EGRL Boosts RNA-Protein Interaction Prediction in Cold-Start Scenarios.

RNA-Protein Interactions (RPIs) are fundamental to cellular functions, but their experimental detection is costly and time-consuming. Deep Learning (DL) methods, particularly Graph Neural Networks (GNNs), offer efficient computational alternatives. However, existing GNN-based approaches often struggle with data sparsity and generalizing to "cold-start" scenarios involving molecules not seen during training. Researchers propose EGRL (Edge Generation-guided Relation-aware Learning) to address these limitations. EGRL incorporates several key components: implicit meta-path learning to capture relational semantics without predefined paths, a multi-relation-aware attention mechanism for adaptive fusion of interaction patterns, and a graph generator that predicts potential "soft" edges. This generator is crucial for supporting cold-start nodes. The framework is jointly trained with a primary task loss and an auxiliary generator loss, culminating in a multi-feature fusion predictor for interaction scoring. Comprehensive evaluations show EGRL achieves competitive overall performance and significantly superior generalization in cold-start settings, demonstrating substantial improvements in AUROC and AUPR for unknown molecules.