acceptodds
Under review as a conference paper at ICLR 2027

Finsler Flow Matching: Dynamics-Aware Geodesic Interpolation for Single-Snapshot Trajectory Inference

74 views

Abstract

Single-cell snapshot data can resolve a continuum of cellular states but do not uniquely determine the dynamics governing transitions between them. However, additional dynamical information can often be encoded in a cell-cell Markov transition kernel. Existing generative approaches for single cell trajectory inference either infer transport only from population marginals, impose a symmetric geometry on the state space, or incorporate directionality through a single velocity vector at each observed state. We introduce Finsler Flow Matching (FFM), a framework for learning continuous stochastic dynamics from discrete Markov transition graphs. We use the first and second moment to construct a Finsler structure motivated by the Freidlin–Wentzell action, where the second moment determines anisotropic accessibility and the first moment introduces a preferred direction of motion. We learn neural approximations of the resulting directed geodesics, use their Finsler cost to construct source-target couplings, and define geometry-aware stochastic conditional paths that can be distilled into a continuous generative process through simulation-free score and flow matching. Across synthetic and single-cell trajectory inference benchmarks, FFM improves recovery of withheld intermediate populations, particularly when the transition dynamics are strongly directional or anisotropic. Our results provide a principled route from discrete transition probabilities to continuous generative dynamics while retaining both directional and diffusive structure.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
flow matching, Finsler geometry, single-cell trajectory inference, optimal transport, Markov chains, stochastic dynamics

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_gYJwOpenReview ↗
Rating
4(+2)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents an optimal transport based trajectory inference algorithm, which can incorporate directional biological preference based on Finsler geometry. It allows the dynamics model to consider biological prior knowledge such as RNA velocity and lineage tree. These leads to more faithful dynamics modeling.

Critical strengths

1.I find the use of Finsler geometry to incorporate directional prior information both elegant and novel.

2.To the best of my knowledge, how to incorporate biological knowledge into the dynamics modeling is a central and critical problem in trajectory inference. I believe this paper is trying to address a good and hard problem, and its method makes lots of sense theoretically.

Critical weaknesses

1.Some definitions are missing, and some details of algorithm implementation and datasets are missing, which makes the paper harder to follow.

2.The real world data experiments are seem inappropriate. I suspect whether its real world data experiments support their claims.

See questions for details.

Questions for authors

Questions on the mathematical and technical details

  1. Although this may seem obvious, it would be helpful to explicitly state that the quantity defined in Eq. (9) (I am not sure what the appropriate terminology is, so let us call it "length" for now) is nonnegative. This is important because this quantity is squared in the definition of the OT cost in Eq. (7). For example, if this quantity could take negative values, squaring it in Eq. (7) would destroy the directional preference induced by the Finsler structure.

  2. The conditional path defined in Eq. (11) appears to be very high-dimensional. Does its optimization suffer from the curse of dimensionality? Approximately how much computational time is required to optimize this term?

  3. How is the Finsler cost in Eq. (13) computed in practice? Do the authors perform numerical integration for every pair of data points?

  4. Based on Eq. (6), I would expect the drift term of the SDE on line 364 to contain only the first two terms. Where does the additional term involving the divergence of CC come from?

  5. What is the "sampling floor" mentioned on line 390? I could not find its definition in either the main text or the appendix.

Major Concerns

My major concerns are as follows.

1. The validity of the temporal information used in the real-data experiments.

To the best of my knowledge, all cells in the Pancreas dataset (Bastidas-Ponce et al., 2019) used in this paper were collected at the same developmental time point (E15.5). What, then, do the three time points t=0,1/2,1t=0, 1/2, 1 in the paper actually represent? I believe that applying the proposed method to single-time-point data is conceptually inappropriate.

A similar issue arises with the erythroid dataset. Although this dataset does contain multiple time points, it is clear from the scVelo documentation that the actual temporal labels do not correspond to the three time points 0,1/2,10, 1/2, 1 presented in the paper.

The authors should provide a detailed description of these real-world datasets in the appendix, particularly explaining how the temporal labels were constructed and what they represent.

2. Potential data leakage through RNA velocity.

I also noticed that both the Pancreas and erythroid datasets were obtained from scVelo (see below). The paper states that RNA velocity information can be incorporated when working with real-world data. However, RNA velocity estimation methods such as scVelo generally rely on the spliced and unspliced counts of the entire cell population, rather than treating each cell independently. I am therefore seriously concerned that incorporating RNA velocity may introduce data leakage. The authors should clarify how RNA velocity was estimated and whether information from the evaluation data was used during this process.

Overall Assessment

I have no major concerns regarding the proposed model, mathematical formulation, or technical details. In fact, I find the use of Finsler geometry to incorporate directional prior information both elegant and novel.

However, for the reasons outlined above, I am concerned that the real-world experiments do not adequately support the paper's claims and may involve some fundamental methodological issues.

If I have misunderstood any aspect of the experimental setup, I would appreciate clarification from the authors, and I would be willing to raise my rating accordingly. Otherwise, I believe that the real-world experiments should be conducted again on more appropriate datasets.

Data source

Pancreas: https://scvelo.readthedocs.io/en/stable/scvelo.datasets.pancreas.html Erythroid: https://scvelo.readthedocs.io/en/stable/scvelo.datasets.gastrulation_erythroid.html#scvelo.datasets.gastrulation_erythroid

5 comments

We thank the reviewer for the careful reading and constructive comments. We agree that several aspects require clarification.

Q1: Non-negativity of Eq. (9). By Cauchy–Schwarz, with A=Cρ−1A=C_{\rho}^{-1},

FP(x,v)≥(∥m∥A2+λ2−∥m∥A)∥v∥A≥0.F_P(x,v) \geq \left(\sqrt{\lVert m\rVert_A^2+\lambda^2}-\lVert m\rVert_A\right)\lVert v\rVert_A \geq 0.

Since λ>0\lambda >0 and CC is positive definite, the inequality is strict v≠0v\neq 0. We will state this explicitly in the manuscript.

Q2: Computational complexity of Eq. (11). On Pancreas, total deterministic FFM runtimes on a V100 are 26/116/277s for d = 2/20/50, respectively. The geodesic phase accounts for 78–86% of runtime.

Q3: Computation of the Finsler cost in Eq. (13). After training the geodesic interpolant, we evaluate the cost of each endpoint pair using numerical quadrature with 8 nodes along the learned path, followed by Sinkhorn iterations. We will add this clarification to the paper.

Q4: Divergence term in the SDE. The additional term arises from the state-dependent diffusion covariance. For a(t,x)=g(t)2Cρ(x)a(t,x)=g(t)^2C_{\rho}(x), the Fokker–Planck equation requires the drift correction 12∇a\frac12\nabla a in addition to 12a∇log⁡pt\frac12 a\nabla\log p_t to preserve the prescribed probability marginals. Since gg depends only on time, this yields 12g(t)2div⁡Cρ(x)\frac12 g(t)^2\operatorname{div}C_\rho(x). We will add the corresponding derivation to the manuscript and clarify its distinction from the state-independent case in Eq. (6).

Q5: Sampling floor. The sampling floor is the empirical W2 distance between two disjoint, subsets of the reference population. We will clarify its definition in the experimental section.

W1: Interpretation of temporal information in real-data experiments . We thank the reviewer for raising this point, we agree that this was not described with sufficient clarity in the manuscript. The reviewer is correct that all Pancreas cells were collected at a single time point (E15.5). This is in fact the setting that FFM was designed for. A primary motivation of our method is to enable trajectory inference from single snapshot experiments. Because differentiation is asynchronous, one snapshot can capture cells along the full trajectory and additional dynamical information, such as RNA velocity, provides the local directionality that would otherwise come from time labels. FFM leverages this information, encoded as a cell-cell transition kernel, to infer continuous trajectories within the snapshot. The values t = 0, 1/2, 1 in our experiments therefore do not denote experimental time points, but rather bins of a single snapshot along an inferred progression coordinate (tertiles of scVelo latent time for Pancreas, diffusion pseudotime for iTracer). The first and last tertiles serve as p0/p1p_0/p_1 , the middle is withheld, and the model must reconstruct it from the endpoints and the dynamical information alone. This is the single-snapshot analogue of leave-one-time-point-out evaluation. For Erythroid, our setup directly follows CurlyFM (Petrović et al., 2025) to allow direct comparison. Their benchmark does not use the embryonic stage labels, but pools the cells and splits them into three quantiles of latent time, which is why our labels do not match the stages in the scVelo documentation. We will edit the manuscript to replace “temporal snapshots” with “pseudotime partitions” and add an appendix section describing the construction for each dataset.

W2: Potential RNA-velocity leakage. The reviewer is right and we thank the reviewer for identifying this important issue. The transition graph was recomputed without withheld cells, but the velocities had been estimated on all cells, which can lead to leakage into the test set. For this response we re-estimated velocity, including the scVelo moments, after excluding the withheld cells. We then repeated the experiments using these refitted velocities. Pancreas:

\hline d & \text{FFM (original/refit)} & \text{Curly-FM (original/refit)} & \text{MFM} & \text{OT-CFM} \\\\ \hline 2 & 1.40/1.60 & 0.95/1.54 & \mathbf{1.47} & 1.76 \\\\ 20 & 10.48/\mathbf{10.95} & 12.19/12.41 & 12.11 & 12.34 \\\\ 50 & 12.57/\mathbf{12.82} & 13.28/13.47 & 12.96 & 13.49 \\\\ \hline \end{array}$$ The results confirm that the original velocity estimates contained information from the withheld cells. Removing this information worsens both velocity-informed methods, particularly for CurlyFM in 2D. With refitted velocities, FFM retains the lowest mean midpoint W2 at d = 20 and 50. At d = 2, MFM becomes the best-performing method. For Erythroid, refitting velocity changes FFM's 2D midpoint W2 from 0.289 to 0.270, and Curly-FM from 0.158 to 0.226. At d = 20/50, all changes are at most 0.012. We will rerun all velocity-based experiments under this protocol and report the refitted values as the main results.

Q1. Thank you.

Q2. I see. This is quite affordable. I have two further questions.

  1. Is there any theoretical or empirical evidence that the path optimization is indeed solved? Or you just try your best to optimize it? (That's OK for me, since the optimization problem seems complicated.) I have read some article like https://arxiv.org/abs/2609.04710. They also tried to learned the interpolant and numerical integrated cost. They mentioned about the low dimensionality of the interpolant makes it capable to learn. However, in your paper, it feels like high dimensional interpolant is also OK to learn. They have reported the error between learned interpolant, cost and the true interpolant , cost on some cases. Can you also do some empirical check? I am very curious about whether high dimension brings difficulty to optimization.

2.116s, 277s seems larger than the total time reported in your table 2?

Q3. Thank you. But isn't 8 too small? 50 or 100 may sound better. Just a comment.

Q4, Q5. Ok, thank you.

W2. Amazing results, that's sufficient to me.

W1. I have major concern about this application setting!! A primary motivation of our method is to enable trajectory inference from single snapshot experiments. To the best of my knowledge, classical trajectory inference were developed to overcome the lack of true physical time. Nowadays, we can measure transcriptomics at different physical time, why should we go back to single snapshot?? I didn't understand.

Pseudotime is not time, it is not even a good analog to time. It is just something like a metric based on cell-cell similarity. I don't think it is proper to use pseudotime as labels. Also, velocity pesudotime, diffusion map based pseudotime, palantir, ……, all of these pseudotime methods were computed on all cells. This brings inevitable data leakage. And, dosn't pseudotime itself something like 'trajectory'? From this perspective, I seriously believe the real data experiments need to be conducted on more proper multi time point datasets.

Their benchmark does not use the embryonic stage labels, but pools the cells and splits them into three quantiles of latent time Like what I mentioned above, in my opinion, that's false.

Overall, I truely appreeciate the idea of introducing the Finsler metric. It is really elegant and promising. But I seriously doubt the experimental setting. If the authors can address W1, I will raise my score to 4, otherwise, I will change it to 1.

Part 1/2 We thank the reviewer for the quick and transparent follow-up, and for their encouraging comments regarding the Finsler construction.

Q2. To test how close we can get to solve this optimization problem we would propose to run and experiment on a sphere Sm⊂RnS^m \subset \mathbb{R}^n for mm and nn large enough, since there we have analytical solutions to geodesic. We will provide details once finished. Furthermore, the formulation for the runtime analysis might have been imprecise out of the number mentioned above 21/99/24s corresponding to the time spent on the geodesic optimization.

W1. We very much agree with the reviewer that the distinction of real time and pseudotime is important and we thank them for emphasizing it. Crucially, pseudotime partitions cannot be interpreted as independent temporal ground truth. We therefore also agree that real-data evaluation on experimental time points would greatly enhance the manuscript and we are running these experiments now:

  • Embryoid body differentiation (Moon et al., 2019): five experimental time points, with the leave-one-time-point-out protocol used by MFM (Kapusniak et al., 2024) and earlier work.

  • Mouse haematopoiesis, LARRY (Weinreb et al., 2020; days 2, 4 and 6): prediction of fate probabilities to compare different kernel constructions, including one derived from the experimental time points.

We will report these results as soon as they are complete. We will integrate them into the manuscript as the major component of the real data evaluation.

Furthermore, we already ran experiments on Mouse erythroid development (Pijuan-Sala et al., 2019) we highlight here the results in PCA d = 20. We evaluated a holdout of experimental time points, using embryonic-stage labels (E7.0–E8.5) instead of latent time. We define $p0p_0 $using E7.0–E7.75 and �1��1� using E8.5, withholding all intermediate-stage cells (E8.0 and E8.25). We compare either samples generated at �2��2� or pooled samples from five equally spaced time points in �3��3� against the withheld population.

\hline \text{Method} & t=1/2 & t\in[1/3,2/3] \\\\ \hline \text{Sampling floor} & 4.08 & 4.08 \\\\ \text{OT-CFM} & 6.72 & 6.66 \\\\ \text{MFM / LAND} & 5.70 & 5.62 \\\\ \text{Curly-FM} & 10.04 & 9.61 \\\\ \text{FFM (velocity)} & 6.39 & 6.40 \\\\ \text{FFM (expression)} & \mathbf{5.47} & \mathbf{5.57} \\\\ \hline \end{array}$$ We additionally performed a leave-one-out evaluation using actual embryonic stages. Each intermediate stage was withheld and predicted at $�4�$from E7.75 → E8.25, and from E8.0 → E8.5. $$\begin{array}{lcccc} \hline \text{Method} & \text{E8.0} & \text{E8.25} \\\\ \hline \text{Sampling floor} & 3.87 & 3.85 \\\\ \text{OT-CFM} & 5.76 & 5.54 \\\\ \text{MFM / LAND} & 5.88 & 5.32\\\\ \text{Curly-FM} & 7.50 & 7.36 \\\\ \text{FFM (velocity)} & 5.61 & 5.46 \\\\ \text{FFM (expression)} & \mathbf{5.59} & \mathbf{5.17} \\\\ \hline \end{array}$$
Reviewer_ykPbOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper tackles trajectory inference from single-cell snapshots. The authors construct a row-stochastic cell-to-cell transition matrix P estimated from different information such as RNA velocity, lineage information, or combinations. FFM aims to generate plausible intermediate distributions while respecting directional information in P. FFM extracts the expected displacement(m) and uncentered second moment(C) from P, and constructs a regularized Randers/Finsler metric. A neural model learns approximate directed geodesics, whose squared energies supply entropic OT pairing costs. Flow matching then learns the resulting paths, with optional score matching and stochastic noise. The metric is motivated by Freidlin–Wentzell minimum-action theory under diffusion-limit assumptions. Experiments include synthetic dataset (Sheet and ForkSheet) and three single-cell datasets (Pancreas, Erythroid, iTracer).

Critical strengths
  • The method is presented intuitively, including the graph-to-geometry construction: the first moment provides directionality and the second indicates transition accessibility. The Randers form is mathematically valid and can incorporate different types of single-cell information.
  • The experiments include relevant baselines (OT-CFM, MFM, Curly-FM) and useful same-information controls: drift-only, fixed-time anisotropic, second-moment-only and one-form ablations, with shared versus separate couplings.
Critical weaknesses
  • The main contribution is a particular Randers metric derived from Markov transition moments, while geodesic learning, OT coupling, and score/flow matching largely build on existing methods. Closely related work (Zweig et al., 2026) has already incorporated directed transition priors into Finsler-based trajectory inference. Although the proposed graph-moment construction is a meaningful distinction, its incremental advantage over existing Finsler-based approaches is not directly established.
  • FFM performs well on Sheet and Pancreas 20D, but its relative advantage varies across datasets and representation dimensions. The benefits of stochastic sampling and different transition-kernel constructions are also inconsistent. The paper attributes its improvements to strongly directional or anisotropic transition dynamics, but does not systematically characterize these properties in the real datasets or relate them to the observed performance differences.
Questions for authors
  • What explains the strong dependence on representation dimensionality? For example, FFM underperforms Curly-FM on Pancreas and Erythroid in 2D but becomes more competitive in higher dimensions. Could the authors provide some intuition, and/or extend the matched-information ablations to settings where FFM is less competitive?
  • How does the proposed graph-derived Finsler geometry compare with closely related approaches, particularly Zweig et al. (2026), and prior-constrained transport frameworks such as WLF (Neklyudov et al., 2024) and GSBM (Liu et al., 2024)? A direct comparison or a more detailed explanation on the practical advantage specific to the proposed construction would be useful.
  • The benefits of stochastic sampling and velocity/lineage-informed transition kernels vary across datasets and representation dimensions. Could the authors clarify when these components are most useful, and whether a consistent default configuration performs competitively without dataset-specific selection?

Neklyudov, Kirill, et al. "A computational framework for solving wasserstein lagrangian flows." arXiv preprint arXiv:2310.10649 (2023).
Liu, Guan-Horng, et al. "Generalized schrödinger bridge matching." International Conference on Learning Representations. Vol. 2024. 2024.

4 comments

Part 1/2 We thank the reviewer for the detailed assessment. We agree that the manuscript did not sufficiently distinguish FFM from related Finsler-based and prior transport methods, and that it attributed improvements to directionality/anisotropy without characterizing these properties in the data. Below we clarify the relation to prior work, report new experiments and specify when stochastic sampling and the different kernel constructions are useful.

W1/Q2: Relation to Zweig et al., WLF and GSBM. We thank the reviewer for this important point and agree that the relation to Zweig et al. (2026) deserves a more detailed discussion. Both methods use Finsler geometry to introduce directionality into geodesic interpolation, but they address different problems and require different inputs. Zweig et al. address the interpolation between populations observed at multiple experimental time points, given a known lineage tree over annotated cell types. Their goal is to keep inferred trajectories consistent with permitted transitions between cell types. FFM addresses trajectory inference within a single snapshot, given a cell-cell Markov transition kernel (e.g. from RNA velocity) and does not require annotations. The goal of FFM is to convert discrete transition probabilities between cells into continuous stochastic dynamics.

Their approach is therefore not applicable to our setting. It requires cell type labels, a lineage tree and time-resolved populations, none of which our problem provides. Their construction carries no directional information within cell types, which is where the continuous differentiation takes place and where RNA velocity is informative. The novelty of FFM is the construction of Finsler geometry directly from the transition kernel: 1) directionality is defined at every cell through the first transition moment, 2) direction and anisotropic accessibility are derived jointly from the same kernel, rather than adding a penalty to a pre-trained symmetric metric, 3) the resulting Randers metric has a Freidlin–Wentzell interpretation as most-probable transition paths (Theorem 2) and defines the noise geometry of a stochastic generative model, whereas Zweig et al. learn deterministic paths. Both WLF (Neklyudov et al., 2024) and GSBM (Liu et al., 2024) are complementary optimization frameworks: once this geometry is defined, both can optimize the resulting transport problem as alternatives to FFM's geodesic-learning and flow-matching pipeline (GSBM we adapt Euclidean kinetic cost). We compared these alternatives using identical graph-derived geometry, parameters, data, and five seeds on Pancreas in 20D and on the Sheet.

\hline d & \text{FFM} & \text{F-WLF} & \text{F-GSBM} & \text{Eucl. WLF/GSBM} \\\\ \hline 2 & \mathbf{0.103} & 0.45 & 0.41 & 0.70/0.69 \\\\ 20 & \mathbf{10.39} & 10.51 & 10.64 & 12.45/12.34 \\\\ \hline \end{array}$$ Both Finsler adaptations improve substantially over their Euclidean counterparts, demonstrating the benefit of the graph-derived geometry independently of the solver. FFM achieves the lowest midpoint error, particularly in 2. Our key construction of a dynamics-informed geometry from transition probabilities generalizes to different optimization frameworks. We will expand the related-work section of the manuscript accordingly. **W2/Q1: Relationship between performance and dimensionality**. We thank the reviewer for this observation and agree that the relationship between data geometry and performance requires further characterization. We added two new analyses to address this. First, we found that part of the 2D result was an artifact. In response to reviewer 3, we re-estimated RNA velocity without the withheld cells. On Pancreas at d = 2, Curly-FM changes from 0.95 to 1.54 and FFM from 1.40 to 1.60, while MFM, which uses no velocity, remains at 1.47. Curly-FM’s 2D advantage was thus largely due to information from the withheld cells. With refitted velocities, no velocity-informed method improves on MFM at d = 2, and FFM remains best at d = 20 and 50. Second, to better characterize the used datasets, we estimated intrinsic dimensionality using the maximum-likelihood estimator (Levina & Bickel, 2004) on the Pancreas dataset. $$\begin{array}{ccc} \hline \text{Ambient dimension} & \text{Intrinsic dimension (MLE)} \\\\ \hline 2 & 1.8 \\\\ 20 & 7.6 \\\\ 50 & 13.0 \\\\ \hline \end{array}$$ In a 2-dimensional embedding the ambient and intrinsic dimensions are nearly the same, so the second moment is close to isotropic. We regard d = 2 to be primarily useful for visualization and d >= 20 as the relevant setting because single-cell analyses, including the construction of graphs and transition kernels, are typically performed in (PCA) representations of this dimensionality. Levina, E., & Bickel, P. J. (2004). Maximum Likelihood Estimation of Intrinsic Dimension. Advances in Neural Information Processing Systems (NeurIPS)

Part 2/2 W2/Q3: Benefits of stochastic sampling and different transition kernel constructions. A major benefit of stochastic sampling in our setting is the evaluation of fate probabilities for unseen cells. We evaluated conditional fate probabilities on Pancreas (d = 20) in the response to Reviewer 1. We found that the fate probability error is 0.103 for stochastic FFM, 0.538 for deterministic FFM and 0.160 for a constant predictor [details in response to Reviewer 1]. Isotropic noise performs similarly (0.106), so the benefit is due to stochasticity, not to the noise geometry. To further investigate how the model recapitulates fate probabilities, we will run experiments on a new dataset (LARRY, Weinreb et al., 2020) and will provide the reviewer with these new results at the latest 16. November. Regarding kernel construction, on iTracer d >= 20, the four kernel variants perform comparatively, so adding lineage information gave little measurable benefit in this evaluation. Nevertheless, we consider the ability to incorporate such information as an advantage of our method. Any measurement that can be expressed as a cell-cell transition kernel, including velocity, lineage tracing, metabolic labelling or their combinations, can be leveraged by FFM through. Whether a given modality improves recovery depends on how informative it is for the task at hand. The iTracer experiment demonstrates that these modalities can be incorporated, but it could not determine when they help. We will further evaluate lineage-informed kernels on the LARRY dataset (Weinreb et al., 2020) and will report the results at the latest on 16. November.

Weinreb, C., Rodriguez-Fraticelli, A. E., Camargo, F. D., & Klein, A. M. (2020). Lineage tracing on transcriptional landscapes links state to fate during differentiation. Science, 367(6479)

W2/Q3 follow-up: LARRY results. We additionally evaluated stochastic sampling and lineage-informed kernels on LARRY (Weinreb et al., 2020). LARRY provides experimentally observed clonal fates, allowing us to evaluate whether the fate probabilities predicted by FFM agree with the differentiated relatives of each Day-2 cell.

For fate prediction, we evaluated Day-2 cells and sampled 64 trajectories per cell (noise width 0.3). We report the centered fate MSE against the observed clonal fates: the mean over cells is subtracted from both predicted and observed fate vectors before computing the squared error. A constant prediction, assigning mean over fates, scores 0.557

At d=20d=20, stochastic sampling reduces the error from 0.591 to 0.457. SF2M performs similarly, with an error of $ 0.468 \pm 0.022 $.

MethodError at d=20d=20
Constant predictor$0.557 $
FFM deterministic$0.591 \pm 0.015 $
FFM stochastic0.457±0.0210.457 \pm 0.021
FFM stochastic, isotropic0.437±0.0180.437 \pm 0.018
SF2M0.468±0.0220.468 \pm 0.022

We also tested whether lineage information improves the transition kernel. In each seed, we split the clones 50/50. The lineage part of the kernel sends a Day-2 cell to the Day-6. We report the correlation between predicted and observed Neutrophil/(Neutrophil + Monocyte) bias on held-out clones (d = 20). “Kernel alone” uses the kernel’s own absorption probabilities, without FFM.

KernelKernel aloneFFM deterministicFFM stochastic
Velocity0.259±0.0250.259 \pm 0.0250.299±0.0270.299 \pm 0.0270.433±0.0550.433 \pm 0.055
Velocity + lineage0.588±0.0140.588 \pm 0.0140.269±0.0290.269 \pm 0.0290.486±0.0500.486 \pm 0.050
Expression−0.008±0.029-0.008 \pm 0.0290.301±0.0420.301 \pm 0.0420.445±0.0510.445 \pm 0.051
Expression + lineage0.593±0.0180.593 \pm 0.0180.272±0.0420.272 \pm 0.0420.486±0.0280.486 \pm 0.028

Lineage makes the kernel itself substantially more predictive than velocity (0.590.59 vs. 0.260.26), but FFM retains only part of this signal: at d=20d=20, adding lineage improves the stochastic model by at most 0.050.05. In contrast, stochastic sampling improves FFM for every tested kernel at d=20d=20.

These results refine our conclusion from iTracer: lineage information can contain substantially more fate information than velocity, but FFM currently captures only part of this additional signal. Stochastic sampling has the larger and more consistent effect, although its benefit is specific to d = 20 in our experiments.

Dear reviewer ykPb,

As the deadline for the reviewer-author discussion is approaching, please check whether the authors have adequately addressed your concerns. If any concerns remain unresolved, please raise them in the discussion.

Best regards,

Your AC

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

The paper introduces Finsler Flow Matching (FFM), a framework for learning continuous stochastic dynamics from discrete Markov transition graphs. Motivated by the Freidling-Wentzell action, they learn the resulting directed geodesics, from their Finsler cost to construct source-target couplings, and define geometry-aware stochastic conditional paths that can be distilled into a continuous generative process through simulation-free score and flow matching. Experiments evaluate intermediate-population recovery on synthetic manifolds and three single-cell datasets.

Critical strengths
  • A principled connection between transition information and trajectory geometry. The method combines directional information from the first transition moment with anisotropic accessibility from the second moment through a regularized Randers metric. This provides a coherent way to incorporate both effects into trajectory inference. The connection to Freidlin–Wentzell theory further motivates the construction, with an interpretation of limiting minimizing curves as most-probable transitions under the stated diffusion-consistency and asymptotic assumptions.

  • Convincing improvements in important benchmark settings. FFM substantially improves intermediate-population recovery on the synthetic Sheet benchmark and 20-dimensional Pancreas. The metric ablation in Figure 4(b) supports the contributions of directionality and anisotropic accessibility. Moreover, the matched-information comparison in Figure 3 provides evidence that the Pancreas improvement is not explained solely by access to second-moment information. These results support the proposed trajectory geometry, although the gains depend on the representation.

Critical weaknesses

The additional benefit of the stochastic component remains unclear. On 20-dimensional Pancreas, the covariance ablations in Figure 5(d–f) show comparable performance for isotropic and anisotropic conditional-path and inference covariances. Adding stochasticity also produces only small or inconsistent changes in intermediate-population recovery across Pancreas representations (Figure 3). Although the ForkSheet experiments evaluate cell-level fate preservation, they do not isolate the contribution of anisotropic noise or demonstrate a consistent advantage for the stochastic variant. Improved marginal recovery is not necessary for stochastic modeling to be useful, but the current evaluation leaves its additional value insufficiently demonstrated. Conditional fate-probability calibration or other statistics sensitive to stochastic dynamics would provide more direct evidence.

The most informative controlled comparisons have limited coverage. The matched-information ablation in Figure 3 is valuable, but it is restricted to 20-dimensional Pancreas. Its generality therefore remains uncertain. In addition, the iTracer comparison omits Curly-FM despite using RNA-velocity information in some FFM variants. Extending the matched-information control to another setting and comparing against Curly-FM on iTracer would more directly establish how consistently the proposed geometry improves on alternatives with comparable dynamical information. Comparisons with [SF]²M [1] or CDC-FM[2] would provide further context for the stochastic extension.


References [1] Simulation-Free Schrödinger Bridges via Score and Flow Matching [2] Carré du champ Flow Matching

Questions for authors
  1. Given the comparable isotropic and anisotropic covariance results in Figure 5(d–f), what benefit should be attributed specifically to the proposed noise geometry? Beyond the existing ForkSheet fate-preservation results, can you provide evidence using conditional fate probabilities or another statistic sensitive to stochastic dynamics?

  2. The matched-information ablation in Figure 3 supports the proposed objective on 20-dimensional Pancreas. Does this advantage persist on another benchmark or representation when the competing objectives receive the same transition moments and share the endpoint coupling?

  3. Why is Curly-FM omitted from the iTracer comparison despite the availability of RNA-velocity information? Could you report this comparison using the same velocity input available to the velocity-based FFM variant?

2 comments

We thank the reviewer for the constructive suggestions. We conducted additional experiments addressing them below.

W1/Q1: Value of the stochastic component. We evaluated conditional fate probabilities on Pancreas (d = 20). Here, the second half of the middle latent time tertile serves as the source, because cells at the original p0 share the same fate probabilities. For each source cell we simulate 64 SDE trajectories, label each endpoint with the majority cell type of its 15 nearest target cells, and use the resulting cell-type frequencies as predicted fate probabilities. Since no ground truth fate probabilities are available in this dataset, we compare them with fate probabilities computed by CellRank to evaluate consistency with the supplied dynamical information. We report MSE after centering each predictor (summed over the four fates, averaged over cells), which measures how well cell-to-cell differences are predicted.

\hline \text{Method} & \text{Centered fate MSE}\downarrow & \text{Endpoint }W_2\downarrow \\\\ \hline \text{Constant predictor} & 0.160 & - \\\\ \text{FFM deterministic} & 0.538 & \mathbf{5.34} \\\\ \text{FFM SDE (anisotropic)} & \mathbf{0.103} & 5.70 \\\\ \text{FFM SDE (isotropic)} & 0.106 & 5.38 \\\\ \text{[SF]}^2\text{M SDE} & 0.129 & 5.45 \\\\ \hline \end{array}$$ In 20D, stochastic FFM at noise width 0.3 achieves a per-cell fate error of 0.103, compared with 0.160 for the constant predictor and 0.538 for deterministic FFM, while retaining comparable endpoint W2 (5.70 vs. 5.34). Isotropic noise performs similarly (0.106). In 50D, meaningful fate variation is only recovered with noise levels that substantially degrade endpoint fidelity. Thus, stochastic sampling captures additional cell-specific fate information in 20D, but we here find no clear benefit specific to anisotropic noise. **W2/Q2: Matched information ablation.** We repeated the fixed-coupling comparison with identical transition moments on Pancreas at all dimensions (midpoint W2). $$\begin{array}{lccc} \hline \text{Method} & d=2 & d=20 & d=50 \\\\ \hline \text{Graph-drift FM} & 1.61 & 11.94 & 13.49 \\\\ \text{Anisotropic fixed-time FM} & 1.57 & 11.57 & 12.74 \\\\ \text{FFM (ours)} & \mathbf{1.43} & \mathbf{10.40} & \mathbf{12.72} \\\\ \hline \end{array}$$ Results averaged over 10 seeds at d = 2 and 5 seeds at d = 20/50. FFM improves on drift-only control in all settings. At d = 2/20/50, FFM obtains midpoint W2 of 1.43/10.40/12.72, compared with 1.57/11.57/12.74 for anisotropic fixed-time FM (Eq. 25) and 1.61/11.94/13.49 for graph-drift FM. The reviewer may also find a controlled ablation in Figure 4 in the Appendix of for the Sheet dataset. These results extend FFM's advantage over drift-only controls, while the advantage over the anisotropic fixed-time objective is clear at d = 2 and 20 but absent at d = 50. **W3/Q3: CurlyFM on iTracer.** We additionally evaluated Curly-FM using the same scVelo velocity information as velocity-informed FFM. $$\begin{array}{lccc} \hline \text{Method} & d=2 & d=20 & d=50 \\\\ \hline \text{OT-CFM} & 1.363 & 15.91 & 18.88 \\\\ \text{MFM} & 1.354 & 15.19 & 18.33 \\\\ \text{Curly-FM (velocity)} & 1.351 & 15.95 & 19.03 \\\\ \text{Curly-FM (P moment)} & 1.350 & 15.83 & 19.03 \\\\ \text{FFM + velocity} & \mathbf{1.104} & \mathbf{12.73} & \mathbf{16.74} \\\\ \hline \end{array}$$ Across d = 2/20/50, Curly-FM obtains W2 of 1.351/15.95/19.03, versus 1.104/12.73/16.74 for FFM + velocity. Using the first moment of the velocity kernel as Curly-FM's input gives similar results (1.350/15.83/19.03). We implemented both baselines under the same training protocol. **W4: Comparison to [SF2M] and CDC-FM.** On Pancreas at d = 20/50, stochastic FFM achieves mid point W2 of 10.45/12.56, versus 12.41/13.49 for [SF]²M and 12.32/13.27 for CDC-FM. These comparisons support the contribution of transition-derived geometry beyond stochasticity alone. For conditional fate probabilities, [SF]²M is the more direct stochastic reference: in 20D Pancreas it achieves per-cell error 0.129 versus FFM's 0.103. Hence, the experiments support the value of the geometric transport construction.

Thank you for the thorough response and the additional experiments. The Curly-FM comparison on iTracer with the same velocity input, the comparisons with [SF]²M and CDC-FM, and the extension of the matched-information ablation to d = 2 and d = 50 address my questions on baseline coverage, and they support the proposed transport geometry. I will keep my positive score of 3 (weak accept).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.