Reproduce “Improving Paratope and Epitope Prediction by Multi-Modal Contrastive Learning and Interaction Informativeness Estimation” on OpenResearch

About this paper

Zhiwei Wang, Yongkang Wang, Wen Zhang

Researchers at Huazhong Agricultural University developed MIPE, a method for paratope and epitope prediction that integrates multi-modal (sequence and structure) data with a contrastive learning strategy and explicitly models antibody-antigen interaction patterns. MIPE achieved an AUPR of 0.741 for single paratope prediction and 0.504 for single epitope prediction, outperforming existing baselines and demonstrating consistent performance even with computationally predicted AlphaFold2 structures.

Key insights

• Combining sequence and structural information through multi-modal contrastive learning significantly enhances the discriminative power of residue representations for binding sites.
• Explicitly modeling antibody-antigen interaction patterns via an Interaction Informativeness Estimation module is crucial for improving the accuracy of paratope and epitope prediction.
• The proposed framework maintains competitive performance even when utilizing computationally predicted protein structures (e.g., from AlphaFold2), expanding its practical applicability beyond experimentally determined structures.

Summary by alphaXiv.