FedEpitope: Parameter-Efficient Federated Protein Language Models for Linear B-Cell Epitope Prediction

FedEpitope adapts ESM-2 with LoRA for linear B-cell epitope prediction, communicating only ~1.11 MB of trainable state while FedAvg improves mean AP, F1, and MCC.

• Why it matters: Linear B-cell epitope data are often distributed across institutions that cannot pool records, limiting model development for vaccine and diagnostic screening.
• What they did: They built a federated framework using LoRA on ESM-2, then simulated FedAvg and FedProx on homology-aware splits across five biologically heterogeneous, organism-partitioned clients.
• The result: FedAvg outperforms all single-client models in mean AP, F1, and MCC, and follow-up tests assess personalization, differential privacy, client heterogeneity, LoRA rank, and generalization to BCPreds and NetBCE.

AI summary of the abstract · may contain errors · not medical advice · read the paper.