Improved modeling of RNA-binding protein motifs in an interpretable neural model of RNA splicing

Sequence-specific RNA-binding proteins (RBPs) play central roles in splicing decisions, but their exact binding locations and activities are difficult to predict. Here, we describe a modular splicing architecture that leverages in vitro-derived RNA affinity models for 79 human RBPs and the annotated human genome to produce improved models of RBP binding and activity. Binding and activity are modeled by separate Motif and Aggregator components that can be mixed and matched, enforcing sparsity to improve interpretability. Standard affinity models yielded reasonable predictions, but substantial improvements resulted from using a new Adjusted Motif (AM) architecture. While maintaining accurate modeling of in vitro binding, training these AMs on the splicing task yielded improved predictions of binding sites in vivo and of splicing activity, using independent crosslinking and massively parallel splicing reporter assay data. The modular structure of our model enables improved generalizability to other species (insects, plants) and to exons of different evolutionary ages.