Graph-pMHC

The MHC class II binding groove is open at both ends, so a 20-residue peptide can sit in it through any of a dozen 9-residue registers, and a mass-spectrometry ligandome records only that the peptide was eluted — never which register did the binding. Every class II presentation model must resolve that ambiguity, and the multilayer perceptrons and sequence encoders that dominate the field absorb it into their weights alongside everything else they learn about which residues touch.

Graph-pMHC, from Genentech, makes both halves explicit. Peptide residues, source-protein flanking residues and an MHC pseudosequence become nodes in a graph whose peptide–MHC edges are read off a structure rather than assumed: predictions from AlphaFold-Multimer across HLA-DR, -DP and -DQ complexes are collapsed into one canonical residue adjacency matrix per gene, so only physically plausible residue pairs exchange information during message passing. The register is handled by enumeration — one graph per candidate binding-core start position, and for multi-allotypic samples one per candidate allotype. All are scored inside the forward pass and the highest-scoring combination supplies the loss, so the alignment is learned by backpropagation rather than imposed by a separate aligner.

The model is the class II member of a Genentech peptide-MHC line whose later entry, Pep2Vec, spans both classes and inherits the source-protein split introduced here.

#Key Features

• Structure-derived adjacency, one matrix per gene: Contacts conserved across three peptides for each of 115 allotypes define a canonical adjacency matrix for HLA-DR, -DP and -DQ, so no per-complex structure prediction is needed at inference.
• Graph enumeration for the binding core: Scoring every candidate register also yields a binding-core assignment as a by-product of prediction.
• Processing signals in the graph: N- and C-terminal flanking residues from the source protein enter through their own edge type, so the model sees the cleavage preferences that determine which peptides are produced at all.
• Pan-allelic and pan-allotypic: The MHC is represented by the pseudosequence of its contact residues, so an allotype absent from the training ligandomes is scored without refitting.
• A clinical antibody immunogenicity benchmark: 109 monoclonal antibodies with anti-drug antibody rates from published trials, paired with a risk rule counting presentable, non-germline binding cores across eight common DR alleles.

#Technical Details

Residues are tokenized into 64-dimensional learned embeddings plus a learned positional embedding, passed through two graph attention layers at 128 dimensions, read out by an AttentiveFP-style attentive GRU, and mapped to a presentation logit by a linear head. Edge features are a 3-dimensional one-hot over intermolecular, intramolecular and flank edges; the adjacency matrices are fixed, never learned. Training ran 30 epochs and 412,331 steps at batch size 64, under 24 hours on a single V100, and the released checkpoint is under 6 MB.

Training aggregates 527,302 peptide:genotype pairs from nine published ligandome studies — 250,643 unique peptides, 408,111 multi-allotypic and 119,191 single-allotypic pairs. The test/train split combines 9mer-overlap control with gene-ontology awareness, holding 9mer overlap to 0.35% while keeping cellular-component, molecular-function and biological-process categories evenly distributed between the halves. On the resulting test set of 57,519 positives, average precision reaches 81.14% against 60.97% for NetMHCIIpan-4.0 and 61.24% for MixMHCIIPred-1.2, with 81.7% ± 0.4% across three protein splits. Ablating the adjacency is what costs most: an empty matrix drops to 66.77%, a fully connected peptide–MHC matrix to 69.00%, and an anchor-residues-only matrix to 79.22%.

#Applications

The authors build the model around two uses: neoantigen selection for CD4+-directed cancer vaccines, and deimmunization of large-molecule drugs, where class II presentation drives the anti-drug antibody responses that compromise the safety of most biotherapeutics. For an antibody, the pipeline enumerates every 12–19mer it can yield, drops binding cores common in human repertoires, counts what survives across eight DR alleles, and returns that count as a risk score an engineer can act on by editing the offending positions. Scanning a million random 15mers also recovers biology: a lysine and proline enrichment just outside the binding core, at positions measured experimentally as protease cleavage signals.

#Impact

Graph-pMHC established that structural inductive bias, supplied cheaply as a fixed contact graph, matters more on this task than model capacity, and its gene-ontology-aware split raises the bar for how peptide-MHC models are evaluated. The authors are candid about the limits. Presentation is not immunogenicity: on an external CD4+ neoantigen set, average precision falls to 23.8%, close to NetMHCIIpan-4.0's 22.5%, and on the antibody set the margin over NetMHCIIpan-4.0 is 0.5% (P = 0.16), which they attribute to not knowing which allele drove each patient's response. The PDB complexes used to check the AlphaFold-derived adjacencies are likely inside AlphaFold's own training data, making that agreement optimistic. Code and the trained checkpoint are public under a Genentech non-commercial license and the full ligandome is on Zenodo under CC BY 4.0, but the repository supports inference only.