PHbinder & PSGM

PHbinder and PSGM are two models released together as a cascaded pipeline for peptide–HLA class I immunology, developed by researchers in the Department of Computer Science at City University of Hong Kong and posted to bioRxiv in June 2025. The pair addresses a structural limitation of the established peptide-MHC tools: methods such as NetMHCpan and MHCflurry score binding affinity between a peptide and a specified allele drawn from a preset list. They therefore cannot answer the question a screening pipeline actually starts with — is this peptide an epitope at all? — and they cannot return a binding profile across the wider HLA-I allelic landscape without being queried allele by allele.

The two models split that problem in half. PHbinder is a binary classifier that takes a peptide sequence alone and predicts whether it is an HLA-I epitope, with no allele supplied as input. PSGM then takes peptides that clear this filter and, rather than scoring them against a fixed allele table, generates candidate 34-residue HLA-I pseudo-sequences conditioned on the peptide, mapping each generated sequence back to known alleles by nearest-neighbor search. Chained together they implement a peptide → epitope determination → HLA-I allele list workflow, positioning the pair as an upstream triage stage that narrows the search space before conventional affinity predictors are run.

Both models build on ESM-2 as a sequence encoder, using the 150M-parameter esm2_t30_150M_UR50D checkpoint. PHbinder adapts it with LoRA; PSGM keeps it frozen and uses its embeddings as the conditioning signal for a generator. The work is a preprint and has not yet been peer reviewed.

#Key Features

• Allele-free epitope screening: PHbinder classifies a peptide as epitope or non-epitope from sequence alone, so candidate peptides can be filtered before any allele hypothesis is committed to.
• Generative allele identification: PSGM produces HLA-I pseudo-sequences for a given peptide instead of scanning a preset allele panel, then resolves them to real alleles through Hamming-distance nearest-neighbor lookup against a 177-allele reference table.
• Parameter-efficient adaptation: PHbinder fine-tunes ESM-2 with low-rank adapters (rank 16, applied to the attention query, key, value, and output projections) rather than updating the full backbone.
• Dual-branch feature fusion: Parallel CNN and Transformer branches capture local motif structure and global sequence context respectively, combined through a cross multi-head attention layer.
• Open code and weights: The implementation is MIT-licensed on GitHub, and trained checkpoints for both models are published on Hugging Face.

#Technical Details

PHbinder embeds peptides (maximum length 16) with the LoRA-adapted ESM-2 encoder, then routes the 640-dimensional token representations through two branches: a 256-channel CNN with kernel size 3, and a 6-layer, 16-head Transformer. A cross multi-head attention module fuses the two representations before a small fully connected head produces the binary label. Training used a curated set of labeled peptides split into roughly 643,000 training, 91,000 validation, and 185,000 test sequences. On the epitope prediction task PHbinder reached 85.12% accuracy, which the authors report as exceeding their benchmark comparisons.

PSGM is a conditional generative adversarial network. Its generator is a 6-layer, 8-head Transformer decoder with 256-dimensional embeddings that autoregressively emits a 34-residue pseudo-sequence conditioned on frozen ESM-2 peptide embeddings; a 3-layer Transformer discriminator supplies the adversarial signal, weighted at 0.3 alongside the reconstruction objective. Generation uses nucleus sampling. Training data comprised roughly 342,000 peptide–allele pairs spanning 121 HLA-I alleles. Generated pseudo-sequences are mapped to alleles by Hamming distance, returning a Top-50 list; across evaluation peptides PSGM 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 PSGM's Top-50 lists were HLA-I alleles that lack experimental verification for those peptides.

#Applications

The pair targets immunopeptidomics and epitope discovery workflows — neoantigen screening for cancer immunotherapy, vaccine candidate selection, and characterization of pathogen-derived epitopes. Because PHbinder needs only a peptide sequence, it can triage large candidate lists from mass spectrometry or in-silico digestion before allele-specific affinity prediction is attempted, and PSGM's Top-50 output gives a short allele panel to carry forward into NetMHCpan or MHCflurry rather than an exhaustive allele sweep.

#Impact

PHbinder and PSGM are a recent preprint from a small academic group, and adoption is correspondingly early. The contribution of interest is methodological: recasting HLA allele identification as a conditional sequence generation problem over pseudo-sequences, rather than as retrieval over a fixed allele vocabulary, which in principle extends to alleles outside the training set. That generality comes with real limits — PSGM's 49.26% average Top-50 coverage means the correct allele is missed roughly half the time, its training data covers 121 alleles, and evaluation is entirely computational. The released MIT-licensed code, training data, and checkpoints make the reported results straightforward to reproduce and extend.