A sequence-based, deep learning model accurately predicts RNA splicing branchpoints

Abstract

Experimental detection of RNA splicing branchpoints, the nucleotide serving as the nucleophile in the first catalytic step of splicing, is difficult. To date, annotations exist for only 16-21% of 3' splice sites in the human genome and even these limited annotations have been shown to be plagued by noise. We develop a sequence-only, deep learning based branchpoint predictor, LaBranchoR, which we conclude predicts a correct branchpoint for over 90% of 3' splice sites genome-wide. Our predicted branchpoints show large agreement with trends observed in the raw data, but analysis of conservation signatures and overlap with pathogenic variants reveal that our predicted branchpoints are generally more reliable than the raw data itself. We use our predicted branchpoints to identify a sequence element upstream of branchpoints consistent with extended U2 snRNA base pairing, show an association between weak branchpoints and alternative splicing, and explore the effects of variants on branchpoints.

Variant effect on splicing regulatory elements, branchpoint usage, and pseudoexonization: Strategies to enhance bioinformatic prediction using hereditary cancer genes as exemplars

“…BP prediction tools (Table 1) have demonstrated poor specificity due to BP motif degeneracy combined with a lack of experimental data to train algorithms (Corvelo, Hallegger, Smith, & Eyras, 2010). BP characterization has lagged far behind that of 5′ and 3′ splice sites because of experimental difficulties in detecting BPs (Paggi & Bejerano, 2018). A large genome‐wide data set of experimentally confirmed BPs (Mercer et al, 2015) has been used to develop the BP prediction tools Branchpointer and LaBranchoR.…”

Assessment of branch point prediction tools to predict physiological branch points and their alteration by variants

“…The deep learning tools, LaBranchoR and RNABS showed the maximum number of common predicted BPs from Ensembl (28.63 %) and from RNA-seq (33.57 %) data. Indeed, these two tools are both based on the same deep learning approach (bidirectional long short-term memory) and used the same sequence length (70 nt) as input [20,21]. By comparison, RNABPS employed a dilated convolution model explaining and showed an improvement of prediction compared to LaBranchoR (73.06 % against 64.77% of accuracy) using the Ensembl data (Table 3).…”

“…This collection of BPs was extended by two further studies: the first used 1.31 trillion reads from 17,164 RNA-seq data sets [16], and the second identified BPs by the spliceosome iCLIP method [17]. Thus, several bioinformatics tools for BP prediction have recently emerged: Branch Point Prediction (BPP) [18], Branchpointer [19], LaBranchoR [20] and RNA Branch Point Selection (RNABPS) [21] (Table 1). Briefly, HSF uses a position weighted matrix approach with a 7-mer motif as a reference (5 nt upstream and 1 nt downstream of the branch point A) (Figure 1).…”

S-CAP extends clinical-grade pathogenicity prediction to genetic variants that affect RNA splicing

“…We also included a feature to capture the change in 3-mer content induced by a variant 53 . Additionally, we included region-specific features, such as a branchpoint disruption term for the 3' intronic region 42 and a 5' cryptic splice site creation term for the 5' intronic region (see Online Methods for a complete description of features).…”

“…The 3' intronic S-CAP model includes a branchpoint feature modeled using LaBranchoR, a bi-directional LSTM (long short-term memory) model trained on the genome sequence surrounding experimentally validated branchpoint sites 42 . Specifically, we used the in silico mutagenesis scores available online (see URLs) as a feature.…”