MFC-PPI: protein-protein interaction prediction with multimodal feature fusion and contrastive learning

Protein-protein interactions (PPIs) are of critical importance in numerous biological processes and disease mechanisms, and the accurate prediction of PPIs is helpful in the comprehension of complex biological systems. In this paper, MFC-PPI, a PPI prediction model based on multimodal feature fusion and contrastive learning, is proposed. The sequential features, structural features, and PPI network features of proteins are extracted and combined for prediction. The contrastive learning is used to compare the subtle difference between the sequential features and structural features. In addition, the feature enhancement module is designed for feature fusion. The comparative experiments on SHS27k and SHS148k datasets demonstrates the excellent performance of MFC-PPI over other state-of-art methods under three partitioning strategies, Random, BFS, and DFS.