Quantitative prediction of enhancer-promoter interactions

Abstract

Recent experimental and computational efforts provided large datasets describing 3-dimensional organization of mouse and human genomes and showed interconnection between expression profile, epigenetic status and spatial interactions of loci. These interconnections were utilized to infer spatial organization of chromatin, including enhancer-promoter contacts, from 1-dimensional epigenetic marks.Here we showed that predictive power of some of these algorithms is overestimated due to peculiar properties of biological data. We proposed an alternative approach, which gives high-quality predictions of chromatin interactions using only information about gene expression and CTCF-binding. Using multiple metrics, we confirmed that our algorithm could efficiently predict 3-dimensional architecture of normal and rearranged genomes.

DeepC: Predicting chromatin interactions using megabase scaled deep neural networks and transfer learning

“…Within promoters and enhancers respectively, we find the ones strongly marked by activity associated histone marks to have a stronger predicted impact. Our observations match with insights from orthogonal methods, that found CTCF binding and cell type specific active chromatin marks 12,13 , and CTCF binding and RNA expression 11 , to be most predictive of chromatin interactions, respectively. Considering the fidelity of the deepC predictions this suggests that the patterns of 3C genome folding seen in methods such as Hi-C arise from the interplay of the activities of enhancers and promoters as well as CTCF bound sites.…”

“…Recent methods also predict chromatin interactions from cell type specific chromatin features, such as CTCF and histone modification ChIP-seq, DNase-seq and RNA-seq. These methods capture the chromatin features between two interaction windows by: averaging them 10,12 ; or by capturing the number, distance and orientation of chromatin features such as CTCF through careful feature engineering 11 . However, these methods are not sequenced based, limiting their ability to map the determinants of chromatin folding at high resolution.…”

“…The average distance stratified Pearson correlation between the Hi-C skeleton and deepC prediction is ~ 0.44 and ~ 0.62 when applying a small mean filter on the noisy skeleton data ( Figure 1D). This is comparable to orthogonal methods that predict chromatin interactions directly from cell type specific chromatin features [11][12][13] , although genome folding prediction from sequence is a much more difficult task. Visually, deepC yields smooth but intricate predictions that resolve the hierarchical nature of TADs and insulated domains at high resolution.…”

“…Methods that use information at the base pair level do not capture large scale continuous context and rather focus on window to window based predictions [8][9][10] . Methods that integrate a large genomic context do so by coarse segregation into genomic features [11][12][13] or polymer beads [14][15][16] , thus compromising on the base resolution. Here we present deepC, a deep neuronal network approach that can accurately predict Hi-C chromatin interactions from DNA sequence at megabase scale.…”

Ketamine’s pharmacogenomic network in human brain contains sub-networks associated with glutamate neurotransmission and with neuroplasticity

“…Table 1A shows that multiple GWAS disease risk SNPs located within the glutamate receptor subnetwork can be annotated as enhancers associated with tobacco smoking status, chronic schizophrenia (ICD diagnostic code F20), and bipolar 1 disorder (ICD diagnostic codes F31.0 -F31.64). (37), column 6, enhancer status determined using annotation from (38,(48)(49)(50)(51)(52)(53)(54)(55)(56)(57)(58)(59)(60), and column 7 using Hi-C data from (46,61,62). eQTL data support the results shown in column 7 (63,64).…”