Deep learning-based modeling of stress-induced epigenomic changes in microglia and prediction of transcriptional networks

Abstract

Stress due to adverse and demanding conditions causes emotional disturbances and increases the risk of mental illness such as depression. Chronic stress, including repeated social defeat stress, activates prefrontal microglia to induce depressive-like behavior in mice. With ChIP-seq analyses, we found that long-term epigenomic changes accompany this microglial activation. However, cis-regulatory interactions among enhancers and promoters for the epigenomic changes remain unsolved. Here we trained a deep learning-based model with our microglial ChIP-seq data segmented into approximately 130k base-pair segments with high accuracy (r = 0.654). In silico mutagenesis in this model revealed genomic regions responsible for stress-induced epigenomic changes at single-nucleotide resolution. Notably, the predicted genomic regions matched nucleosome-free regions predicted from microglial ATAC-seq data that had not been used to train the model. We identified transcription factor binding motifs enriched in the predicted genomic regions and pairs among them in proximity. Our deep learning-based epigenomic analyses offer a novel method to predict chromatin interactions at single-nucleotide resolution even with limited sample sizes and pave the way for elucidating transcriptional networks underlying health and diseases, including stress and depression.