EpitopeTransfer

Linear B-cell epitopes are the short peptide stretches an antibody actually binds, and predicting them is a routine first step in designing immunodiagnostics, subunit vaccines, and therapeutic antibodies. Almost all published predictors are generalists: they are trained on pooled epitope data from every organism with entries in public databases, which means their representations are dominated by a handful of intensively studied pathogens. That bias is invisible when a tool is benchmarked on the same pooled data, and painfully visible when it is applied to a neglected or newly emerging pathogen.

EpitopeTransfer, developed by researchers at the University of Bristol, the Universidade de Brasília, and the Universidade Federal de Minas Gerais, reframes the problem as domain adaptation guided by evolutionary relatedness. Rather than train one model for everything, it fine-tunes a protein language model on epitope data from a higher-level taxon — a genus, family, or order that contains the pathogen of interest — and then uses that taxon-adapted embedder to build a compact classifier for the specific lower-level target. The premise is that epitope-relevant sequence signal is shared along phylogeny, so data from relatives is more informative than an equal volume of data from distant organisms. The framework was first posted as a preprint in April 2025.

Unlike the structure-based Patch-Centric Conformational B-Cell Epitope Predictor or the Structure-Fused Immunogenic B-Epitope Predictor, EpitopeTransfer works from sequence alone and makes taxonomic scope, rather than structural context, the axis of specialization.

#Key Features

• Phylogeny-informed embedder fine-tuning: The protein language model is retrained on epitope data from a higher-level taxon before feature extraction, so its representation is tuned to the evolutionary neighborhood of the target.
• Whole-protein feature extraction: Features are computed by passing full antigen sequences through the tuned embedder and then selecting the labelled peptide positions, which preserves non-local context that peptide-only embedding discards.
• Lightweight downstream classifier: A random forest is fit on the extracted features for each target taxon, keeping per-taxon model development cheap and the resulting predictors easy to inspect.
• Nineteen ready-to-run pathogen models: The repository ships trained checkpoints covering viral, bacterial, and eukaryotic pathogens, so predictions for a covered taxon require no retraining.
• Controlled attribution of the gains: Two internal baselines isolate the contribution of fine-tuning itself from the contribution of choosing phylogenetically close data for it.

#Technical Details

Both ESM-1b and ESM-2 were tested as base embedders at the 650M-parameter scale; the analysis of variance found no significant effect of embedder choice, and ESM-2 was selected for the headline comparisons. Epitope data was assembled from IEDB exports with quality control and label-conflict resolution applied before splitting. Nineteen target taxa were evaluated, spanning viruses (Filoviridae, Orthopoxvirus, influenza A, Lentivirus, Mononegavirales and others), bacteria (Mycobacterium tuberculosis, Escherichia coli, Pseudomonas aeruginosa, Chlamydia trachomatis, Enterobacteriaceae and others), and eukaryotes (Plasmodium falciparum, Toxoplasma gondii, Schistosoma mansoni, Onchocerca volvulus).

Against an unmodified ESM embedder, the fine-tuning step raised mean AUC by 0.032 (p = 0.004); against an identical pipeline fine-tuned on deliberately distant taxa, the phylogeny-aware version gained 0.029 AUC (p = 0.0105) and 0.107 MCC (p = 0.0004), which is the paper's evidence that relatedness — not fine-tuning alone — drives the improvement. Compared with four external predictors, mean AUC gains were 0.090 over BepiPred 3.0, 0.120 over EpitopeVec, 0.130 over EpiDope, and 0.132 over epitope1D, all statistically significant. Individual models range widely: Filoviridae reaches AUC 0.96 and MCC 0.61, E. coli 0.91 and 0.44, C. trachomatis 0.83 and 0.57, and P. falciparum 0.79 and 0.41.

#Applications

The intended user is a group working on one pathogen rather than epitope prediction in general — screening an antigen for candidate immunodiagnostic peptides, prioritizing regions for a subunit vaccine, or narrowing a target list before peptide synthesis and ELISA. For the nineteen covered taxa the shipped checkpoints run directly on new sequences. For an uncovered pathogen, the same pipeline can be re-executed provided a phylogenetically adjacent higher-level taxon has enough curated epitope data, which the authors note is the binding constraint for fungal pathogens.

#Impact

EpitopeTransfer's contribution is less a single model than a demonstration that taxonomic structure is usable supervision for protein language model adaptation, tested with a factorial design and internal baselines that most tools in this area omit. Its results are uneven: twelve of nineteen datasets improved against every external baseline, nine by more than 0.05 AUC, but M. tuberculosis was worse than all four competitors at AUC 0.52 — a dataset where every method performed near chance — and B. pertussis and S. mansoni also landed near 0.53. Coverage is limited to the taxa modelled here, and the framework has not been extended to conformational or T-cell epitopes. The code and 38 per-taxon checkpoints are released under GPL-3.0 and archived on Zenodo; the preprint carries a CC BY-NC-ND license.