iterative-lsi-dimensionality-reduction

iterative-lsi-dimensionality-reduction

Summary

Iterative latent semantic indexing (LSI) is a dimensionality reduction technique that jointly compresses scATAC-seq chromatin accessibility peaks and scRNA-seq gene expression data into a unified low-dimensional embedding for integrated multimodal analysis. This approach enables visualization, clustering, and downstream analysis of paired single-cell multiome datasets.

When to use

When you have aligned paired scATAC-seq and scRNA-seq data from the same cells (multiome data) and need to create a single reduced-dimension coordinate space that integrates both chromatin accessibility and gene expression signals for joint clustering, trajectory analysis, or visualization.

When NOT to use

• Input data are unimodal (scATAC-seq only or scRNA-seq only) — use standard LSI or PCA on the single modality instead.
• Cells in the peak matrix and gene expression matrix are not aligned or do not represent the same cells — addIterativeLSI requires matched cell identities across modalities.
• Peak matrix and expression matrix have already been independently dimensionality-reduced without joint integration in mind — joint LSI requires the raw or lightly-processed feature tables as input.

Inputs

• ArchR project object with imported scATAC-seq peak matrix (from importFeatureMatrix)
• ArchR project object with appended scRNA-seq gene expression matrix (from addGeneExpressionMatrix)
• Aligned cell barcodes across both modalities

Outputs

• ArchR project object with computed iterative LSI dimensions
• Latent semantic indexing coordinates integrating both scATAC-seq and scRNA-seq variance
• Input for addCombinedDims to produce unified reduced-dimension embedding

How to apply

After ingesting the scATAC-seq peak matrix via importFeatureMatrix and appending the scRNA-seq gene expression matrix via addGeneExpressionMatrix to the same ArchR project object, execute addIterativeLSI on the combined project to perform latent semantic indexing jointly across both modalities. This function computes iterative dimensionality reduction that treats accessibility peaks and gene expression features together, producing latent dimensions that capture variance explained by both data types. The resulting LSI coordinates form the basis for downstream integration via addCombinedDims, which generates a unified embedding. Key rationale: iterative LSI avoids the bias of treating modalities separately and ensures that the reduced-dimension space reflects the joint signal of chromatin and transcriptomic state.