iResNetDM: Interpretable deep learning approach for four types of DNA modification prediction

Abstract

MotivationDespite the development of several computational methods to predict DNA modifications, two main limitations persist in the current methodologies: 1) All existing models are confined to binary predictor which merely determine the presence or absence of DNA modifications, constraining comprehensive analyses of the interrelations among varied modification types. While multi-class classification models for RNA modifications have been developed, a comparable approach for DNA remains a critical need. 2) The majority of previous studies lack adequate explanations of how models make decisions, relying on the extraction and visualization of attention matrices which identified few motifs, and do not provide sufficient insight into the model decision making process. ResultIn this study, we introduce iResNetDM, a deep learning model that integrates ResNet and self-attention mechanisms. To the best of our knowledge, iResNetDM is the first model capable of distinguishing between four types of DNA modifications. It not only demonstrates high performance across various DNA modifications but also unveils the potential capabilities of CNN and ResNet in this domain. To augment the interpretability of our model, we implemented the integrated gradients technique, which was pivotal in demystifying the models decision-making framework, allowing for the successful identification of multiple motifs. Importantly, our model exhibits remarkable robustness, successfully identifying unique motifs across different modifications. Furthermore, we compared the motifs discovered in various modifications, revealing that some motifs share significant sequence similarities which suggests that these motifs may be subjected to different types of modifications, underscoring their potential importance in gene regulation. Contactzeruiyang2-c@my.cityu.edu.hk