PHbinder

PHbinder is a binary classifier that decides whether a peptide is an HLA class I epitope from its sequence alone, built by researchers in the Department of Computer Science at City University of Hong Kong and posted to bioRxiv in June 2025. It was released together with PSGM, a generative model that proposes which alleles an epitope is likely to bind; chained, the two implement a peptide → epitope determination → HLA-I allele list workflow.

The established peptide-MHC tools work in the opposite direction. Methods such as NetMHCpan and MHCflurry score binding affinity between a peptide and a specified allele drawn from a preset list, so they cannot answer the question a screening pipeline actually starts with — is this peptide an epitope at all? PHbinder is built for precisely that upstream filter: no allele is supplied as input, so candidate lists from mass spectrometry or in-silico digestion can be narrowed before any allele hypothesis is committed to.

The model adapts ESM-2 — the 150M-parameter esm2_t30_150M_UR50D checkpoint — with low-rank adapters rather than updating the full backbone, then routes its representations through parallel convolutional and Transformer branches. The work is a preprint and has not yet been peer reviewed.

#Key Features

• Allele-free epitope screening: Classification runs from peptide sequence alone, so peptides can be filtered without committing to an allele hypothesis or scanning a preset allele panel.
• Parameter-efficient adaptation: Low-rank adapters of rank 16 are applied to the attention query, key, value, and output projections of the frozen encoder, leaving the backbone weights untouched.
• Dual-branch feature fusion: A convolutional branch captures local motif structure while a Transformer branch captures global sequence context, and a cross multi-head attention module combines the two before classification.
• Open code, data, and weights: The implementation is MIT-licensed on GitHub, the labeled peptide splits are committed alongside it, and trained checkpoints are published on Hugging Face.

#Technical Details

Peptides up to 16 residues are embedded with the LoRA-adapted encoder, producing 640-dimensional token representations. Those feed two parallel branches — a 256-channel CNN with kernel size 3, and a 6-layer, 16-head Transformer — which a cross multi-head attention module fuses before a small fully connected head emits 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 the model reached 85.12% accuracy, which the authors report as exceeding their benchmark comparisons. Metric code in the repository computes accuracy, F1, recall, precision, MCC, and AUC over the held-out split.

#Applications

The model targets immunopeptidomics and epitope discovery — neoantigen screening for cancer immunotherapy, vaccine candidate selection, and characterization of pathogen-derived epitopes. Because it needs only a peptide sequence, it can triage the large candidate lists that mass spectrometry and in-silico digestion produce, cutting the volume that reaches allele-specific affinity prediction. Peptides that clear the filter are the intended input to PSGM, which returns a short allele panel to carry forward into NetMHCpan or MHCflurry rather than an exhaustive allele sweep.

#Impact

This is a recent preprint from a small academic group, and adoption is correspondingly early. Its contribution is to treat epitope determination as a standalone, allele-free step rather than a by-product of allele-specific affinity scoring — a reframing that matters most at the top of a screening funnel, where the allele context is not yet known. The limits are those of the evaluation: a single curated dataset, entirely computational validation, and no peer review to date. The MIT-licensed code, committed training data, and published checkpoints make the reported results straightforward to reproduce and extend.