Supervised discovery of interpretable gene programs from single-cell data.

Area of Science:

• Computational Biology
• Genomics
• Immunology

Background:

• Factor analysis is crucial for understanding single-cell gene expression by identifying gene programs.
• Existing matrix factorization methods often suffer from technical artifacts and lack interpretability.

Purpose of the Study:

• To develop an algorithm, Spectra, that overcomes limitations of current methods by combining user-defined and novel gene programs.
• To improve the interpretability and accuracy of factor analysis in gene expression data.

Main Methods:

• Spectra integrates user-provided gene sets and cell-type labels as prior biological information.
• It models cell type explicitly and uses a gene-gene knowledge graph with a penalty function to guide factorization.
• The algorithm detects novel programs alongside user-defined ones to explain expression covariation.

Main Results:

• Spectra outperforms existing methods in tumor immune contexts.
• It identifies factors changing under immune checkpoint therapy.
• The algorithm successfully disentangles CD8+ T cell tumor reactivity and exhaustion, explains macrophage state changes, and reveals cell-type-specific immune metabolic programs.

Conclusions:

• Spectra offers a more interpretable and robust approach to factor analysis for single-cell gene expression data.
• The algorithm provides significant advancements in analyzing complex biological systems, particularly in cancer immunology.
• Spectra enhances the discovery of biologically relevant gene programs and cellular states.