iResNetDM: An interpretable deep learning approach for four types of DNA methylation modification prediction

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Existing computational models for DNA methylation prediction are limited to binary classification, hindering analysis of inter-type modification relationships.
• Current models lack interpretability, relying on attention matrices that offer limited insight into decision-making processes and motif discovery.

Purpose of the Study:

• To address limitations in DNA methylation prediction by framing it as a multi-class classification problem.
• To develop an interpretable deep learning model for predicting multiple DNA methylation modification types simultaneously.

Main Methods:

• Introduction of iResNetDM, a novel deep learning architecture integrating Residual Networks (ResNet) with self-attention mechanisms.
• Application of the integrated gradients technique to enhance model interpretability and elucidate decision-making processes.
• Comparative analysis of discovered motifs across different methylation modifications to identify sequence similarities and regulatory implications.

Main Results:

• iResNetDM is the first model capable of distinguishing between four types of DNA methylation modifications.
• The model demonstrates strong performance and effectively captures relationships between different DNA methylation modification types.
• Integrated gradients successfully identified multiple DNA sequence motifs, revealing unique patterns for different modifications and highlighting potential regulatory roles.

Conclusions:

• The study establishes multi-class classification as a viable approach for DNA methylation prediction, offering a more comprehensive analysis.
• iResNetDM provides an interpretable framework for understanding DNA methylation prediction, facilitating the discovery of biologically relevant motifs.
• Identified sequence similarities among motifs suggest cross-talk between different DNA methylation modifications, impacting gene regulation.