Abstract

Cell-fate dynamics are defined not only by where cells end, but by when fate programs emerge, diverge, and reshape population structure. Much recent effort has been devoted to reconstructing these dynamics from temporally resolved single-cell RNA-seq and spatial transcriptomics. However, it is difficult because these measurements only provide destructive snapshots instead of continuous molecular histories. On the other hand, a diverse set of lineage-tracing methods are now available which could serve as temporal constraints to connect the dots. The challenge is how to utilize and unify these diverse ancestry records in the reconstruction of a continuous cell state dynamics. Here we introduce PhyloFM, a lineage-constrained flow-matching framework that maps heterogeneous lineage information onto a cell-fate-anchor topology and learns a continuous velocity field together with relative cell-abundance changes on a topology-regularized latent geometry. From the same integrated trajectories, PhyloFM predicts population transport, future fate probabilities, commitment timing and branch-resolved projected gene- and module-level dynamics. We tested PhyloFM across five lineage-tracing settings spanning C. elegans embryogenesis, LARRY hematopoiesis, pandaCREST ventral-midbrain development, zebrafish heart regeneration and spatial eTracer tumours. In C. elegans, our method improved population transport by 17.3% and lineage-grounded dynamic error by 16.3% relative to state-of-the-art dynamical baselines, with transport error reduced by 26.6% relative to additional lineage-aware control. Independent EPIC GFP reporter traces support the inferred timing of proneural, ciliated-neuron and late-selector programs, with 67% of matched reporter genes showing Pearson r > 0.6. On the spatial eTracer tumour benchmark, PhyloFM increased clone-grounded fate accuracy by 70.4% relative to the best spatial baseline. Together, these analyses show that diverse lineage records can serve as biological constraints for reconstructing continuous, interpretable molecular fate dynamics from single-cell snapshots.