PatchEpi: Patch-Aware Equivariant Learning Improves Structure-Based Epitope Prediction

Accurate prediction of B-cell epitopes is essential for antibody design and vaccine development, yet remains fundamentally challenging. A major but often overlooked limitation of existing predictors is their implicit assumption that epitope identity can be decomposed into independent residue-level signals. In contrast, antibody recognition is governed by spatially contiguous surface patches whose functional identity emerges from coherent three-dimensional geometry rather than isolated residues.

Here, we reformulate structure-based B-cell epitope prediction as a surface-patch learning problem and introduce PatchEpi, a patch-aware geometric deep learning framework for antigen-intrinsic epitope modeling. It employs a patch-aware attention encoder and boundary-contrastive objectives to align the learning patch signal with the biological reality of antibody binding. PatchEpi integrates patch embeddings, fine-tuned ESM2 representations, and an equivariant message-passing network to model the 3D topology of antigen surfaces. To enable rigorous evaluation, we created a homology leakage-controlled split by reclustering the ANABAG dataset. Across multiple benchmarks, PatchEpi consistently outperforms residue-centric and graph-based state-of-the-art methods. Structural analyses further show that the model learns coherent surface patches that closely match experimentally resolved antibody footprints. These results demonstrate that accurate epitope prediction benefits primarily from reformulating the task around surface-patch learning. Patch-aware learning provides a principled and biologically grounded pathway toward more reliable structure-based epitope detection.

The code and related resources are available at GitHub (https://github.com/wsicheng739/PatchEpi) and Zenodo (https://doi.org/10.5281/zenodo.20366444).

Supplementary data are available at Bioinformatics online.