# Validation & Extension Note — EoE Signature Robustness + Complete HLA Panel **Two computational refinements strengthening prior findings (no wet-bench / no GPU required).** --- ## 1. Leave-One-Cohort-Out (LOCO) Robustness of the 567-Gene Meta-Signature **Question:** How many of the 567 high-confidence signature genes are driven by any single cohort, versus reproducible across the whole set of 9 bulk cohorts? **Method:** The DerSimonian-Laird random-effects meta-analysis was re-run 9 times, each dropping one cohort (per-study SE = |log2FC/t|; high-confidence = FDR<0.05 AND |pooled log2FC|≥1 AND concordant direction in ≥6 studies). The full-data re-derivation reproduced **exactly 567 genes and 25,654 pooled genes**, confirming faithful method reproduction before the drop experiments. **Results:** - **399/567 genes (70.4%) are "core-robust"** — they remain high-confidence in all 9 leave-one-out re-derivations. Saved as `eoe_meta_signature_loco_robust.csv`. - **476/567 (84%) are "stable"** (survive ≥7/9 drops). - **91/567 (16%) are "fragile"** (depend on ≥3 cohorts; survive ≤6/9). - Signature stability varies by cohort: dropping GSE148381 removes 118 genes (79% retained), whereas dropping GSE228083 removes only 1 (99.8% retained) — identifying GSE148381 as the most influential single cohort. **Canonical markers:** 17 of 18 canonical EoE markers are core-robust (9/9), including **every protein-design lead (CCL26, IL1RL1, SIGLEC6, POSTN)** and **every GWAS-risk gene (CAPN14, DSG1)**. The sole exception is **FLG (5/9 — fragile)**, which independently rediscovers the previously-flagged cohort-specific direction reversal in the recurrence cohort GSE278888. The LOCO analysis found this caveat without being told to look for it. **Interpretation:** The core disease signature and all downstream drug targets rest on genuinely reproducible cross-cohort biology, not single-study artifacts. Deliverables: `loco_robustness.png`, `loco_gene_robustness.csv`, `loco_cohort_summary.csv`. --- ## 2. Complete HLA-DRB1 Allele Panel (13 alleles) with Cross-Predictor Validation **Question:** The Phase-4 epitope panel characterized only 5 HLA-DRB1 alleles (IEDB netMHCIIpan became unresponsive before more could be run); the other common alleles were left "indeterminate." Complete the common-allele panel. **Method:** IEDB remained unresponsive this session, so predictions used **mhcnuggets** (a locally-installed, trained LSTM MHC-II binding-affinity predictor) across all 13 common European HLA-DRB1 alleles × 2,424 unique 15-mers from the 10 allergen proteins = **31,512 predictions**. To justify substituting a second predictor, mhcnuggets was **cross-validated against the netMHCIIpan-4.1 results on the 5 shared alleles.** **Cross-predictor concordance (`predictor_concordance.csv`):** - Mean Spearman ρ = **0.79** (log-IC50 rank correlation), range 0.74–0.84 across the 5 shared alleles. - Both predictors **independently agree on the biological ranking**: DRB1*01:01 most promiscuous, DRB1*03:01 by far the least — the ordering that anchors patient stratification. - Absolute strong-binder thresholds differ between tools (Jaccard 0.20–0.62), so counts are not identical, but the rank agreement validates using mhcnuggets to extend the panel. **Complete 13-allele burden (`hla_allele_burden_13.csv`, `hla_panel_13allele.png`):** - Panel now covers **76.8% of the European DRB1 allele pool** (~59% of patients have both alleles characterized). - **Common + high-burden:** DRB1*01:01 (9.3% freq, 733 strong binders) — the primary recruitment-relevant allele. - **Common + genuinely low-burden (now measured, not assumed zero):** DRB1*03:01 (11.3%, 55), DRB1*13:01 (6.2%, 152), DRB1*08:01 (2.8%, 43). - **Rare + high-burden:** DRB1*14:01 (1.5%, 928), DRB1*10:01 (1.0%, 734) — high per-allele load but low population impact. **Extended stratification (`patient_epitope_burden_tiers_13.csv`):** 91 diplotypes tiered by combined burden — Tier A (high) ≈21.5%, Tier B ≈38.9%, Tier C (low) ≈39.6% of the characterized-allele diplotype space. **Key correction locked in:** the earlier framework's assumption that non-panel alleles carry "zero" epitopes is now replaced with real predictions. Some (DRB1*14:01) are in fact high-burden; others (DRB1*03:01, *08:01) are confirmed low. This removes the last piece of unmeasured assumption from the patient-stratification design. **Caveat:** mhcnuggets and netMHCIIpan are both sequence-based binding-affinity predictors; neither establishes immunodominance or T-cell reactivity, which remain the Phase-1 ex vivo validation task. --- **Files:** loco_robustness.png, loco_gene_robustness.csv, loco_cohort_summary.csv, eoe_meta_signature_loco_robust.csv, hla_panel_13allele.png, hla_allele_burden_13.csv, allergen_allele_matrix_13.csv, patient_epitope_burden_tiers_13.csv, predictor_concordance.csv, mhcnuggets_all_alleles.csv. > **Frequency-source caveat:** The European DRB1 frequencies used here are approximate literature values (order-of-magnitude, for platform-coverage illustration only) and were not fetched from a database in this session. They match the current in-session `hla_allele_frequencies.csv` but differ from a prior-session version of that table; because eligibility is HLA-agnostic and per-patient (see recruitment doc v4.0), these frequencies are used only to illustrate that the predictor generalizes across common alleles — never to select, rank, or exclude patients. Any production use should genotype each patient directly rather than rely on population frequencies.