What is scATAC-seq used for?

It maps chromatin accessibility genome-wide in individual cells, which you use to find cell-type-specific enhancers and promoters, infer regulatory heterogeneity within a tissue, and estimate transcription factor activity from motif accessibility. It's the regulome counterpart to scRNA-seq's transcriptome readout, and pairing the two is how most multiomics studies link accessibility changes to expression changes.

Why can't I just run PCA on scATAC-seq data like I would for scRNA-seq?

The peak-by-cell matrix is sparse and near-binary, not continuous, so PCA's assumptions don't hold and the top components usually end up tracking total fragments per cell instead of biology. TF-IDF normalization followed by LSI (SVD on the TF-IDF matrix) is the standard substitute, implemented in Signac as `RunTFIDF()` + `RunSVD()`.

Signac vs ArchR: which should I use?

Both run the same core pipeline (TF-IDF/LSI, clustering, UMAP) and both are actively used alongside tools like MAESTRO, Cicero, and cisTopic for more specialized steps. Signac keeps you inside the Seurat/R object model, useful if your scRNA-seq work is already there; beyond that, check each project's current docs and any recent benchmark before committing, since no authoritative head-to-head is settled here.

What does a scATAC-seq fragments file contain?

Cell Ranger ATAC outputs a `fragments.tsv.gz` with one line per unique fragment, giving its genomic coordinates and the cell barcode it came from. This file, not the aligned BAM, is what most downstream tools (Signac, MACS3's FRAG parser) expect as input.

Can scATAC-seq pinpoint exact transcription factor binding sites?

Not reliably on its own. Tn5 transposase cuts both nucleosome linker DNA and true open regulatory regions, and it carries its own sequence-cutting bias, which together can produce motif footprints that mimic TF binding without real occupancy. Treat scATAC-seq peaks as accessibility evidence, and corroborate specific binding claims with orthogonal data.

Glossary · Single-Cell and Spatial

scATAC-seq

The assay tells you where chromatin is open in each cell, not what any transcription factor is doing there, mixing up the two wrecks your interpretation before you've run a single test.

By Ming "Tommy" Tang, Director of Bioinformatics in Big Pharma · Reviewed September 2026 · 2 min read

Also: single-cell ATAC, Signac, ArchR

Definition

scATAC-seq (single-cell Assay for Transposase-Accessible Chromatin using sequencing) measures chromatin accessibility genome-wide in individual cells. Tn5 transposase cuts open DNA and tags it with sequencing adapters in one step, so read density marks regions that aren't wrapped by a nucleosome or occupied by a large protein complex. Per cell you get a sparse, near-binary count across tens of thousands of peaks rather than a continuous signal, and that sparsity is why scATAC-seq analysis diverges from scRNA-seq at normalization, dimensionality reduction, and clustering.

You meet scATAC-seq the moment someone hands you a Cell Ranger ATAC output, a fragments.tsv.gz file and a peak-by-cell matrix, and asks for clusters and cell types, the same deliverables you'd produce from scRNA-seq. The temptation is to reuse the scRNA-seq playbook: normalize, PCA, UMAP, Louvain. That playbook breaks here, and where it breaks decides whether your clusters reflect real cell states or just how many fragments each cell happened to get sequenced.

scATAC-seq (single-cell ATAC-seq) is the standard way to map chromatin accessibility at single-cell resolution. It's also the entry point to a small, opinionated toolchain, Signac in R, ArchR, MACS3 for peak calling, each with defaults built around the assay's specific failure modes: sparsity, binary counts, and an enzyme that isn't as selective as you'd like.

Why it matters

Get the normalization step wrong and the downstream biology is fiction. Human scATAC-seq datasets typically have 10², 10⁴ cells with only 10³, 10⁵ reads each, producing a peak matrix that's too sparse and binary for PCA's assumptions. Run PCA on the raw matrix and the first component usually tracks total fragments per cell, so your clusters split by sequencing depth, not cell type, the same failure mode that corrupts depth-imbalanced scRNA-seq batches, except here it's the default outcome rather than an edge case.

The standard fix, implemented in Signac as RunTFIDF() followed by RunSVD(), is TF-IDF normalization then LSI (SVD on the TF-IDF matrix). TF-IDF downweights peaks open in nearly every cell and upweights peaks that mark specific states, which is the signal you actually want before reducing dimensions. Skip it, and you can burn a week chasing a "batch effect" that's really a normalization bug.

Where people get it wrong

Two mistakes recur. First, treating Tn5 insertions as a direct readout of transcription factor binding. Tn5 tags nucleosome linker DNA and true TF-bound regions without distinction, and it carries its own sequence-cutting bias that can fabricate motif "footprints" resembling TF occupancy where none exists. If the question is precise binding-site location, scATAC-seq alone isn't the right assay, it reports general accessibility, not footprint-grade resolution.

Second, Signac vs ArchR gets picked by habit rather than by requirement. Both implement the same TF-IDF/LSI/clustering/UMAP core, alongside more specialized tools like MAESTRO, Cicero, and cisTopic for co-accessibility and topic modeling. Signac stays inside the Seurat/R object model, which matters if your scRNA-seq work already lives there. The research behind this entry doesn't include a current runtime or accuracy benchmark between the two, so before committing to one for a large project, check each tool's current documentation and any recent published comparison rather than trusting an old blog post's verdict.

A concrete example

A minimal Signac workflow on a peak-by-cell matrix: filter low-signal peaks, normalize with TF-IDF, reduce dimensions with LSI instead of PCA, then embed. LSI component 1 is conventionally excluded from downstream steps when it correlates with sequencing depth.

library(Signac)

# keep peaks accessible in enough cells
atac <- FindTopFeatures(atac, min.cutoff = 10)

# TF-IDF normalization, then LSI (SVD on the TF-IDF matrix)
atac <- RunTFIDF(atac)
atac <- RunSVD(atac, assay = "peaks", reduction.name = "lsi", n = 50)

# drop LSI component 1 if it correlates with depth, then embed
atac <- RunUMAP(atac, reduction = "lsi", dims = 2:30)

Related terms

Questions people ask

What is scATAC-seq used for?: It maps chromatin accessibility genome-wide in individual cells, which you use to find cell-type-specific enhancers and promoters, infer regulatory heterogeneity within a tissue, and estimate transcription factor activity from motif accessibility. It's the regulome counterpart to scRNA-seq's transcriptome readout, and pairing the two is how most multiomics studies link accessibility changes to expression changes.
Why can't I just run PCA on scATAC-seq data like I would for scRNA-seq?: The peak-by-cell matrix is sparse and near-binary, not continuous, so PCA's assumptions don't hold and the top components usually end up tracking total fragments per cell instead of biology. TF-IDF normalization followed by LSI (SVD on the TF-IDF matrix) is the standard substitute, implemented in Signac as RunTFIDF() + RunSVD().
Signac vs ArchR: which should I use?: Both run the same core pipeline (TF-IDF/LSI, clustering, UMAP) and both are actively used alongside tools like MAESTRO, Cicero, and cisTopic for more specialized steps. Signac keeps you inside the Seurat/R object model, useful if your scRNA-seq work is already there; beyond that, check each project's current docs and any recent benchmark before committing, since no authoritative head-to-head is settled here.
What does a scATAC-seq fragments file contain?: Cell Ranger ATAC outputs a fragments.tsv.gz with one line per unique fragment, giving its genomic coordinates and the cell barcode it came from. This file, not the aligned BAM, is what most downstream tools (Signac, MACS3's FRAG parser) expect as input.
Can scATAC-seq pinpoint exact transcription factor binding sites?: Not reliably on its own. Tn5 transposase cuts both nucleosome linker DNA and true open regulatory regions, and it carries its own sequence-cutting bias, which together can produce motif footprints that mimic TF binding without real occupancy. Treat scATAC-seq peaks as accessibility evidence, and corroborate specific binding claims with orthogonal data.

Compare · Seurat vs Scanpy: Which One Should You Use?

Sources

Single-cell ATAC-seq Signal Extraction and Enhancement with SCATE — Tn5-based assay definition and typical dataset scale/sparsity
Computational Analyses and Challenges of Single-cell ATAC-seq — six-component analysis workflow and tool landscape (ArchR, Signac, MAESTRO, Cicero, cisTopic)
Cell Ranger ATAC Fragments File Documentation — fragments.tsv.gz format
MACS3 callpeak Documentation — MACS3 FRAG parser and peak-calling command on 10x fragment files