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

Area of Science:

• Single-cell biology
• Genomics
• Computational biology

Background:

• Single-cell transcriptomics reveals cellular diversity but faces measurement precision challenges.
• Existing foundation models often process cells independently, limiting their contextual understanding.

Purpose of the Study:

• To introduce S tack, a novel foundation model for single-cell transcriptomics.
• To leverage contextual information from neighboring cells for improved cellular representations.
• To enable general-purpose in-context learning for single-cell data analysis.

Main Methods:

• Trained S tack on 149 million human single cells using tabular attention.
• Developed a framework where cells act as guiding examples at inference time.
• Applied S tack to create the Perturb Sapiens atlas of perturbed cells.

Main Results:

• S tack significantly improves zero-shot performance on downstream tasks compared to baseline methods.
• The model enables in-context learning from unlabeled cells for predicting condition effects.
• Generated Perturb Sapiens, a comprehensive atlas of perturbed human cells across tissues and cell types.

Conclusions:

• S tack offers a powerful new modeling framework for single-cell biology.
• The model unlocks general-purpose in-context learning capabilities, advancing the analysis of cellular phenotypes.
• Perturb Sapiens provides a valuable resource for studying cellular responses to perturbations.