Pan-allelic transformer for peptide-MHC class I presentation that deconvolves multi-allelic ligandome data inside the model itself.
No providers recorded yet. Browse all providers
Mass spectrometry of an HLA eluate returns a list of peptides and nothing else. The donor carried three to six MHC class I allotypes, all contributing to that list, and the experiment never records which allotype held which peptide. Most presentation models handle this by running a separate deconvolution heuristic first and then training on its assignments as though they were measurements. The second gap is worse: only presented peptides are reported, so negatives must be invented by sampling the proteome, and a model that trains hard on those labels learns to reject peptides that are genuinely presented but were never sequenced.
HLApollo, from Genentech, closes both gaps inside the model. All of a sample's MHC-I pseudosequences are pushed through the network in parallel alongside the peptide, and only the highest-scoring peptide-allotype pairing contributes to the loss, so allele assignment is re-estimated after every batch instead of being fixed beforehand. Against invented negatives it uses negative-set switching: 500 disjoint candidate negative sets are prepared and a new one is drawn each epoch, so no falsely presumed negative is seen often enough to be memorized. Because the MHC enters as sequence rather than as an allele identifier, one trained checkpoint scores arbitrary peptide-allele pairs, including allotypes with no training data.
The work appeared as a bioRxiv preprint in December 2022 and in Nature Communications in December 2024. It is the class I member of a Genentech line of peptide-MHC models that also includes Graph-pMHC for class II and the later, broader Pep2Vec.
The concatenated N-flank, peptide and C-flank pass through one transformer encoder layer while the MHC pseudosequence passes through its own, with weights shared across a sample's allotypes; the joined pair then runs through four more encoder layers, all 400-dimensional with 16 heads. A beginning-of-sequence token pools the pair into a fixed-size vector regardless of peptide length and is the only input to the classification head. Training used mhcDB, a purpose-built SQLite store of 953,693 unique peptide-genotype tuples covering 305,646 peptides and 171 HLA-I alleles across 347 genotypes, drawn from 22 published studies and IEDB, with matched bulk RNA-seq for 43.8% of it. The flagship ensembles 10 models, each trained for 115 epochs in 7.7 hours on eight V100 GPUs. On a held-out test set at a 1:99 positive-to-negative ratio the ensemble reached 78.7% average precision, 12.65% above the next-best model evaluated; it gained 4.1% AP over NetMHCpan-4.1-EL on an immunogenicity set and recovered 7.9% more peptides than SHERPA-EL on held-out tissue studies. A feedforward network keeping the flanks, negative-set switching, in-model deconvolution and ensembling still lost 14.5% AP.
The target application is individualized neoantigen therapy, where a patient's mutations must be narrowed to the few peptides their HLA genotype will display. HLApollo ranks those candidates, and the untrained-allele accuracy estimate tells a clinician whether the prediction can be trusted for a given genotype rather than leaving rare alleles silently unsupported. The same coverage analysis tells groups assembling ligandome data which alleles would most repay collecting next.
HLApollo shows that the two preprocessing heuristics the field had accepted — deconvolving multi-allelic data upfront and freezing a sampled negative set — are both avoidable, and that removing them accounts for much of the remaining headroom on presentation prediction. Treating allele coverage as a quantity to be predicted and reported, rather than assumed, addresses an equity problem in neoantigen vaccine development, where training ligandomes skew toward alleles common in European ancestries. Release is limited: Genentech distributes a compiled inference executable under an academic license confined to non-commercial research at a single site, and training code and weights are not published. The benchmark datasets and source data are separately available under CC BY 4.0.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.