scYeast: a Biological-knowledge-guided Foundation Model on Yeast Single-Cell Transcriptomics

Pre-trained large models have emerged as a pivotal technological approach for foundational cell modeling. However, existing deep learning-based foundational models for cells have predominantly focused on human or murine systems, with a relative scarcity of research on model microorganisms such as Saccharomyces cerevisiae. Furthermore, these models often exhibit limitations in the integration of biological prior knowledge. To address this gap, we introduce scYeast, the first foundational cell model specifically designed for yeast that deeply integrates biological priors. scYeast features an innovative asymmetric parallel architecture that embeds transcriptional regulatory prior information directly into the Transformer's attention mechanism, thereby systematically incorporating and leveraging established biological knowledge during model training. After large-scale pre-training on single-cell transcriptomics data from yeast, scYeast demonstrates robust generalization capabilities and strong biological interpretability. It can perform zero-shot tasks, such as inferring specific regulatory relationships and resolving critical cell states, functional types, and developmental trajectories. Moreover, by constructing fine-tuning networks, scYeast excels at tasks including cell type identification, doubling time prediction, and forecasting responses to gene perturbations. Further leveraging transfer learning, scYeast can be adapted to other multi-omics data, such as proteomics, expanding its application boundaries in systems biology research. scYeast not only provides a novel tool for fundamental research in yeast single-cell biology but also establishes a new paradigm for the organic fusion of foundational models with prior knowledge, laying a solid groundwork for synthetic and systems biology.