Differential analysis of alternative splicing events in gene regions using residual neural networks

Several computational methods for the differential analysis of alternative splicing (AS) events among RNA-Seq samples typically rely on estimating isoform-level gene expression. However, these approaches are often error-prone due to the interplay of individual AS events, which results in different isoforms with locally similar sequences. Moreover, methods based on isoform-level quantification usually need annotated transcripts. In this work, we leverage the ability of deep learning networks to learn features from images and propose deepSpecas, a novel method for event-based AS differential analysis between two RNA-Seq samples. Our method does not rely on isoform abundance estimation, neither on a specific annotation. deepSpecas employs an image embedding scheme to represent the alignments of the two samples on the same region and utilizes a residual neural network to predict the AS events possibly expressed within that region. To our knowledge, deepSpecas is the first deep learning approach for performing an event-based AS analysis of RNA-Seq samples. To validate deepSpecas, we also address the lack of high quality AS benchmark datasets. For this purpose, we manually curated a set of regions exhibiting AS events. These regions were used for training our model and for assessing the predictions of our method. Our results highlight that deepSpecas achieves higher precision at the expense of a small reduction in sensitivity. The tool and the manually curated regions are available at https://github.com/sciccolella/deepSpecas.