Inferring RNA Sequence Specificities from Protein Sequences to Characterize Post-transcriptional Regulation in Eukaryotes

Most RNA-binding proteins (RBPs) find their targets through unique binding preferences towards specific RNA sequence or RNA sequence-structure patterns, called specificities. Recently, we used RNAcompete, an in vitro binding assay, to measure the RNA sequence specificities of 174 RBPs. Combined with previous measurements, we established the largest collection of RBP sequence specificities to date, containing 381 RBPs from 33 eukaryotes, from protists to humans. RNA sequence specificities are nearly always conserved if RBP sequences share more than 70% identity in their binding regions (70%-rule). However, only half of the RBPs sharing between 30% and 70% sequence identity recognize similar RNA, limiting the confidence in predictions from sequence identity. To increase the number of RBPs with a confidently inferred specificity from our measured data, I developed a computational method, called joint protein-ligand embedding (JPLE), which jointly embeds amino acid 5-mers and RNA sequence specificities into a joint latent space. The joint latent embedding of an RBP sequence can be approximated from protein sequence features alone and enables reconstructions of RNA sequence specificity, prediction of RNA binding similarity, and identification of important binding regions in the protein sequence. JPLE doubles the number of RBPs with confidently inferred RNA sequence specificities compared to predictions with the 70%-rule. I embed RBPs from 690 eukaryotes in JPLE’s latent space, confidently reconstruct specificities for 29,000 RBPs, and cluster RNA sequence specificity groups (RSSGs). I use these RSSGs to estimate the number of distinct RSSGs for all eukaryotic RBPs with RRM and KH domains. I confirm that RRM containing RBPs recognize highly diverse sequences, covering every possible 7-mer, while KH containing RBPs recognize only 40% of all possible 7-mer sequences. Moreover, I determine the last common ancestor of the RBPs in an RSSG and derive the evolutionary rate with which new RNA sequence specificities evolve in different eukaryotic clades. Lastly, to demonstrate the utility of this new resource to investigate post-transcriptional regulation, I combine 101 inferred RNA sequence specificities from Arabidopsis thaliana with RNA-seq from 69 plant tissues and I identify RBPs that regulate mRNA stability through interactions with the 3’UTR.Ph.D