Biologically informed deep learning to query gene programs in single-cell atlases.

Area of Science:

• Computational Biology
• Genomics
• Artificial Intelligence

Background:

• Large-scale single-cell atlases offer detailed cell state information.
• Deep learning enables rapid analysis of new single-cell datasets by mapping them to reference atlases.
• Current mapping methods lack biological interpretability, failing to connect with known concepts like genes or pathways.

Purpose of the Study:

• To introduce expiMap, a biologically informed deep learning architecture for single-cell reference mapping.
• To enable the interpretation of single-cell data analysis using biologically meaningful components.
• To improve the interpretability of integrative single-cell analysis.

Main Methods:

• Developed expiMap, a deep learning architecture for single-cell reference mapping.
• ExpiMap learns to map cells into 'gene programs', refining existing and discovering novel ones.
• Simultaneously learned gene program activities and refined program definitions.

Main Results:

• ExpiMap provides a higher level of interpretability compared to existing methods.
• The method successfully maps cells into biologically understandable gene programs.
• Demonstrated expiMap's effectiveness in analyzing single-cell perturbation responses across tissues and species.
• Applied expiMap to resolve treatment responses in COVID-19 patients across cell types.

Conclusions:

• ExpiMap offers a powerful, interpretable approach for single-cell reference mapping.
• The biologically informed gene programs enhance understanding of cellular states and responses.
• ExpiMap has broad applications in analyzing perturbation responses and clinical data, including COVID-19 patient treatments.