SpliceSCANNER: An Accurate and Interpretable Deep Learning-Based Method for Splice Site Prediction

Abstract

The identification of splice sites is significant to the delineation of gene structure and the understanding of complicated alternative mechanisms underlying gene transcriptional regulation. Currently, most of the existing approaches predict splice sites utilizing deep learning-based strategies. However, they may fail to assign high weights to important segments of sequences to capture distinctive features. Moreover, they often only apply neural network as a ‘black box’, arising criticism for scarce reasoning behind their decision-making. To address these issues, we present a novel method, SpliceSCANNER, to predict canonical splice sites via integration of attention mechanism with convolutional neural network (CNN). Furthermore, we adopted gradient-weighted class activation mapping (Grad-CAM) to interpret the results derived from models. We trained ten models for donor and acceptor on five species. Experiments demonstrate that SpliceSCANNER outperforms state-of-the-art methods on most of the datasets. Taking human data for instance, it achieves accuracy of 96.36% and 95.77% for donor and acceptor respectively. Finally, the cross-organism validation results illustrate that it has outstanding generalizability, indicating its powerful ability to annotate canonical splice sites for poorly studied species. We anticipate that it can mine potential splicing patterns and bring new advancements to the bioinformatics community. SpliceSCANNER is freely available as a web server at http://www.bioinfo-zhanglab.com/SpliceSCANNER/.