Joint Prediction of Branchpoint and Splice Acceptor Sites in pre-mRNA Transcripts Using Sequence to Sequence Networks

Computational efforts to understand the process of pre-mRNA splicing have led to the development of various tools that can separately predict the locations of branchpoints and splice sites. Since experimental studies have shown that the majority of branchpoints are distributed just upstream of the 3' splice site, it should be possible to use this information to jointly predict the locations of the branchpoint and splice acceptor site. Here, we propose a deep neural network based sequence-to-sequence learning solution that can label each nucleotide of an input sequence as being a branchpoint or a splice site or neither. We demonstrate how mutation maps can be generated to interpret our model and use them to show that our network learns the interdependence between the branchpoint and the splice site. An analysis of these maps also reveals that our model is able to learn about other influential sequence features such as the polypyrimidine tract and alternative splice sites and make predictions of mutations in these regions.