Abstract

The mode of action for many known functional non-coding RNAs involves RNA-RNA interactions. Identifying interactions between novel ncRNAs and its targets is thus a promising way of assigning potential functional roles to non-coding RNAs. Whereas the experimental determination of RNA-RNA interactions is possible, it can be greatly accelerated and informed by computational predictions. Only a few existing tools for predicting intermolecular basepairs utilize evolutionary information. We provide a solution to this problem via INTERRACT, which is a general method that can be combined with any existing tool, using machine learning to improve energy-based predictions with evolutionary information. We demonstrate improvement to the results of four existing prediction algorithms using INTERRACT, along with the first comparative evaluation of these state-of-the-art algorithms. We show that incorporating evolutionary information via INTERRACT is superior to existing tools that utilize conservation. We confirm that the usage of a support vector machine in INTERRACT allows for speed and scalability, assigns reliability scores to predictions, and allows the analysis of interaction features importance. We evaluate our method on a comprehensive and diverse data set of the two different biological classes of RNA-RNA interactions (sRNA-mRNA and snoRNA-rRNA) comprising 97 pairs of full-length transcripts, experimentally validated interacting basepairs, and corresponding multiple sequence alignments.