# EoE pMHC-II Tolerance Therapeutics — Design Report

**Antigen-specific peptide–MHC-II reagents to tolerize the food-reactive pathogenic effector-Th2 clone in eosinophilic esophagitis, targeting the top three triggers: wheat, dairy, and soy.**

*Computational protein-design deliverable for the EoE hackathon. All binding values are computational priors; the dairy anchor is the one experimentally validated epitope. Not a validated therapeutic and not medical advice — see §7.*

---

## 1. Rationale

EoE is a food-antigen-driven, CD4⁺/Th2-mediated disease. The disease-causing agent is a specific food **peptide presented on a specific MHC-II molecule to a specific TCR** — the pMHC–TCR triad. Blocking downstream cytokines (dupilumab, anti-IL-5) controls inflammation but is not antigen-specific and not curative. An antigen-specific reagent that carries the offending food peptide on its restricting MHC-II can, in principle, **silence only the disease-driving clone** and leave the rest of the repertoire intact — the teplizumab disease-modification principle, but with antigen-specific safety.

This design complements the two existing EoE pMHC deliverables in this project:
- the **food-trigger diagnostic** (`eoe_food_trigger_dx_design.md`) — names the trigger from a blood draw;
- the **pMHC review** (`eoe_review_part3_pmhc.md`) — lays out the therapeutic logic.

What was missing, and what this report adds, is the **therapeutic construct itself** — engineered, structurally modeled, and GPU-validated.

## 2. Target epitopes (one shared allele)

All three food leads were selected on **HLA-DRB1\*07:01** — the same allele that restricts the tetramer-validated EoE dairy TCR — so a single DR7 backbone serves all three foods by swapping only the peptide cassette.

| Food | Protein (UniProt) | Epitope 15-mer | 9-mer core | Region | IC50 (nM) | Status |
|---|---|---|---|---|---|---|
| **Dairy** | β-casein (P02666, CSN2) | `KIHPFAQTQSLVYPF` | `FAQTQSLVY` | precursor aa59–78 | 16.8 | **tetramer-validated** (Dilollo/Spergel/Hill 2025, eoeTCR-4) |
| **Wheat** | α-gliadin (P04722) | `IHNVVHAIILHQQQQ` | `HNVVHAIIL` | mature aa208–222 | 35.0 | computational prior (Cys-free frame) |
| **Soy** | β-conglycinin α′ (P25974, Gly m 5) | `AYPFVVNATSNLNFL` | `FVVNATSNL` | mature aa344–358 | 50.3 | computational prior |

**Selection notes.** The dairy anchor is deliberately the Hill/Spergel β-casein aa59–78 epitope — the only food epitope in EoE confirmed by MHC-II tetramer and TCR transduction. IEDB netMHCIIpan resolved its DR7 core to `FAQTQSLVY` (the proline-rich "mature aa59–78" reading frame is a non-binder, IC50 >1500 nM — a numbering-convention trap this design checked explicitly). For wheat and soy the leads were filtered to exclude signal-peptide/leader stretches and high-hydrophobicity windows (both overscore in MHC prediction but are not physiologically presented), and the wheat lead was reframed to a cysteine-free window to avoid a synthesis/disulfide liability while preserving the `HNVVHAIIL` core.

## 3. Construct designs — 3 formats × 3 foods = 9 constructs

Every construct is a **single polypeptide** built on one shared, validated backbone:

```
[food peptide 15-mer] – L1 – [HLA-DRβ1β2 ectodomain, DRB1*07:01] – L2 – [HLA-DRα1α2 ectodomain] – [C-terminal format tag]
```
- **L1** = `GGGGSGGGGSGGS` — tethers the peptide N-terminally so it seats in the groove (covalent pMHC-II, the standard single-chain topology).
- **L2** = `(GGGGS)₄` — flexible β→α interchain linker.
- DR chains are authentic mature ectodomains: **DRα** from P01903, **DRβ** from **P13761 (DR-7 β chain)**.

The three **formats differ only in the C-terminal tag**, which sets the tolerance mechanism:

| Format | C-terminal tag | Length | Mechanism | Precedent |
|---|---|---|---|---|
| **Soluble single-chain** | His6 | 449 aa | Signal-1 without costimulation → clonal **anergy** | soluble pMHC / altered-peptide-ligand tolerance |
| **Nanoparticle (multivalent)** | His6 + `GGGSC` free thiol | 454 aa | Maleimide conjugation to a nanoparticle → multivalent display → **Tr1 regulatory expansion** | Navacim platform (Santamaria) |
| **Tetramer (depleting)** | `GGGGS` + AviTag + His6 | 469 aa | BirA biotinylation → streptavidin tetramer arming a depleting payload → **clone removal** | pMHC-II tetramer clone tracking/depletion |

Full amino-acid sequences: `eoe_pmhc_constructs.fasta`. Per-construct spec: `eoe_pmhc_construct_spec.csv`.

**Why this architecture.** One backbone, one expression/purification process, one QC assay — only a ~15-residue cassette changes between foods, and only the C-terminal tag changes between formats. This is the key manufacturability property: a nine-member panel collapses to *one platform + a peptide library*, and the same peptide library is shared with the diagnostic front-end (theranostic).

## 4. Structural validation (GPU co-folding)

The three pMHC-II:peptide complexes and the flagship dairy single-chain construct were co-folded on an A100 GPU with **ESMFold2-Fast** (Biohub, 2026), 10 loops × 68 sampling steps × 5 diffusion samples, best-of-5 by ipTM.

| Model | mean pLDDT | pTM | ipTM | Peptide residues in groove |
|---|---|---|---|---|
| Dairy pMHC-II:peptide | 0.86 | 0.89 | **0.87** | 15/15 |
| Wheat pMHC-II:peptide | 0.88 | 0.90 | **0.90** | 15/15 |
| Soy pMHC-II:peptide | 0.88 | 0.90 | **0.89** | 15/15 |
| Dairy single-chain construct | 0.82 | 0.83 | — (monomer) | groove intact |

All three complexes fold with high interface confidence (ipTM 0.87–0.90) and **all 15 peptide residues in contact with the DR groove**. Per-residue burial profiles confirm the predicted 9-mer core is deeply seated (59–64 heavy-atom contacts across the P1–P9 pockets) with the characteristic C-terminal flank protruding — canonical class-II geometry, not surface docking. The single-chain construct folds with the groove intact (pLDDT 0.82), confirming the linker topology is compatible with a seated peptide.

Structures (open in the 3D viewer): `dairy_pmhc.pdb`, `wheat_pmhc.pdb`, `soy_pmhc.pdb`, `dairy_singlechain.pdb`. Metrics: `eoe_pmhc_foldback_metrics.csv`.

## 5. Strategy figure

`eoe_pmhc_strategy_figure.png` — (a) the 3-format × 3-food strategy matrix; (b) presentation load per food; (c) selected core affinity; (d) the shared-backbone architecture; (e–g) the co-folded DR7 grooves with each food peptide's 9-mer core highlighted in the P1–P9 pockets.

## 6. Per-food, per-format spec sheets

Each entry below is a ready-to-order single-chain ORF. Peptide cassette in **bold-conceptual** terms is the only per-food variable; tag is the only per-format variable.

### Dairy — β-casein `FAQTQSLVY` (tetramer-validated anchor)
- **Single-chain** (449 aa): `KIHPFAQTQSLVYPF` + backbone + His6 — lead tolerance candidate.
- **Nanoparticle** (454 aa): + `GGGSC` free thiol for iron-oxide NP conjugation → Tr1 expansion.
- **Tetramer** (469 aa): + AviTag → SA tetramer, doubles as the diagnostic staining reagent.

### Wheat — α-gliadin `HNVVHAIIL` (computational prior)
- Single-chain / Nanoparticle / Tetramer as above with the `IHNVVHAIILHQQQQ` cassette.
- **Caveat:** gliadin is HLA-DQ-associated in celiac; here the DR7-restricted frame is a prediction only, and wheat is the documented false-positive of presentation-vs-pathology (§7).

### Soy — β-conglycinin `FVVNATSNL` (computational prior)
- Single-chain / Nanoparticle / Tetramer with the `AYPFVVNATSNLNFL` cassette.

## 7. The guardrail: presentation ≠ pathology, and prior art

**This is the rule that governs the whole design.** A binding prediction says a peptide *can* be presented on DR7 — it does **not** say that peptide drives disease. Disease requires an expanded, pathogenic effector-Th2 clone specific for that peptide. Two hard consequences:

1. **Only the dairy epitope is validated.** β-casein aa59–78/DRB1\*07:01 is confirmed by tetramer and TCR transduction (the eoeTCR-4 clone). The wheat and soy epitopes are **computational priors** — the constructs are correctly built and fold well, but whether a pathogenic clone against them exists in any given patient is an open functional question. Wheat specifically is the documented false-positive: a patient can present gliadin on a celiac-risk allele yet tolerate wheat entirely.

2. **Every construct is a hypothesis until a functional readout confirms its target clone exists.** The next step for wheat and soy is not synthesis — it is the AIM/tetramer functional assay (the diagnostic arm) to confirm a food-reactive peTh2 clone is present before a tolerance reagent against it is meaningful.

**Prior art.** The functional-T-cell-assay-to-name-food paradigm and the EoE food-specific TCR platform are foundational work of the **Hill–Spergel group and are patent-protected** (Dilollo/Spergel/Hill 2025, *JACI*, doi:10.1016/j.jaci.2025.01.008; Cianferoni/Ruffner/Spergel 2017, doi:10.1016/j.anai.2017.11.006). This design operates downstream of that platform. Any translation would need to engage that IP.

## 8. Scope and disclaimer

This is a **research/hackathon design specification**, not a validated therapeutic and not medical advice. MHC-II binding rests on computational predictions (netMHCIIpan/mhcnuggets); DR predictions are better benchmarked than DQ but remain priors. Structural models are co-folding predictions, not experimental structures. Antigen-specific tolerance carries real risks (anaphylaxis on peptide exposure, incomplete anergy, off-target immunomodulation) that only preclinical and clinical work can address. Any therapeutic decision for a real patient must be made by qualified clinicians.

---

## Deliverables
- `eoe_pmhc_therapeutic_cores.csv` — epitope core design table
- `eoe_pmhc_constructs.fasta` — 9 full construct sequences
- `eoe_pmhc_construct_spec.csv` — per-construct architecture/length
- `dairy_pmhc.pdb`, `wheat_pmhc.pdb`, `soy_pmhc.pdb` — co-folded pMHC-II:peptide complexes
- `dairy_singlechain.pdb` — folded flagship single-chain construct
- `eoe_pmhc_foldback_metrics.csv` — fold-quality metrics
- `eoe_pmhc_strategy_figure.png` — strategy figure