single-cell-atac-seq-dimensionality-reduction

single-cell-atac-seq-dimensionality-reduction

Summary

Reduce the dimensionality of scATAC-seq data using iterative latent semantic indexing (LSI) in ArchR to enable downstream clustering, visualization, and trajectory analysis. This skill is essential for handling the high-dimensional, sparse peak-by-cell matrices typical of single-cell chromatin accessibility data.

When to use

Apply this skill after loading and preprocessing raw scATAC-seq data into an ArchR project object when you need to compute low-dimensional embeddings for clustering, UMAP/tSNE visualization, or integrated multi-omic analysis. Use it as a prerequisite before trajectory analysis (Monocle3 or Slingshot), gene expression matrix integration, or any downstream analysis that requires dimensionality-reduced cell representations.

When NOT to use

• Input data is already a low-dimensional embedding or has been previously dimensionality-reduced by another method—apply this skill only on raw or minimally processed peak-by-cell matrices.
• scATAC-seq data has not been quality-filtered or peak-called; LSI performance depends on upstream preprocessing.
• Analysis goal is restricted to single-cell gene expression only without chromatin accessibility component; use gene expression-specific dimensionality reduction methods instead.

Inputs

• ArchR project object (with imported scATAC-seq peak matrix)
• Peak-by-cell count matrix (sparse, typically filtered for quality control)

Outputs

• ArchR project object with LSI embedding (addIterativeLSI result)
• Low-dimensional cell representation (LSI dimensions)
• Optional: combined LSI embedding (if using addCombinedDims for multiome data)

How to apply

Invoke the addIterativeLSI function on your ArchR project object to compute iterative latent semantic indexing across the peak-by-cell matrix. This function performs dimensionality reduction while accounting for sparsity inherent in scATAC-seq data. For paired scATAC-seq and scRNA-seq analysis, compute LSI embeddings for both modalities separately, then use addCombinedDims to integrate the reduced dimensions from both datasets into a unified embedding space suitable for joint clustering and downstream analysis.