Researchers have developed Stack, a single-cell foundation model designed to overcome the limitations of current transcriptomic tools by utilizing cellular context to improve biological discovery. Unlike previous models that function as simple denoisers, Stack employs a transformer-based architecture that captures both inter-cellular and intra-cellular relationships across large cell sets. This framework enables in-context learning, allowing the model to predict how cell populations will respond to novel conditions through a process known as cell prompting. By training on nearly 190 million cells, the model achieves superior performance in tasks like perturbation effect prediction and batch integration without needing dataset-specific fine-tuning. The authors utilized these capabilities to generate Perturb Sapiens, a comprehensive virtual atlas detailing simulated responses to drug and cytokine treatments across the whole human organism. Ultimately, Stack provides a generative platform for exploring counterfactual cell states, potentially accelerating the identification of new therapeutic targets.

References:

• Stack: In-Context Learning of Single-Cell Biology