ChromBPNet iPSC-EC Differentiation Pipeline

Deep learning-based chromatin accessibility analysis (ChromBPNet) across 5 iPSC-to-EC differentiation timepoints (d0–d4), followed by motif discovery and TF activity analyses.

---

Repository structure

pipeline/          ChromBPNet workflow: preprocessing -> training -> motif discovery
analysis/          Downstream analyses: QC, motif annotation, gene-motif linking, TF motif enrichment in cNMF programs
data/              Input data: BAMs, peak BED files, scE2G links
data/folds/        Cross-validation fold definitions (JSON) used by the pipeline

Large output directories (gitignored):

results/preprocessing/        Preprocessed peaks, non-peaks, genome reference
results/bias_models/          Trained Tn5 bias models
results/full_models/          Trained ChromBPNet models (5 days x 5 folds)
results/contrib_scores/       Fold-averaged contribution scores + MoDISco results
results/compendium/           Unified motif compendium (MotifCompendium clustering)
results/finemo_unified/       Cross-day comparable motif hit calls (Fi-NeMo)
results/predictions/          Predicted accessibility BigWig tracks
results/plots/                QC and analysis output plots

---

Quick start

1. Configure paths

Edit pipeline/config.sh:

• core_path is auto-detected from the script location; override only if needed
• Edit pipeline/dataset_config.sh for days, folds, and input peak/fragment paths
• Place hg38.fa, hg38.chrom.sizes, and blacklist.bed.gz in genome/
• #SBATCH partition/QOS lines are generic (normal, gpu); adjust them for your cluster
• Set CONDA_INIT, CONDA_ENV, finemo_conda, motif_compendium_conda to your conda paths (envs in envs/)
• Set ref_db_meme to the path of MotifCompendium-Database-Human.meme.txt

2. Run the pipeline

All scripts are in pipeline/. Run from that directory:

cd pipeline/

# Stage 1: Preprocess peaks and non-peaks
bash 01.preprocess_peaks.sh
bash 02.preprocess_nonpeaks.sh

# Stage 2: Train bias models, select per fold, QC, train full model
sbatch 03.0.train_bias_model.sh
bash 03.1.select_bias.sh            # after 03.0; update fold_bias_suffix in dataset_config.sh
sbatch 03.2.qc_selected_bias.sh
sbatch 04.0.train_full_model.sh
sbatch 04.1.qc_run_full_model.sh    # optional model QC
sbatch 04.2.generate_predictions.sh # optional predicted-accessibility BigWigs

# Stage 3: Contribution scores, fold averaging, MoDISco
sbatch 05.get_contrib_scores.sh
sbatch 06.average_contrib_scores.sh
sbatch 07.contribs_to_bigwig.sh
sbatch 08.run_modisco.sh

# Stage 4: Unified motif compendium and hit calling (counts head)
sbatch 09.motif_compendium.sh
sbatch 10.run_finemo_unified.sh
sbatch 11.postprocess_finemo.sh
# Profile-head variants: 09.1.motif_compendium_profile.sh,
# 10.1.run_finemo_unified_profile.sh, 11.1.postprocess_finemo_profile.sh

See pipeline/README.md for full documentation.

3. Run downstream analyses

Scripts in analysis/ are numbered in run order:

cd analysis/
python 00.chrombpnet_model_qc.py      # model and data QC
python 0.0.motif_id_annotation.py     # TF annotation of compendium motifs
python 0.2.finemo_hit_qc.py           # CWM-similarity filtering of hits
sbatch 0.3.hits_to_bed.sh             # IGV tracks
python 1.0.gene_motif_linking.py      # TF scores linked to genes via scE2G
python 2.tf_enrichment_in_programs.py # TF motif enrichment in cNMF programs
python 3.chrombpnet_figures.py        # figures for programs of interest

See analysis/README.md for full documentation.

---

Dependencies

Tool	Purpose
ChromBPNet	Bias-factorised accessibility model
TF-MoDISco (modiscolite)	Motif discovery from contribution scores
MotifCompendium	Cross-day motif clustering and JASPAR annotation
Fi-NeMo	Motif hit calling from contribution scores
scanpy / muon	Single-cell data handling
pybedtools	Genomic coordinate operations

Conda environments used:

• chrombpnet: main pipeline (ChromBPNet, modiscolite)
• finemo: Fi-NeMo hit calling
• motif_compendium: MotifCompendium clustering

---

References

Genome: hg38 Motif annotation: MotifCompendium-Database-Human (download from https://github.com/kundajelab/MotifCompendium)