Enhancer-Gene Linking - Usage Guide

Overview

Predict which gene each distal enhancer regulates using ABC, ENCODE-rE2G, HiChIP, or Cicero. Choose the right method by available data: ATAC + H3K27ac + Hi-C/Micro-C enables ABC and ENCODE-rE2G; ATAC alone falls back to Cicero. Validate predictions against CRISPRi-FlowFISH gold-standard. Build cell-type-specific regulatory maps for fine-mapping causal SNPs to target genes or for therapeutic target discovery.

Prerequisites

# ABC pipeline (Snakemake-based; install its conda env)
git clone https://github.com/broadinstitute/ABC-Enhancer-Gene-Prediction
mamba env create -f ABC-Enhancer-Gene-Prediction/workflow/envs/abcenv.yml
conda activate abc-env

# ENCODE-rE2G
git clone https://github.com/EngreitzLab/ENCODE_rE2G
pip install snakemake

# HiChIP loop calling
conda install -c bioconda samtools bedtools
# HiC-Pro and FitHiChIP are not on bioconda: install HiC-Pro via its environment.yml
# (github.com/nservant/HiC-Pro) and FitHiChIP from github.com/ay-lab/FitHiChIP (Docker/Singularity)

# Cicero (atac-seq/co-accessibility skill)
BiocManager::install(c('cicero', 'monocle3', 'GenomicInteractions'))

Inputs: - ATAC-seq bigWig (RPGC-normalized) - H3K27ac ChIP-seq bigWig (matched to ATAC; RPM-IP normalized) - Hi-C/Micro-C contact matrix (cooler or hic format) at 5-10 kb resolution - Gene annotation BED (RefSeq protein-coding TSSs) - ENCODE blacklist v2

Quick Start

Tell your AI agent what you want to do: - "Run ABC on K562 ATAC + H3K27ac + Micro-C; threshold at ABC.Score >= 0.02" - "Run ENCODE-rE2G with the K562 pre-trained model; report scores >= 0.5" - "I have ATAC but no Hi-C. Use ABC's average HiC fallback OR fall back to Cicero co-accessibility" - "Cross-validate ABC predictions against published Fulco 2019 CRISPRi-FlowFISH ground truth" - "Combine ABC + ENCODE-rE2G + HiChIP -- intersection is high-confidence" - "Score the variant effect at fine-mapped GWAS SNPs at predicted enhancers (link to deep-learning-atac)"

Example Prompts

Standard ABC

"Run ABC on K562 with cell-type-matched Micro-C at 5 kb resolution. Pre-filter promoter regions from candidate enhancers. Threshold ABC.Score >= 0.02 for high-confidence predictions."

ENCODE-rE2G

"Use ENCODE-rE2G with the GM12878-trained model on my B-cell ATAC. Report predictions with ENCODE-rE2G.Score >= 0.5 and effective HiC; cross-check against ABC."

No Hi-C Fallback

"I only have ATAC + H3K27ac, no Hi-C. Run ABC with the average HiC fallback (10 ENCODE cell types pooled) and acknowledge degraded performance in methods. Or fall back to Cicero co-accessibility."

Method Combination

"Run both ABC and ENCODE-rE2G; intersect with HiChIP H3K27ac loops at FDR < 0.05. Report the triple-overlapping enhancer-gene pairs as high-confidence and the union as exploratory."

CRISPRi Validation

"Compare my ABC predictions in K562 against the Fulco 2019 CRISPRi-FlowFISH catalog. Report sensitivity and specificity at threshold ABC.Score >= 0.02."

Variant-to-Gene

"I have a GWAS lead SNP in a non-coding region. Identify the candidate enhancer (ATAC + H3K27ac peak overlap), then use ABC to find the most likely target gene."

What the Agent Will Do

1. Verify available inputs (ATAC, H3K27ac, Hi-C, gene annotations)
2. Choose method by data availability (decision tree above)
3. For ABC: define candidate enhancers (non-promoter ATAC peaks), compute Activity (ATAC * H3K27ac), Contact (Hi-C), apply ABC formula
4. For ENCODE-rE2G: choose pre-trained model by cell-type proximity; run Snakemake pipeline
5. For HiChIP: call loops at FDR < 0.05 + contact count >= 5; intersect with ABC
6. Cross-validate against CRISPRi-FlowFISH ground truth where available
7. Threshold at recommended cutoffs (ABC >= 0.02; rE2G >= 0.5)
8. Report intersection of methods as high-confidence; union as exploratory
9. Document cell-type proxies if exact match not available

Decision Quick Reference

Threshold Quick Reference

Tips

• ABC requires matched Hi-C from the same cell type; cross-celltype Hi-C performs degraded but is documented as an acceptable fallback (Fulco 2019).
• ENCODE-rE2G's logistic regression weights are cell-type-specific; using a wrong cell type's weights produces calibration errors. Match by tissue similarity.
• ABC's Activity term combines ATAC and H3K27ac as the geometric mean of normalized signals; both must be on the same scale (use RPGC for ATAC, RPM-IP for H3K27ac).
• Cicero co-accessibility is NOT enhancer-gene linking. Cicero predicts co-varying peak pairs; ABC predicts target genes. Concordance is ~30-50%.
• HiChIP H3K27ac directly measures enhancer-promoter loops; orthogonal to accessibility-based methods.
• For high-confidence reporting, agree across two methods. Single-method calls are exploratory.
• CRISPR enhancer-screen ground-truth catalogs (Fulco 2019 K562 FlowFISH, Gasperini 2019 K562, Schraivogel 2020 K562 TAP-seq) are primarily K562 with ~thousands of enhancer-gene pairs; the ENCODE-rE2G combined CRISPR benchmark is what spans ~10 cell types. Benchmark predictions against these.
• For GWAS variant interpretation, combine with deep-learning-atac (chromBPNet) variant effect predictions; SNP at predicted enhancer + ABC target gene + chromBPNet effect = strong functional hypothesis.
• ENCODE 4 published pre-computed ENCODE-rE2G predictions for many cell types -- check availability before retraining.
• EpiMap and GeneHancer are useful baselines but cell-type-agnostic; do not use as primary calls for cell-type-specific biology.

Related Skills

• atac-seq/co-accessibility - Cicero (ATAC-only fallback)
• atac-seq/atac-peak-calling - Enhancer candidate generation
• atac-seq/consensus-peakset - Fixed-width enhancer regions
• atac-seq/deep-learning-atac - chromBPNet variant effect prediction
• atac-seq/single-cell-atac - Per-cell-type scATAC inputs
• hi-c-analysis/loop-calling - Hi-C / Micro-C contact prediction
• hi-c-analysis/contact-pairs - Hi-C / Micro-C input
• chip-seq/peak-calling - H3K27ac peaks
• gene-regulatory-networks/scenic-regulons - TF -> target downstream