Stack: In-Context Learning of Single-Cell Biology
Journal:
bioRxiv
Published Date:
Jun 8, 2026
Abstract
Foundation models trained on single-cell transcriptomic data offer the promise of identifying and predicting the diversity of cellular phenotypes across species, diseases, and other biological conditions. However, the current models are limited to their supervised training conditions and tasks, which limits their utility for biological discovery. Here, we present STACK, a foundation model trained on 149 million uniformly preprocessed human single cells that leverages tabular attention to generate representations for each cell informed by the cells in its context. STACK offers substantial improvements for downstream tasks in the zero-shot setting compared to baselines, whether they are zero-shot, fine-tuned, or trained from scratch on the target dataset. STACK can perform in-context learning from unlabeled cells representing arbitrary conditions, such as a chemical perturbation or a different donor, and predict the effect of those conditions on a target cell population without requiring data-specific fine-tuning. We apply STACK to generate Perturb Sapiens, the first human whole-organism atlas of perturbed cells, spanning 28 tissues, 40 cell types, and 892 drug, cytokine, and genetic perturbations. We validated subsets of Perturb Sapiens using in vitro stimulation profiles. STACK uniquely empowers prioritization of donor-specific perturbation effects, a capability we validated in our newly collected DiseasePert-3M data, comprising T cells from 40 donors across 14 diseases, stimulated with 11 cytokines. Overall, STACK presents a new modeling framework where cells themselves act as guiding examples at inference time, unlocking general-purpose in-context learning capabilities for single-cell biology.
