Deep identifiable modeling of single-cell atlases enables zero-shot query of cellular states.

With the emerging single-cell RNA-seq datasets at atlas levels, the potential of a universal model built on existing atlas that can extrapolate to new data remains unclear.A fundamental yet challenging problem for such a model is to identify the underlying biological and batch variations in a zero-shot manner, which is crucial for characterizing scRNA-seq datasets with new biological states. In this work, we present scShift, a mechanistic model that learns batch and biological patterns from atlas-level scRNA-seq data as well as perturbation scRNA-seq data. scShift models genes as functions of latent biological processes, with sparse shifts induced by batch effects and biological perturbations, leveraging recent advances of causal representation learning. Through benchmarking in holdout real datasets, we show scShift reveals unified cell type representations as well as underlying biological variations for query data in zero-shot manners, outperforming widely-used atlas integration, batch correction, and perturbation modeling approaches. scShift enables mapping of gene expression profiles to perturbation labels, and predicts meaningful targets for exhausted T cells as well as a list of diseases in the CellxGene blood atlas.