scPACE: A Personalized Autoencoder for Cell Evolution in Single-Cell Transcriptomics

Single-cell studies increasingly profile large cohorts of individuals, measuring thousands of cells in each donor. These datasets contain information at two levels: variation among individual cells and among donors. Analyzing both levels requires methods that can learn cellular representations, infer cell and donor states, and account for cell types that may be missing in some donors. Existing methods partially address these tasks. We introduce scPACE, a hierarchical personalized variational autoencoder handling these tasks. scPACE learns a cell representation, a cell state, and then aggregates cells within each type of cell to infer a donor state. It also imputes missing cell types with confidence estimates. Across simulated cohorts and a developing mouse visual-cortex dataset, scPACE recovered held-out cell and donor states accurately while achieving efficient cell-type imputation and representation. scPACE provides a unified framework for learning phenotypes across cellular and donor scales in cohort-scale single-cell transcriptomics.