strong
binders vastly outnumber class-I (212/91 vs 28/13) — consistent with CD4-dominant biology. **Antigen-
directed (pMHC/TCR) therapy should lead with class-II/CD4** (tolerance induction to dominant food
epitopes); class-I/CD8 is a mechanistic sub-study of the IFN-γ epithelial-injury finding, not a
co-equal pillar. See food_antigen_pmhc_review.md for the full rationale and literature.

## 6. Binder design specifications (Step 5) & validation protocol (Step 6)

Per-target design briefs (framework, CDR-targeting strategy, KD goal, selectivity screen) authored
for all three leads. De novo backbone generation (RFdiffusion/BindCraft) and designed-complex
refolding/scoring (ESMFold2-Fast → Chai-1, with explicit ipTM/pAE/DockQ/epitope-recapitulation
acceptance thresholds) are specified as a ~1 GPU-day pipeline, executable once a remote GPU host is
added.

## 7. Companion diagnostics (Step 9)

Universal companions (CCL26/eotaxin-3, ALOX15/15-HETE — ~99% addressable all endotypes; SIGLEC6+
mast fraction also broad, 80→92% mild→severe, and the most homogeneous target I²=37) vs precision
selectors (NTRK2/BDNF, CPA3 — steepest severe-endotype enrichment, e.g. NTRK2 84→100%). SIGLEC6's
value is mechanistic specificity (mast) and biomarker consistency, not severe-endotype restriction.
Serum assays (sST2,
periostin, BDNF) enable non-invasive selection; biopsy assays (SIGLEC6 IHC, 15-HETE lipidomics)
build on standard EoE histology.

## Deliverables & GPU-pending items

**Delivered (CPU):** roster + structures, interface analysis, differentiation map, ESM epitope
analysis, design briefs, antigen-presentation/pMHC analysis, developability screen, companion-Dx
specs, this report + dossier.

**Pending GPU host (Customize → Compute):** de novo binder backbones (Step 5 extension) and
designed-binder structure validation (Step 6). Protocol spec ready in binder_validation_protocol.md.

**Pending wet-lab / tools:** MSLN EoE surface-expression validation; MHC-II (CD4/ADA) prediction
(netMHCIIpan); prospective clinical linkage of endotypes.

## Key caveats
- No experimental human SIGLEC6 structure (all 13 RCSB candidates were Siglec-5/7/8 or bacterial); design uses the AlphaFold model.
- CCL26 epitope is small and its single escape-risk residue lies in the docking surface — CCR3-antagonist route hedges this.
- pMHC analysis is exploratory (no EoE-specific antigen ground truth; food-antigen hypothesis).
- Binder sequences do not yet exist (GPU-deferred), so the designed-binder immunogenicity screen is specification-tied.