Abstract
BACKGROUND: RNA-binding proteins (RBPs) play diverse roles in eukaryotic RNA processing. Despite their pervasive functions in coding and noncoding RNA biogenesis and regulation, elucidating the sequence specificities that define protein-RNA interactions remains a major challenge. Recently, CLIP-seq (Cross-linking immunoprecipitation followed by high-throughput sequencing) has been successfully implemented to study the transcriptome-wide binding patterns of SRSF1, PTBP1, NOVA and fox2 proteins. These studies either adopted traditional methods like Multiple EM for Motif Elicitation (MEME) to discover the sequence consensus of RBP's binding sites or used Z-score statistics to search for the overrepresented nucleotides of a certain size. We argue that most of these methods are not well-suited for RNA motif identification, as they are unable to incorporate the RNA structural context of protein-RNA interactions, which may affect to binding specificity. Here, we describe a novel model-based approach--RNAMotifModeler to identify the consensus of protein-RNA binding regions by integrating sequence features and RNA secondary structures. RESULTS: As an example, we implemented RNAMotifModeler on SRSF1 (SF2/ASF) CLIP-seq data. The sequence-structural consensus we identified is a purine-rich octamer 'AGAAGAAG' in a highly single-stranded RNA context. The unpaired probabilities, the probabilities of not forming pairs, are significantly higher than negative controls and the flanking sequence surrounding the binding site, indicating that SRSF1 proteins tend to bind on single-stranded RNA. Further statistical evaluations revealed that the second and fifth bases of SRSF1octamer motif have much stronger sequence specificities, but weaker single-strandedness, while the third, fourth, sixth and seventh bases are far more likely to be single-stranded, but have more degenerate sequence specificities. Therefore, we hypothesize that nucleotide specificity and secondary structure play complementary roles during binding site recognition by SRSF1. CONCLUSION: In this study, we presented a computational model to predict the sequence consensus and optimal RNA secondary structure for protein-RNA binding regions. The successful implementation on SRSF1 CLIP-seq data demonstrates great potential to improve our understanding on the binding specificity of RNA binding proteins.