📄 / / profiles / r-and-d / raw_evidence / OpenAI4S / skills / bioskills / bio-atac-seq-single-cell-atac / references / usage-guide.md

Single-Cell ATAC-seq - Usage Guide

Single-Cell ATAC-seq - Usage Guide

Overview

End-to-end single-cell ATAC-seq processing and analysis. Covers ecosystem selection (Signac, ArchR, SnapATAC2), per-cell QC thresholds (looser than bulk because of per-cell sparsity), doublet detection (AMULET, ArchR, scDblFinder), TF-IDF + LSI dimensionality reduction (skipping the depth-correlated first component), per-cluster pseudobulk peak calling, and 10X Multiome RNA+ATAC integration via WNN. Includes per-tool failure modes and reconciliation between ecosystems.

Prerequisites

# Signac ecosystem
BiocManager::install(c('GenomicRanges', 'EnsDb.Hsapiens.v86', 'BSgenome.Hsapiens.UCSC.hg38'))
install.packages(c('Signac', 'Seurat'))

# ArchR ecosystem
remotes::install_github('GreenleafLab/ArchR', ref='master', repos=BiocManager::repositories())
ArchR::installExtraPackages()

# SnapATAC2 (Python)
pip install snapatac2

# Doublet detection: AMULET is a GitHub release (jar + AMULET.sh; needs numpy/pandas/scipy/statsmodels + Java 8+)
# from github.com/UcarLab/AMULET, or use the R amulet() implementation in the scDblFinder Bioconductor package

BiocManager::install('scDblFinder')

Inputs (from 10X Cell Ranger ATAC or Multiome): - outs/fragments.tsv.gz (and .tbi index) - outs/filtered_peak_bc_matrix.h5 (Signac path) - outs/singlecell.csv (per-barcode metadata)

Quick Start

Tell your AI agent what you want to do: - "Process 10X scATAC with Signac: TF-IDF + LSI (dims 2:30, skip depth) + UMAP + Leiden" - "Use ArchR for a 200K-cell dataset because Signac will run out of memory" - "Run AMULET for ATAC-specific doublet detection; cross-validate with ArchR's doublet score" - "Build per-cluster pseudobulk peaks with MACS3, then iterative-overlap consensus" - "Integrate 10X Multiome via WNN with PCA on RNA and LSI(2:30) on ATAC" - "Re-filter cellranger output at fragment count >= 1000 AND TSS enrichment >= 4 because cellranger's cell calling is too lenient"

Example Prompts

Standard Signac Pipeline

"Load 10X scATAC h5 + fragments file into Signac. Compute NucleosomeSignal and TSSEnrichment. Filter cells: peak fragments 1000-20000, FRiP >= 15%, blacklist ratio < 0.05, nucleosome signal < 4, TSS enrichment > 4. Then RunTFIDF -> RunSVD -> RunUMAP(dims=2:30) -> FindNeighbors(dims=2:30) -> FindClusters(algorithm=4, resolution=0.5)."

Large Dataset with ArchR

"150K-cell scATAC dataset. Use ArchR: createArrowFiles with minTSS=4 and minFrags=1000, addDoubletScores, filterDoublets, addIterativeLSI on TileMatrix, addClusters Leiden, addUMAP. Then addReproduciblePeakSet per cluster."

SnapATAC2 Python Workflow

"Use SnapATAC2: import fragments.tsv.gz via snap.pp.import_fragments with hg38 chrom_sizes, compute tsse, filter min_counts=1000 and min_tsse=4, add_tile_matrix bin_size=500, select_features 250k, spectral, leiden, umap."

Doublet Reconciliation

"Run AMULET (collision-based) and ArchR's doublet score in parallel; flag cells called doublet by both as high-confidence; report flagging rate per cluster."

Per-Cluster Pseudobulk DA

"Aggregate cells per cluster into pseudobulk BAMs, run MACS3 callpeak per cluster, build iterative-overlap consensus across clusters, then DESeq2 differential between cluster A and B (atac-seq/differential-accessibility)."

Multiome WNN

"I have Multiome data. Run RNA workflow (LogNormalize -> PCA -> dims 1:30) and ATAC workflow (TF-IDF -> SVD -> dims 2:30) separately, then FindMultiModalNeighbors and RunUMAP nn.name='weighted.nn' for the joint embedding."

Cell Type Annotation

"Compute gene activity scores with Signac::GeneActivity (or ArchR getGeneScore); transfer cell type labels from a reference scRNA-seq atlas via Seurat::FindTransferAnchors."

What the Agent Will Do

Verify input is Cell Ranger ATAC or Multiome output (chemistry version matters)
Re-filter cellranger cell calls at fragment >= 1000 AND TSS enrichment >= 4
Compute per-cell QC: NucleosomeSignal, TSS enrichment, FRiP, mt fraction, blacklist ratio
Run doublet detection (AMULET primary; ArchR or scDblFinder confirmatory)
Choose ecosystem (Signac for Multiome / Seurat user; ArchR for large; SnapATAC2 for Python)
TF-IDF + LSI/spectral; SKIP depth-correlated first component (dims 2:30)
UMAP + Leiden clustering
Per-cluster pseudobulk peak calling (MACS3 with ATAC parameters)
Iterative-overlap consensus across clusters
Gene activity scores for cell type annotation
Integration with paired scRNA-seq if Multiome (WNN)
Optional: trajectory analysis (ArchR getTrajectory or Cicero)

Per-Cell QC Quick Reference

Metric	Pass	Reject
Fragment count	3000-50000	< 1000 or > 80000
TSS enrichment	>= 4	< 2
FRiP	>= 0.15	< 0.05
Mt fraction	< 0.05	> 0.20
Nucleosome signal	< 4	> 10
Doublet score	< 0.5 (AMULET)	> 0.7

Per-cell thresholds are looser than bulk; cells are sparse by design.

Ecosystem Quick Reference

Use case	Ecosystem
10X Multiome (RNA + ATAC)	Signac (Seurat integration)
Large dataset (>100K cells)	ArchR (Arrow files; memory-efficient)
Python ecosystem	SnapATAC2
Trajectory analysis	ArchR getTrajectory; Cicero (Signac)
ATAC-only standard analysis	Either Signac or ArchR

Doublet Detection Quick Reference

Tool	Method	When to use
AMULET	Collision-based	Primary for ATAC; orthogonal to clustering
ArchR addDoubletScores	Synthetic + LSI projection	Built into ArchR; auto-thresholds
scDblFinder	Synthetic + classifier	RNA-developed; works on Signac/SCE

Run two; intersection of flagged cells is high-confidence doublet set.

Tips

The single most common bug: including LSI component 1 in clustering. It correlates with depth, not biology. Always use dims=2:30.
Cell Ranger ATAC's cell calling is lenient. Always re-filter at fragment >= 1000 AND TSS enrichment >= 4.
Per-cell thresholds are looser than bulk; the population aggregate is what matters statistically.
AMULET is ATAC-specific (collision detection from 2-allele biology); pair with ArchR or scDblFinder for orthogonal evidence.
Multiome WNN equal-weights modalities by default; ATAC's sparseness can dominate noise. Inspect weights and adjust if needed.
Per-cluster pseudobulk peak calling needs >= 200 cells per cluster; aggregate small clusters or use union peaks.
Gene activity scores (Signac::GeneActivity / ArchR getGeneScore) are approximate but useful for cross-modality cell-type transfer.
chromVAR / AddMotifs MUST run after peakset finalized; otherwise motif annotations become stale.
For trajectory analysis, ArchR's getTrajectory is more polished than Cicero; Cicero is better for cis-regulatory inference.
Plant / non-model organisms need custom EnsDb / TxDb / BSgenome objects; build with txdbmaker or AnnotationDbi.
10X Multiome ATAC peakset MUST be called from the Multiome ATAC fragments, not transferred from a separate scATAC dataset.

atac-seq/atac-qc - Bulk QC adapted for per-cell
atac-seq/atac-peak-calling - Pseudobulk peak calling
atac-seq/consensus-peakset - Per-cluster consensus
atac-seq/differential-accessibility - Pseudobulk DA
atac-seq/motif-deviation - chromVAR for per-cell TF activity
atac-seq/footprinting - scprinter for sc footprinting
atac-seq/co-accessibility - Cicero / SCENIC+
single-cell/preprocessing - General sc patterns
single-cell/clustering - Cluster definition
single-cell/cell-annotation - Automated reference-based label transfer
single-cell/multimodal-integration - Multiome integration
single-cell/scatac-analysis - Cross-reference single-cell ATAC-specific patterns