SCMPPI: Supervised contrastive multimodal framework for predicting protein-protein interactions

Accurate prediction of protein-protein interactions (PPIs) is crucial for understanding cellular functions and disease mechanisms, yet it is hindered by the cost of experimental methods and limitations in existing computational approaches, particularly in learning robust multimodal representations and false-negative suppression. To address these challenges, we propose SCMPPI—a novel supervised contrastive multimodal framework. SCMPPI effectively leverages a rich set of sequence-based embeddings (AAC, DPC, and ESMC-CKSAAP) with network topology (Node2Vec embeddings) for comprehensive protein representation. Furthermore, it incorporates an enhanced supervised contrastive learning strategy featuring a negative sample filtering mechanism. This strategy regularizes the latent space by enforcing semantic consistency among interacting pairs, thereby significantly improving the discriminative power of the model. This comprehensive approach enables SCMPPI to achieve superior prediction performance. Extensive experiments on five benchmark datasets demonstrate its state-of-the-art performance and excellent cross-species generalization, consistently outperforming baselines. Successful applications in CD9 and Wnt networks establish SCMPPI as a promising tool for multimodal biological data analysis. The code is accessible at https://github.com/xshengrui/SCMPPI. © 2026 Elsevier B.V.