WALLE++

A B-cell epitope is the patch of an antigen's surface that an antibody actually contacts, and predicting it is central to vaccine design, diagnostics and therapeutic antibody engineering. The task resists sequence-based methods for a structural reason: epitopes are conformational, defined by residues distant in sequence but adjacent in three dimensions, and the same antigen presents different epitopes to different antibodies. Classical predictors ignore that second point, scoring antigen residues on general immunogenicity without reference to the antibody in question, which caps how useful their output is for a specific engineering problem.

WALLE++ addresses the antibody-aware formulation. Developed by Mansoor Ahmed, Sarwan Ali, Avais Jan, Murray Patterson at Georgia State University and Imdad Ullah Khan at the Lahore University of Management Sciences, it extends WALLE, the reference method released with the AsEP antibody-specific epitope prediction benchmark. Both represent an antibody-antigen complex as a bipartite graph and recast epitope prediction as link prediction between antibody and antigen residues; WALLE++ systematically re-examines what representation should sit on the graph nodes.

That question turns out to matter. The paper's finding is that the choice of protein language model embedding materially changes prediction quality, and that mixing model families — an antibody-specific language model for the antibody chains and a general protein language model for the antigen — beats using either family throughout. The work appeared as a preprint in February 2025 and was published in the ICCABS 2025 proceedings.

#Key Features

• Antibody-conditioned prediction: The epitope is predicted for a given antibody rather than as a generic antigen property, which makes the output usable when engineering one specific binder.
• Bipartite graph formulation: Antibody residues and antigen residues form two node sets, and epitope identification becomes link prediction across the partition, so the model reasons about contacts rather than labeling residues independently.
• Mixed language model embeddings: ProtTrans and ESM-2 supply antigen representations while the antibody-specific AntiBERTy supplies antibody representations, exploiting the fact that antibody sequence statistics differ sharply from the general proteome.
• Frozen encoders, light trainable head: The language models act as fixed feature extractors and only a graph convolutional network is trained, keeping the model small enough for a modest benchmark without overfitting.
• Embedding ablation across six representations: ESM-2, ESM-IF1, ProtBert, AntiBERTy, BLOSUM62 and one-hot encoding are compared head to head, making the representation choice an empirical result rather than an assumption.

#Technical Details

Each antibody-antigen complex from the AsEP benchmark — 1,723 filtered structural complexes, the largest curated collection of its kind — is converted into a bipartite graph whose nodes carry per-residue embeddings and whose edges encode residue contacts. A graph convolutional network propagates information across this structure and scores candidate antibody-antigen residue links; residues on the antigen side of predicted links constitute the epitope.

The best configuration pairs AntiBERTy embeddings for antibody chains with general protein language model embeddings for the antigen, reaching an AUC-ROC of 0.650, precision 0.281, recall 0.457, F1 0.348 and Matthews correlation coefficient 0.263 on the AsEP evaluation splits. That improves on the WALLE baseline, which recorded an MCC of 0.210, and on EpiPred, ESMBind, MaSIF-site and EpiGraph. These are absolute numbers worth reading carefully: antibody-specific epitope prediction remains an unsolved problem, and a precision below 0.3 means most predicted epitope residues are wrong even in the best configuration. The released repository contains training and inference code, configuration files and baseline implementations, but the checkpoint directory ships empty, so reproducing the reported numbers requires retraining from scratch.

#Applications

The realistic use is candidate narrowing rather than definitive annotation. A group with an antibody sequence and an antigen structure can rank antigen surface residues by predicted contact probability, focusing alanine scanning, escape mapping or structural work on a smaller region, and the same ranking suggests which antigen positions are vulnerable to escape mutation. Because the framework is modular in its embeddings, it also functions as a testbed: a group with a new antibody language model can substitute it into the node featurization and measure the effect on a standard benchmark.

#Impact

WALLE++'s value is more diagnostic than performative. By holding the graph architecture fixed and varying only the representation, it isolates how much of epitope-prediction performance comes from the protein language model rather than the network, and establishes that domain-matched embeddings — antibody models for antibodies — carry real signal general models miss. That is a reusable finding for the many interface-prediction methods now built on frozen language model features. The result also documents how far the field has to go: an AUC-ROC of 0.650 is an improvement on a hard benchmark, not a solution to it, and generalization to unseen epitope groups remains the open problem AsEP was built to expose. The repository carries no LICENSE file, leaving reuse terms undefined despite the code being public.