RT-SRFM：面向发育单细胞时间序列的潜动力学框架

Single-cell time-series data provide a powerful opportunity to model developmental progression, but existing methods often focus separately on state transition, population growth, or fate inference. Here, we present RT-SRFM, a staged latent dynamical framework for developmental single-cell transcriptomic time series. The framework integrates optimal-transport-guided velocity learning, decomposed population growth modeling, shape refinement, and supervised biological readouts within a shared latent space. It is trained through an eight-stage curriculum to improve local transport accuracy, mass calibration, fate consistency, and long-horizon rollout stability. We evaluate the model on zebrafish embryogenesis and an independent sea urchin dataset. Across both systems, it achieves competitive transport performance while improving population mass calibration and fate/segment prediction compared with representative optimal transport, neural dynamics, and fate-mapping baselines. Ablation experiments further show that shape refinement, mass calibration, growth decomposition, fate warmup, and multi-step rollout repair contribute to model stability and predictive performance. These results suggest that the proposed framework provides a unified and interpretable approach for modeling developmental dynamics from single-cell time-series snapshots.