# Patient Recruitment & Personalized Epitope Platform Design for pMHC Therapy
## EoE Allergen-Specific Immunotherapy — Phase 4A Study

**Version:** 4.0 (inclusive platform design — HLA-agnostic eligibility, per-patient epitope prediction)
**Status:** Recruitment Protocol
**Data provenance:** mhcnuggets (13 HLA-DRB1 alleles, 31,512 predictions) cross-validated against IEDB netMHCIIpan-4.1 (5 shared alleles, mean Spearman ρ=0.79)

---

## 1. DESIGN PRINCIPLE: A PLATFORM FOR EVERY PATIENT

**This therapy is designed to be available to any EoE patient, regardless of HLA genotype.** HLA-DRB1/DQ/DP allele frequencies vary substantially across ancestral populations. Any design that made a fixed set of "high-burden" alleles an eligibility criterion would systematically under-serve patients whose alleles are common in non-European populations — encoding an inequity into both the trial and the product. We reject that design.

Instead, the therapeutic is a **personalized epitope platform**:

1. **Eligibility is HLA-agnostic.** Every consented EoE patient is eligible. HLA type is never an inclusion/exclusion criterion.
2. **The epitope panel is computed per patient.** A patient's own HLA genotype (all class-II loci: DRB1, DRB3/4/5, DQ, DP) is entered into the predictor, which returns *that individual's* personalized set of high-affinity allergen epitopes. The pMHC reagents for that patient are then built on their own MHC-II alleles.
3. **Predicted burden is a mechanistic covariate, never a gate.** We record each patient's predicted epitope load to test whether clinical response tracks with it — an internal mechanistic check — but this never affects who is enrolled or treated.
4. **Scale is the point.** The prediction + reagent-selection step is fully computational and runs per patient in minutes; the assay workflow (Task 3) is designed to run at scale so a bespoke epitope panel can be generated for each individual.

The allele-level results below are therefore presented as **platform-coverage evidence** ("can we compute a meaningful panel for the alleles patients actually carry?"), explicitly **not** as a patient-prioritization scheme.

---

## 2. PLATFORM COVERAGE: DOES PER-PATIENT PREDICTION GENERALIZE?

The predictor (mhcnuggets) has trained models for **136 HLA class-II alleles** — 50 HLA-DRB1 (39 unique two-field), plus DRB3/4/5, 28 DQA1 × 38 DQB1 (DQ heterodimers), and 8 DPA1 × 18 DPB1 (DP). IEDB netMHCIIpan covers thousands more. In practice, essentially any patient's class-II genotype can be turned into a personalized epitope panel.

To confirm the prediction is meaningful across the alleles patients carry, we characterized the 13 most common HLA-DRB1 alleles against the 10-protein allergen panel (2,424 unique 15-mers → 31,512 predictions):

| HLA-DRB1 | Strong binders (IC50<500 nM) | Approx. freq (EUR ref) |
|---|---|---|
| *14:01 | 928 | 1.5% |
| *10:01 | 734 | 1.0% |
| *01:01 | 733 | 9.3% |
| *16:01 | 442 | 1.4% |
| *07:01 | 356 | 12.8% |
| *09:01 | 338 | 1.2% |
| *15:01 | 324 | 13.0% |
| *04:01 | 304 | 8.9% |
| *11:01 | 254 | 5.8% |
| *12:01 | 176 | 1.6% |
| *13:01 | 152 | 6.2% |
| *03:01 | 55 | 11.3% |
| *08:01 | 43 | 2.8% |

**What this shows:** every characterized allele presents *some* allergen epitopes (range 43–928 strong binders), and the count varies ~20-fold across alleles. This variation is exactly why the panel must be **personalized** — a patient's reagents are built on their own alleles so that a low-DRB1 allele (e.g. *03:01) is complemented by that patient's second DRB1 allele and their DQ/DP loci, rather than the patient being excluded. The frequency column is included only to document that the predictor generalizes across common alleles; it is not used to rank or select patients.

> **Note on ancestry and frequencies.** The reference frequencies above are approximate illustrative European values (not fetched from a database in this session; they differ from a prior-session frequency table and should be treated as order-of-magnitude only). A production platform computes each patient's panel from their actual genotype and does not rely on any population reference. Validation should explicitly include cohorts of diverse ancestry, and additional common alleles from African, East Asian, South Asian, and Indigenous populations should be characterized on the same footing (the predictor already supports them). This is a stated equity requirement of the platform, not an optional extension.

Cross-predictor validation (`predictor_concordance.csv`): mhcnuggets vs netMHCIIpan-4.1 on the 5 shared alleles gives mean Spearman ρ=0.79 on log-IC50, with both tools independently agreeing on the allele ranking — supporting use of the predictor to build per-patient panels.

---

## 3. RECRUITMENT (ALL-COMERS)

1. Screen and consent EoE patients on clinical criteria only (histology, symptoms) — **no HLA screening for eligibility**.
2. Baseline endoscopy + biopsy (peak intraepithelial eosinophils, EREFS).
3. HLA class-II genotyping (clinical PCR-SSO/SSP) — used **only** to build each patient's personalized epitope panel and as a mechanistic covariate.
4. Generate the individualized epitope panel and patient-specific pMHC reagents.
5. Enroll to the total target (Section 4). Aim for an ancestrally diverse cohort that reflects the EoE patient population; monitor enrollment demographics as an equity metric.

---

## 4. STATISTICAL DESIGN (NO PATIENT EXCLUDED)

**Primary analysis:** regression of clinical/immunological response on each patient's **individualized predicted epitope burden**, across the entire enrolled cohort. This tests the platform's mechanistic hypothesis (does response track with personalized epitope load?) using every patient rather than comparing pre-selected groups.

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

**Recommended enrollment: ~60 patients**, powered to detect a burden–response correlation of r≈0.35 while giving margin for covariate adjustment (ancestry, baseline severity, atopic status).

**Secondary (descriptive, no exclusion):** patients naturally spanning high vs low personalized burden can be compared post hoc (tertiles arising *within* the all-comers cohort; ~25/group by an unpaired-t heuristic at d=0.8). Nobody is screened out to form these strata — they emerge from the enrolled population.

(Supersedes v3.0, which used a Tier A vs Tier C selected-enrollment design. That approach is retracted because HLA-based selection would inequitably restrict access.)

---

## 5. SECONDARY BIOMARKERS (Task 3 assay)

Allergen-specific CD4+ proliferation (SI), IFNγ/IL-5/IL-13 (Luminex), CD25+HLA-DR+ activation and PD-1 (flow), esophageal Th2 mRNA and eosinophil count — each read against the patient's own personalized epitope panel.

---

## 6. SUCCESS CRITERIA

- **Primary:** significant positive association between individualized predicted epitope burden and treatment response across the full cohort (regression slope > 0, p<0.05), with clinically meaningful response (≥50% peak-eosinophil reduction) in a substantial fraction of patients.
- **Platform:** a valid personalized epitope panel and patient-specific reagent set is generated for ≥95% of enrolled patients, across diverse HLA types and ancestries.
- **Equity:** enrolled-cohort ancestry distribution reflects the EoE patient population; response is not confined to any single ancestry or allele group.

---

## 7. REFERENCES (retrieved via OpenAlex)

- Arias Á et al. (2014) *Gastroenterology* — dietary-intervention histologic remission in EoE. doi:10.1053/j.gastro.2014.02.006
- Kagalwalla AF et al. (2006) *Clin Gastroenterol Hepatol* — six-food elimination diet.
- Rothenberg ME (2001) *J Allergy Clin Immunol* — EoE pathogenesis. doi:10.1067/mai.2001.120095
- Mulder DJ et al. (2011) *Am J Pathol* — MHC class II expression/antigen presentation by human esophageal epithelium. doi:10.1016/j.ajpath.2010.10.027

*Prediction methods:* mhcnuggets (trained LSTM class-II predictor) and IEDB netMHCIIpan-4.1 (Nielsen lab) — analysis tools, not literature references.

---

**Corrections in v4.0:** (i) eligibility is now HLA-agnostic — every EoE patient qualifies; (ii) epitope panels are computed per patient from their own genotype rather than screened against a fixed allele list; (iii) predicted burden is a mechanistic covariate, not an enrollment gate; (iv) the selected Tier A vs Tier C comparison (v3.0) is retracted as inequitable and replaced with an all-comers burden–response regression; (v) explicit equity requirements added for ancestral diversity and multi-population allele characterization.