scYeast: a biological-knowledge-guided foundation model on yeast single-cell transcriptomics

Area of Science:

• Single-cell transcriptomics
• Yeast biology
• Foundational models

Background:

• Large-scale pre-trained models are crucial for cell modeling but often overlook model organisms like yeast.
• Existing models inadequately incorporate biological prior knowledge.
• There is a need for specialized models for yeast single-cell data.

Purpose of the Study:

• Introduce scYeast, the first foundational cell model for yeast single-cell transcriptomics.
• Effectively embed biological priors into a deep learning model for yeast.
• Enhance the analysis and interpretation of yeast single-cell data.

Main Methods:

• Developed scYeast using a novel asymmetric parallel architecture.
• Infused transcriptional regulatory information into the Transformer's attention mechanism.
• Pre-trained scYeast on large-scale yeast single-cell transcriptomics data.

Main Results:

• scYeast demonstrates strong generalization and biological interpretability.
• Achieved success in zero-shot tasks like inferring regulatory relationships.
• Showed high performance in cell state classification, growth doubling time prediction, and gene perturbation response prediction after fine-tuning.

Conclusions:

• scYeast is a valuable tool for yeast single-cell biology research.
• Presents a novel framework for integrating foundational models with biological priors.
• Accelerates discovery in yeast synthetic and systems biology and offers a replicable framework for other organisms.