Activity-by-Contact model of enhancer specificity from thousands of CRISPR perturbations

Abstract

Mammalian genomes harbor millions of noncoding elements called enhancers that quantitatively regulate gene expression, but it remains unclear which enhancers regulate which genes. Here we describe an experimental approach, based on CRISPR interference, RNA FISH, and flow cytometry (CRISPRi-FlowFISH), to perturb enhancers in the genome, and apply it to test >3,000 potential regulatory enhancer-gene connections across multiple genomic loci. A simple equation based on a mechanistic model for enhancer function performed remarkably well at predicting the complex patterns of regulatory connections we observe in our CRISPR dataset. This Activity-by-Contact (ABC) model involves multiplying measures of enhancer activity and enhancer-promoter 3D contacts, and can predict enhancer-gene connections in a given cell type based on chromatin state maps. Together, CRISPRi-FlowFISH and the ABC model provide a systematic approach to map and predict which enhancers regulate which genes, and will help to interpret the functions of the thousands of disease risk variants in the noncoding genome.We defined Activity (A) as the geometric mean of the read counts of DHS and H3K27ac ChIP-Seq at an element E, and Contact (C) as the normalized Hi-C contact frequency between E and the promoter of gene G (see Methods). (The ABC score performed similarly across a range of data preprocessing parameters, and when defining Activity using other combinations of measurements of chromatin accessibility, histone modifications, and nascent transcription, see Methods, Fig. S6,S7,S8).The ABC model performed remarkably well, and much better than alternatives, at predicting DE-G connections in our CRISPR dataset. The quantitative ABC score correlated with the experimentally measured relative effects of candidate elements on gene expression (Spearman ρ for regulatory DE-G pairs = -0.68 Fig. 3C). Binary classifiers based on thresholds on the ABC score substantially outperformed existing predictors of enhancer-gene regulation. For example, when we used an ABC threshold corresponding to 70% recall, the predictions had 63% precision, and the area under precision-recall curve (AUPRC) was 0.66, compared to 0.36 for predictions based solely on genomic distance (Fig. 3A).

Quantitative prediction of enhancer-promoter interactions

“…formation of loops, is important indicator of promoter-enhancer connectivity, but cannot be solely used to distinguish functional interactions. In accord with this, recent large-scale CRISPR assay of promoter-enhancer connections 28 suggested quantitative "contact-by-activity" model of PE interaction. In this model, enhancer impact is quantitative and proportional to both promoter-enhancer proximity and enhancer activity.…”

Inherited Causes of Clonal Hematopoiesis of Indeterminate Potential in TOPMed Whole Genomes

“…Predicted enhancers were identified by the activity-by-contact model. 20 Predicting enhancer-gene regulation for TET2.…”

“…We used the Activity-by-Contact (ABC) model to predict which enhancers regulate which genes in CD34+ hematopoietic progenitor cells as previously described 20 , with minor modifications as follows.…”

“…We used a chromatin-state model to predict gene-enhancer pairs in hematopoietic stem cells and aggregated rare variants in these enhancer elements by target gene. 20 There were 2453 sets of enhancer regions with at least 10 individuals carrying a predicted damaging variant (alpha=2.03 x 10 -5 , see Methods). One set of variants in HAPLN1 enhancers exceeded a p-value threshold of p<0.05 after Bonferroni-correction (OR: 6.8, p=1.96 x 10 -5 , Supplementary Table 10).…”

MAUDE: Inferring expression changes in sorting-based CRISPR screens

“…For instance, CRISPR interference (CRISPRi) and CRISPR activation (CRISPRa) have been used to respectively repress or activate elements in a targeted cis-regulatory region and identify enhancers of a gene's expression 2, 4, 8, 9 . In each case, we estimate how each guide affects the targeted (or reporter) gene's expression by sorting cells into discrete expression bins, and measuring the guide's abundance in the bins. Such screens help dissect enhancer regulatory logic 5,8 , and identify the genetic variation most likely to contribute to common human disease 9 .…”