Forecasting Cell State Futures from Static Snapshots
Erpai Luo, Haoxiang Gao, Haiyang Bian, Shuguang Peng, Yaru Li, Chen Li, Minsheng Hao, Mo Chen, Yuli She, Lei Wei, Kai Liu, Xuegong Zhang
• bioRxiv
CellTempo is an autoregressive generative AI model designed to forecast long-range cell state evolution trajectories directly from static single-cell transcriptomic snapshots. Pretrained on scBaseTraj a dataset of over 48 million multi-step cellular transitions it reconstructs data-driven developmental potential landscapes. Crucially, CellTempo distinguishes transient transcriptional spikes from sustained long-term cell-fate commitment under genetic and chemical perturbations, aiding targeted cell engineering.
Single-Cell Genomics
Autoregressive Foundation Models
CellTempo
Waddington Landscape
Cell-Fate Engineering
Full Abstract
Forecasting Cell State Futures from Static Snapshots
Authors: Erpai Luo, Haoxiang Gao, Haiyang Bian, Shuguang Peng, Yaru Li, Chen Li, Minsheng Hao, Mo Chen, Yuli She, Lei Wei, Kai Liu, Xuegong Zhang
Journal: bioRxiv
Published: July 2026
Autoregressive Trajectory Forecasting in Latent Space
Single-cell RNA sequencing effectively captures static transcriptomic snapshots, but this leaves long-range developmental trajectories and cell-fate transitions difficult to predict mathematically. CellTempo addresses this gap by formulating cell-state evolution as a generative forecasting problem, using a causal Transformer backbone to predict future discrete cell representations autoregressively from a single initial cell state.
To train CellTempo, the researchers constructed scBaseTraj, the largest single-cell trajectory dataset to date, built by integrating RNA velocity and pseudotime methods. It contains over 48 million multi-step trajectories spanning 71 tissues derived from 90+ million cells.
Key Capabilities & Architectural Workflow
Two-Stage Architecture: A VQ-VAE tokenizer first compresses high-dimensional transcriptomes into 26 discrete latent tokens (via a 512-codebook size). A 12-layer causal Transformer (49.64M parameters) then models the trajectory-level sequences.
Waddington Potential Landscape Reconstruction: CellTempo quantifies state plasticity by computing the entropy of its predictive token distribution. Low entropy corresponds to committed lineage basins (deep potential), while high entropy highlights multipotent branching points (e.g., HSCs).
Long-Term Perturbation Horizon: Unlike existing tools that only predict immediate, short-term transcriptional responses, CellTempo propagates perturbation effects forward in time. In iPSC-to-OPC differentiation models, it successfully distinguished transient marker spikes (e.g., from Urea/Glycerol) from persistent, lineage-committed OPC specification driven by compounds like XAV939 and Go6983.
Benchmark Performance
Temporal Accuracy: Achieved a 0.94 Pearson correlation between real and generated mean pseudotime values in developmental trajectories.
Lineage Topology: Recovered PAGA cluster transition graphs with 0.78 shortest-path correlation and 0.84 root-distance correlation, significantly outperforming traditional delta-extrapolation baselines.
One-Line Takeaway: CellTempo establishes a predictive "world model" for single-cell biology, proving that static transcriptomes encode sufficient temporal dynamics for accurate, long-range fate forecasting.