MKGR Predicts Protein Interactions in Cold-Start Scenarios.

Predicting protein-protein interactions (PPIs) is a fundamental task in functional genomics, crucial for understanding disease mechanisms and developing new drugs. A particularly challenging scenario arises when attempting to predict interactions for "cold-start" proteins, which are those without any previously observed PPI edges during model training. In such cases, models relying solely on network topology often lack sufficient context. This research introduces MKGR, a multimodal representation learning framework specifically designed to address cold-start PPI prediction. MKGR integrates two primary sources of information. First, it uses a region-aware protein sequence encoding to extract contextual representations from structurally informed sequence regions. Second, it incorporates information from four protein-centered biomedical knowledge graphs, including protein-drug, protein-disease, protein-miRNA, and protein-lncRNA associations. Graph attention encoders are employed to learn modality-specific protein embeddings from these sparse biomedical associations. To enhance learning, a bridge reconstruction objective regularizes the graph learning process by recovering shared protein-entity associations. Furthermore, a pair-level gating module adaptively combines the evidence from both sequence and graph modalities for each candidate protein pair. Experiments conducted on two benchmark datasets, under both novel-old and novel-novel cold-start settings, demonstrated that MKGR consistently surpassed competitive sequence, network, and knowledge-graph baselines across various evaluation metrics, including ACC, F1, AUC, AUPR, and MCC.