Base-resolution binding profile prediction of proteins on RNAs with deep learning.

RNA-binding proteins play crucial roles in various RNA-associated biological processes, which are closely linked to cellular function and disease. Based on CLIP-seq data, the existing deep learning methods are developed to predict protein-RNA interactions. However, CLIP-seq relies on gene expression, which varies significantly across cells. Existing methods are typically trained on peak-associated binding sites and implicitly defined non-binding sites, without considering the cell-specific expression profiles. Given the dynamic nature of protein-RNA interactions, these methods struggle to accurately predict the binding nucleotides and strength of proteins on RNAs across cell lines. Therefore, this study proposes a novel deep learning-based method, iDeepB, designed to predict the proteins binding profile on RNAs at base resolution by integrating cell-line-specific gene expression profiles. iDeepB first constructs expression-aware benchmark datasets based on cell-specific RNA-seq and eCLIP-seq data, which is used to train a hybrid deep network with multi-head attention, enabling the prediction of protein binding profiles, analysis of binding motif syntax composition, and quantification of functional effects of genome mutations related to human diseases. Comprehensive evaluation on the newly developed benchmark datasets demonstrates that iDeepB outperforms existing methods in predicting protein binding profile on RNAs.