Skip to content

ENCODE-rE2G

Summary

ENCODE-rE2G is a supervised predictive model designed to identify enhancer-gene (E-G) regulatory interactions in the human genome. Developed by the Engreitz Lab, it integrates cell-type-specific chromatin activity and physical contact features to predict which distal genetic elements regulate specific target genes. The model achieves state-of-the-art performance across CRISPR perturbation, eQTL, and GWAS benchmarking datasets, providing a comprehensive resource of over 92 million regulatory links across 1,458 human biosamples (as of 2026-07-15).

Overview

Identifying the target genes of transcriptional enhancers is a major challenge in human genetics, particularly for interpreting the functional impact of noncoding disease-associated variants. ENCODE-rE2G addresses this by using supervised machine learning trained directly on gold-standard CRISPR perturbation datasets (such as CRISPRi-FlowFISH and Perturb-seq) rather than relying solely on physical contact or correlation proxies.

The framework consists of two main classifier variants: 1. ENCODE-rE2G: A general model that relies on features computable from cell-type-specific DNase-seq data combined with a tissue-averaged Hi-C contact map, making it widely applicable across hundreds of biosamples. 2. ENCODE-rE2GExtended: An expanded model utilizing 45 features derived from additional assays (such as H3K27ac ChIP-seq, cell-type-specific Hi-C, and ChIA-PET) available in Tier 1 cell types.

Model Features

The standard ENCODE-rE2G model uses a logistic regression classifier with eight selected features: - Chromatin State: Quantitative DNase-seq signals at both the enhancer element and promoter. - 3D Contact: Contact frequency derived from a reference Hi-C Megamap (averaged across 65 ENCODE datasets). - Activity-by-Contact: The Activity-by-Contact (ABC) Model score, which multiplies enhancer activity and 3D contact frequency. - Genomic Position: Distance from the element to the promoter and number of intervening genes/promoters. - Promoter Class: Whether the gene is expressed uniformly (housekeeping) or cell-type-specifically. - Nearby Enhancer Activity: The cumulative activity of other elements within 5 kb of the candidate element, capturing synergistic enhancer-enhancer interactions.

Benchmarking and Performance

ENCODE-rE2G was systematically evaluated against a gold-standard dataset of 10,356 element-gene pairs tested with CRISPR in K562 cells: - CRISPR Validation: Achieved an Area Under the Precision-Recall Curve (AUPRC) of 0.66, outperforming the standard ABC model (AUPRC = 0.56). - eQTL Linking: Demonstrated 25-fold enrichment of fine-mapped GTEx eQTL variants in predicted enhancers. - GWAS Target Gene Identification: Achieved 68% precision in linking lead variants of 197 blood-trait credible sets to causal target genes.

As a Benchmark for Sequence-Only Models

The ENCODE–rE2G CRISPRi dataset has become the reference benchmark for enhancer–gene linking, including for models that use no assay data at all. Evaluated zero-shot against it, AlphaGenome — which predicts DNase, H3K27ac and contact signal from sequence rather than measuring them — outperformed Borzoi, particularly for enhancers beyond 10 kb from the target TSS, and came within 1% auPRC of the supervised ENCODE-rE2G (extended) model that was trained on this task and cell line.[2] It also strongly outperformed both the DNase-based ENCODE-rE2G model and a distance-to-TSS baseline.[2]

The two approaches compose rather than compete: adding AlphaGenome-derived features to ENCODE-rE2G (extended) produced a new state of the art across all distance-to-TSS categories.[2] This dataset also supplies the empirical justification for the 1-Mb context window used by sequence models — 99% (465 of 471) of validated enhancer–gene pairs fall within 1 Mb.[2]

Code and Data Availability

The ENCODE-rE2G model, Snakemake workflow, and predictions are available at the following repositories: - Model Code: GitHub - ENCODE_rE2G - Predictions: Available via the ENCODE Portal (as of 2026-07-15).

See Also

Citations

  • [1] Gschwind, A. R. et al. (2026). An encyclopedia of human enhancer–gene regulatory interactions. Nature. DOI: 10.1038/s41586-026-10781-4. Source paper: s41586-026-10781-4.pdf
  • [2] Avsec, Ž. et al. (2026). Advancing regulatory variant effect prediction with AlphaGenome. Nature, 649, 1206–1217. DOI: 10.1038/s41586-025-10014-0. Source paper: s41586-025-10014-0.pdf Supports: the zero-shot and feature-augmented benchmark results on the ENCODE-rE2G CRISPRi dataset, and the 465/471 enhancer-pair-within-1-Mb figure. Location: "Competitive enhancer-gene linking"; "Unifying DNA sequence-to-function model"; Fig. 4j.