# Celiac Disease Antigen-Epitope / pMHC Stream — Consolidated Report

**Proof-of-transfer:** porting the EoE antigen-epitope/pMHC pipeline to celiac disease, with a TG2-deamidation extension and — uniquely for celiac — a benchmark against experimentally-mapped immunodominant epitopes.

---

## 1. Summary

We re-ran the project's disease-agnostic antigen-epitope pipeline on celiac disease, its intended second indication. The pipeline transferred cleanly: gluten antigen set → UniProt sequences → TG2 deamidation → paired native/deamidated 15-mer windows → HLA-DQ binding prediction → per-HLA burden → benchmark → HLA-resolved panel. Because celiac has a defined antigen (gluten), a near-absolute HLA restriction (DQ2.5/DQ8/DQ2.2), and an experimentally-mapped epitope truth set, we could do something EoE never permitted: **measure whether the method recovers known biology.** It does — but the benchmark also surfaced a concrete calibration lesson that changes how the output must be read.

**Three headline results:**
1. **Deamidation drives presentation, directionally.** Modeling TG2 Gln→Glu editing, **55 peptide windows cross the strong-binder threshold only after deamidation and 0 are lost** — a strictly one-way effect, with 48/55 on DQ2.5. This reproduces the central molecular event of celiac pathogenesis from first principles.
2. **Relative allele-specificity is recovered; absolute IC50 calibration is not.** Every DQ2.5-restricted known epitope (11/11) is predicted to prefer DQ2.5 over the other alleles (binomial p = 5.6×10⁻⁶ for the 11/11 subset; p = 2.1×10⁻⁴ across all 13 epitopes with a restricting-allele prediction, 11/13 correct), and all 3 tested immunodominant epitopes prefer their correct allele. But at the standard IC50 < 500 nM cutoff, recall of known epitopes is **0%** — the canonical DQ2.5-glia-α1a scores 1885 nM. netMHCIIpan systematically under-scores HLA-DQ affinity; the pipeline is trustworthy for *ranking and allele assignment*, not for absolute strong/weak calls on DQ.
3. **DQ2.5 dominates the predicted burden**, as expected for a disease where ~90% of patients carry it: 142 strong deamidated binders vs 20–23 (DQ8) and 12–16 (DQ2.2).

---

## 2. Antigen set (Step 1)

20 proteins, 7,380 residues, spanning the full clinically-relevant gluten families across all three toxic grains — following the EoE lesson that a source must be represented by its complete family, not a token protein:
- **Wheat (13):** α/β-gliadin ×5, γ-gliadin ×3, ω-gliadin ×1, LMW-glutenin ×2, HMW-glutenin ×2
- **Barley (5):** B-, C-, D-, γ-hordeins
- **Rye (2):** ω-, γ-secalins

Four candidate accessions were caught and replaced during QC (three returned empty sequences; one — Q94IL7 — was a *poplar potassium transporter* mislabeled as secalin). Deliverables: `celiac_gluten_antigen_set.csv`, `celiac_gluten_antigens.fasta`.

## 3. Ground-truth epitope reference (Step 2)

31 HLA-DQ-restricted immunodominant gluten epitopes transcribed verbatim from the Sollid nomenclature (Sollid et al. 2012, *Immunogenetics* 64:455–460, doi:10.1007/s00251-012-0599-z; Update 2020, doi:10.1007/s00251-019-01141-w): 23 DQ2.5, 4 DQ8, 3 DQ8.5, 1 DQ2.2. 29 fall in families covered by our antigen set (2 oat/avenin epitopes excluded); 26/29 native cores are present verbatim in the fetched sequences, confirming the set spans the immunodominant regions. Listed cores are the TG2-**deamidated** forms (E); native counterparts reconstruct E→Q. Deliverable: `celiac_known_epitopes.csv`.

## 4. TG2 deamidation model (Step 3)

Implemented the QXP specificity rule (Q→E where a proline sits at +2 and not at +1; Fleckenstein/Vader/Dorum consensus). Validated against the Sollid-annotated deamidation sites: **recovers 24/31 (77%) of annotated sites, 19/26 epitopes fully**, including all α-gliadin and ω-gliadin immunodominant anchors (e.g. PQPQLPY→PQPELPY). The 7 misses are non-canonical sites (e.g. glutenin FSQQQQP motifs) known to be hard to predict from sequence alone. Applied genome-wide: **464 predicted sites across 20 antigens** (6.4/100 aa); ω-secalin and C-hordein carry the highest density. Deliverables: `celiac_deamidation_map.csv`, `celiac_deamidation_map.png`.

## 5. HLA-DQ binding prediction (Step 4)

netMHCIIpan (IEDB `netmhciipan_ba`) over all 20 antigens × {native, deamidated} × {DQ2.5 (DQA1\*05:01/DQB1\*02:01), DQ8 (DQA1\*03:01/DQB1\*03:02), DQ2.2 (DQA1\*02:01/DQB1\*02:02)} = **41,664 predictions**. Two ω-gliadin/C-hordein sequences containing unknown (X) residues were handled by windowing out X-containing 15-mers. Deliverable: `celiac_epitope_panel.csv` (+ raw checkpoint `iedb_predictions.json`).

## 6. Burden and the deamidation shift (Step 5)

Using deduplicated unique-core counting (42 unique deamidated strong cores vs 181 raw per-allele windows — avoiding the double-counting the pipeline warns about):
- **DQ2.5 burden dominates:** 142 strong deamidated binder-windows vs 20–23 (DQ8), 12–16 (DQ2.2).
- **Deamidation shift is directional:** paired native→deamidated comparison over 20,832 windows, 58% improve, Wilcoxon p ≈ 0; **55 windows gain strong-binder status after deamidation, 0 lose it** (48 on DQ2.5).

Deliverables: `celiac_epitope_burden.csv`, `celiac_deamidation_shift.csv`, `celiac_epitope_burden.png`.

## 7. Benchmark against known epitopes (Step 6) — the validation unique to celiac

| Metric | Result | Interpretation |
|---|---|---|
| Recall @ IC50 < 500 nM | **0%** (0/29) | Absolute threshold fails on DQ — see calibration note |
| Recall @ top-10% rank within allele | 11% | Weak even by rank at the immunodominant core |
| DQ2.5-restricted epitopes preferring DQ2.5 | **11/11 (100%)**, binomial p = 5.6×10⁻⁶ (subset); 11/13 all epitopes, p = 2.1×10⁻⁴ | Relative allele assignment works |
| Immunodominant epitopes preferring correct allele | 3/3 | " |
| Deamidation lowers IC50 of known epitope | 11/13 (85%), median 1.57× | Deamidation modeling improves recovery |
| DQ2.5-glia-α1a (most immunodominant) best IC50 | **1885 nM** | Above 500 nM cutoff — the calibration problem in one number |

**The calibration finding (important, and generalizable):** netMHCIIpan is far less benchmarked on HLA-DQ than on HLA-DR, and it systematically under-predicts DQ affinity. Celiac's known epitopes are potent T-cell activators yet score as *weak* by IC50. Consequently: **use within-allele rank and relative allele-preference, not a fixed IC50 threshold, when interpreting DQ predictions.** This is the celiac analogue of the EoE HLA-false-positive lesson — a predictor caveat that must travel with the output. Deliverables: `celiac_epitope_benchmark.csv`, `celiac_allele_specificity.csv`, `celiac_epitope_benchmark.png`.

## 8. HLA-resolved panel (Step 7)

A 45-entry panel combining the top-12 predicted binders per HLA (novel hypotheses) with the 9 known epitopes located in our windows (each carrying its actual predicted rank), tagged `entry_type` ∈ {known_immunodominant, known_epitope, novel_predicted}. Two figures: a family×HLA best-IC50 heatmap (every family scores lowest on DQ2.5) and a DQ2.5 epitope-load bar flagging which families carry a known epitope. Deliverables: `celiac_epitope_map.csv`, `celiac_epitope_map.png`.

---

## 9. The guardrail: presentation ≠ pathology

**Every number in this report is a computational prior, not a diagnosis.** A binding or burden prediction says a peptide *can be presented* on a given HLA-DQ molecule; it does not establish that the peptide drives disease in a given person. In celiac the presentation→pathology link is far tighter than in EoE (near-absolute HLA restriction, a defined antigen, and — for the known epitopes — decades of functional T-cell confirmation). But for any *novel* predicted binder in this panel, the specificity-defining evidence remains a **functional readout**: gluten-specific CD4⁺ T-cell activation by HLA-DQ–gluten tetramer or AIM assay on patient cells, ideally after gluten challenge. Rank the panel to decide which peptide pools to build and test first; do not name a novel epitope "pathogenic" from prediction alone.

A second, celiac-specific guardrail: HLA-DQ2.5 is carried by ~25–40% of the general Western population but only ~3% develop celiac disease. HLA presentation is **necessary but not sufficient** — burden is a risk/capacity axis, never a diagnosis. Any diagnostic or tolerance-reagent use must pair prediction with serology (anti-TG2) and functional confirmation.

## 10. Caveats and next steps

- **DQ prediction is approximate** (§7); treat magnitudes as relative, not absolute.
- **TG2 model** captures canonical QXP sites (77% of annotated); non-canonical sites are under-called.
- **Prior art:** the gluten-specific-TCR and HLA-DQ2.5:gluten-complex therapeutic space is active and partly patented (e.g. DONQ52-class presentation blockers, Chugai; NCT05425446). Any translation-bound claim needs a named-investigator prior-art search before it is made.
- **Natural next steps into the design streams:** (a) fold the top DQ2.5:deamidated-gluten complexes (AF2/Boltz) as inputs for a presentation-blocking binder, mirroring the DONQ52 modality; (b) cross-reference the predicted panel with public gluten-specific TCR repertoires; (c) proceed to the omics stream to find the villous-damage biology that a "protect the villi from accidental gluten" therapeutic would target.

---

### Deliverables index
| File | Content |
|---|---|
| celiac_gluten_antigen_set.csv / .fasta | 20-protein gluten antigen set |
| celiac_known_epitopes.csv | 31 Sollid-nomenclature ground-truth epitopes |
| celiac_deamidation_map.csv / .png | 464 predicted TG2 sites + density figure |
| celiac_epitope_panel.csv | 41,664 DQ2.5/DQ8/DQ2.2 predictions (native+deamidated) |
| celiac_epitope_burden.csv / celiac_deamidation_shift.csv / .png | per-family/HLA burden + paired shift |
| celiac_epitope_benchmark.csv / celiac_allele_specificity.csv / .png | recall, allele-specificity, deamidation-recovery |
| celiac_epitope_map.csv / .png | HLA-resolved panel (validated + novel), heatmap + load |