lemur

lemur fits a latent embedding multivariate regression (LEMUR) model to multi-condition single-cell data to enable differential expression analysis using a continuous low-dimensional latent space parameterization of cell type and state diversity.

Key Features:

• Condition alignment: The parametric model integrates data from different experimental conditions to align them within a shared latent space.
• Log fold change prediction: The model predicts how each cell's gene expression would change as a function of the conditions and its position in latent space, generating predicted log fold changes between conditions for all cells.
• Adaptive neighborhood identification: The model identifies compact neighborhoods of cells with consistent differential expression that adapt to the underlying gene expression changes, rather than relying on statically defined clusters.
• Pseudobulked differential expression testing: For identified neighborhoods, a pseudobulked differential expression test is conducted to assess which genes are significantly changed.

Scientific Applications:

• Tumor subpopulation drug response analysis: Identification of tumor subpopulations with distinct drug responses in multi-condition sc-RNA-seq data.
• Developmental biology: Analysis of the interplay between cell state and developmental time, as demonstrated in zebrafish embryos.
• Spatial transcriptomics: Discovery of cell state x environment interactions in spatial single-cell studies, as demonstrated in a study of plaques in Alzheimer's disease.

Methodology:

LEMUR is a generative model that provides a parametric description of single-cell data using a continuous low-dimensional latent space parameterization of cell type and state diversity, operating without or before commitment to discrete categorization.