• Velocity-Guided Representation Learning for Single-Cell Trajectory Inference

Single-cell RNA sequencing has mapped cellular heterogeneity at atlas scales. However, current single-cell foundation models treat cells as independent, static snapshots, failing to capture the continuous, regulatory dynamics of cellular evolution. Here, I present a generative framework that bridges RNA velocity with temporal diffusion to simulate directed cellular trajectories. By enforcing splicing dynamics as an explicit inductive bias, I demonstrate a velocity-aware encoder, scVI-Velo, significantly outperforms standard variational autoencoders in representation learning and trajectory inference across a curated 10-million-cell corpus. Leveraging this scale, I also utilize velocity-guided Monte Carlo sampling to construct scBaseTrajectory, a foundational database of high-quality cellular trajectories. Finally, I outline an exploratory framework for scLDM-Velo, a proposed temporal latent diffusion model that frames single-cell foundation modeling as the generation of continuous cell trajectories rather than the embedding of independent cell states. Ultimately, I demonstrate RNA velocity provides critical temporal context conventionally ignored by single-cell foundation models, enabling more accurate trajectory inference and generation.