MsipNet: a multi-scale representation learning framework for predicting protein-RNA interaction

Protein-RNA interactions (PRIs) are fundamental to post-transcriptional regulation, influencing key processes such as RNA splicing, stability, translation, and broader cellular functions. Understanding these interactions is essential for uncovering gene regulatory mechanisms and discerning how disease-associated mutations disrupt cellular homeostasis, thereby providing insights that can inform the development of novel therapeutic strategies. Consequently, accurate identification of PRIs represents a critical link between basic biological research and biomedical applications. To advance PRI prediction, we introduce MsipNet, a multi-scale representation learning framework that integrates global and local RNA sequence features with structural information via a multimodal learning strategy. MsipNet employs a hybrid architecture combining Long Short-Term Memory (LSTM) networks with U-shaped convolution-dilated convolution (UCDC) modules, enabling fine-grained feature refinement and improving prediction accuracy. This design facilitates the capture of intricate binding patterns while maintaining high computational efficiency. Experimental evaluations show that MsipNet consistently outperforms eight state-of-the-art (SOTA) methods across 42 RNA-binding proteins (RBPs) from six cell lines, demonstrating superior performance in predicting binding preferences. Furthermore, MsipNet reliably identifies biologically validated binding motifs and exhibits strong generalizability when applied to unseen data. Collectively, these findings position MsipNet as a robust and interpretable tool for PRI prediction and functional mutation prioritization, with broad potential for mechanistic studies and biomedical applications.