Sparse snapshots become continuous tissue: ST-FateFlow reconstructs gene expression and spatial position together
Synopsis
Time-resolved spatial transcriptomics yields only a few destructively sampled time points, and cells cannot be paired across them, so gene expression and tissue organization at intermediate stages go unmeasured. ST-FateFlow learns gene-expression and spatial-coordinate velocity fields together by flow matching to interpolate unobserved time points, and separately trains a time-conditioned potential to order differentiation progressions. Across ten interpolation settings on ARTISTA, MOSTA and PRISTA4D it has the lowest gene-expression MMD, and it is best in 26 of 30 dataset-time-metric spatial combinations.
Interpretation
Gene-expression distributions and tissue spatial structure at unobserved time points can be filled in together, not just expression profiles predicted. Earlier temporal models infer population dynamics mainly in gene-expression space, while spatial alignment and spatiotemporal models align sections or model continuous change without a directly evaluable differentiation potential; STORIES learns differentiation trajectories and a potential landscape but is not designed to generate continuous joint gene-spatial trajectories at unobserved stages. Across ten held-out interpolation settings on ARTISTA, MOSTA and PRISTA4D, ST-FateFlow has the lowest gene-expression MMD in all ten; it is best in 26 of 30 dataset-time-metric spatial combinations, and on the three-dimensional PRISTA4D it beats stVCR on all three spatial metrics at all six time points.
The reconstruction branch and the potential branch are trained separately and share the same cross-time pairs, but the potential is not required to equal the velocity field. Local, high-dimensional gene-spatial velocities do not by themselves assign each cell a comparable position along differentiation, so a separate potential branch is needed. The potential branch trains its negative gradient to follow earlier-to-later transition directions inferred from adjacent snapshots; on ARTISTA at 10 DPI, progenitor-like wntEGC and reaEGC populations sit at higher potential, differentiated mpEX and dpEX at lower potential, and rIPC2 lies between reaEGC and dpEX.
Perspective
Evaluation covers three single-cell-resolution spatial transcriptomic datasets, mouse embryonic organogenesis, planarian whole-body regeneration and axolotl telencephalon regeneration, with both two-dimensional sections and three-dimensional reconstruction; potential values compare cells within the same developmental time rather than acting as fate probabilities or physical energy.
Results depend on the quality of cross-time coupling and spatial registration and may be affected by batch effects or missing intermediate states; learning the potential requires extra gradient computation over high-dimensional gene-spatial states, raising computational cost; genes such as SFRP1 and CEBPD are only candidates associated with high-potential progenitor states, not direct regulatory or causal evidence for axolotl neuronal differentiation.
