Section frames stop masquerading as cell motion: Horizon strips rigid drift from tissue dynamics and reconstructs four spatial transcriptomics time courses more accurately
Synopsis
Spatial transcriptomics time courses are built from snapshots of different cells recorded in different coordinate frames, so rotations and translations between sections can be misread as cell movement. Horizon defines tissue dynamics only up to a shared rigid motion, removing the best-fitting translation and rotation with the same projector in training and generation while jointly estimating relative pose and cell correspondence. On four real spatial transcriptomics series it reconstructs tissue geometry, gene expression and cell abundance with lower error than WFR-FM and stVCR, and interpolates held-out developmental stages.
Interpretation
Whole-section rotation and translation are no longer counted as cell motion; the model learns tissue deformation, molecular-state change and changes in cell abundance. Earlier dynamic optimal transport and flow matching described dynamics in the recorded coordinates, charging the spatial cost for rotations and translations unrelated to biology, distorting cell correspondence during matching and producing rigid drift when the learned velocity was integrated. On a simulated tissue benchmark with hidden applied poses, worsening the perturbation from canonical to severe raises Horizon's direct spatial sliced Wasserstein distance only from 0.111 to 0.115, while WFR-FM grows more than fivefold and stVCR more than doubles.
The method first jointly estimates the relative pose and cell correspondence of consecutive sections, then regresses velocity and growth along closed-form paths between matched cells. This joint estimate, qOET, optimizes one shared pose together with a coupling that allows mass creation and destruction in Wasserstein–Fisher–Rao geometry; the pose is shared by all cell pairs, while one source cell's mass may split among several targets. On four real series (axolotl brain regeneration, chicken heart development, serial Drosophila embryo sections, zebrafish embryogenesis, each with four or five snapshots), Horizon has lower mean error than WFR-FM and stVCR on all seven metrics and all four datasets, and lower molecular error than STORIES.
Generation recomputes the projector at every step, subtracting any residual whole-tissue rigid component from the network output so the trajectory does not rotate as a whole. The training loss does not penalize the rigid component, which can carry 2% to 37% of the network output's spatial kinetic energy, so aligning sections in advance without projecting during generation is not enough. Without projection in generation the trajectory rotates by up to and direct spatial sliced Wasserstein distance grows up to 6.8-fold; paired over nine severe variants, adding the projector lowers direct spatial sliced Wasserstein distance from 0.127 to 0.116.
Perspective
The demonstrated scope is four spatial transcriptomics time courses (axolotl, chicken, Drosophila, zebrafish, each with four or five snapshots) plus a simulated tissue benchmark with hidden pose perturbations; the model is fitted separately to each time course, and generation starts from the first snapshot without later observations or endpoint correspondences.
The quotient formulation also removes any genuine collective rigid tissue motion indistinguishable from a change of frame, so such motion cannot be recovered from these observations alone; when a section is small and nearly round, as chicken D4, several poses are almost equally good and the pose is weakly identified by the snapshots; generation uses a fixed set of carriers, which can limit spatial coverage during rapid expansion; and both coupling construction and attention are quadratic in the number of cells.
