Abstract
Time course single-cell RNA sequencing (scRNA-seq) enables researchers to probe expression dynamics at the resolution of individual cells. Analyzing this data poses several challenges, including deconvolving the contributions of time and cell type, discriminating true dynamics from batch effects, and inferring per-cell dynamics despite cells being destroyed at each time point. We present VIST (Variational Inference for Single cell Time series), a deep-learning algorithm that deconvolves single-cell time series into time-dependent and time-independent components. Using both synthetic and real scRNA-seq data, we show that VIST constructs biologically meaningful latent spaces, removes batch effects, and generates realistic single-cell time series. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s13059-025-03874-2.