scATAC-seq

Start a scATAC-seq analysis

In a project, click New session and choose scATAC-seq as the kind of analysis. The sidebar marks it with a peak inside a cell. For bulk ATAC-seq (one library per sample) choose ATAC-seq instead.

What to upload

Per sample, what Cell Ranger ATAC leaves in its outs folder (Cell Ranger ARC's ATAC half reads the same), any one of:

• The fragments, with the metrics and the peaks (best): fragments.tsv.gz, singlecell.csv (ARC: per_barcode_metrics.csv), which says which barcodes are cells, and peaks.bed. Every cell's insertions are counted in the peaks, and its quality measured on the way.
• The peak matrix: filtered_peak_bc_matrix.h5 or the filtered_peak_bc_matrix folder (matrix.mtx, barcodes.tsv, peaks.bed). Add the fragments too for the quality measures, the doublet test and the coverage tracks; without them the quality is Cell Ranger's metrics, if you add those.
• An .h5ad of cells by peaks (SnapATAC2, or ArchR and Signac objects exported), its cells' metadata kept.

Several samples: drop each sample's outs folder (the folder's name is the sample's), or start each file's name with its sample (PBMC1_fragments.tsv.gz, as GEO names them), or upload a .zip. When the samples' peaks differ, as Cell Ranger calls them per sample, they are merged where they overlap (Signac's reduce) and every sample is counted again in the merged peaks. With fragments and no peaks at all, the genome is cut into 500 bp tiles, as ArchR and SnapATAC2 do. Without the metrics or a matrix, the cells are called from their fragments (Cell Ranger's order-of-magnitude rule).

Optional: a GTF or GFF of the genes, for a genome UCSC does not have. The tabix index (.tbi) and the BAM are not needed: the fragments hold what is read.

From GEO: type a series number in Import from GEO on the session's start screen. The lookup offers each sample's fragments, metrics, peak matrix and peaks.

What the upload measures

One pass over each sample's fragments counts the insertions (both ends of every fragment, as Cell Ranger counts cut sites) in each peak, and for every cell:

• Fragments, and the fraction of insertions in peaks;
• TSS enrichment, ArchR's per-cell definition: insertions per base within 50 bp of a transcription start site over insertions per base 1.9 to 2 kb away;
• nucleosome signal, Signac's: fragments of one nucleosome (147–294 bp) over nucleosome-free ones;
• blacklist ratio: insertions in ENCODE's blacklist over insertions in peaks;
• AMULET's overlaps (Thibodeau et al. 2021): the regions where three or more of the cell's fragments overlap. A nucleus has two copies of each locus, so a third means a second nucleus, or a repeat.

Remove Doublets uses ArchR's simulated doublets unless you choose AMULET. AMULET leaves out RepeatMasker's repeats, which baSeq does not have: on 10x PBMCs its count followed each cell's depth and flagged a quarter of the cells, where 10x expects under 1%. The step says so when AMULET finds far more doublets than expected.

Gene activity is counted at the same time: the insertions over each protein-coding gene's body and 2 kb upstream. The cells' fragments are kept beside the analysis (only the cells', a file per chromosome), so Coverage at Genes can draw any locus per cluster later.

From cells to cell types

Accessibility in a cell is close to binary, so the reduction is LSI rather than PCA: TF-IDF weighs each peak by how rare it is across cells, and a truncated SVD finds the main patterns. The first component almost always follows how deeply a cell was sequenced; TF-IDF and LSI draws each component's correlation with depth and leaves out those above 0.75. Harmony, the neighbours, the UMAP and Leiden clustering then run on what is left, as Signac's tutorials do. Name the clusters by their gene activity markers, by hand or automatically; gene activity follows expression, but more loosely than RNA.

Comparing conditions

Differential Accessibility sums each sample's cells of the chosen clusters and compares the samples with DESeq2, so the samples, not the cells, are the replicates (Squair et al. 2021): each condition needs two samples or more. Marker Peaks compares clusters cell by cell. Set each sample's condition in the metadata step; pick the organism there too, which decides the genome the motifs are read from.

Motifs and transcription factors

TF Activity is chromVAR (Schep et al. 2017), which was made for single cells: each motif's accessibility in each cell against background peaks of the same GC content and accessibility, as a z-score. The motifs are JASPAR 2024's CORE non-redundant matrices for the organism's group. Motif Enrichment tests each cluster's marker peaks against GC-matched background peaks. Related factors share motifs, so a hit names a family as much as a factor.

Not yet

FASTQ files (aligning the reads into fragments), Multiome's RNA and ATAC together, label transfer from a single-cell RNA-seq analysis, peak-to-gene links and footprinting are not part of scATAC-seq yet.

Step reference