Cell-GraphCompass: modeling single cells with graph structure foundation model.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Human cells rely on complex gene regulatory networks for function.
• Existing foundation models for single-cell transcriptomics often overlook intrinsic biological features by imposing sequential gene structures.
• There's a need for models that integrate prior biological knowledge effectively.

Purpose of the Study:

• To introduce Cell-GraphCompass (CGCompass), a pioneering foundation model for single-cell analysis.
• To develop a graph-based pre-training approach for modeling genes and cells.
• To leverage diverse gene-related information for enhanced biological network representation.

Main Methods:

• Constructed cell graphs using three types of gene information as node features.
• Incorporated gene relationships as edge features from three perspectives.
• Pre-trained the CGCompass model on over 50 million human cells.
• Fine-tuned the model for tasks including batch integration, cell type annotation, and gene perturbation prediction.

Main Results:

• CGCompass demonstrated commendable performance across various single-cell analysis tasks.
• The graph pre-training approach effectively modeled gene and cell relationships.
• The model successfully integrated prior biological knowledge into single-cell analysis.

Conclusions:

• CGCompass offers a practical architecture for foundation models in single-cell analysis.
• Graph pre-training is a viable strategy for incorporating prior knowledge.
• The model advances the deciphering of gene regulatory networks.