Title: Improving Protein-Protein Interaction Prediction via Hierarchical Motif-Driven Multimodal Embeddings

Original: arXiv:2606.02629v1 Announcement Type: Cross Abstract: Protein-protein interactions (PPIs) play a crucial role in numerous biological functions. Nevertheless, current methods for predicting PPIs face two significant drawbacks: they neglect the hierarchical structure of proteins, especially meso-scale motifs that are vital for regulating these interactions, and they do not adequately combine sequence, structural, and functional data. To overcome these challenges, we introduce MMM-PPI, a Hierarchical Motif-based Multi-Modal protein Encoder designed for PPI prediction. This approach generates PPI embeddings through a bottom-up, multi-modal process spanning three distinct scales. At the micro-level, it encodes features from three residue modalities. At the meso-level, a new multimodal motif encoder groups residues into spatially aware motif embeddings. At the macro-level, a multimodal protein encoder synthesizes these motifs into protein embeddings by simultaneously assessing motif significance and inter-modal correlations. The pre-trained encoder is ready for immediate use in large-scale PPI prediction tasks. Comprehensive evaluations across various PPI datasets demonstrate that MMM-PPI surpasses current state-of-the-art multi-label PPI prediction models, especially in difficult data splits and scenarios with scarce data. The source code is available at https://github.com/yzf-code/MMM-PPI.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC