The genome is organised via CTCF/Cohesin binding sites, which partition chromosomes into 1-5Mb topologically associated domains (TADs), and further into smaller sub-domains (sub-TADs). Here we examined in vivo an ~80kb sub-TAD, containing the mouse α-globin gene cluster, lying within a ~1Mb TAD. We find that the sub-TAD is flanked by predominantly convergent CTCF/cohesin sites which are ubiquitously bound by CTCF but only interact during erythropoiesis, defining a self-interacting erythroid compartment. Whereas the α-globin regulatory elements normally act solely on promoters downstream of the enhancers, removal of a conserved upstream CTCF/cohesin boundary extends the sub-TAD to adjacent upstream CTCF/cohesin binding sites. The α-globin enhancers now interact with the flanking chromatin, upregulating expression of genes within this extended sub-TAD. Rather than acting solely as a barrier to chromatin modification, CTCF/cohesin boundaries in this sub-TAD delimit the region of chromatin to which enhancers have access and within which they interact with receptive promoters.

Predicting the impact of non-coding genetic variation requires interpreting it in the context of 3D genome architecture. We have developed deepC, a transfer learning based deep neural network that accurately predicts genome folding from megabase-scale DNA sequence. DeepC predicts domain boundaries at high-resolution, learns the sequence determinants of genome folding and predicts the impact of both large-scale structural and single base pair variations.

Mammalian gene expression patterns are controlled by regulatory elements, which interact within topologically associating domains (TADs). The relationship between activation of regulatory elements, formation of structural chromatin interactions and gene expression during development is unclear. Here, we present Tiled-C, a low-input chromosome conformation capture (3C) technique. We use this approach to study chromatin architecture at high spatial and temporal resolution through in vivo mouse erythroid differentiation. Integrated analysis of chromatin accessibility and single-cell expression data shows that regulatory elements gradually become accessible within pre-existing TADs during early differentiation. This is followed by structural reorganization within the TAD and formation of specific contacts between enhancers and promoters. Our high-resolution data show that these enhancer-promoter interactions are not established prior to gene expression, but formed gradually during differentiation, concomitant with progressive upregulation of gene activity. Together, these results provide new insight into the close, interdependent relationship between chromatin architecture and gene regulation during development.