Proteins interact with several RNA types to facilitate a broad spectrum of cellular functions. However, the underlying interaction data are sparse, and existing methods for predicting RNA-binding residues (RBRs) in protein sequences are almost exclusively RNA type-agnostic, limiting their utility. To this end, we introduced RNAdetector, a sequence-based method that accurately predicts messenger RNA-, ribosomal RNA-, small nuclear RNA (snRNA)-, and transfer RNA-binding residues and type-agnostic RBRs. RNAdetector employs an innovative deep transformer network architecture and transfer learning, which together produce a large boost to predictive performance and minimize cross-predictions, defined as incorrectly predicting wrong types of RBRs. Moreover, our design has a low computational footprint and produces accurate predictions in about 9 s per protein, facilitating analysis of large collections of proteins. A comparative evaluation on a low-similarity test dataset showed that RNAdetector provided substantially more accurate RNA-type-specific predictions, a much shorter runtime, additionally covered snRNA, and significantly reduced cross-predictions compared to the only other RNA-type-specific predictor that is cross-prediction prone. Moreover, we empirically showed that RNAdetector's predictions of type-agnostic RBRs are modestly more accurate than those generated by several representative predictors of RNA-type-agnostic RBRs and RNA-binding proteins. RNAdetector is available as a convenient web server at http://biomine.cs.vcu.edu/servers/RNAdetector/.
Hu et al. (Thu,) studied this question.