# Phase 4: Prospective pMHC-Directed T-Cell Therapy Framework for EoE
## Consolidated Report — Allergen Epitope Mapping to IND-Track Preclinical Plan

**Project:** EoE therapeutic target discovery & protein design (Life Sciences Hackathon)
**Phase 4 scope:** Expand food-allergen epitope mapping to 8 categories; design a prospective, inclusive T-cell-targeted pMHC therapy platform from computational epitopes through an IND-enabling preclinical plan.

---

## Executive Summary

Phase 4 builds a prospective, antigen-directed therapy framework on top of the Phase 1–3 finding that EoE is a food-antigen-driven, CD4+/Th2 disease with epithelial MHC-II antigen presentation. It delivers: (1) an **MHC-II binding-affinity epitope panel** across 10 food-allergen proteins and 13 HLA-DRB1 alleles; (2) an **inclusive, HLA-agnostic recruitment design** built on per-patient personalized epitope prediction; (3) an **IND-ready T-cell reactivity assay SOP**; (4) a **mechanism specification** selecting pMHC-II multimer-induced anergy as the lead approach, with a DOI-verified literature synthesis; and (5) a **four-phase preclinical roadmap** with go/no-go gates. Two validation analyses harden the foundation: a **leave-one-cohort-out (LOCO) robustness test** of the upstream 567-gene signature, and a **cross-validated 13-allele HLA panel**.

**Design principle — a platform for every patient.** Because HLA allele frequencies vary strongly across ancestral populations, this framework deliberately does **not** use HLA type as an eligibility criterion. Every EoE patient is eligible; each patient's own class-II genotype drives a personalized epitope panel computed on demand, and predicted burden is used only as a mechanistic covariate — never as an enrollment gate. The goal is a scalable platform that develops a bespoke therapeutic for any individual, regardless of ancestry or HLA.

**Method integrity note.** An initial version of the epitope panel used a hand-coded position-preference heuristic that produced identical score distributions across unrelated allergens; it was invalid and has been fully retracted. Epitope results come from **IEDB netMHCIIpan-4.1** (5 alleles) and the cross-validated **mhcnuggets** predictor (all 13 alleles). All literature citations were retrieved from OpenAlex with verified DOIs.

---

## 1. Allergen Epitope Panel (Task 1)

**Allergen set (10 proteins, 8 food categories):** milk (β-lactoglobulin P02754, α-S1-casein P02662), wheat (α-gliadin P18573), egg (ovomucoid P01005), soy (glycinin G5 P04347, β-conglycinin P25974), peanut (Ara h 2 Q6PSU2), fish (parvalbumin Q90YK9), shellfish (tropomyosin M1H607), tree nut (walnut Jug r 1 P93198). Sequences verified against UniProt; annotated FASTA in `allergen_sequences.fasta`.

**Core panel (netMHCIIpan-4.1 BA mode, 15-mer windows, 5 HLA-DRB1 alleles):**

| Metric | Value |
|---|---|
| Total predictions | 12,140 |
| Strong binders (IC50 < 500 nM) | 2,287 |
| Unique binding cores (global, deduplicated) | 470 |
| Cross-allergen shared cores | 0 (all allergen-specific) |

**Allergen epitope load (strong binders):** soy β-conglycinin 551, soy glycinin 526, milk β-lactoglobulin 283, wheat gliadin 246, milk α-S1-casein 230, shellfish tropomyosin 116, egg ovomucoid 109, fish parvalbumin 93, walnut 78, peanut Ara h 2 55.

Deliverables: `allergen_epitope_panel.csv`, `allergen_epitope_load.csv`, `hla_allele_prediction_summary.csv`, `strong_epitope_binders.csv`, `epitope_core_matrix.csv`, `allergen_epitope_heatmap.png`, `epitope_load_by_allergen.png`.

---

## 2. Inclusive Recruitment & Personalized Epitope Platform (Task 2)

**Eligibility is HLA-agnostic — every EoE patient qualifies.** The therapeutic is a personalized platform: a patient's own HLA class-II genotype (DRB1, DRB3/4/5, DQ, DP) is entered into the predictor, which returns *that individual's* epitope set; pMHC reagents are built on the patient's own alleles. Predicted burden is recorded as a **mechanistic covariate**, never an enrollment gate.

**Platform coverage.** The predictor spans **136 HLA class-II alleles** (50 DRB1, plus DRB3/4/5, 28 DQA1 × 38 DQB1, 8 DPA1 × 18 DPB1); netMHCIIpan covers thousands more — so essentially any patient's genotype yields a meaningful panel. The 13 most common HLA-DRB1 alleles were characterized against the allergen panel (31,512 mhcnuggets predictions): every allele presents allergen epitopes (43–928 strong binders, ~20-fold range), which is precisely why panels must be personalized rather than patients screened against a fixed list.

**Statistical design (no patient excluded).** Primary analysis is a regression of clinical/immunological response on each patient's individualized predicted burden, across the full cohort:

| Effect (Pearson r) | Required n (α=0.05, power=0.80) |
|---|---|
| 0.50 | 30 |
| 0.40 | 47 |
| 0.35 | 62 |
| 0.30 | 85 |

**Recommended enrollment ≈60 patients** (powers r≈0.35 with covariate margin). An explicit equity requirement: the enrolled cohort should reflect the ancestral diversity of the EoE population, and common alleles from non-European populations should be characterized on the same footing.

Deliverables: `recruitment_stratification_rationale.md` (v4.0), `hla_allele_burden_13.csv`, `patient_epitope_burden_tiers_13.csv`, `predictor_concordance.csv`, `hla_panel_13allele.png`.

---

## 3. T-Cell Reactivity Assay (Task 3)

An IND-ready SOP (`t_cell_assay_protocol.md`): PBMC/biopsy isolation; 14-day allergen-specific CD4+ expansion; proliferation ([³H]-thymidine SI); cytokines (IFNγ/IL-5/IL-13, Luminex); flow phenotyping (CD25/HLA-DR/PD-1); antigen-specific cloning and TCR sequencing; QC, controls, troubleshooting, reporting templates. Peptides are drawn from **each patient's personalized panel**. This assay is both the Phase-1 validation readout and the clinical pharmacodynamic biomarker, and is designed to run at scale (one bespoke panel per patient).

---

## 4. Mechanism Specification (Task 4)

Four mechanisms with a decision tree:
1. **pMHC-II multimer anergy** — antigen-specific, protein reagent, reversible → **recommended for Phase 4A** (lowest development risk).
2. **pMHC particle + deletion signal** — deletion, but complex CMC.
3. **CTLA-4 costimulation blockade** — precedented but antigen-nonspecific (adjunct only).
4. **Engineered T-cell (CAR-Treg/TCR)** — most selective, highest regulatory burden (later phase).

Rationale rests on the signal-1-without-costimulation tolerance principle and antigen-specific tolerance-induction literature, cited with retrieved DOIs (`pmhc_tcr_mechanism_review.md`).

---

## 5. Preclinical Roadmap (Task 5)

Four phases with go/no-go gates (`preclinical_workflow_roadmap.md`):
- **Phase 1 (0–9 mo):** ex vivo human validation — Gate A: response tracks with individualized burden + reagent engagement.
- **Phase 2 (6–18 mo):** reagent scale-up + in vivo proof-of-concept — Gate B: ≥50% tissue-eosinophil reduction, specific, protective immunity preserved.
- **Phase 3 (15–30 mo):** IND-enabling GLP tox / PK-PD / CMC — Gate C: safe margin + manufacturable.
- **Phase 4 (30–42 mo):** Phase 1b/2a all-comers trial (~60 patients; burden–response regression).

---

## 6. Validation & Extension (this session)

**6.1 Signature robustness (LOCO).** The upstream 567-gene meta-signature was re-derived 9 times, each dropping one cohort (DerSimonian-Laird; the full re-run reproduced exactly 567 genes, confirming method fidelity). **399/567 genes (70.4%) are core-robust** (high-confidence in all 9 drops); 84% survive ≥7/9. **17 of 18 canonical markers are core-robust**, including every protein-design lead (CCL26, IL1RL1, SIGLEC6, POSTN) and every GWAS-risk gene (CAPN14, DSG1). The single exception — FLG (5/9) — independently rediscovers its known cohort-specific direction reversal. Deliverables: `loco_robustness.png`, `loco_gene_robustness.csv`, `loco_cohort_summary.csv`, `eoe_meta_signature_loco_robust.csv`.

**6.2 Complete HLA panel + cross-predictor validation.** IEDB was unresponsive, so the panel was extended from 5 to 13 HLA-DRB1 alleles with mhcnuggets, **cross-validated against netMHCIIpan-4.1 on the 5 shared alleles** (mean Spearman ρ=0.79 on log-IC50; both tools independently agree on the allele ranking). Panel coverage rose to ~76.8% of the European DRB1 pool (using approximate literature frequencies for illustration only — not fetched this session, and differing from a prior-session table; eligibility is HLA-agnostic, so this figure never gates enrollment). Crucially, this **replaced the earlier "zero-burden" assumption** for uncharacterized alleles with real predictions — some (DRB1*14:01) are high-burden, others (DRB1*03:01, *08:01) confirmed low. Full write-up: `validation_extension_note.md`.

---

## 7. Limitations & Honest Caveats

1. **Epitopes are computational.** Binding-affinity predictors (netMHCIIpan, mhcnuggets) do not establish immunodominance or T-cell reactivity; ex vivo validation (Phase 1) is the required confirmation.
2. **13 DRB1 alleles characterized against the allergen panel so far;** the predictor supports 136 class-II alleles (incl. DQ/DP), and per-patient panels use the patient's full genotype. Broader multi-ancestry characterization is a stated equity requirement.
3. **Allele frequencies are approximate European reference values,** used only to document platform coverage — never to select or rank patients.
4. **No wet-lab or in vivo data exist yet** — this is a prospective framework.
5. **pMHC-directed CD4+ therapy is early-stage;** the roadmap's escalation ladder exists because Mechanism 1 durability is unproven.

---

## 8. Deliverable Index

| Task | File | Description |
|---|---|---|
| 1 | allergen_sequences.fasta | 10 allergen proteins, UniProt-verified |
| 1 | allergen_epitope_panel.csv, strong_epitope_binders.csv | 12,140 netMHCIIpan predictions; 2,287 strong |
| 1 | allergen_epitope_load.csv, hla_allele_prediction_summary.csv, epitope_core_matrix.csv | summary tables |
| 1 | allergen_epitope_heatmap.png, epitope_load_by_allergen.png | visualizations |
| 2 | recruitment_stratification_rationale.md (v4.0) | inclusive platform recruitment design |
| 2 | hla_allele_burden_13.csv, patient_epitope_burden_tiers_13.csv, allergen_allele_matrix_13.csv | 13-allele panel |
| 2 | predictor_concordance.csv, hla_panel_13allele.png | cross-predictor validation |
| 3 | t_cell_assay_protocol.md | IND-ready T-cell assay SOP |
| 4 | pmhc_tcr_mechanism_review.md | mechanism spec + literature synthesis |
| 5 | preclinical_workflow_roadmap.md | 4-phase roadmap |
| 6 | validation_extension_note.md, loco_robustness.png, loco_gene_robustness.csv, loco_cohort_summary.csv, eoe_meta_signature_loco_robust.csv, mhcnuggets_all_alleles.csv | LOCO + HLA validation |
| — | eoe_phase4_pmhc_therapy_package.md | this report |

All computational epitope claims trace to netMHCIIpan-4.1 / mhcnuggets output; all citations trace to OpenAlex-retrieved DOIs. **Eligibility is HLA-agnostic; the platform is designed for every patient.**