#37 · Structural bio & protein design

Structure-aware models and continual training for epitope and TCR-peptide binding

The common thread is learned models that map sequence or 3D structure onto immune recognition outcomes. EpiRanha couples ESM-2 residue embeddings with an E(n)-equivariant graph neural network over 3D structures to score and rank epitope similarity, and is reported to outperform TM-align. MsgaBpred predicts B-cell epitopes from sequence alone by combining AlphaFold3-predicted structures, multi-scale graph convolutions, attention, and ESM-C embeddings. ContinualTCR addresses shifting epitope data in TCR-peptide binding prediction by combining reservoir replay with EWC regularization, reducing forgetting 62.9% relative to fine-tuning. What is new is the pairing of structure-aware representation learning with continual training strategies for evolving epitope data.