Overview of branchpoints and their genome-wide prediction using LaBranchoR. (A) Branchpoints play a key role in 3′SS recognition and are essential to the mechanism of splicing. (B) RNA sequencing reads that span a 5′SS-branchpoint junction implicate branchpoints. The first part of these reads (yellow) align upstream of a branchpoint and the second part of these reads (red) align downstream from a 5′SS. In this way, the downstream end of the first part of the read marks the branchpoint. (C) Cartoon of information flow in a bidirectional LSTM. The RNA sequence upstream of a 3′SS is input to the model, and a predicted probability of being a branchpoint is outputted for each nucleotide. (D) Model performance on held-out test sets in comparison with two existing methods. Model performance is defined as the fraction of 3′SSs where the highest scoring position overlaps with an experimentally determined branchpoint. Each cluster of bars indicates the performance on a different test set: “High Conf” refers to Mercer et al. (2015) high confidence sites, “Low Conf” refers to the complete set of Mercer et al. (2015) sites, “Taggart” refers to the Taggart et al. (2017) predictions. The “<5” test sets count the highest scoring position as overlapping with an experimental branchpoint if there is an experimental branchpoint <5 nt from it. (*) Significant difference with P < 1 × 10−14 by a two-sided Fisher exact test. Local sequence context (E), PhyloP conservation as a function of position relative to a branchpoint averaged across all branchpoints (F), and position relative to 3′SS (G) for predicted and experimentally determined branchpoints.
