Pep2Vec

A mass-spectrometry ligandome is not a clean list of presented peptides. Some entries come from an allele the engineered cell line was supposed to have lost; some are MHC class I peptides that survived a class II pulldown; some are protein termini, cheap to produce and therefore over-observed. A model with enough capacity fits all of them and then scores similar peptides confidently — which, in a personalized cancer vaccine with a handful of slots, means spending one on a peptide the patient's immune system never sees, with nothing in a single presentation score to audit against.

Pep2Vec, from Genentech, answers that by making the model's internal representations the deliverable rather than a by-product. Each input — the peptide, its flanking residues, its source protein, and the presenting MHC allele — is mapped into its own named vector space, and the spaces are combined by fixed algebraic operations rather than a learned mixing layer: peptide, flank and source protein vectors are added elementwise into an Integrated Peptide Vector, that vector is multiplied elementwise by the MHC Vector, and a small multilayer perceptron emits the presentation logit. Because the combination is elementwise, each space remains independently readable, and the Peptide Vector can be clustered, browsed and filtered on its own.

That turns a scoring model into a data-auditing instrument. Graph-based clustering of the Peptide Vector space, a k-nearest-neighbour mutual-information novelty score, and a linear support vector machine trained on the frozen Peptide Vectors together identify four contaminant classes — misannotated alleles, class I peptides inside class II ligandomes, protein-terminus peptides, and low-complexity peptides — covering 77,475 peptides, 5.0% of roughly 1.5 million. The October 2024 preprint inherits the held-out source-protein split the same group developed for Graph-pMHC.

#Key Features

• Named vector spaces per input: Peptide, Flank, Source Protein and MHC each get a dedicated representation, so any one can be inspected without disentangling it from the rest.
• Binding-core inductive bias: The peptide transformer's first two layers use a 9-residue sliding attention mask — the length of an MHC class II binding core — and the final layer passes only the BOS row. After calibration, the largest attention value marks the binding core start position, learned without explicit supervision.
• Pan-allelic by construction: The MHC representation is a PCA over ESM-2 embeddings of all 15,175 allotypes deposited in IMGT plus one linear layer, so allele coverage does not depend on which alleles carry training data.
• Source protein context: A residue-averaged ESM-2 embedding of the source protein organizes the Source Protein Vector space by gene expression and by protein-protein interaction neighbourhoods, neither of which is an input.
• A contaminant taxonomy, not just a score: The four classes are enumerated, quantified and published as peptide lists, and a variant trained on the filtered data separates positives from negatives more cleanly.

#Technical Details

The peptide submodule runs three transformer encoder layers at a model dimension of 512; the flanking-residue submodule is identical with one layer. Source protein and MHC representations both start from the final layer of esm2_t36_3B_UR50D, residue-averaged to 2,560 dimensions. Training ran 60 epochs at batch size 3,072, about four hours per model on one A100. Ten models from different seeds were distilled into one at double the dimension, giving the 1,024-dimensional latent vector the released inference binary emits alongside the logit and predicted binding core.

Training data comprises 738,427 class I peptide:allele pairs (254,256 unique peptides, 171 alleles, lengths 8–12) and 801,102 class II peptide:genotype pairs (380,087 unique peptides, 83 alpha and beta chains, lengths 9–25), against source-gene decoy pools of roughly 46 and 129 million. Test proteins are disjoint from training proteins, at 1:99 and 1:9 positive:negative ratios approximating natural presentation rates, holding 84,916 class I and 88,398 class II positives. Four external immunogenicity sets, including a clinical anti-drug-antibody set of 43 positives, are scored by the same checkpoint without refitting; on that set Pep2Vec reaches 56% positive predictive value at 80% sensitivity. Joint class I and II training, the authors note, yields limited benefit to predictive performance and is done for interpretability.

#Applications

The immediate uses are neoantigen ranking for individualized cancer vaccines and deimmunization of large-molecule drugs, where a class II presentation score predicts anti-drug antibody risk. Groups assembling immunopeptidomics corpora can also screen them for the four contaminant classes before training, and the interactive dashboard shipped with the code lets an analyst navigate the peptide latent space to find submotifs, allele misannotations and artifacts in their own data.

#Impact

Pep2Vec reframes interpretability in peptide-MHC modelling as a data-quality tool rather than a post-hoc explanation, and the contaminant lists it produced are reusable independently of the model — a framing that follows FDA guidance on explainability for clinical AI. Access is constrained: the release is a compiled inference binary under an academic license limited to internal teaching and non-commercial academic research, with source code and weights withheld. The work remains a preprint.