Deep learning predicts boundaries of topologically associated domains in human cell lines using CTCF, RAD21, and 12 epigenetic signals

Genomes are organized into self-interacting units called Topologically Associated Domains or TADs. The boundaries of TADs are characterized by enriched CTCF binding sites and active gene and chromatin activity. Deciphering the locations of boundaries is key to understanding the relationship between genome organization and gene expression. Detection of these boundaries is mainly based on interaction frequency matrices from costly Hi-C experiments. Recent studies suggest that epigenetic features, CTCF, and the Cohesin complex are more important than sequence features that characterize TAD boundaries. The patterns exhibited by these features in TAD boundaries are also unknown. In this paper, we propose deep learning models that rely on 14 features comprising of CTCF, Cohesin complex, and 12 epigenetic features to predict TAD boundaries in human cell lines. By exploring different architectures of popular deep learning models that incorporate dense layer, convolution layer, long short-term layer, and residual blocks, we find that a model consisting of four convolution layers followed by a dense layer performs the best in predicting TAD boundaries. Using this model and by computing saliency maps of TAD boundaries, we uncover four main motifs that capture recurring patterns of the 14 features (12 epigenetic, CTCF, and RAD21 (subunit of the Cohesin Complex)) enriched in TAD boundaries.