Mammalian chromatin is tightly compacted inside the 3D space of the nucleus in a highly ordered manner. Deformation of the 3D chromatin structure plays a profound role in gene regulation and function, and associates with many pathological conditions like cancer.  Many proteins have been shown to play a role in regulating the 3D chromatin structure, of which CCCTC-binding zinc-finger protein (CTCF) is among the most prominent and best described. CTCF binds to DNA by recognizing a specific 19bp sequence and can bring one binding locus into contact with other distal CTCF binding sites, thus forming a chromatin interaction network inside the nucleus. Although this network has been considered largely invariant among different cell-types, we find that it exhibits extensive cell-type-specific interactions that contribute to cell identity. Here, we present Lollipop, a machine-learning framework, which predicts CTCF-mediated long-range interactions using genomic and epigenomic features. Using ChIA-PET data as a benchmark, we demonstrate that Lollipop accurately predicts CTCF-mediated chromatin interactions both within and across cell types, and outperforms other methods based only on CTCF motif orientation. Predictions are confirmed computationally and experimentally by Chromatin Conformation Capture (3C). Moreover, our approach identifies other determinants of CTCF-mediated chromatin wiring, such as gene expression within the loops. Our study contributes to a better understanding of the underlying principles of CTCF-mediated chromatin interactions and their impact on gene expression.

The work is published in Nature Communications, 2018 (https://www.nature.com/articles/s41467-018-06664-6).