TransHLA

Designing a vaccine or a personalised cancer immunotherapy starts with a shortlist of peptides that the immune system might actually see. Getting there normally means running allele-specific binding predictors — NetMHCpan, MHCflurry, MixMHCpred — across every HLA allele a target population carries, which is expensive, requires knowing the alleles in advance, and answers a slightly different question: whether a peptide binds a given molecule, not whether it is ever presented at all. TransHLA, from Shuaicheng Li's group at City University of Hong Kong, reframes the task as a property of the peptide alone. Given a sequence, it predicts whether that peptide is presented as an epitope, with no allele supplied.

That framing makes it a filter rather than a replacement. Ranking a proteome or a tumour's mutated peptidome against hundreds of alleles is combinatorially painful; one allele-agnostic pass that discards peptides no HLA molecule is likely to present shrinks the candidate set before the expensive per-allele step runs. Two models handle the two presentation pathways, because class I and class II epitopes differ in length and in binding-groove geometry: TransHLA_I covers 8–14 residue peptides, TransHLA_II the longer 13–21 residue range.

The architecture is a probe on a frozen protein language model rather than a model trained from scratch. Peptides are embedded with ESM-2, whose predicted contact maps supply a structural view alongside the sequence representation, and a hybrid transformer-plus-residual-CNN head reads both. The work was published in GigaScience in 2025. It sits alongside catalog entries attacking neighbouring problems by different means, including mhcSFM, EpitopeTransfer, and the Structure-Fused Immunogenic B-Epitope Predictor.

#Key Features

• Allele-agnostic screening: The model takes only a peptide sequence, so it applies when HLA typing is unavailable or when the goal is a broadly reactive peptide rather than one matched to a specific allele.
• Dual sequence and structure input: ESM-2 supplies both a per-residue sequence embedding and a predicted residue-residue contact map, and the ablation shows the sequence embedding is by far the larger contributor.
• Separate class I and class II heads: Two checkpoints handle the distinct length distributions of the two pathways instead of forcing one padded representation to serve both.
• Ready for direct inference: Both checkpoints load through the standard Hugging Face AutoModel interface and return an epitope probability plus a peptide representation.
• Neoepitope evaluation: Beyond held-out IEDB data, the model is tested against TCR-validated neoepitopes and peptides from CEDAR, VDJdb, ImmuneCode, and dbPepNeo2.0.

#Technical Details

Training data was curated from IEDB, with redundancy removed by CD-HIT at 0.8 identity, yielding balanced sets of 459,442 positive and 459,442 negative class I peptides and 312,245 of each for class II, split 7:1:2. The data spans 258 distinct HLA class I alleles and 227 class II alleles. Peptides are padded to fixed length (16 for class I, 23 for class II), embedded by ESM-2 into 1,280-dimensional per-residue vectors, and passed through a multi-head attention module in parallel with a convolutional branch of 256-channel blocks reading the contact map.

On the IEDB test split, TransHLA reaches 84.72% accuracy, 84.59 F1, and 91.95 AUC for class I, and 79.94% accuracy, 81.07 F1, and 88.14 AUC for class II — improvements of 0.97 and 2.53 accuracy points over the strongest sequence-classification baselines, TextCNN, TextRCNN, DPCNN, and RNN-ATTs. The ablation is informative about where the signal lives: replacing the ESM-2 sequence embedding with a randomly initialised one drops class I accuracy to 73.37% and class II to 67.52%, while removing the contact-map branch costs about one point. On the external neoepitope sets, TransHLA attains higher specificity and precision than the allele-specific tools it is compared against, which achieve higher recall by calling far more peptides positive.

#Applications

The natural place for TransHLA is the front of a vaccine or neoantigen pipeline: score every candidate peptide from a pathogen proteome or a patient's mutation calls, discard those unlikely to be presented, and pass the survivors to allele-specific binding and immunogenicity tools. Because it needs no HLA typing, it suits population-scale vaccine design where target alleles are heterogeneous, and triage where typing data is not yet available. The peptide representations it emits can also feed downstream immunogenicity or TCR-recognition models.

#Impact

TransHLA's contribution is the decomposition — separating "is this peptide presentable at all" from "which allele presents it" — and it has found real uptake, with the class I checkpoint drawing heavy traffic on the Hugging Face Hub. The limits are worth stating: the model predicts presentation, not immunogenicity, and it inherits the allele and pathogen biases of IEDB, where a minority of alleles dominate the deposited data. Class II performance trails class I, reflecting the harder problem of an open-ended binding groove. The code is MIT-licensed, and the two checkpoints are published without a stated license.