scE2G: linking enhancers to their target genes from single-cell data

Introducing scE2G: a model to link enhancers to their target genes using single-cell data, with state-of-the-art performance across multiple perturbation benchmarks.

scE2G tackles the challenge of building cell-type-specific maps of enhancer–gene regulation. If we can do this well, we can build enhancer maps across thousands of human cell types from emerging single-cell atlases — to interpret genetic variants and understand gene regulation in tissues that are only practically accessible to single-cell methods.

Schematic: single-cell multiome or ATAC-seq data from healthy or diseased tissues, drawn from resources like the Human Cell Atlas, is run through scE2G to produce separate enhancer–gene maps for cell types A, B and C, in which the same risk variant connects to different genes. — The goal: run scE2G on single-cell data from any tissue and get a separate enhancer–gene map for every cell type in it.

Two models, trained on CRISPR ground truth

Using a gold-standard CRISPR perturbation dataset of more than 10,000 evaluated element–gene pairs, we trained two logistic regression models: scE2G (ATAC) and scE2G (Multiome), which use single-cell ATAC-seq or single-cell multiomic ATAC and RNA-seq data respectively.

Three-stage method schematic. First, single-cell multiome data is reduced to a pseudobulk fragment file and an RNA count matrix. Second, candidate peak–gene pairs are annotated with an ABC score, chromatin accessibility, gene density, Kendall correlation between peak accessibility and gene expression, and promoter class. Third, these features enter the scE2G model, trained by logistic regression on CRISPRi-validated enhancers in K562 cells. — How scE2G works: derive features for each candidate peak–gene pair from the single-cell data, then score them with a classifier trained on CRISPRi-validated enhancers.

Benchmarking

In systematic benchmarking against CRISPR perturbations, fine-mapped eQTLs, and GWAS variant–gene associations, the scE2G models outperform existing single-cell models and distance-derived baselines. We applied and extended the benchmarking pipelines built for ENCODE-rE2G.

Precision–recall curves for eighteen numbered predictors on the CRISPR benchmark. scE2G Multiome traces the highest curve, followed by scE2G ATAC, then the ARC-E2G score and Kendall correlation; existing single-cell tools including SnapATAC, Signac, Cicero, FigR, SCENT, SCENIC+, DIRECT-NET and ArchR fall well below, as do distance baselines. — scE2G (Multiome) and scE2G (ATAC) are the top two predictors, ahead of every existing single-cell method tested and of distance-based baselines.

For the published version we added a benchmark on a fully held-out CRISPR dataset — 189 positives among 4,175 tested pairs from five cell types, none of them used in training. scE2G continues to show state-of-the-art performance, reinforcing that it generalizes well.

Bar chart of weighted area under the precision–recall curve on a held-out CRISPR benchmark of 189 positives and 4,175 tested pairs from five cell types. scE2G Multiome is highest, then scE2G ATAC, then distance to TSS and ABC, then eight other single-cell methods, all lower. — Added during revision: performance on a fully held-out CRISPR dataset from five cell types the model never saw.

What the model learns

Key features in scE2G include the ABC score, the Kendall correlation between peak accessibility and gene expression, and whether the gene is “ubiquitously expressed.”

Notably, the Kendall correlation improves long-range predictions and appears to detect stochastic transcriptional bursting — which is why it needs genuinely single-cell resolution rather than aggregated meta-cells.

Top: three schematic designs — single cells from one cell type, meta-cells aggregated from one cell type, and single cells from multiple cell types. Bottom: grouped bar chart of model performance stratified by enhancer-to-transcription-start-site distance, where the single-cell scE2G Multiome model retains the largest advantage in the most distal 100 kilobase to 1 megabase bin. — Single-cell resolution matters most for long-range links: in the most distal distance bin, the single-cell model keeps an advantage the meta-cell and correlation-only versions lose.

How much data do you need?

scE2G has robust performance for cell types with at least 2 million total ATAC fragments and 1 million RNA UMIs — about 200–400 cells from a typical 10x Multiome experiment in a tissue. That is a low enough bar to apply it to most annotated cell types in a well-powered atlas.

Applying it across cell types

We applied scE2G to more than 40 cell types from PBMCs, BMMCs and pancreatic islets, validating that its predictions reflect expected patterns of cell-type specificity — predictions cluster by lineage, and the enrichment of fine-mapped eQTLs rises as the prediction’s cell type more closely matches the eQTL’s.

Left: a symmetric matrix of pairwise similarity between scE2G predictions across roughly forty blood, bone marrow and pancreatic islet cell types, showing blocks of high similarity within erythroid, lymphoid, myeloid and pancreatic lineages. Right: violin plots showing enrichment of fine-mapped eQTLs over common variants rising from unmatched, to same lineage, to same cell type. — Predictions group by lineage (left), and eQTL enrichment increases as the prediction's cell type matches the eQTL's more closely (right).

For example, we identify cell-type specific links for SPTA1 — a red blood cell membrane gene — in erythroblasts and normoblasts, but not in the progenitors upstream of them.

Genome browser view of the SPTA1 locus across five bone marrow cell types with ATAC-seq tracks and scE2G prediction arcs. Dense arcs connecting distal elements to SPTA1 appear in erythroblasts and normoblasts, fewer in proerythroblasts, and none in haematopoietic stem cells or granulocyte–monocyte progenitors; two fine-mapped blood-trait variants are marked. — Predicted enhancers for *SPTA1* appear in erythroblasts and normoblasts and are absent from upstream progenitors.

Prioritizing causal genes and cell types for disease variants

We also integrate scE2G predictions with orthogonal information to prioritize causal genes and cell types for noncoding variants associated with complex traits. For example, we nominate regulatory interactions linking INPP4B and IL15 to lymphocyte counts in T cells — and the risk allele is predicted to strengthen a TP73 binding motif.

A complete chain for one variant: the enhancer link, the gene-level priority score, and the transcription factor motif the risk allele alters.

We have also applied scE2G to identify and validate disease-related enhancers in the human coronary artery and the fetal heart.

Which model should you use?

scE2G complements the ENCODE-rE2G model published in Nature:

Bulk DNase-seq, ATAC-seq, H3K27ac and/or Hi-C → run ENCODE-rE2G
Single-cell ATAC-seq, or multiomic ATAC + RNA → run scE2G

Same pipeline as the preprint, so if you have single-cell ATAC or multiome data you can run scE2G on it today. We look forward to hearing how it works on your dataset.

This was a wonderful collaboration led by co-first authors Maya Sheth and Wei-Lin Qiu, with Robin Andersson and Jesse Engreitz, and made possible by support from the Novo Nordisk Foundation, the IGVF Consortium, CZI, NSF and NHGRI.

Read the paper in Nature Genetics (58, 1941–1952) · Browse predictions across hundreds of cell types at e2g.stanford.edu · Code on GitHub.

This write-up is adapted from Wei-Lin Qiu’s thread on the preprint (Bluesky, X) and Jesse Engreitz’s thread on the published paper (Bluesky, X).

Figures from Sheth, Qiu et al., Nature Genetics 58, 1941–1952 (2026), CC BY 4.0; some panels cropped.