scYeast

Single-cell foundation models have transformed how transcriptomes are analyzed, but nearly all of them target human or mouse systems, leaving important model organisms underserved. Yeast (Saccharomyces cerevisiae) is a workhorse of synthetic and systems biology, yet it had no dedicated single-cell foundation model. scYeast, developed in Hongzhong Lu's group at Shanghai Jiao Tong University and posted to bioRxiv in August 2025, is the first foundational cell model tailored to yeast, and it is designed from the outset to exploit the rich prior knowledge available for this organism.

Rather than treating gene expression as an unstructured token sequence, scYeast uses an asymmetric parallel architecture that injects transcriptional regulatory information directly into the transformer's attention mechanism. This lets established regulatory-network knowledge shape how the model attends across genes during pretraining on large-scale yeast single-cell transcriptomics. The result is a model that combines strong generalization with biological interpretability.

By pairing a domain-specific pretraining corpus with explicit biological priors, scYeast offers a template for building foundation models for other organisms where curated knowledge exists but data are comparatively scarce.

#Key Features

• Yeast-specific pretraining: The first single-cell foundation model built for yeast, pretrained on large-scale yeast single-cell transcriptomics rather than transferred from mammalian data.
• Knowledge-guided attention: An asymmetric parallel architecture infuses transcriptional regulatory relationships directly into transformer attention, coupling learned representations with established biology.
• Zero-shot capabilities: Without fine-tuning, scYeast infers regulatory relationships and identifies critical cell states, reflecting the biological priors baked into training.
• Fine-tuning across tasks: After fine-tuning it performs cell type classification, growth doubling-time prediction, and gene perturbation response prediction.
• Cross-omics transfer: Transfer learning adapts the model to other omics layers such as proteomics, broadening its use beyond transcriptomics.

#Technical Details

scYeast is a transformer-based foundation model whose distinguishing feature is an asymmetric parallel design that embeds transcriptional regulatory priors into the attention computation. Genes are represented with 200-dimensional gene2vec embeddings, and gated fusion layers integrate the knowledge-guided and expression-driven signals; a knowledge-graph-informed variant is provided alongside a baseline. The public repository ships pretrained transcriptome checkpoints (with and without the knowledge graph) and a proteome checkpoint, and supports downstream tasks including gene regulatory network analysis, growth-rate prediction, ribosome-profiling prediction, and protein-turnover prediction. The code is released under an MIT license, while the preprint itself is posted under a CC BY-NC-ND license and has not yet completed peer review.

#Applications

scYeast is aimed at yeast systems and synthetic biologists who want to mine single-cell data for regulatory structure and phenotype. Its zero-shot regulatory inference and cell-state identification can guide hypothesis generation, while fine-tuned heads support concrete predictions such as classifying cell types, estimating growth doubling times, and anticipating responses to genetic perturbation. Through transfer learning it can also be brought to bear on proteomic datasets, making it a flexible analysis hub for yeast big-data workflows.

#Impact

By demonstrating that a domain-specific corpus combined with explicit regulatory priors yields a performant, interpretable foundation model, scYeast extends the single-cell foundation model paradigm beyond mammalian systems and sets a reference point for model-organism modeling. Its knowledge-guided attention offers a concrete strategy for injecting curated biology into transformer architectures, and its cross-omics adaptability suggests a path toward unified multi-omics analysis in yeast. As an openly licensed but not-yet-peer-reviewed preprint, its broader adoption will be shaped by community benchmarking and the continued growth of yeast single-cell datasets.