acceptodds
Under review as a conference paper at ICLR 2027

Horizontal Flow Matching for Single-Cell Spatiotemporal Dynamics

Abstract

Learning the spatiotemporal dynamics of cells from spatial transcriptomics requires connecting snapshots of different cells acquired in different coordinate frames. A rotation or translation of a section can appear as cell movement and distort the inferred correspondence. We introduce horizontal flow matching, which learns changes in tissue organization, molecular state, and cell abundance up to shared rigid motion. The dynamics follow a continuity-reaction equation whose action assigns no cost to shared rigid motion, and the resulting horizontal projector removes this component from spatial prediction errors during learning and from predicted velocities during generation. An equivalent endpoint problem jointly estimates correspondence and relative pose in Wasserstein-Fisher-Rao geometry and yields analytic paths whose velocity and growth serve as regression targets. Regressing these targets recovers the horizontal dynamics under ideal population conditions. Our model, HORIZON, uses an equivariant transformer to predict velocity and growth from the current tissue, so that a tissue can be generated continuously from a single initial snapshot. On a simulated tissue observed in perturbed coordinate frames, alignment improves reconstruction and projection reduces rigid drift. On four biological time courses, HORIZON attains lower reconstruction error than state-of-the-art methods across spatial, molecular, and abundance metrics, with consistent gains when interpolating held-out snapshots.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
optimal transport, Schrödinger bridge, trajectory inference, single-cell

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_a1cXOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

In this submission the authors propose Horizon, an FM algorithm that regresses onto an adapted version of the Wasserstein-Fisher-Rao equation -- the adaptation accounts for artifacts that distorts the alignment of temporally ordered spatial transcriptomics slides. Using closed-form WFR targets, the method can thus be trained efficiently and can be used to align temporally ordered ST sections. Using multiple datasets for evaluation, Horizon outperforms the baselines on temporal tissue reconstruction tasks.

Critical strengths

Relevance: As is clear not least from the references in the paper, studying WFR is currently of interest to the community (a missing reference to add evidence to the relevance of the topic is [1]), and flow matching models are increasingly being applied to spatial transcriptomics data (another missing reference in my opinion is [2], which to my knowledge is the first paper to apply FM to ST data. Thus, adapting the WFR equation to alignment of temporally evolved ST data is a highly relevant problem to solve.

Exposition and validation There are some non-crucial references missing in my opinion, but overall the landscape appears to have been quite extensively covered and sensible baselines are used in the experiments. Although I did not scrutinize the equations of the closed-form bridge in §4.2, I appreciate the inclusion of the equations in the main text.

[1] Ying et al., (2026) "Beyond Continuity: Simulation-free Reconstruction of Discrete Branching Dynamics from Single-cell Snapshots" [2] Kviman et al. (2026) "Multi-marginal flow matching with adversarially learnt interpolants"

Critical weaknesses

The only weakness I could identify was that the writing and presentation style was slightly on the opaque, jargon side. That is, there are multiple instances where I had some issues understanding or parsing sentences, which overall made it difficult to properly understand the proposed method. A way to circumvent this would be to give some examples and dummy down presentations. As an instance of where this would be nice, consider the transition from Eq. 1 to Eq. 2 in Sec. 3.2: ideally I would like to understand the difference between these two equations, but I am only provided with wordings such as orbits, rigid motions/frames. Surely the adaptation of Eq. 1 can be explained in practical, exemplifying terms. This would help the presentation of the paper.

Minor:

Comparing Eq 1 and 2, the Z\mathcal{Z} got lost.

"SE" undefined in Sec. 3.2.

Questions for authors

What are the implications of the adaptation of the WFR in Eq. 2 vs. Eq. 1 in terms of runtime and implementation complexity?

Reviewer_MWaTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the issue in longitudinal spatial transcriptomics where inconsistent coordinate systems across tissue slices lead to rigid-body motion being misidentified as actual cellular movement. To resolve this, the authors propose "horizontal flow matching." By incorporating a "cost-free" rigid-body reference frame into the Wasserstein-Fisher-Rao (WFR) dynamics, the authors derive a closed-form "horizontal projector." They prove (Theorem 2) that this dynamic problem is equivalent to a static WFR problem with optimally aligned endpoints. This allows for the use of qOET (quadratic Optimal Entropic Transport) to jointly solve for the rigid-body pose and the cell-coupling, while enabling simulation-free training by leveraging the closed-form "bridge" from WFR-FM. The horizontal projector is utilized both in the training loss and for generative integration; the authors demonstrate that it recovers the optimal path at the population level. The resulting model, HORIZON, employs an SE(2)-equivariant transformer to generate a complete trajectory from a single initial tissue slice, outperforming both WFR-FM and stVCR on synthetic benchmarks and four real-world datasets.

Critical strengths
  • The core theoretical contribution bridges rigid-body invariance with simulation-free training. Theorem 2 establishes that a WFR dynamic problem—permitting cost-free rigid-body motion—is equivalent to a static WFR problem with optimally aligned endpoints. This enables the direct reuse of the closed-form bridge from WFR-FM. In contrast, previous approaches (such as stVCR) have relied on NeuralODEs, which require simulation during the optimization process.

  • The horizontal projector is derived from the optimal solution for the reference frame within the action functional. In 2D, it takes a simple, elegant form based on the mass-weighted centroid velocity and angular momentum/moment of inertia. The same operator is used for both the training loss and generative integration, supported by Lemma 3 and Theorem 4, with the underlying assumptions and scope of applicability clearly articulated.

  • The optimal transport cost in qOET corresponds precisely to the action along the bridge path, with the optimal value equaling the cost of the dynamic problem. Consequently, the estimated poses represent the configuration that minimizes the "effort" required for cellular dynamics, rather than being determined by independent registration criteria. Furthermore, pose updates are weighted by the couplings; since newly appearing or disappearing cells contribute little to this weighting, they do not skew the alignment.

Critical weaknesses
  • qOET is non-convex; for approximately symmetric structures (such as the ring stage of zebrafish development), there may exist local optima corresponding to different orientations. The paper does not report the consistency of solutions obtained from different initial poses.
  • The "cone-plane bridge" is derived from WFR-FM and earlier studies, and Theorem 2 is a generalization of existing results on rotation-invariant optimal transport (OT). It is recommended that the specific technical contributions—such as the handling of translations, unbalanced masses, representation dimensions, and the existence of an optimal pose—be explicitly clarified.
  • The "horizontal projector" mentioned in the title yields only a marginal reduction in drift during synthetic ablation studies (from 0.127 to 0.116), with virtually no impact on shape accuracy.
Questions for authors

See my weaknesses.

Reviewer_okSsOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The research team studies spatial transcriptomic time-series data in which snapshots contain different cells and are measured in different coordinate frames. To account for unknown cell correspondences and differences in rotation and translation, the authors formulate unbalanced tissue dynamics on a quotient space that identifies tissue configurations differing by a shared rigid transformation. They establish an equivalence between the dynamic formulation and an endpoint problem that jointly estimates alignment and correspondence through optimal entropy transport. The method uses analytic WFR bridges to train a mass-aware SE(2)-equivariant transformer, applying horizontal projection to spatial prediction errors during training and to predicted spatial velocities during generation. The experiments evaluate robustness to coordinate-frame perturbations, reconstruction of four biological time courses, and interpolation at held-out interior time points.

Critical strengths
  1. The question the authors are trying to solve is important and well motivated. The model's framework is also clear; it uses the same horizontal structure across the transport formulation, the regression objective, and the integrated velocity. This makes the connection between the geometry and the implemented learning rule clear.
  2. The experiments in the paper are well set up. Specifically, I like the controlled frame-perturbation experiment.
  3. The leave-one-timepoint-out study supports the spatial component of the method
Critical weaknesses
  1. The WFR teacher construction includes mass splitting and pure creation, whereas generation evolves a fixed set of carriers from the initial snapshot (lines 352–356; Appendix B.1). The paper acknowledges this restriction, but its practical effect is not quantified. For example, chicken starts with 147 carriers while the final observation contains 1,967 points (Table 3). Could the authors report how much teacher mass uses creation or splitting, and whether increasing the initial carrier support changes the spatial and molecular results?
  2. The holdout protocol excludes the omitted snapshot from coupling construction and neural training (lines 1454–1457), but the preprocessing is described at the level of the full time course (lines 1230–1234). Were gene selection, normalization statistics, PCA and spatial scaling refit within each fold? This matters for interpreting Table 14 as prediction at an unobserved timepoint. Since growth is a central output, I would also like to see mass error for these same held-out targets.
  3. The synthetic ablation convincingly separates the two mechanisms, but it also shows that registration alone accounts for most of the shape improvement: quotient spatial SW is 0.039 without projection and 0.042 for the full model (Table 1). The projector improves direct drift metrics, as the authors acknowledge. The real-data ablation removes alignment and projection together (Table 9), so it does not establish the additional benefit of the horizontal objective on biological series. A registered, unprojected control would clarify whether horizontal learning improves stability or other practically relevant outcomes beyond registration.
Questions for authors

As above

Comments on the paper

1 comment

Dear Reviewers,

Now that the authors have responded, please review what they have said and give consideration to whether they have adequately addressed your concerns.

-The AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.