single-cell-epigenomics-peak-analysis

single-cell-epigenomics-peak-analysis

Summary

Identifies and characterizes chromatin accessibility peaks in single-cell ATAC-seq data using spectral embedding and peak-calling algorithms. This skill enables discovery of cell-type-specific regulatory regions and their enrichment for transcription factor motifs.

When to use

Apply this skill when you have preprocessed single-cell ATAC-seq fragment files or count matrices and need to identify open chromatin regions (peaks) to support downstream differential accessibility analysis, motif discovery, or regulatory network inference. Use it after BAM-to-fragment conversion and cell filtering but before comparing accessibility between cell populations.

When NOT to use

• Input is already a curated set of consensus peaks from bulk ATAC-seq or ChIP-seq; skip to motif enrichment or annotation.
• Single-cell data lacks sufficient sequencing depth (~5,000 fragments per cell minimum); peak calling will be unreliable.
• Analyzing bulk ATAC-seq or RNA-seq data; use bulk peak callers (MACS2, ENCODE pipeline) instead.

Inputs

• BAM or fragment files (TSV or gzipped format)
• Cell barcodes and metadata (cell-type annotations or cluster assignments)
• Reference genome (optional; for peak annotation)

Outputs

• Peak count matrix (.h5ad AnnData object with peaks × cells)
• Peak coordinates (BED format or interval table)
• Spectral embedding coordinates (low-dimensional representation)
• Differential accessibility results (peak IDs, log2-fold-change, p-values)

How to apply

Begin by constructing a tile matrix or peak matrix from fragment files using pp.make_tile_matrix or pp.make_peak_matrix. Apply dimension reduction via matrix-free spectral embedding (tl.spectral) to embed cells in a low-dimensional space, which enables clustering and visualization without materializing the full count matrix. Perform peak calling using tl.macs3 (or merge peaks across cell types with tl.merge_peaks) to define consensus peak sets. For peaks identified as differentially accessible via tl.diff_test, validate peak quality by checking for non-zero counts and reasonable distribution of peak widths. The spectral embedding is scalable to >10 million cells and supports integration with downstream tools (Scanpy, peak annotation) via AnnData format.