scChronos: Hierarchical Temporal Context Learning for Missing Snapshot Prediction in Single-Cell Time Series

Single-cell time-series atlases provide discrete cross-sectional snapshots of continuous developmental processes, making missing-snapshot prediction essential for reconstructing trajectories and dynamic programs. However, destructive scRNA-seq profiling prevents tracking the same cells across time, yielding unpaired population snapshots and making missing-snapshot prediction challenging. Existing methods typically rely on explicit dynamical propagation or transport-based coupling between snapshots, which may ignore the statistical constraints of the overall population distribution and be sensitive to sampling noise and cell-to-cell variability.

We introduce scChronos, a hierarchical temporal context learning framework that formulates missing-snapshot prediction as target-conditioned population distribution prediction rather than individual-cell propagation. Unlike cell-centric propagation approaches, scChronos explicitly captures dual-level dynamics: (i) Within each observed snapshot, scChronos compresses cell populations into cluster-representative cell representations with a transcriptomic encoder, preserving population information while reducing noise; (ii) Across reference snapshots, the dual-branch temporal context integration module fuses information from all available and nearby time points into a temporal representation, which is then decoded with learnable cell-generation seeds to predict the target distribution. Extensive experiments demonstrate that scChronos improves both interpolation-style recovery and extrapolation-style forecasting compared with state-of-the-art methods. Beyond reducing distributional discrepancy, scChronos better preserves developmental topology and maintains cell-type-specific marker and pathway programs.

The code and data are available at https://github.com/yjs193/scChronos and https://doi.org/10.5281/zenodo.20933283.