MCTASmRNA: A deep learning framework for alternative splicing events classification

Alternative splicing (AS) plays crucial post-transcriptional gene function regulation roles in eukaryotic. Despite progress in studying AS at the RNA level, existing methods for AS event identification face challenges such as inefficiency, lengthy processing times, and limitations in capturing the complexity of RNA sequences. To overcome these challenges, we evaluated 10 AS detection tools and selected rMATS for dataset construction. We then developed a multi-scale convolutional and Transformer-based model (MCTASmRNA) to classify AS events in mRNA sequences without relying on a reference genome. To handle the problem of large intra-class and small inter-class difference in AS event sequences, we incorporated an efficient channel attention mechanism and designed a new joint loss function to optimize MCTASmRNA training. MCTASmRNA outperformed baseline models, with an accuracy improvement and exhibited enhanced cross-species generalizability. This model provides valuable support for AS research across different organisms. Future work will focus on optimizing and expanding the model to further explore the complex mechanisms underlying AS.