henhao
Sign in
GenomicsbioRxiv
Preprint

An encyclopedia of enhancer-gene regulatory interactions in the human genome

Andreas R. GschwindiD, Kristy S. MualimiD, Alireza KarbalaygharehiD, Maya U. ShethiD, Kushal K. DeyiD, Evelyn Jagoda, Ramil NurtdinoviD, Xi WangiD, Anthony S. Tan, Hank JonesiD, X. Rosa MaiD, David YaoiD, Joseph NasseriD, Žiga AvseciD, Benjamin T. JamesiD, Muhammad S. ShamimiD, Neva C. DurandiD, Suhas S.P. RaoiD, Ragini MahajaniD, Benjamin R. DoughtyiD, Kalina AndreevaiD, Jacob C. UlirschiD, Kaili FaniD, Elizabeth M. PereziD, Tri C. NguyeniD, , J EngreitziD

Abstract

Identifying transcriptional enhancers and their target genes is essential for understanding gene regulation and the impact of human genetic variation on disease1-6. Here we create and evaluate a resource of >13 million enhancer-gene regulatory interactions across 352 cell types and tissues, by integrating predictive models, measurements of chromatin state and 3D contacts, and large-scale genetic perturbations generated by the ENCODE Consortium7. We first create a systematic benchmarking pipeline to compare predictive models, assembling a dataset of 10,411 element-gene pairs measured in CRISPR perturbation experiments, >30,000 fine-mapped eQTLs, and 569 fine-mapped GWAS variants linked to a likely causal gene. Using this framework, we develop a new predictive model, ENCODE-rE2G, that achieves state-of-the-art performance across multiple prediction tasks, demonstrating a strategy involving iterative perturbations and supervised machine learning to build increasingly accurate predictive models of enhancer regulation. Using the ENCODE-rE2G model, we build an encyclopedia of enhancer-gene regulatory interactions in the human genome, which reveals global properties of enhancer networks, identifies differences in the functions of genes that have more or less complex regulatory landscapes, and improves analyses to link noncoding variants to target genes and cell types for common, complex diseases. By interpreting the model, we find evidence that, beyond enhancer activity and 3D enhancer-promoter contacts, additional features guide enhancer-promoter communication including promoter class and enhancer-enhancer synergy. Altogether, these genome-wide maps of enhancer-gene regulatory interactions, benchmarking software, predictive models, and insights about enhancer function provide a valuable resource for future studies of gene regulation and human genetics.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
119
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 1 (13 Nov 2023) · CC BY-ND 4.0

ENCODE-rE2G, a classifier trained on K562 CRISPR perturbations, maps 13,455,443 enhancer-gene interactions across 352 biosamples, matching or exceeding prior models on CRISPR, eQTL, and GWAS benchmarks.

Earlier enhancer-gene predictors often lacked accuracy estimates, were not systematically compared on common benchmarks, or required many cell types or assays, leaving no ENCODE-wide map for linking noncoding variants to genes. The work supplies benchmarks built from genetic perturbations (CRISPR, eQTL, GWAS) rather than physical contacts, a CRISPR-trained classifier needing minimally DNase-seq plus averaged Hi-C, and maps across 352 biosamples.

Key results

  1. ShownTraining a logistic regression directly on CRISPR outcomes gave the best classification of regulatory element-gene pairs in K562, beating the Activity-by-Contact (ABC) model, published predictors such as EpiMap, GraphReg and Enformer, and closest-gene or 3D-contact baselines.

    K562 CRISPR benchmark (574 comparator models; hold-one-chromosome-out cross-validation): ENCODE-rE2G vs ABC (DNase, averaged Hi-C) AUPRC 0.63 vs 0.55, precision at 70% recall 54% vs 46% (Pbootstrap = 0.0001), Pearson R -0.44 vs -0.37 (P = 7.16 x 10^-9); ENCODE-rE2G^Extended vs runner-up ABC (DHS*H3K27ac, ENCODE Hi-C) AUPRC 0.76 vs 0.61, precision 70% vs 54%. · Supervised classification of enhancer-gene…; Systematic benchmarking of ENCODE-rE2G

  2. ShownAlthough trained only in K562, ENCODE-rE2G transferred to other cell types: it matched top models for eQTL enrichment, recovered more fine-mapped eQTL-gene links across GTEx tissues, and linked blood-trait GWAS credible sets to known genes with the highest precision.

    GM12878 predictions vs GTEx lymphoblastoid eQTLs (PIP > 50%, n=273): enrichment 28 at 15% recall, on par with top models; 11 other GTEx tissues: mean enrichment 6.5 (3.5-11.4), recall 16% (10-22%). Blood GWAS: highest variant enrichment, 10.6-fold (P = 1 x 10^-4); for 197 credible sets, precision 61% vs 52% for DNase/averaged-Hi-C ABC (P = 8 x 10^-4), recall 35% each. · Systematic benchmarking of ENCODE-rE2G; Figure legends

  3. ShownApplied to all biosamples, the model predicts mostly short-range, highly biosample-specific links with few elements per gene; genes with many predicted enhancers lean toward cell-type-specific processes, those with few toward housekeeping functions.

    13,455,443 links in 352 biosamples (0.76% of the genome); median 31.8% of links <10 kb and 86.8% <100 kb; 3.99 elements per gene, 1.6 genes per element; 26.4% of links in a single biosample. Genes with mean >=5 enhancers enriched for angiogenesis (2.87-fold, P = 3.60 x 10^-4); genes with <=1.5 for Golgi transport (2.92-fold, P = 1.44 x 10^-14). · An encyclopedia of enhancer-gene regulatory…; Genome-wide maps of regulatory interactions; Figure legends

  4. ShownIntersecting the top two ENCODE-rE2G and top two Polygenic Priority Score (PoPS) genes per locus raised precision for known causal genes over either method alone and nominated target genes for credible sets across 94 UK Biobank traits, including loci missed by the earlier ABC-Max atlas.

    Precision 68% (1.2x and 2.0x the top ENCODE-rE2G or PoPS gene alone, respectively), above ABC-Max at comparable recall; genes assigned to 3,528 of 13,012 credible sets covering 5,851 variants (3.7-fold excess vs 9.2 million 1000 Genomes controls), including 1,811 sets absent from ABC-Max. rs875741 (mean corpuscular hemoglobin) linked to CPEB4 only in hematopoietic progenitors. · Linking noncoding variants to target genes and…; Figure legends

  5. ShownBeyond activity and 3D contact, promoter class and nearby-enhancer features carry predictive signal: ubiquitously expressed genes less often have distal regulators, and at MYC in K562, nearby enhancer pairs combine super-additively while distant pairs act near-additively.

    Ubiquitously expressed genes were 5-fold less likely to have distal regulators in K562 CRISPR data (P = 6.2 x 10^-45); adding 3 promoter-class features to ABC raised AUPRC by 0.044; removing nearby-enhancer features cut AUPRC by 0.0385. All 21 MYC enhancer pairs were super-additive (BH P < 0.001; 16-37% deviation at 3-89 kb vs 2-10% at 107 kb-1.79 Mb). · Molecular features guiding enhancer-gene…; Ubiquitously expressed genes are less sensitive…; Super-additive functions of distal enhancers

Methods, data and reproducibility

Logistic regression classifiers fit to 10,411 K562 CRISPR-tested element-gene pairs (472 positives) with hold-one-chromosome-out cross-validation, using 13 features from cell-type-specific DNase-seq plus Hi-C averaged over 35 biosamples (ENCODE-rE2G) or 47 multi-assay features (Extended), benchmarked on CRISPR, eQTL and GWAS and applied to 352 biosamples.

  • Harmonized CRISPR perturbation data (mostly CRISPRi) in K562 cells, re-analyzed from Nasser 2021, Gasperini 2019 and Schraivogel 2020 screens, with per-pair power estimates · human (K562 erythroleukemia cell line) · n = 10,411 element-gene pairs (472 positive, 9,938 negative) · ENCSR998YDI
  • ENCODE DNase-seq profiles of cell types, tissues and contexts used as cell-type-specific model input and for genome-wide predictions · human · n = 352 biosamples
  • ENCODE Hi-C averaged across diploid biosamples for 3D contact features; cell-type-specific ENCODE Hi-C and in situ Hi-C in K562 used for comparisons · human · n = 35 diploid biosamples (averaged map)
  • ENCODE 1D chromatin experiments in K562 (DNase-seq, ATAC-seq, histone and transcription factor ChIP-seq) substituted as ABC activity inputs · human (K562) · n = 513 experiments
  • Fine-mapped distal noncoding GTEx eQTLs from lymphoblastoid cell lines (LCLs) plus 11 other GTEx tissues matched to ENCODE biosamples · human (GTEx) · n = >30,000 fine-mapped eQTLs; 273 LCL variants with PIP > 50% · Synapse syn52264240
  • SuSIE fine-mapped distal noncoding GWAS variants (PIP > 0.1) for UK Biobank traits; blood-trait credible sets with coding-supported genes used as gene-linking truth · human (UK Biobank) · n = 29.2K variants, 94 traits, 13,012 credible sets; 197 truth sets
  • Pairwise combinatorial CRISPRi-FlowFISH targeting all pairs of 7 enhancers near MYC in K562-KRAB-dCas9 cells · human (K562) · n = 21 enhancer pairs · ENCSR443VTK
  • Enhancer perturbations (CRISPRi, genetic deletions, natural variants) with H3K27ac ChIP-seq readout at nearby enhancers, including CRISPRi at the MYC locus · GSE225157
Methods & tools
ENCODE-rE2G and ENCODE-rE2G^Extended (logistic regression), Activity-by-Contact (ABC) model, EpiMap, GraphReg, Enformer, EPIraction, CIA, Polygenic Priority Score (PoPS), SuSIE fine-mapping, CRISPRi-FlowFISH, DNase-seq, H3K27ac ChIP-seq, Hi-C, ChIA-PET, chromHMM and ENCODE cCRE catalog
Code
https://github.com/karbalayghareh/ENCODE-rE2Ghttps://github.com/EngreitzLab/CRISPR_comparisonhttps://github.com/EngreitzLab/eQTLEnrichmenthttps://github.com/EngreitzLab/ABC-Max-pipelinehttps://github.com/broadinstitute/ABC-Enhancer-Gene-Predictionhttps://github.com/argschwind/ENCODE_CRISPR_data
Data & accessions
http://www.encodeproject.org/https://github.com/EngreitzLab/CRISPR_comparison/blob/main/resources/crispr_datahttps://www.synapse.org/#!Synapse:syn52264240https://www.synapse.org/#!Synapse:syn52234396https://www.finucanelab.org/data/
Reagents & environment
Analyses in hg38 (GRCh38); wet-lab perturbations used K562 cells (K562-KRAB-dCas9 for pairwise CRISPRi-FlowFISH) and CRISPR-Cas9 enhancer knockouts; no software versions, packages or containers are stated.
Experimental conditions
No hardware or runtime reported; assay parameters include a CRISPRi-FlowFISH readout for pairwise MYC perturbations, K562 ENCODE Hi-C at 2 billion reads and 5-kb resolution, and 1000 bootstrap iterations for AUPRC error ranges.
Reproducibility gaps
Accessions for ENCODE inputs, predictions and comparator outputs are given only in supplementary tables (S2, S8, S9, S10); no software versions, containers, hardware, or runtimes are reported.
Limitations (5)
  • CRISPR training data have relatively few positives, biased choice of tested enhancers and genes, no separation of cis from trans effects, mostly K562 origin, and weak power for small effects (e.g., <25%); predictions may inherit these biases. · stated by the authors
  • eQTL benchmarks come mostly from tissues lacking cell-type resolution and can miss small effects; the authors call for larger, unbiased, cell-type-resolved CRISPR and eQTL datasets to train better models. · stated by the authors
  • Super-additivity is defined on a linear expression scale, and the authors caution that these observations need careful interpretation because earlier studies using higher-order phenotypes reached conflicting conclusions. · stated by the authors
  • Direct pairwise enhancer tests come from one locus (7 MYC enhancers, 21 pairs, K562); the distance dependence of synergy at other genes rests on indirect evidence from tiling effect sums and H3K27ac cross-effects. · noted in this summary
  • Per-biosample averages differ between sections (62,509 vs 38,225 interactions), and CRISPR pair totals vary by analysis (10,411, 10,375, 10,292), so exact counts should be checked against the supplementary tables. · noted in this summary

Who should read:Read this if you need cell-type-resolved enhancer-gene links for variant interpretation, want a CRISPR/eQTL/GWAS benchmark for enhancer-gene models, or study how promoter class and enhancer synergy shape regulation.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISSheth '26DeGroat '26Fulco '19Nasser '20Sathian '26Gao '19
72% semantic matchGenetics
Maya U. Sheth, Wei‐Lin Qiu, X. Rosa Ma+20 more·Nature Genetics·Aug 1, 2026· 29 citations