Antigen epitope pipeline

Antigen-epitope / pMHC pipeline

A reusable, disease-agnostic pipeline for turning a set of candidate antigens into an HLA-resolved epitope map: antigen set -> UniProt sequences -> peptide windows -> MHC-II/I binding prediction -> per-HLA burden -> personalized panel. Built for EoE (food allergens x HLA-DR/DQ) but the antigen set and HLA panel are parameters — swap them for celiac (gluten x DQ2.5/DQ8), T1D (islet autoantigens), IBD (microbial/self-antigens), or any class-II-restricted disease.

kernel.py ships the pure-compute helpers (fetch_uniprot_fasta, sliding_windows, summarize_binders, unique_core_count). The MHC-prediction step is env-specific and described below.

The one guardrail that matters: presentation != pathology

A binding prediction says a peptide CAN be presented on a given MHC. It does NOT say that peptide drives disease. Disease requires an expanded, pathogenic effector-T-cell clone specific for that peptide. Binding/burden is only a PRIOR — which pools to build and test first. The specificity-defining evidence is always a FUNCTIONAL readout (antigen-specific Th2/peTh2 by AIM assay or tetramer on patient cells). This was learned concretely in EoE: a patient's DQ2.2 presented wheat gluten at high burden, yet the patient tolerated wheat (negative functional response, negative tTG-IgA). Never name a trigger, or claim pathogenicity, from binding prediction alone. Report HLA-presentation findings as risk/capacity, not disease.

Workflow

1. Define the antigen set and HLA panel (the two swappable parameters)

• Antigens: list UniProt accessions for the disease's candidate proteins. Represent each source by its FULL clinically-relevant family, not a token protein (EoE milk lesson: omitting beta-casein CSN2/P02666 mis-ranked the patient's #1 trigger; complete casein family fixed it).
• HLA panel: population-representative class-II alleles (DRB1*01:01, 03:01, 04:01, 07:01, 15:01; DQ heterodimers) or a patient's own typing for a personalized panel. For an inclusive platform, treat burden as a mechanistic covariate, never an enrollment gate — pair with a functional (HLA-agnostic) AIM readout.

2. Fetch sequences

# python tool
recs = {acc: fetch_uniprot_fasta(acc) for acc in accessions}   # (header, seq)

3. Generate peptide windows

windows = {acc: sliding_windows(seq, k=15) for acc, (h, seq) in recs.items()}  # class-II 15-mers

4. Predict MHC binding (env-specific — pick one)

• netMHCIIpan via IEDB (http://tools-cluster-interface.iedb.org/tools_api/mhcii/, POST, method=netmhciipan_ba for IC50 nM). Run as a background cell over the antigen x allele grid; retry timeouts with a longer per-request timeout. This is the recommended primary predictor.
• mhcnuggets (local, in a TF env, e.g. mhc): class-II on DRB1 alleles, class-I via mhcflurry. Good for extending to many alleles offline; cross-validates with netMHCIIpan at Spearman rho ~= 0.79 on shared DR alleles.
• Class-I: mhcflurry (pan-allele, 9-mers) — use DEFAULT_CLASSI_K=9. Collect predictions into a flat list of dicts: {allele, antigen, peptide, core, ic50}.

5. Aggregate burden (pure helpers)

summary = summarize_binders(rows)         # per_allele / per_antigen / total strong (IC50<500 nM)
n_cores = unique_core_count(rows)         # distinct strong cores — NOT the per-cell sum

Gotcha: the true number of unique binding cores dedups cores shared across alleles; summing per-allele-per-antigen cells double-counts (the EoE "864 vs 470" bug). Use unique_core_count.

6. Rank, but confirm functionally

Rank antigens by strong-binder burden to prioritize pools — then STOP and route to the functional assay (see the guardrail). The ranking is a hypothesis, not a diagnosis.

Deliverables to produce

• <disease>_epitope_panel.csv — all predictions (allele x antigen x peptide x IC50)
• <disease>_epitope_load.csv — per-antigen / per-allele strong-binder burden
• an epitope heatmap + per-antigen load bar (use figure-style)
• a personalized panel if patient HLA is supplied

Provenance / honesty rules

• All predictions are computational PRIORS; state this in every output.
• DQ predictions are less benchmarked than DR — flag magnitudes as approximate.
• For any translation-bound claim, do a named-investigator prior-art search (the EoE food-specific-TCR paradigm is patented — Dilollo/Spergel/Hill 2025, doi:10.1016/j.jaci.2025.01.008).