Illuminating cell states by a comprehensive and interpretable single cell foundation model.

Area of Science:

• Computational Biology
• Genomics
• Artificial Intelligence

Background:

• Single-cell sequencing generates complex data, posing challenges for AI models due to sparsity and heterogeneity.
• Current AI foundation models struggle with practical application in cell biology due to data limitations and interpretability issues.

Purpose of the Study:

• To introduce CellVQ, an AI framework designed to overcome limitations in single-cell data analysis.
• To enhance the generalizability, representation, and interpretability of AI models for single-cell data.

Main Methods:

• Incorporated a large-scale dataset (68 million cells) and 500 million model parameters for robust pretraining.
• Developed a Single-Cell Discretization (SCD) module to represent cell embeddings and address data heterogeneity.
• Introduced CellVQ-Graph for integrating multimodal data into a knowledge graph for biological discovery.

Main Results:

• CellVQ demonstrated superior performance across various downstream tasks compared to existing methods.
• The SCD module effectively transformed high-dimensional single-cell data into an interpretable 'cell code'.
• CellVQ-Graph facilitated the discovery of novel biological insights with clear explanations.

Conclusions:

• CellVQ offers a generalizable and interpretable AI solution for the cell biology community.
• The framework effectively addresses key challenges in single-cell data analysis, paving the way for advanced biological discovery.