Deep Learning for ATAC-seq - Usage Guide

Overview

Sequence-based neural networks for ATAC-seq: chromBPNet (bias-corrected per-base accessibility), BPNet (foundational profile model), scBasset (per-cell sequence model), Enformer (long-context Transformer), and Borzoi (multi-tissue chromatin + RNA). Covers in silico variant effect prediction at GWAS / rare-variant SNPs, motif discovery via DeepLIFT + TF-MoDISco, deep bias correction beyond TOBIAS k-mer model, and per-cell-type accessibility prediction for unobserved states.

Prerequisites

# Core deep learning frameworks
pip install tensorflow==2.13 torch torchvision torchaudio

# Tools
pip install chrombpnet bpnet-lite tangermeme modisco-lite captum kipoi

GPU with >= 16 GB VRAM recommended for training (A100/V100); CPU acceptable for inference with pre-trained models. Training data: deduplicated BAM, peaks/non-peaks BED, reference FASTA, chrom sizes.

Quick Start

Tell your AI agent what you want to do: - "Score 100 GWAS SNPs for chromatin effect using a pre-trained chromBPNet K562 model" - "Run chromBPNet pipeline (bias + accessibility) on a new ATAC dataset" - "Compute DeepLIFT contributions and run TF-MoDISco-lite to discover de novo motifs" - "Use chromBPNet bias-corrected bigWig as input to TOBIAS instead of ATACorrect" - "Run Enformer pre-trained on a 196 kb window to predict variant effects in distal regulation" - "Train scBasset on my scATAC for per-cell accessibility prediction"

Example Prompts

Variant Effect Prediction

"I have 50 GWAS lead SNPs in autoimmune loci. Use chromBPNet pre-trained models for the closest cell types (CD4 T, B cells), score ref vs alt with tangermeme marginal prediction, and report log2FC magnitudes. SNPs with abs(log2FC) > 1 are strong-effect candidates."

chromBPNet Training

"Train a per-cell-type chromBPNet on my 80M-read primary T cell ATAC. Generate splits, train bias model from non-peaks, train accessibility model with bias correction, then use the corrected bigWig for downstream footprinting."

TF-MoDISco Motif Discovery

"Compute DeepLIFT contributions for all peaks in my chromBPNet model output, run TF-MoDISco-lite with sliding_window_size=20 and target_seqlet_fdr=0.05, then check motif clusters against JASPAR for known TFs."

chromBPNet Bias Correction Replacement

"Use chromBPNet's bias-corrected bigWig output as input to TOBIAS ScoreBigwig (formerly FootprintScores), skipping TOBIAS ATACorrect. Compare aggregate CTCF footprint between chromBPNet-corrected and TOBIAS-corrected to verify both produce clean V-shapes."

Enformer Variant Effect

"For a SNP in a non-coding distal element, use Enformer pre-trained model: predict for ref and alt 196 kb windows centered on the SNP, report differential predictions across HepG2/K562/GM12878 tracks for cell-type-specific effects."

scBasset for scATAC

"Train scBasset on my 50k-cell scATAC-seq dataset; use the per-cell projection layer to identify cells with anomalous accessibility patterns and compare cluster discrimination to chromVAR."

What the Agent Will Do

Determine if classical pipelines suffice, or whether DL is genuinely needed (variant effect, cross-cell-type, fine-grained motif)
Check for pre-trained model availability at encodeproject.org/atac-seq before training new
Verify input data: dedup, peaks, chrom sizes, reference FASTA
Prepare splits (train / val / test chromosomes)
Train bias model first (from naked-DNA control if available; otherwise non-peak background)
Train accessibility model with bias correction
For variant effect: tangermeme marginal prediction at SNPs of interest
For motif discovery: DeepLIFT contributions -> TF-MoDISco-lite clustering
For sc state interpolation: scBasset training on metacells
Document model versions and pre-trained model sources for reproducibility

Tool Decision Quick Reference

Goal	Tool
Score GWAS SNPs in known cell types	Pre-trained chromBPNet + tangermeme
Score GWAS SNPs in novel cell types	Train chromBPNet first
Long-range distal regulation variants	Enformer or Borzoi (196 kb context)
Motif discovery from new ATAC data	chromBPNet + DeepLIFT + TF-MoDISco-lite
Bias correction for footprinting	chromBPNet bias-corrected bigWig as TOBIAS input
Per-cell accessibility prediction	scBasset
Cross-cell-type accessibility	Enformer pre-trained

Compute Quick Reference

Task	Hardware	Wall time
chromBPNet training (1 cell type)	A100, 80 GB	~24 h
chromBPNet inference 1M variants	A100	~4 h
Enformer pre-trained inference	V100+	~30 min / 100k variants
Borzoi training	A100, 250 GB	~7 days
scBasset training (10k cells)	V100, 32 GB	~12 h
TF-MoDISco on 1M peaks	CPU 32 cores	~6 h

Tips

For most ATAC analysis, classical pipelines (MACS3 + DiffBind + TOBIAS) remain primary. Deep learning is justified when variant interpretation, cross-cell-type prediction, or fine motif discovery are the goal.
Pre-trained models exist for ENCODE cell types (K562, GM12878, HepG2, HUVEC, IMR90 ATAC). Check the chromBPNet model zoo before training.
Bias model and accessibility model are trained sequentially; the bias must be trained first because the accessibility model uses bias-corrected residuals.
chromBPNet's bias model is mostly cell-type-invariant. A K562 bias model used on T cells works with degraded performance; better than nothing.
DeepLIFT vs Integrated Gradients give different attribution patterns. Use DeepLIFT (chromBPNet default) unless documented reasons exist for IG.
Variant effect prediction window matters: short windows (~1-2 kb) for proximal regulation; Enformer/Borzoi (196 kb) for distal.
TF-MoDISco-lite (its interface is TF-MoDISco v2) is the maintained version; the original kundajelab/tfmodisco (v0.x) is for replication only.
scBasset's per-cell projection is unstable below 100 cells per cluster; aggregate cells before training.
Marginal effects (ref vs alt at the SNP only) are different from in silico mutagenesis (saturation across all positions). Choose by question.
Borzoi is newer than Enformer with better chromatin + RNA joint prediction; use for variant effects on RNA via chromatin linkage.
For high-confidence variant prediction, agree across two approaches (chromBPNet + Enformer/Borzoi). Single-tool calls are exploratory.

atac-seq/atac-peak-calling - Classical input pipeline
atac-seq/footprinting - Use chromBPNet bias correction as TOBIAS alternative
atac-seq/motif-deviation - chromVAR vs scBasset for sc TF activity
atac-seq/single-cell-atac - scBasset integration
atac-seq/enhancer-gene-linking - Variant effect feeds enhancer scoring
atac-seq/allele-specific-accessibility - DL predictions vs observed allelic imbalance
causal-genomics/fine-mapping - Downstream use of variant effect scores
machine-learning/biomarker-discovery - General ML patterns
gene-regulatory-networks/scenic-regulons - Motif discovery + TF networks