Interpretable deep learning strategy for accurate hierarchical RNA-protein interaction prediction

RNA-protein complexes are essential for many biological processes, including fundamental translation and gene regulation. However, computational prediction of RNA-protein structures with unknown templates and limited multiple sequence alignments remains challenging, often resulting in discrepancies between the predicted interfaces and those observed experimentally. Accurate prediction of RNA-binding preferences can shed light on the overall framework for precise complex prediction. Here, we introduce ZHMolSite, a general framework that uses an interpretable learning strategy to identify binding nucleotides on RNA molecules hierarchically. ZHMolSite outperforms state-of-the-art methods, including AlphaFold3, showing absolute improvements of 4.5%-13.2% and 6.5%-6.7% in identifying interface RNA-binding sites on two independent test sets, respectively. Without relying on structural templates or generating MSAs, ZHMolSite also provides predicted interacting-nucleotide information that effectively guides RNA-protein docking and improves the recovery of native-like interfaces. We believe the approach behind ZHMolSite will contribute to solving the ongoing challenge of accurately predicting RNA-protein structures.