HydRA: Deep-learning models for predicting RNA-binding capacity from protein interaction association context and protein sequence

RNA-binding proteins (RBPs) control RNA metabolism to orchestrate gene expression, and dysfunctional RBPs underlie many human diseases. Proteome-wide discovery efforts predict thousands of novel RBPs, many of which lack canonical RNA-binding domains. Here, we present a hybrid ensemble RBP classifier (HydRA) that leverages information from both intermolecular protein interactions and internal protein sequence patterns to predict RNA-binding capacity with unparalleled specificity and sensitivity using support vector machine, convolutional neural networks and transformer-based protein language models. HydRA enables Occlusion Mapping to robustly detect known RNA-binding domains and to predict hundreds of uncharacterized RNA-binding domains. Enhanced CLIP validation for a diverse collection of RBP candidates reveals genome-wide targets and confirms RNA-binding activity for HydRA-predicted domains. The HydRA computational framework accelerates construction of a comprehensive RBP catalogue and expands the set of known RNA-binding protein domains. Overall design: To characterize and validate the uncharacterized known RBPs (HSP90A, YWHAG, YWHAH, YWHAE, and YWHAZ) and candidate RBPs (INO80B, PIAS4, NR5A1, ACTN3, and MCCC1) predicted by HydRA, we V5-tagged and overexpressed them in HEK293 cell lines and implement eCLIP experiments. The same procedure was also applied on truncated proteins of INO80B, PIAS4, NR5A1, ACTN3, and MCCC1 where the HydRA predicted RNA-binding domains are deleted to validate the functions of these predicted RNA-binding domains.