Abstract RNA velocity provides a powerful scope for understanding cell state dynamics by modeling spliced and unspliced mRNA captured by single-cell or spatial transcriptomic technologies. However, prevailing methods rely on restrictive kinetic assumptions often fail in the presence of heterogeneous kinetic regimes, which is common in tissues of complex biological systems. These limitations hinder both accurate inference and interpretability, particularly in spatial contexts with kinetic mixing. Here, we present STEER (Spatial-Temporal Explainable Expert model for RNA-velocity inference), a flexible and interpretable framework that integrates spatially informed graph-attention auto-encoder with a kinetically guided mixture-of-experts architecture. STEER disentangles kinetically and/or spatially mixed populations by assigning cells to expert-defined regimes, and infers cell-gene-specific kinetic rates with cell-level latent time. Benchmarking STEER on synthetic and challenging real-world systems, demonstrates its robust performance and enhanced interpretability. Particularly, STEER reveals spatiotemporally complementary immunoregulatory programs at the maternal–fetal interface of mouse uterus. Overall, STEER provides a generalizable and explainable framework for decoding nuanced spatio‑temporal dynamics in complex tissues, offering insight into tissue morphogenesis, lineage specification, and tumor progression.
Liu et al. (Sat,) studied this question.