ACCESS-ATAC

Analysis code for ACCESS-ATAC, an assay that measures chromatin accessibility two ways from a single library:

• ATAC signal — Tn5 insertion (cut) sites, derived from fragment ends with a +4 / −5 offset.
• ACCESS signal — per-base DddSs cytosine deaminase edits, read off alignments as C→T (forward strand) and G→A (reverse strand) mismatches against the reference.

Because both signals come from the same molecules they can be compared directly, and that comparison is what most of this repository does.

The top panel is the assay: Tn5 and the DddSs deaminase act on open chromatin in the same reaction, Tn5 fragmenting the DNA and the deaminase converting exposed cytosines, so one library carries both readouts. The bottom panel is the downstream analysis, covered here by 01_process_access/ (genome-wide accessibility and TF footprinting), 03_accessbpnet/ (sequence-to-function modelling, the ACCESSBPNet panel) and 04_single_cell_access_atac/ (the single-cell panel). 02_tfbs_prediction/ is not depicted.

Contents

Four analysis pipelines, each self-contained and each with its own README documenting how to run it and which paths need adapting.

Pipeline	What it does
01_process_access/	Concurrent ACCESS-ATAC in HepG2 and K562: QC, filtering, depth matching, signal tracks, peak calling and TF footprint quantification, compared against ENCODE ATAC-seq. Start here.
02_tfbs_prediction/	Benchmark of TF binding site prediction from ACCESS-ATAC signal versus sequence alone. Builds the training data, trains a CNN classifier and evaluates it.
03_accessbpnet/	ChromBPNet applied to ACCESS-ATAC: bias model, ChromBPNet model, contribution scores, TF-MoDISco, marginal footprints and variant effect prediction.
04_single_cell_access_atac/	Single-cell ACCESS-ATAC in mouse airway cells: barcode correction, fragment generation, clustering and annotation (ArchR / Seurat), and TF footprinting.

Reproducing the analysis

Within each pipeline, scripts run in the numeric order of their filenames; notebooks are run interactively at the point where their number falls. The shell scripts are SLURM job scripts, submitted with sbatch.

Each pipeline bundles the Python it needs — under scripts/, model/ or plotting/ — so no pipeline reaches into another's code. Those bundled scripts are argparse command-line tools that import their siblings by bare module name (from utils import ...). That resolves against the directory the script itself lives in, so they are invoked by path from the pipeline directory:

cd 04_single_cell_access_atac
python scripts/fragment_to_bw_atac.py \
    --input_fragments fragments.tsv.gz \
    --chrom_size_file genome.chrom.sizes \
    --out_dir out --out_name sample_atac

Shared conventions: paired --out_dir / --out_name arguments produce {out_dir}/{out_name}.{ext}; coordinates are 0-based half-open (BED convention); output directories are not created automatically.

The pipelines are chained — 02 and 03 consume signal tracks produced by 01 — and all of them read reference data and raw input from paths that are hard-coded for the cluster they were run on. Each pipeline's README lists those paths in a "paths that must be adapted" table.

Dependencies

Python: numpy, scipy, pandas, polars, pyranges, pysam, pyBigWig, pyfaidx, pybedtools, numba, torch, scikit-learn, matplotlib, seaborn, biopython, tqdm, plus h5py, hdf5plugin and logomaker for the ChromBPNet evaluation notebooks.

R (single-cell analysis in 04_single_cell_access_atac/, 13 of its notebooks use an R kernel): ArchR, Seurat, Signac, rtracklayer, BSgenome.Mmusculus.UCSC.mm39, ggplot2, dplyr, tidyr, tibble, cowplot, pheatmap, ComplexUpset, openxlsx, future.

External tools: samtools, bedtools, deeptools, MACS2, UCSC wigToBigWig and bedGraphToBigWig, RGT for motif matching, and Nextflow for the alignment step of pipeline 04.

Pipelines 01 and 03 build their signal tracks with deamtools bam2bw. deamTools is a separate tool for deaminase-based assays, developed alongside this work and still under active development (docs).

Training the TFBS classifier (02_tfbs_prediction/04_train.sh) and the ChromBPNet models in 03_accessbpnet/ require a CUDA device.

Third-party code

03_accessbpnet/chrombpnet/ is a copy of ChromBPNet (MIT, Copyright 2019 Kundaje Lab) with local modifications, taken from the access branch of lzj1769/chrombpnet. Its LICENSE is included unchanged, and the modifications are described in 03_accessbpnet/chrombpnet/README.md.