PSGM

PSGM is a conditional generative adversarial network that, given a peptide, emits candidate HLA class I pseudo-sequences and maps them back to known alleles. It comes from the Department of Computer Science at City University of Hong Kong and was posted to bioRxiv in June 2025 alongside PHbinder, the allele-free epitope classifier that supplies its input.

The design choice that distinguishes it is treating allele identification as generation rather than retrieval. Established tools such as NetMHCpan and MHCflurry score a peptide against a specified allele drawn from a preset list, so producing a binding profile across the wider HLA-I allelic landscape means querying allele by allele. PSGM instead generates 34-residue pseudo-sequences conditioned on the peptide and resolves each one to a real allele by nearest-neighbor search. Because the output space is the pseudo-sequence rather than a fixed allele vocabulary, the formulation extends in principle to alleles outside the training set.

The peptide conditioning signal comes from ESM-2 — the 150M-parameter esm2_t30_150M_UR50D checkpoint — kept frozen and used purely as an embedding source. The work is a preprint and has not yet been peer reviewed.

#Key Features

• Generative allele identification: Rather than scanning a preset allele panel, the model produces pseudo-sequences for a given peptide, then resolves them to real alleles through Hamming-distance nearest-neighbor lookup against a 177-allele reference table.
• Frozen encoder as conditioning: The protein language model backbone is not fine tuned; its peptide embeddings are consumed directly as the conditioning signal for the generator.
• Adversarial training with a reconstruction anchor: A Transformer discriminator supplies the adversarial signal at a weight of 0.3, sitting alongside a reconstruction objective rather than replacing it.
• Shortlist output, not a single call: Generation returns a Top-50 allele list, sized to be carried forward into a conventional affinity predictor instead of being treated as a final answer.

#Technical Details

The generator is a 6-layer, 8-head Transformer decoder with 256-dimensional embeddings that autoregressively emits a 34-residue pseudo-sequence, conditioned on the frozen peptide embedding. A 3-layer Transformer discriminator provides the adversarial signal. Generation uses nucleus sampling.

Training data comprised roughly 342,000 peptide–allele pairs spanning 121 HLA-I alleles, deduplicated and split 80/10/10. Generated pseudo-sequences are mapped to alleles by Hamming distance against a 177-allele table, returning a Top-50 list; across evaluation peptides the model achieved 49.26% average coverage of the true alleles within that list. In an orthogonal check with MHCflurry, 63% of the highest-affinity binding partners found within those Top-50 lists were HLA-I alleles that lack experimental verification for the peptides in question.

#Applications

The intended position is immediately downstream of epitope triage in immunopeptidomics and epitope discovery workflows — neoantigen screening for cancer immunotherapy, vaccine candidate selection, and characterization of pathogen-derived epitopes. Peptides that clear PHbinder's filter arrive without an allele hypothesis attached, and the Top-50 output supplies one: a short panel to run through NetMHCpan or MHCflurry, replacing an exhaustive allele sweep. The orthogonal MHCflurry result also suggests a secondary use in hypothesis generation, surfacing peptide–allele pairs that have no experimental record.

#Impact

The methodological contribution is the reframing itself — casting HLA allele identification as conditional sequence generation over pseudo-sequences rather than retrieval over a fixed allele vocabulary. That generality comes with real limits, and the preprint is candid about them: 49.26% average Top-50 coverage means the correct allele is missed roughly half the time, training data covers 121 alleles, and evaluation is entirely computational with no wet-lab confirmation. Adoption is early, as expected for a recent preprint from a small academic group. The MIT-licensed code, committed training data, and published checkpoints make the reported numbers straightforward to reproduce.