NetMHCpan-4.1 4.1
The reference pan-allele predictor of MHC class I antigen presentation: one small neural network covers more than 11,000 MHC molecules because the groove itself is part of the input — a 34-residue pseudosequence next to the peptide.
At a glance
What it does
Takes a peptide (8–14 aa) and an HLA allele and answers whether that pair will appear on the cell surface. Output is a %Rank — the peptide's position against a background of random natural peptides for that specific allele (0.5% strong binder, 2% weak) — plus the predicted 9-mer binding core and its offset.
Notable uses in oncology
The default first filter in neoantigen pipelines for individualised mRNA cancer vaccines: it narrows thousands of mutated peptides to the few dozen worth testing on the patient's own cells.
Tasks, data types and cancers
Architecture
Pseudosequence
- 34 positions chosen from crystal structures: within 4 Å of the peptide AND polymorphic across alleles. An allele stops being a name and becomes a 34-letter word, which is what makes prediction possible for alleles with zero measurements.
Variable length
- Everything is reduced to a 9-mer core by searching over insertions and deletions and keeping the best-scoring alignment; the model reports which core it chose.
Two heads, one shared layer
- Affinity measurements and eluted ligands each have their own output neuron but share the hidden layer, so the wide allele coverage of in-vitro data and the true presentation signal of mass spectrometry train one representation.
MOTIF DECONVOLUTION (NNAlign_MA).
- After 20 warm-up iterations on single-allele data, each peptide from a multi-allele sample is assigned to the best-scoring allele among the ones that sample actually carries, with score standardisation so a 'generous' allele cannot take everything.
Training data
13,245,212 training data points — but only about 850,000 are laboratory measurements (binding affinities plus MS-eluted ligands). The rest are random natural peptides from UniProt assumed to be negatives, roughly 94% of the set: mass spectrometry never reports true negatives, so they have to be invented.
Linked datasets
- training IEDB — Immune Epitope Database Open download — binding-affinity measurements and eluted-ligand deposits
- training Mono-allelic HLA class I peptidome (Sarkizova / Abelin) Open download — mono-allelic MS data give unambiguous allele labels
- training HLA Ligand Atlas Open download — multi-allelic tissue peptidomes, deconvolved during training
- pretraining IPD-IMGT/HLA Database Open download — allele sequences from which the 34-residue pseudosequences are cut
- benchmark IEDB automated benchmark (MHC class I) Open download
Evaluation
| Benchmark / dataset | Metric | Value | External validation | Source |
|---|---|---|---|---|
| MS-eluted ligand prediction vs NetMHCpan-4.0 | PPV / AUC | clearly improved (paper) | yes | Source |
| true T-cell epitopes vs NetMHCpan-4.0 | AUC | comparable — consistent gain only for HLA-B and HLA-C | yes | Source |
| vaccinia virus epitopes in mice (220 peptides tested experimentally) | epitopes recovered in the top 0.04% of predictions | over half of the most immunogenic epitopes; 90% required going to 1.3–1.5% of the list | yes | Source |
| IEDB automated benchmark, eight years | ranking among top methods | leading methods statistically indistinguishable | yes | Source |
How to run
Download requires an academic licence form; the binary is not on PyPI or Hugging Face. Read %Rank, not the raw score — and remember %Rank is a rank, not a probability: 0.1% is not ten times more likely to be presented than 1%.
Regulatory status and intended use
Regulatory status is quoted from the source linked above and can change. Research-use-only models must not be used for clinical decisions.
Limitations and bias
- Answers presentation, not immunogenicity: a well-presented peptide may still meet no matching T-cell receptor.
- All negative training examples are assumed, never observed.
- Inherits mass-spectrometry bias: cysteine-containing peptides are 5–10× under-detected, and solvent conditions alone shifted detected ligands by more than twofold for HLA-A*02 while dropping HLA-A*30 by 25%.
- Expression is not an input: the model does not know whether the source gene is transcribed in that tumour.
- Rare alleles (especially HLA-C) are predicted by analogy, and the antibody used to collect data itself prefers some genes.
Sources
This page is educational — it is not medical advice and does not replace consultation with an oncologist. Diagnostic and treatment decisions are made solely by specialist physicians.