Towards Interpretable Multitask Learning for Splice Site and Translation Initiation Site Prediction

Abstract

AbstractIn this study, we investigate the effectiveness of multi-task learning (MTL) for handling three bioinformatics tasks: donor splice site prediction, acceptor splice site prediction, and translation initiation site prediction. As the foundation for our MTL approach, we use the SpliceRover model, which has previously been successful in predicting splice sites. While providing benefits such as efficient resource utilization, reduced complexity, and streamlined model management, our findings show that the newly introduced MTL model performs comparably to the SpliceRover model trained separately for each task (single-task models), with a slight decrease in specificity, sensitivity, F1-score, and Matthews Correlation Coefficient (MCC). However, these differences are statistically insignificant (the specificity decreased with 0.0081 for acceptor splice site prediction and the MCC decreased with 0.0264 for TIS prediction), emphasizing the comparable performance of the MTL model. We further analyze the effectiveness of our MTL model using visualization techniques. The outcomes indicate that our MTL model effectively learns the relevant features associated with each task when compared to the single-task models (presence of nucleotides with a higher contribution to donor splice site prediction, polypyrimidine tracts in the upstream of acceptor splice sites, and the Kozak sequence). In conclusion, our results show that the MTL model generalizes well across all three tasks.