From correlation to contact: how enhancers really find their genes

These are the papers sitting underneath the contact work I’ll be doing at Institut Curie later this year, so I presented them together: the first gives the principle, the second gives the data, and the third gives the engine that turns both into something I can actually run on my atlas. They all attack the same weakness, that correlation tells you a peak and a gene behave alike, but not that they’re physically connected, and not which gene an enhancer actually controls.

The gap: correlation isn’t contact

In my atlas (SCARLET, Objective 1), each accessible peak is linked to a target gene by correlating accessibility with gene activity across cells. That’s a guess dressed up as a link. An enhancer can regulate up to five genes; a gene can be controlled by up to fourteen elements; enhancers routinely skip the nearest gene. Correlation-only or distance-only linkage assigns the wrong target often enough that no reviewer should trust it.

Paper 1, Activity-by-Contact (Fulco et al., 2019)

Fulco and colleagues showed, with thousands of direct CRISPR measurements, that to predict which enhancer regulates which gene you need both how active the enhancer is and how often it contacts the gene.

First they showed the existing rules fail. Measuring 3,863 element–gene pairs by CRISPR in K562:

• “Everything within 100 kb” gives 13% precision (most nearby elements do nothing).
• “Nearest gene” gives 47% precision, 37% recall (the closest gene is often wrong).
• “Hi-C loops alone” gives 29% precision, 4% recall (contact alone misses almost everything).

Then the measurement that makes the paper a ground truth: CRISPRi-FlowFISH. KRAB-dCas9 plus a guide library (~one guide per cell) silences one element per cell; cells are labelled for one gene’s RNA by FISH, sorted into bins, and the guides sequenced, reading off how silencing each element shifted expression. 30 genes, every element within 450 kb, 3,863 distal pairs.

The model is one multiplication

An enhancer’s predicted effect on a gene is its activity × contact, divided by the same summed over all nearby elements, i.e. its share of the gene’s regulation. Activity is the geometric mean of accessibility and H3K27ac; contact is normalised Hi-C frequency (approximable from averaged Hi-C or even distance). Two enhancers, equal activity, but one contacts the promoter 3× as often, it takes three-quarters of the regulation. That multiplication is the whole idea.

Scored against the CRISPR links (AUPRC): activity alone 0.22, contact alone 0.29, distance 0.39, and the full ABC model 0.65. Neither half is sufficient; it’s the product that works. And it generalises to variants, ABC correctly linked 13 of 16 trait-associated variants to their known target gene (81% recall vs 56% for nearest-gene), including the classic rs12740374 → SORT1 in liver.

Paper 2, the 3D genome of real tumours (Yost et al., 2025)

Where Fulco is the principle, Yost is the template, contact data in actual tumours, including breast, with the copy-number correction and pipeline I’ll use at Curie.

They built an enhancer connectome with H3K27ac HiChIP (which measures activity and 3D contact at once) across 69 primary tumours, 15 cancer types, 665,000 enhancer-anchored loops. The breast cohort is the deepest, 14 samples, ≥3 per molecular subtype (basal, HER2, luminal A/B), each with matched whole-genome sequencing, RNA-seq and single-cell ATAC.

Three results I leaned on:

• Contact is subtype-informative. Megabase compartments and TADs are largely shared across cancers; it’s the fine-scale 2D loop signal that cleanly separates cancer types and, within breast, basal from non-basal. Cancer-type loops are enriched for expected TF motifs, biological, not noise. That’s the assumption my whole subtype-specific atlas rests on.
• Most oncogenes are enhancer-driven, not copy-driven. Across 110 oncogenes, >70% are better explained by enhancer activity than by copy-number gain (MYC by dynamic enhancer rewiring; KRAS one of the few copy-driven exceptions). So the gene a non-coding element controls depends on its wiring, and wiring is what contact data resolve.
• Copy-number correction is not optional. Amplified regions produce more reads, inflating both apparent activity and apparent contact, call loops naively and you’ll call amplifications. Yost correct with ploidy-aware copy-number from matched WGS before clustering, using HiC-Pro and FitHiChIP. Doing that normalisation properly, in person, is the irreplaceable reason for the residency.

They even go from contact to mechanism at named loci, an ESR1 enhancer ~9 kb away whose contact tracks ESR1 expression, and non-coding point mutations (MECOM promoter, FGFR1 enhancer) nominated by comparing allele frequency in contact data against the genome. That’s the logic of my Objectives 2–3, demonstrated in primary tumours.

Paper 3, SCEG-HiC (Liang et al., 2026), the engine

This is the paper that turns the first two into something I can actually run. Fulco told me contact matters; Yost gave me the tumour contact data. But my atlas is single-cell, scATAC-seq only, and neither paper tells me how to put a contact prior inside a single-cell enhancer-gene method. SCEG-HiC does exactly that, and it was published this year.

The gap it fills is precise. My peak-to-gene links come from accessibility co-varying between an enhancer and a promoter across cells. That single-cell correlation is noisy, full of indirect links (an enhancer and promoter look connected only because a third element drives both), and carries no contact information at all.

SCEG-HiC injects a Hi-C contact prior straight into single-cell linking via a Hi-C-weighted graphical lasso. It estimates partial correlations, not raw ones, which is what removes the indirect links, and the Hi-C prior sets a different penalty for every enhancer-gene pair: strong contact gets a low penalty (the link survives), little contact gets a high penalty (the link is shrunk to zero). It’s the same activity-times-contact intuition, expressed as a statistical prior, and crucially it runs on scATAC-seq alone, which is exactly my atlas.

What convinced me it’s real biology, not dressed-up distance:

• It beats the field broadly. Across 10 datasets (human and mouse, paired and scATAC-only) vs

  10 methods, validated against independent cell-type Hi-C and eQTLs. In CD4 T cells it reached AUPRC 0.683 vs Hi-C and 0.219 vs blood eQTL, top on both, with validation data independent of the prior.

• It’s not just distance. The true Hi-C prior beats both no prior and a shuffled, distance-preserving prior. Links gained from the prior are strongly eQTL-enriched; links lost are few and lack functional support.
• It does variant-to-gene in disease. Applied to COVID-19 immune cells it recovered AP-1/C-EBP networks in inflammatory monocytes, and mapped the fine-mapped non-coding variant rs71327024 to CCR1 (supported by monocyte Hi-C and blood eQTL) where a correlation-only method (Cicero) linked the wrong gene. It validated all 20 of its SNP-to-gene links against Hi-C, vs ~40% for Cicero.

How the three fit together

ABC gives the principle: contact is required to assign an enhancer to a gene. Yost gives the data: breast tumour contact maps, copy-number-corrected and subtype-resolved. SCEG-HiC gives the engine: the published method that injects a Hi-C prior into single-cell links. All three feed one contact prior that re-scores the weakest layer of my atlas, peak-to-gene linkage, turning a correlation into a contact-supported link. And the same links, scored for variant effect with AlphaGenome, become the seed for the non-coding driver work in Objectives 2 and 3. Principle, data, engine, then the bridge forward.