Deciphering epigenomic code for cell differentiation using deep learning

Abstract

Epigenomic markers, such as histone modifications, play important roles in cell fate determination and type maintenance during cell differentiation. Although genomic sequence plays a crucial role in establishing the unique epigenome in each cell type produced during cell differentiation, little is known about the sequence determinants that lead to the unique epigenomes of the cells. Here, using a dataset of six histone markers measured in four human CD4+ T cell types produced at different stages of T cell development, we showed that two types of highly accurate deep convolutional neural networks (CNNs) constructed for each cell type and for each histone marker are a powerful strategy to uncover the sequence determinants of the various histone modification patterns in difference cell types. We found that sequence motifs learned by the CNN models are highly similar to known binding motifs of transcription factors known to play important roles in CD4+ T cell differentiation. Our results suggest that both the unique histone modification patterns in each cell type and the different patterns of the same histone marker in different cell types are determined by a set of motifs with unique combinations. Interestingly, the level of shared few motifs learned in the different cell models reflect the lineage relationships of the cells, while the level of few shared motifs learned in different histone marker models reflect their functional relationships. Furthermore, using these models, we can predict the importance of the learned motifs and their interactions in determining specific histone marker patterns in the cell types.