Modeling cross-cell type cis-regulatory patterns via hierarchical deep neural network and gene expression prediction

Linking cis-regulatory sequences to target genes has been a long-standing challenge due to the intricate nature of gene regulation. Here, we present a hierarchical deep neural network, CREaTor, to decode cis-regulatory mechanisms across cell types by predicting gene expression from flanking candidate cis-regulatory elements (cCREs). With attention mechanism as the core component in our network, we can model complex interactions between genomic elements as far as 2Mb apart. This allows a more accurate and comprehensive depiction of gene regulation that involves cis-regulatory programs. Testing with a held-out cell type demonstrates that CREaTor outperforms previous methods in capturing cCRE-gene interactions spanning varying distance ranges. Further analysis suggests that the performance of CREaTor may be attributed to its ability to model regulatory interactions at multiple levels, including higher-order genome organizations that govern cCRE activities and cCRE-gene interactions. Together, this study showcases CREaTor as a powerful tool for systematic study of cis-regulatory programs in different cell types involved in normal development processes and diseases. ### Competing Interest Statement The authors have declared no competing interest.