Detection of alternative splicing: deep sequencing or deep learning

Abstract

Alternative splicing enables the expression of a variety of isoforms coding for functionally diverse proteins from a single gene. RNA sequencing (RNAseq) has become the state-of-the-art tool for comprehensive profiling of the transcriptome. While for exploration of alternative splicing events still a high read depth is needed, for RNAseq from cells infected by virus a reliable detection of alternative splicing with low number of reads is necessary. There are a range of computational tools which predict alternative splicing events from RNAseq data, but only a small number actively try to improve performance on low read depth data such as JCC and Deep Splice. We compare those tools' performance on subsampled 50M read data regarding their precision and recall in retrieving alternative splicing events from a ground truth of 500M read data from four patients with dilated cardiomyopathy. As a comparison we use SpliceAI to extract acceptor and donor probabilities from DNA sequence. The tools were found to show high precision but very poor recall. On the basis of our benchmark findings we highlight the need to develop new methods specialized on low read depth RNAseq data.