Patch-Centric Conformational B-Cell Epitope Predictor

Identifying which residues on an antigen surface will be recognized by antibodies — the conformational B-cell epitope problem — is a central bottleneck in antibody discovery and vaccine design. Experimental epitope mapping is slow and expensive, and most computational tools are trained on generic protein-protein interfaces that transfer poorly to the distinctive geometry and chemistry of antibody-antigen recognition. This work, from Yinghao Wu's group at the Albert Einstein College of Medicine and posted to bioRxiv in October 2025, introduces a patch-centric framework that predicts epitopes directly on antigen structures.

The core idea is to decompose an antigen surface into "patches," each defined as a triad of neighboring residues that captures the smallest local unit encoding both shape and chemistry. Rather than engineering hand-crafted descriptors, the primary model represents each patch with a protein language model: it averages ESM-2 residue embeddings over the triad and scores the result with a compact multilayer perceptron. A convolutional baseline that consumes a hand-crafted 15x20 feature matrix (amino-acid identity, secondary structure, solvent accessibility, and shape index) is provided for comparison.

By reframing epitope prediction as patch classification and leaning on pretrained protein representations, the method converts language-model features into interpretable epitope-likelihood maps over an antigen, offering a practical aid for prioritizing regions for antibody engineering and vaccine design.

#Key Features

• Patch-centric representation: Antigen surfaces are decomposed into residue triads, the smallest local unit that jointly encodes shape and chemistry, giving fine-grained, interpretable predictions across the whole surface.
• Protein language model features: Each patch is embedded by averaging ESM-2 residue vectors, avoiding hand-crafted descriptors and letting a small MLP do the classification.
• Ensemble aggregation: Patch-level scores are aggregated to per-residue epitope likelihoods using an ensemble over all cross-validation folds, improving robustness.
• Interpretable epitope maps: Outputs are likelihood maps over the antigen that localize binding regions, supporting antigen prioritization and rational design.

#Technical Details

The predictor was trained with five-fold cross-validation on 1,151 antibody-antigen complexes drawn from AbDb. At the patch level the ESM-2 plus MLP model markedly outperforms the convolutional baseline, reaching F1 around 0.986 and ROC-AUC around 0.998. Aggregating patch scores to residues via the five-fold ensemble yields robust residue-wise performance (ROC-AUC 0.689 plus or minus 0.072) versus the CNN (0.548 plus or minus 0.018). Benchmarked against widely used sequence- and structure-based epitope tools on AbDb, the language-model approach achieves the best summary metrics (ROC-AUC 0.67, PR-AUC 0.56) while retaining full coverage of all antigens. On five external complexes unseen during development, it generalizes with ROC-AUC 0.663 and qualitatively localizes binding regions. The work is a preprint and has not yet completed peer review; it is released under a CC BY-NC-ND license.

#Applications

The model is aimed at immunologists and antibody engineers who need to prioritize candidate epitopes on an antigen before committing to experimental mapping. Because it operates directly on antigen structures and produces per-residue likelihood maps, it can guide the selection of surface regions for antibody generation campaigns, inform immunogen design in vaccine development, and help interpret why particular surface patches are antigenic.

#Impact

Conformational epitope prediction has lagged behind linear-epitope and general interface prediction because antibody-antigen interfaces are governed by subtle, non-contiguous surface features. By pairing a patch-level structural decomposition with pretrained ESM-2 embeddings, this work demonstrates that protein language model representations transfer usefully to the antibody-antigen recognition problem and can outperform both hand-crafted convolutional baselines and established epitope tools on held-out antigens. As a preprint without an accompanying public code release, its near-term influence will depend on independent reproduction, but it adds to a growing body of evidence that PLM features are a strong foundation for immunoinformatics.