← Về thư mục
📄 / / profiles / r-and-d / raw_evidence / OpenAI4S / skills / bioskills / bio-chip-seq-chip-deep-learning / references / usage-guide.md

Deep Learning for ChIP-seq - Usage Guide

Overview

Apply base-resolution deep learning models to ChIP-seq, ChIP-nexus, and CUT&RUN data for variant-effect prediction, motif syntax discovery, and sequence-only signal prediction. Covers BPNet (Avsec 2021), chromBPNet (Pampari et al 2024 bioRxiv), EnFormer (Avsec 2021), DeepSEA, and the JASPAR 2026 Deep Learning collection (1259 precomputed BPNet ChIP models). Embeds in silico mutagenesis (counterfactual variant prediction), TF-MoDISco motif discovery from attribution scores, and model-ensemble uncertainty estimation.

Prerequisites

# chromBPNet (modern; handles ChIP and ATAC)
pip install chrombpnet
# or conda for full env: conda install -c bioconda chrombpnet

# BPNet (canonical for ChIP-nexus)
pip install bpnet

# EnFormer
pip install enformer-pytorch

# TF-MoDISco
pip install tfmodisco-lite

# DeepLIFT / SHAP for attribution
pip install shap

# CUDA-capable GPU recommended (16+ GB VRAM for training)
# CPU inference possible for precomputed models

Quick Start

Tell the agent what to do: - "Train chromBPNet on my CTCF ChIP-seq peaks for variant effect prediction" - "Predict the effect of a SNP on TF binding using in silico mutagenesis with chromBPNet" - "Run TF-MoDISco on attribution scores from a trained BPNet model to discover motif syntax" - "Use a precomputed JASPAR 2026 BPNet model to score a list of GWAS variants for TF binding effects" - "Apply EnFormer for variant effect prediction (196 kb input window, ~100 kb effective receptive field; captures distal regulatory effects)" - "Compare chromBPNet and EnFormer variant predictions for a list of fine-mapped SNPs"

Example Prompts

Variant effect prediction

"For a list of fine-mapped SNPs from GWAS, predict the effect on TF binding using chromBPNet. Report |log2_fc| > 1 as strong-effect, 0.3-1 as moderate, <0.3 as weak. Concordance with EnFormer increases confidence."

Train chromBPNet from scratch

"Train chromBPNet on FOXA1 ChIP-seq in HepG2 cells. First train a bias model on non-peak regions, then train the main model. Use ENCODE-matched non-peak background."

TF-MoDISco motif discovery

"Compute DeepLIFT attribution scores from a trained BPNet model, then run TF-MoDISco to discover motif syntax including cooperative TF interactions. Compare to PWM motifs from MEME-ChIP."

Use JASPAR precomputed

"Download the BPNet model for FOXA1 from JASPAR 2026 Deep Learning collection. Use it to predict variant effects on FOXA1 binding for my list of SNPs without training."

Cross-model validation

"Predict the same variant effect with chromBPNet and EnFormer. If they agree (both |log2_fc| > 1 in same direction), high-confidence functional variant. If they disagree, low-confidence."

Long-range prediction

"Use EnFormer's long-range model (196 kb input window, ~100 kb effective receptive field) to predict whether a distal enhancer SNP affects target gene expression. Native EnFormer outputs include CAGE-seq signal at TSS."

Ensemble uncertainty

"Train 5 chromBPNet replicate models with different seeds. Report ensemble mean + std for each variant. High std indicates model extrapolation; low std with consistent prediction is high-confidence."

What the Agent Will Do

  1. Choose model based on goal:
  2. Single TF, ChIP-nexus data: BPNet
  3. TF or histone mark, regular ChIP, with bias factorization: chromBPNet
  4. Long-range variant effects: EnFormer
  5. Pre-trained models without training cost: JASPAR 2026 deep-learning collection
  6. For training:
  7. Verify ≥5000 high-confidence peaks per TF
  8. Train bias model on non-peak background
  9. Train main model with cross-validation folds (chromBPNet uses 5 folds by default)
  10. Monitor training: Spearman correlation between predicted and observed profiles should be ≥0.6
  11. For variant prediction:
  12. Encode ref and alt sequence (2114 bp window for chromBPNet)
  13. Predict counts and profile for each
  14. Compute log2_fc = log2(alt_counts / ref_counts)
  15. Apply ensemble for uncertainty
  16. For motif discovery:
  17. Compute attribution scores (DeepLIFT or SHAP)
  18. Run TF-MoDISco with appropriate parameters
  19. TOMTOM the discovered motifs against JASPAR / HOCOMOCO for known matches
  20. Identify motifs not in known DBs as candidate novel syntax
  21. Output:
  22. Variant-effect TSV with predicted log2_fc per variant
  23. Attribution score plots (sequence logo of importance)
  24. TF-MoDISco motif logos
  25. Document: model version, training data, hyperparameters, ensemble size, validation metrics

Tips

Troubleshooting

Training accuracy low (Spearman < 0.5)

  1. Insufficient peaks -> need ≥5000 high-confidence
  2. Wrong bias model -> train on matched assay
  3. Library quality poor -> check chipseq-qc upstream
  4. Hyperparameters -> check chromBPNet recommended config

Variant predictions noisy / unstable

  1. Single model; train ensemble of 5-10 with different seeds
  2. Variant outside training distribution; extrapolation; experimental validation needed
  3. Variant in low-quality region (low mappability, repeat); predictions unreliable

TF-MoDISco produces empty motif list

  1. FDR too strict; lower target_seqlet_fdr to 0.10
  2. Attribution scores too noisy; train model longer; use ensemble attribution
  3. Window size mismatch; adjust sliding_window_size

GPU out of memory

  1. Reduce batch_size
  2. Use gradient checkpointing
  3. For EnFormer, use the 5x smaller architecture variant

import chrombpnet fails

TensorFlow version mismatch. Use the conda env from chrombpnet repo with pinned tf==2.13.

EnFormer predictions for short variant don't capture biology

Variant >100 kb from any annotated regulatory element; effect may be too distal or not captured by training data. Combine with chromBPNet for local effects.

Chromosome naming mismatch with model

JASPAR / EnFormer models expect specific genome assembly and chromosome conventions. Verify hg38 chromosome naming (chr1 vs 1).