# Index-Case Calibration Against Elimination-Diet Ground Truth

*Post-hoc validation of the personalized epitope panel and the food-trigger diagnostic concept against the patient's blinded, clinically-established food triggers.*

## The ground truth (revealed after the computational work)

The index patient completed a formal elimination diet with endoscopic follow-up:
- **Triggers (drive disease): MILK and SOY.**
- **Tolerated: wheat and the rest of the Top-8** — the patient stays in histologic remission while eating wheat.
- **Family history of celiac** (sibling + uncle), but the patient's own **tTG-IgA screening is negative** (no celiac disease).

This was withheld during panel development to keep the prediction unbiased. The reveal is the real test.

## Result: 2 of 2 triggers rank top-2 after correction; 1 instructive false positive

| Food | Original panel rank | **Corrected panel rank** | Ground truth | Call |
|---|---|---|---|---|
| **Milk** (complete casein family) | #3 (181) | **#1 (720)** | **TRIGGER** | ✓ true positive |
| **Soy** | #1 (526) | **#2 (526)** | **TRIGGER** | ✓ true positive |
| Wheat | #2 (204) | #3 (204) | tolerated | ✗ **false positive** |
| Fish / Egg / Peanut / Tree nut / Shellfish | low | low | tolerated | ✓ true negatives |

**Both clinically-confirmed triggers are the two highest-ranked foods** once the milk pool is complete. The five tolerated non-wheat foods all rank low (true negatives). The one discordance — wheat — is the most scientifically informative result of the whole exercise.

## Why milk was initially under-ranked — and the fix that came from Hill et al.

The original panel represented milk with only **α-S1-casein + β-lactoglobulin** and **omitted β-casein (CSN2)**. Reading Dilollo, Spergel, Hill et al. 2025 (*JACI*, doi:10.1016/j.jaci.2025.01.008) — who identified the first validated EoE food-specific TCR, **eoeTCR-4 recognizing β-casein aa59–78 restricted to HLA-DRB1\*07:01** — exposed that gap. Adding the complete casein family (β-, αS2-, κ-casein) + α-lactalbumin added **539 strong binders**, moving milk from #3 to #1.

**Independent predictor validation:** my mhcnuggets model, run on β-casein, calls the **exact Hill epitope window (aa59–78) as a strong DR7 binder** — 5 strong 15-mers spanning that region, best IC50 = 50 nM, core positioned where the tetramer-validated TCR engages. The predictor was correct; only the panel's protein coverage was incomplete. This is both a fix and an external validation.

## Why wheat is a false positive — the central lesson

Wheat gliadin ranked #2 by *presentation* (204 strong binders, 88 on DQ2.2). The patient **tolerates wheat**. The prediction was not wrong about what it measured — DQ2.2 genuinely can present gluten peptides (this patient carries the celiac-risk heterodimer) — but **presentation capacity is not pathology**:

- **Binding prediction answers:** *can this patient's MHC-II display peptides from this food?* (For wheat/DQ2.2: yes.)
- **Disease requires additionally:** a peripheral **pathogenic effector-Th2 (peTH2)** clone specific for those peptides, actually expanded and driving tissue inflammation. For wheat in this patient: **absent** — consistent with negative tTG-IgA despite celiac family history. The patient has the presentation *hardware* for gluten but has not mounted a pathogenic T-cell *response* to it.

This is exactly why the diagnostic's call must come from the **functional Th2 readout**, not the binding panel. The panel is the pre-test prior (which pools to build); the functional assay (Th2-AIM / tetramer, as in Hill's platform) is the evidence that separates a true trigger (milk, soy → peTH2 present) from a merely-presentable food (wheat → no peTH2). Had the assay been run functionally, wheat would be expected to score negative despite its high presentation rank.

## Implications for the diagnostic design

1. **Panel completeness is critical.** Represent each food by its full clinically-relevant allergen set (all caseins + whey for milk, not a token protein), or the pre-test ranking misorders true triggers. The β-casein omission would have de-prioritized the patient's #1 trigger.
2. **Never call a trigger from binding alone.** The wheat false positive is the proof: a presentation-only test would wrongly flag wheat and send the patient into an unnecessary gluten elimination. The functional Th2/peTH2 step is not optional — it is the specificity-defining component.
3. **The HLA-celiac-risk finding must be reported as risk-not-disease.** This patient's DQ2.2 + celiac family history + negative tTG-IgA is precisely the "presentation without pathology" pattern; the report should say *gluten-presentation capacity present, no evidence of gluten-driven disease* — and monitor, not eliminate.
4. **Confusion-matrix readout:** after correction, sensitivity 2/2 for triggers in the top-2; specificity 5/6 among tolerated foods (wheat the lone false positive), and that FP is explained mechanistically rather than being noise.

## Caveats
- Numbers are strong-binder counts from mhcnuggets class-II predictions; the DQ2.2 prediction uses mhcnuggets' closest trained model (DQB1\*03:01), so DQ2.2 magnitudes are approximate.
- "Ground truth" here is the clinical elimination-diet result as reported; no functional T-cell assay was run in this exercise, so the wheat-negative expectation is a mechanistic prediction, not a measurement.
- This remains a research/hackathon calibration, not a validated diagnostic. See the main design doc disclaimer.