A transcriptomics-native foundation model for universal cell representation and virtual cell synthesis.

Area of Science:

• Computational Biology
• Genomics
• Artificial Intelligence

Background:

• Current single-cell foundation models often use language architectures that disregard transcriptomic data distributions, leading to suboptimal performance compared to specialized methods.
• This limitation hinders accurate analysis and interpretation of complex single-cell datasets.

Purpose of the Study:

• To introduce xVERSE, a novel transcriptomics-native foundation model designed to overcome the limitations of existing models.
• To enhance single-cell data analysis through batch-invariant representation learning and probabilistic generation of expression profiles.

Main Methods:

• xVERSE employs a transcriptomics-native architecture, integrating batch-invariant representation learning with probabilistic expression profile generation.
• The model was evaluated against leading foundation models, batch-effect correction methods, and spatial imputation techniques.

Main Results:

• xVERSE achieved superior performance in representation learning, outperforming leading foundation models by 17.9% and batch-effect correction methods by 11.4%.
• It demonstrated a 34.3% improvement over the second-best spatial imputation method and synthesized virtual cells indistinguishable from real data (AUROC ≈ 0.5).
• xVERSE enabled accurate clustering and marker detection in small datasets, resolving rare cell types with as few as four cells, and improved cross-modality prediction generalizability.

Conclusions:

• xVERSE represents a transformative framework for single-cell data analysis, offering enhanced capabilities beyond conventional models.
• Its ability to preserve biological heterogeneity, diminish batch effects, and generate high-fidelity virtual cells unlocks new analytical possibilities, particularly for rare cell type identification and data augmentation.