mhcmatch — Peptide–MHC presentation & cross-reactivity

Peptide–MHC presentation, cross-reactivity, and motif tools — the applied peptide–MHC layer on top of the seqtree fuzzy-search substrate. mhcmatch productionizes the reference seqtree.pmhc methodology (anchor-masked TCR-facing homology, presentation-aware E-values, allele guessing) and adds a pseudosequence-based cross-allele diffusion model that rescues rare alleles by borrowing from groove-similar frequent ones.

The mathematical/statistical theory is in appendix/mhcmatch.tex; the development plan is in ROADMAP.md.

What it does (v0)

1. MHC restriction & presentation — rank presenting alleles for a peptide (single / set / all, human & mouse), flag non-binders, and scan a whole protein for presented peptides.
2. Large-scale similarity search — find similar peptides across big sets / proteomes, either by same-MHC binding (presentation signature) or similar TCR recognition (anchor-masked, TCR-facing); neoantigen molecular mimicry with per-allele E-values.
3. Anchor / TCR-facing split — decompose a peptide into anchor and TCR-facing parts (X masks).
4. Near-exact source lookup — find the self peptide a neoantigen derives from + its parent protein / mutated position, against a reference proteome.
5. Motif logos — per-allele information-content logos with length distributions.
6. Pseudosequence diffusion — allele similarity, clustering, and kernel-shrinkage pooling over 34-mer groove pseudosequences (rare-allele rescue).
7. Quantitative affinity (IC50 nM) — a pan-allele Potts / direct-coupling model (peptide×pocket couplings, fit on measured IEDB IC50) predicts nM affinity and the neoantigen-fitness differentials — Łuksza amplitude A = Kd_WT/Kd_MT and the differential agretopicity index — for MHC-I and MHC-II, human and mouse. Optional structure-based MJ ΔΔG via the [structure] extra (tcren).

Install

bash setup.sh            # repo-local .venv + editable install (uses sibling ../seqtree if present)
bash setup.sh --tests    # + pytest
bash setup.sh --logo     # + logomaker/matplotlib for rendering logos

Quickstart

import mhcmatch

# build from the isalgo/pmhc_data table (full or shortlist tier)
store = mhcmatch.Store.from_pmhc("pmhc_full.tsv.gz", species="human")

store.restriction("NLVPMVATV")                  # ranked presenting alleles + binder flags
store.is_binder("NLVPMVATV", "HLA-A*02:01")
store.scan_protein(my_protein, cls="mhc1")       # presented peptides in a protein
store.decompose("NLVPMVATV", cls="mhc1")         # (tcr_facing, presentation) with X masks

# similarity at scale
mhcmatch.search.search("NLVPMVATV", big_peptide_set, mode="tcr")   # TCR-facing homologs
mhcmatch.search.find_mimics("EAAGIGILTV", self_set, bacterial_sets={...})

# near-exact source of a neoantigen
pm = mhcmatch.Proteome.from_hf("human")          # auto-fetched from HF (or .from_fasta(<your FASTA>))
pm.find_source("NLVPMVATV", max_subs=1)

# pseudosequence allele similarity + rare-allele diffusion
ps = mhcmatch.Pseudoseq("mhc1")
ps.neighbors("HLA-A*02:01", candidates=store.alleles("mhc1"))

# diffusion-powered forward scorer (rescues rare alleles by borrowing from groove-neighbours)
am = store.anchor_model("mhc1")          # learned anchor weights + bounded-prior shrinkage
am.score("NLVPMVATV", "HLA-A*02:01")     # anchor log-odds; am.score(..., raw=True) disables borrowing

# footprint (which core positions) and background (the log-odds null) tune the model to the question:
store.anchor_model("mhc1", footprint="adaptive")             # anchors for rare alleles, full core otherwise
store.anchor_model("mhc1", background="proteome")            # presentation null (is it presented at all?)
store.anchor_model("mhc1", background="ligand")              # specificity null (which allele? — default)

# calibrated, cross-allele-comparable output (NetMHCpan %Rank_EL analogue + P(present) + band)
for r in store.restriction("NLVPMVATV", cls="mhc1", calibrated=True):
    print(r.allele, r.rank, r.p_present, r.band)             # e.g. HLA-A*02:01  1.6  0.98  weak

mhcmatch.logo.motif(store, "HLA-A*02:01", "mhc1")

# quantitative affinity + neoantigen-fitness differentials (Potts model, vendored weights)
aff = store.affinity_model("mhc1")
aff.predict_ic50("NLVPMVATV", "HLA-A*02:01")            # -> ~64 nM
aff.amplitude("NLVPMVATL", "NLVPMVATV", "HLA-A*02:01")  # Kd_WT/Kd_MT (Łuksza eq. 9) -> ~2.05
aff.dai("NLVPMVATL", "NLVPMVATV", "HLA-A*02:01")        # differential agretopicity (log10 ratio)
store.affinity_model("mhc2").predict_ic50("PKYVKQNTLKLAT", "HLA-DRB1*15:01")   # MHC-II, core auto-located

Command line

mhcmatch decompose NLVPMVATV                                  # anchor / TCR-facing split (no data)
set -x MHCMATCH_PMHC /path/to/pmhc_data                       # or pass --pmhc to each command
mhcmatch restriction NLVPMVATV --allele 'A*02:01' --diffuse   # allele name auto-resolved; rare-aware
mhcmatch restriction NLVPMVATV --calibrated                   # + %rank, P(present), binding band
mhcmatch scan my_protein.fasta --correction bh                # presented windows, BH-FDR controlled
mhcmatch source MKTAYIAKW --proteome human                    # HF name auto-fetched (or a FASTA path)
mhcmatch logo 'HLA-A*02:01'
mhcmatch affinity NLVPMVATV --allele 'A*02:01' --wt NLVPMVATL   # IC50 nM + amplitude A=Kd_WT/Kd_MT + DAI

Data

• Reference ligands: the public HF dataset isalgo/pmhc_data (full / shortlist tiers). Store.from_pmhc() auto-fetches pmhc/pmhc_<tier>.tsv.gz on first use (cached by huggingface_hub) — no manual download, which is what lets the container/nextflow deploy bootstrap with no pre-staged data. Override with a local copy via Store.from_pmhc(path=...) or $MHCMATCH_PMHC.
• Pseudosequences: 34-mer groove pseudosequences vendored in src/mhcmatch/data/ (see its PROVENANCE.md).
• Reference proteomes: the human (UP000005640) and mouse (UP000000589) UniProt proteomes — plus pathogen proteomes for mimicry — live in the same HF dataset. Proteome.from_hf("human") / mhcmatch source --proteome human auto-fetch them (cached), or mhcmatch bootstrap --proteome human,mouse to pre-fetch. Pass your own FASTA to Proteome.from_fasta to override.

Benchmark vs NetMHCpan

Benchmarks live in a separate repo. bench/ moved to 2026-mhcmatch-benchmark — the head-to-head harness, the bench/results/*.md tables referenced throughout, and their provenance notes. Paths like bench/results/... below resolve there, not here.

A reproducible head-to-head against NetMHCpan-4.2b and NetMHCIIpan-4.3i lives in bench/compare/ (results in bench/results/compare_*.md, provenance and caveats in bench/compare/SOURCES.md). It compares the two tools on the same per-(peptide, allele) task, stratified by allele rarity, with AUROC / AUPRC / PPV@k, bootstrap CIs and paired significance. Headline results (shortlist tier, human, seed 0):

• Allele-specificity (which allele presents a peptide — the restriction problem): mhcmatch beats NetMHCpan on MHC-I medium and frequent alleles on AUROC, AUPRC and PPV@k (all p < 0.001; e.g. frequent AUPRC 0.850 vs 0.769). Rare MHC-I is a wash (+0.008 AUROC, p = 0.41).
• Presented-vs-random screening (background="proteome"): mhcmatch beats NetMHCpan on MHC-I frequent alleles (AUROC p < 0.001, AUPRC 0.881 vs 0.846, p = 0.001). Medium and rare are a wash — the deltas sit inside the CI. This task is much easier for both tools (every AUROC ≥ 0.97).
• MHC-II: NetMHCIIpan leads on medium and frequent alleles (p < 0.001); rare is a wash (Δ AUROC +0.013 specificity / −0.015 screening, both inside the CI). Read these with compare/SOURCES.md in hand: NetMHCIIpan trained on essentially all public IEDB eluted-ligand data, so the in-corpus medium/frequent strata are contaminated in its favour and the rare/zero-shot axis is the fair one.
• Speed: mhcmatch scores ~68× faster (pure Python, ~195k peptide-allele scores/s).

python bench/compare/run_compare.py --cls mhc1 --decoy-mode hard   --background ligand    # specificity
python bench/compare/run_compare.py --cls mhc1 --decoy-mode random --background proteome  # screening

Quantitative affinity (Potts head)

The affinity head is benchmarked head-to-head against NetMHCpan-4.2 −BA / NetMHCIIpan-4.3 −BA on held-out measured IEDB IC50 (bench/affinity/; provenance in bench/affinity/SOURCES.md). Metric: median per-allele Spearman ρ against measured log-IC50, and AUROC at the 500 nM binder threshold, on the same held-out (peptide, allele) pairs. Honest numbers (per-allele held-out split, seed 0):

NetMHCpan/IIpan lead on this eval, but the gap is inflated by train/test overlap we cannot undo: both tools trained on much of IEDB, so the held-out pairs are in-sample for them and out-of-sample for mhcmatch. On truly unseen alleles (leave-20-alleles-out, zero training rows for the allele) the Potts model still generalizes — MHC-I common orphan ρ ≈ 0.57 — because its peptide×pocket couplings interpolate across groove-similar alleles. The affinity head is a compact, dependency-light linear model (numpy-only dot product at predict time, ~µs/peptide) and gives the WT-vs-mutant ratio (amplitude / DAI) that percentile ranks cannot express.

python bench/affinity/train_potts.py --cls mhc1 --alpha 40                 # MHC-I head-to-head
python bench/affinity/train_potts.py --cls mhc2 --species all --alpha 40   # MHC-II, human + mouse

Status

Beta (v0.4). Affinity (IC50 nM) + neoantigen amplitude/DAI shipped for MHC-I/II, human & mouse; optional structure-based MJ ΔΔG via the [structure] extra. See ROADMAP.md for what's next (order-k Markov / covariance null, a learned reranker for rare-allele screening, full-tier + temporal cluster sweeps, and the stability/cleavage/immunogenicity predictors).