Building an encyclopedia of which enhancers regulate which genes

Four years in the making, and a huge team effort. This is a paper where the author list cannot convey everyone’s contributions.

The question: which enhancers regulate which genes in the human genome? Enhancers can act over long genomic distances, regulate multiple genes, and have cell-type specific effects. We need a map — to look up the enhancers for any given gene, and to understand noncoding DNA variants.

So we set out to use the vast ENCODE resource to try to build one. It took seven ingredients: reference data, perturbations, benchmarking, modeling, resource building, discovery — and team science enabling all of it.

Reference data

We began with the data resource ENCODE has built over twenty years, which is invaluable for learning gene regulation and developing methods. It has taken a huge effort to generate and uniformly process these data.

We were particularly interested in DNase-seq and hundreds of ChIP-seq experiments to measure enhancer activity, and Hi-C to estimate 3D contacts. The challenge: how can we use these rich observational datasets to learn which enhancers regulate the expression of which genes?

Perturbations

Until recently we lacked gold-standard data on which enhancers regulate which genes. We and others have now developed CRISPR screens to collect exactly that. Here we integrated three datasets with roughly 10,000 well-powered examples to learn the rules of enhancer–gene regulation.

Benchmarking

At the start, everyone on the team had a favorite model for predicting enhancer–gene regulation. We needed to compare their performance. So Andreas Gschwind built a benchmarking pipeline using the CRISPR data in K562 cells, building on work from Joseph Nasser. We also built pipelines using eQTL data (Maya Sheth) and GWAS gold-standard genes (Kushal Dey, Kristy Mualim), so that maps could be evaluated across additional cell types and tissues.

Exactly which benchmarks to use took a good deal of thought: how to compute power to define positive and negative examples in CRISPR experiments, which metrics to use, how to aggregate and merge regions, how to threshold the predictors.

Then we calculated and compared all of the methods. Kristy Mualim, Anthony Tan and Rosa Ma updated ABC. Benjamin James updated EpiMap. David Kelley computed Enformer predictions. Ramil Nurtdinov contributed a new model, EPIraction, and updated the classic peak–gene correlation approach.

Comparing all of these models to the CRISPR data, performance varied. ABC performed best — and notably, different models did better or worse on different subsets of the data. That got us thinking about how to combine features across them.

Modeling

The solution was a logistic classifier trained to predict the CRISPR data. Alireza Karbalayghareh developed the first version, which took 46 input features including many of the other models — the ENCODE-rE2G “Extended” model. It beat previous models by a wide margin.

With more exploration and feature engineering we developed the main ENCODE-rE2G model, which uses the ABC score plus twelve other features — distance, activity, contact — that can be derived from DNase-seq data in a given cell type, plus a reference cell-type average Hi-C.

Resource building

ENCODE-rE2G can therefore be applied to any cell type with DNase-seq data: hundreds of cell types and tissues in ENCODE. Kristy Mualim analyzed all of the ENCODE DNase-seq data, computed and QC’d hundreds of ABC maps, and fed these into the model. Working with Alireza Karbalayghareh, Evelyn Jagoda, Maya Sheth and Andreas Gschwind, the pipeline was applied across hundreds of ENCODE datasets.

In the end we identified more than 13 million predicted enhancer–gene regulatory interactions across 352 biosamples.

Discovery

With a resource linking enhancers to target genes, we explored what we could learn about the mechanisms underlying enhancer–gene specificity, the global properties of gene regulation, and the genetics of complex disease.

Which features does the model actually need? ENCODE-rE2G selects the ABC score as its top feature — nice. And removing the ABC score from the full model reduces performance, even though activity and contact are still included individually.

Contact. Comparing different ways of estimating contact, the new cell-type specific ENCODE Hi-C datasets give a significant boost over using a function of distance, or cell-type averaged Hi-C.

Activity. Andreas Gschwind compared hundreds of ENCODE tracks that could be used to estimate enhancer activity, and found that H3K27ac ChIP-seq and DNase-seq are in fact the best assays. ATAC is notably worse.

Promoter class. The next most important feature. Evelyn Jagoda showed that ENCODE-rE2G learns that promoters of housekeeping genes are less sensitive to activation by distal enhancers — consistent with previous results from Drew Bergman and from the van Steensel lab.

Enhancer–enhancer interactions. The model learned that enhancer clusters increase regulatory interactions. Maya Sheth, David Yao and Hank Jones tested this and showed that combinatorial perturbations to MYC enhancers do have super-additive effects on gene expression.

Correlation. The common heuristic — correlation of enhancer activity with gene expression — did not do well at predicting the CRISPR data. It only works for the small subset of enhancer–gene pairs where both are cell-type specific, and misses many other true pairs.

Evelyn Jagoda and Kristy Mualim then explored the regulatory landscapes of different protein-coding genes. Genes with different regulatory landscapes tend to have distinct molecular functions: more enhancers means more cell-type specific, fewer enhancers means housekeeping. Which makes sense.

Kushal Dey combined ENCODE-rE2G with PoPS to prioritize genes at GWAS loci. A practical tip from that: enhancer–gene regulatory interactions are useful but not sufficient to identify causal genes at GWAS loci. You also need to know the function of the gene.

The future

The formula for this project — reference data, perturbations, benchmarking, modeling, resource building, discovery, and team science — is going to be useful for a whole host of other problems in gene regulation and human genetics.

With more CRISPR data incoming, other hard problems like predicting the targets of transcription factors and variants might become far easier.

Read the paper in Nature · Predictions are public as part of the ENCODE resource · Code on GitHub.