Chromatin interactions are two or more distal genomic regions that come into close spatial proximity, which can lead to the aberrant expression of oncogenes in cancer. High-throughput methods require large amounts of material to detect chromatin interactions (eg. Hi-C and ChIA-PET), which have not been widely applied to large cohorts of cell lines or clinical samples. However, large cohorts of RNA-Seq data of clinical samples are available from public. Can we predict chromatin interactions of large cohorts of cell lines or clinical samples from RNA-Seq data? To do this, we developed AI4Loop, a bidirectional long short-term memory (BiLSTM) network model that integrates multi-scale gene expression information, and showed that exclusively RNA-seq information is sufficient to identify cell type-specific chromatin interactions. AI4Loop exhibits robust performance and generalization of chromosome-split strategies across different cell types and samples, and its predicted key chromatin interactions can successfully distinguish Acute Myeloid Leukemia (AML) samples from normal samples. With AI4Loop, we discovered new patterns of gene-gene interactions (GGI), creating a unified view of chromatin interactions at the gene level. At inference time, AI4Loop can be used to study the perturbations of chromatin interactions after different drug treatments, thereby guiding the selection of potential cancer drugs.
 

chromatin interaction, deep learning, cancer, gene expression