Single-cell RNA sequencing (scRNA-seq) allows scientists to analyze gene expression at the level of individual cells, offering insights into how tissues function and respond to disease. However, analyzing this data is complex due to noise, technical artifacts, and high dimensionality. Researchers from the University of Toronto, have developed a new tool called sciRED (Single-Cell Interpretable REsidual Decomposition) to help interpret these datasets more effectively.

scRNA-seq generates vast amounts of data, but distinguishing meaningful biological signals from technical artifacts is a challenge. Existing factorization methods help separate different components of the data, but interpreting these components remains difficult. sciRED aims to improve this process by refining factor analysis and identifying key biological patterns.

sciRED overview

A sciRED comprises four main steps: factor discovery, factor interpretation, factor evaluation and biological characterization. B In the factor discovery step, a Poisson generalized linear model is applied to the data to remove technical covariates, followed by extraction of residuals and factorization using PCA. The resulting score and loading matrices are then rotated for enhanced interpretability. The score matrix represents the projection of the original data onto the new factor space, illustrating the relationship between cells and factors. Each entry in this matrix reflects how much each cell contributes to the factors. The loading matrix contains the weights or coefficients that define the factors as linear combinations of the original genes. These weights can be used to rank genes according to their contribution to each factor, facilitating further interpretation of the factors. Models and algorithms are color-coded in blue, while input, output, and intermediate data elements are represented as white boxes. C Factor interpretation uses an ensemble classifier to match factors with given covariates, generating a Factor-Covariate Association (FCA) table. Covariate-matched factors are identified by thresholding FCA scores based on the distribution of all FCA scores. Unannotated factors may capture novel biological processes or other covariates. D Factor-interpretability scores (FIS) are computed for each factor. E The top genes and enriched pathways associated with a selected factor are identified for manual interpretation.

sciRED enhances the interpretation of scRNA-seq data by:

• Removing confounding technical effects, making biological signals clearer.
• Using mathematical rotations to improve how factors are understood.
• Mapping factors to known biological features, such as sex differences or immune responses.
• Detecting unknown factors, which could reveal new biological insights.
• Determining genes and pathways associated with each factor.

Applying sciRED to various datasets revealed important biological insights, including:

• Sex-specific gene expression patterns in kidney cells.
• Strong vs. weak immune responses in peripheral blood mononuclear cells (PBMCs).
• Reduction of background RNA contamination in liver tissue, allowing identification of genetic variations and rare cell types.
• Gene programs linked to tissue structure, such as anatomical zonation in the human liver.

sciRED provides a powerful, interpretable approach for analyzing complex scRNA-seq datasets. By improving how scientists extract and interpret biological signals, this tool can aid in discovering new cell types, understanding disease mechanisms, and advancing precision medicine.

As scRNA-seq continues to expand, tools like sciRED will be essential for unlocking the full potential of these datasets. Future applications could include identifying novel drug targets, improving tissue engineering, and enhancing our understanding of cellular diversity in health and disease.

Availability – The Python package for sciRED, along with example case scenarios and the analysis code used in this manuscript, is freely accessible at https://github.com/BaderLab/sciRED.