Using Convolutional Neural Network to Study the Regulatory Relationship Between DNA Methylation and Gene Expression

In this work, we propose a novel framework based on convolutional neural networks to integrate information of DNA methylation and DNA sequences for studying their relationship with gene expression. Using the data from 8 cell types, we show that the trained models provide higher explained variation of gene expression from DNA methylation (up to 30%) compared to methods that only consider one genomic data type or from one region (up to 12%). In addition, the extracted sequence motifs from the filters of the first convolutional layers reveal the involvement of regulatory modulators including DNA Methyltransferase 1 (DNMT1) and other transcription factors, shedding light on the complex regulatory mechanisms of DNA methylation. We found that more than half of the 20 most enriched motifs correspond to regulators that are well supported by previous publications and ChIP-Seq transcription factor binding datasets. Our work has demonstrated the usefulness of convolutional neural networks in revealing complex relationships from omics data.