acceptodds
Under review as a conference paper at ICLR 2027

GraphFM: Graph-Constrained Flow Matching for Dynamics Reconstruction from Temporal Snapshots

51 views

Abstract

Reconstructing continuous population dynamics from sparse, unpaired temporal snapshots is a fundamental yet underdetermined problem in many scientific applications. Many vector fields can reproduce the same finite collection of temporal marginals. Existing flow-matching methods reduce this ambiguity through assumptions on endpoint couplings, interpolation geometry, or temporal smoothness, but typically treat the variables within each state as unstructured. This overlooks known interaction structure in systems such as gene-regulatory and road networks. We introduce Graph-Constrained Flow Matching (GraphFM), which incorporates a variable-interaction graph into both conditional interpolation and vector-field learning. GraphFM parameterizes an endpoint-constrained graph neural interpolant whose message-passing structure restricts each variable to depend only on its graph-defined receptive field. To exploit multiple observed marginals, we train the interpolant with a hierarchical objective combining distributional matching at observed time points, multi-anchor trajectory regularization induced by a multi-marginal coupling, and local temporal smoothness. The derivatives of the learned trajectories then supervise a graph-aware velocity field defined over the same interaction graph. We evaluate GraphFM on synthetic dynamics, singlecell trajectory inference, spatial-transcriptomic interpolation, and traffic-state imputation. GraphFM improves or matches strong multi-marginal flow-matching baselines across these settings, while graph-corruption experiments demonstrate the value of informative structural priors. Our results show that variable-level interaction structure provides a effective inductive bias for reconstructing dynamics from temporal snapshots.

Primary area:
learning on time series and dynamical systems
Keywords:
Flow Matching, Trajectory Inference, Graph Neural Networks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DYzROpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper use graph constraint to regularize the flow matching based dynamics inference models. In classical dynamics inference algorithms such as OT-CFM, the velocity net is parameterized by a MLP. To take varaible correlations such as gene-gene regulations into account, the authors use a GNN, whose graph structure is obtained by some prior knowledge such as GRN, to replace the MLP. This technique advances the model performance on several datasets.

Critical strengths

The paper tackles with an important problem. Its method is also clear and easy to follow. I'd like to say this paper is well written.

Critical weaknesses

The new technique is quite natural and easy. (But turns out consistently useful. So that is not a big problem.)

My main concerns are 1.The baselines are quite easy, some critical baselines are missing.

2.The results in table1 and table2 are unmatched, which makes me wonder if there is any artifact.

Questions for authors

1.I am so surprised that only knowing the graph structure without the explicit form of the dynamical system can consistently improve the inferenced dynamics. Can author give some intuitive interpretation?

Main concerns

2.What is the definition of graph corruption? There are only experimental results, but no experiment details. Since this is a central ablation study, this definition is necessary.

3.Baselines are OT-MFM, OT-MMFM, OT-ALI-FM. First, the most basic baseline OT-CFM is missing. Second, these baselines and graph corruption ablations are not enough convincing for the claim that the improvment is gained through using the graph information. I advice the authors to add OT-CFM as a baseline. Also, I advice the authors to add naive GraphFM baselines where the graph are chosen to be (1) random graph (2) full connected graph (3) no graph (which is repalcing the GNN with MLP). Since GraphFM introduces several losses other than just graph constraint, I believe these baselines can further support the necessity of the graph constraints.

4.Why the last row of table1 is not consistent to the first row of table2? Aren't they the same experiment? I give rate 1 mainly due to this question. If there are any misunderstanding, I will raise my score.

2 comments

Question 1:

We thank the reviewer for this insightful question. Although the graph does not specify the functional form of the underlying dynamical system, it provides valuable information about dependencies among variables. By restricting irrelevant dependencies, such structural information can reduce the hypothesis space and introduce a useful inductive bias, potentially improving dynamics inference, particularly when observations are limited or noisy.

Our motivation is partly inspired by RegVelo [1], which incorporates gene regulatory network (GRN) priors into cellular dynamics inference. Biological processes are inherently complex, and their exact governing equations are often unknown. Although RegVelo assumes a specific form of RNA kinetics and parameterizes regulatory interactions using a nonlinear function, such formulations are still approximations of the underlying biological processes, for which alternative models, such as Hill-type regulation, also exist [2]. Importantly, RegVelo demonstrates that incorporating GRN priors can improve the inference of cellular dynamics within its assumed kinetic framework. This provides supporting evidence that structural information can offer valuable inductive biases even when the exact underlying regulatory mechanisms are not fully characterized.

Inspired by this observation, GraphFM explores the role of graph priors in a more general setting without assuming a specific mechanistic form of the governing dynamics. Rather than specifying how variables interact, we use the graph to constrain the dependencies of the interpolation model, and learning the corresponding dynamical behavior from data. Specifically, our message-passing GNN imposes an (L)-hop receptive field, such that each variable's interpolated state depends only on the endpoint variables within its graph-defined neighborhood.

[1] Wang, W., et al. RegVelo: Gene-regulatory-informed dynamics of single cells. Cell, 189(12), 3773–3800.e44, 2026.

[2] Tyson J J, Laomettachit T, Kraikivski P. Modeling the dynamic behavior of biochemical regulatory networks[J]. Journal of theoretical biology, 2019, 462: 514-527.


Question 2:

Thanks for your comments. We have made this point clearer. In Table 2, graph corruption is defined as replacing a fraction r of the directed edges in the original graph. Specifically, for a graph with E edges, we randomly remove \mathrm(rE) existing edges and add the same number of previously absent directed edges, uniformly sampled without replacement from all previously absent directed node pairs (excluding self-loops). Thus, the number of edges remains fixed. We use r\in{0,0.3,0.6,0.9,1.0}, where r=0 denotes the original graph and r=1 has no edges in common with it. The sampled graphs have no self-loops and are required to remain strongly connected. For each corruption level, the same graph is used across three training seeds. All conditions in this ablation use the SWD-only training setting.


Question 4:

Thanks for your comment. We have made our experimental setting clearer in our revised manuscript.. The first row of Table 2 reports the results obtained using SWD as the sole loss function, rather than the full loss configuration used in Table 1 (SWD + linear regularization + second-order smoothing). This explains the discrepancy between the two tables. Importantly, all experiments in Table 2 were conducted under the same loss configuration (SWD only), ensuring a fair comparison that isolates the effect of graph structure without interference from other regularization terms.

To further ensure transparency and reproducibility, we have made our implementation publicly available. The reviewer can verify the experimental settings and reproduce the results through our code repository: https://anonymous.4open.science/r/GraphFM-80E4A/

Question 3:

We thank the reviewer for the valuable suggestions. Following these suggestions, we have conducted additional experiments to further evaluate the effectiveness of GraphFM and the contribution of graph information.

  1. Additional Baseline Comparisons

We have added OT-CFM as an additional baseline, as suggested by the reviewer. We have also included OTP-FM, following the recommendation of another reviewer. The comparison with OT-CFM is presented below.

Dataset / SettingMetric ↓GraphFMOT-CFMOTP-FM(W2Inf)
Synthetic (Table 1)MMD0.04570 ± 0.007260.08392 ± 0.003410.12461 ± 0.00170
Synthetic (Table 1)SWD0.08199 ± 0.011090.12809 ± 0.005160.24327 ± 0.00316
EB (Table 3)EMD1.3711 ± 0.05321.3839 ± 0.00421.4971 ± 0.0312
Multi (Table 3)EMD7.9472 ± 0.06048.4602 ± 0.02219.5370 ± 0.2852
HT206B1 (Table 4)EMD7.5191 ± 0.08927.8663 ± 0.00307.5491 ± 0.0239
Traffic (Leave-One-Out, Table 5)EMD34.930 ± 0.13735.882 ± 0.14337.497 ± 0.112
Traffic (Leave-One-Out, Table 5)MAE2.837 ± 0.0103.055 ± 0.0153.028 ± 0.006
Traffic (Leave-One-Out, Table 5)MSE34.714 ± 0.23735.959 ± 0.21239.983 ± 0.271
Traffic (Joint Missing, Table 5)EMD34.992 ± 0.03435.994 ± 0.12037.439 ± 0.186
Traffic (Joint Missing, Table 5)MAE2.844 ± 0.0093.036 ± 0.0093.014 ± 0.018
Traffic (Joint Missing, Table 5)MSE34.884 ± 0.11636.373 ± 0.18439.968 ± 0.395

As shown above, GraphFM consistently outperforms OT-CFM across all evaluated settings, further demonstrating the effectiveness of our framework.

  1. Additional Graph Structure Ablations

To further isolate the contribution of graph information, we have conducted additional ablation experiments in both single-cell trajectory inference and traffic data imputation, considering the three settings suggested by the reviewer: (1) random graph, (2) fully connected graphs, and (3) no graph, where the GNN is replaced with an MLP. For each dataset, we generated an independent directed random graph. For each realization, we sampled without replacement from all ordered pairs of distinct nodes, keeping the number of edges exactly equal to that of the original graph. Edges present in the original graph were eligible for sampling. We did not constrain node degrees or connectivity. Each sampled graph was fixed across all held-out time points and five training seeds.

Setting (EMD ↓)Original Graph (GraphFM)Random GraphFully Connected GNNNo Graph (MLP)
EB single-cell1.371 ± 0.0531.345 ± 0.0331.379 ± 0.0081.349 ± 0.008
Multi single-cell7.947 ± 0.0608.135 ± 0.2588.020 ± 0.0178.210 ± 0.032
Traffic (LOO)34.930 ± 0.13735.043 ± 0.09836.311 ± 0.08835.008 ± 0.031
Traffic (Joint)34.992 ± 0.03435.152 ± 0.07436.273 ± 0.13035.128 ± 0.090

The results show that using the original graph achieves the best performance in three out of four settings, outperforming random graph, fully connected graphs, and the MLP-based variant. The exception is the EB dataset, where the random graph yield slightly better results.

Overall, these additional ablations provide further empirical evidence that incorporating meaningful graph structure can improve the performance of GraphFM, beyond the benefits of the other components in our framework.

Reviewer_p8NmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors introduce GraphFM which is a graph-constrained flow matching framework. The idea is to introduce a variable-interaction graph in both a learnt interpolant and in a velocity field learnt with FM. In the experiments, the auhtors show that their method improves held-out population reconstruction on synthetic data, one of two single-cell datasets, tumor sections, and traffic data.

Critical strengths

The experimental section evaluates the method on quite diverse datasets, so the breadth is good. The authors provide a synthetic corruption experiment which shows the sensitivity with respect to the supplied graph.

Critical weaknesses

The graph corruption experiment is tested only on synthetic data, but it should also be conducted on the real-world datasets. There is also differences in model architectures between GraphFM and the baselines (for exemple the architecture of the velocity field). So the comparison is not really fair. The implementation samples one intermediate pseudo-triplet for the synthetic and single-cell experiments. Therefore the experiments do not show how much the multi-anchor extension itself helps.

Questions for authors

See the weaknesses

3 comments

The graph corruption experiment is tested only on synthetic data, but it should also be conducted on the real-world datasets.

We thank the reviewer for this valuable suggestion. Following this suggestion, we have extended our graph ablation experiments from synthetic data to real-world datasets, covering both single-cell trajectory inference and traffic data imputation. Specifically, we replace the original graphs derived from external data with (1) random graph, (2) fully connected graphs, and (3) no graph, where the GNN is replaced with an MLP. The random graph is generated by uniformly sampling directed edges without replacement from all possible node pairs, excluding self-loops. The results are summarized below.

Table 1

SettingOriginal Graph (GraphFM)Random GraphFully Connected GNNNo Graph (MLP)
EB single-cell EMD ↓1.371 ± 0.0531.345 ± 0.0331.379 ± 0.0081.349 ± 0.008
Multi single-cell EMD ↓7.947 ± 0.0608.135 ± 0.2588.020 ± 0.0178.210 ± 0.032
Traffic (LOO) EMD ↓34.930 ± 0.13735.043 ± 0.09836.311 ± 0.08835.008 ± 0.031
Traffic (Joint) EMD ↓34.992 ± 0.03435.152 ± 0.07436.273 ± 0.13035.128 ± 0.090

As shown above, the original graph achieves the best performance in three out of four settings, with the EB dataset being the exception. These results provide further empirical evidence that incorporating meaningful graph structure can improve the performance of GraphFM on real-world datasets.

There is also differences in model architectures between GraphFM and the baselines (for exemple the architecture of the velocity field). So the comparison is not really fair.

We thank the reviewer for raising this important point. To address this concern, we conducted additional experiments in which we replaced the GNN-based velocity field in GraphFM with an MLP, while keeping the graph-based interpolator unchanged. We denote this variant as GraphFM (MLP velocity). This allows us to evaluate the contribution of graph information in the interpolation process independently of its use in the velocity field. The results are presented below, where bold indicates the best performance and ** indicates the second-best performance.

Table 2

SettingMetric ↓GraphFMGraphFM (MLP velocity)OT-MFMOT-MMFMOT-ALI-CFMOT-CFMOTP-FM
Synthetic (Table 1)MMD0.04570 ± 0.00726**0.03586 ± 0.000860.09400 ± 0.090400.09341 ± 0.001800.05757 ± 0.001200.08392 ± 0.003410.12461 ± 0.00170
Synthetic (Table 1)SWD0.08199 ± 0.01110**0.06285 ± 0.001120.14101 ± 0.125000.14393 ± 0.002800.08896 ± 0.001380.12809 ± 0.005160.24327 ± 0.00316
EB (Table 3)EMD1.3711 ± 0.05321.3400 ± 0.00561.3621 ± 0.0127**1.3842 ± 0.00921.5248 ± 0.03371.3839 ± 0.00421.4971 ± 0.0312
Multi (Table 3)EMD7.9472 ± 0.06048.1799 ± 0.0688**8.5482 ± 0.23058.4844 ± 0.01108.8042 ± 0.08668.4602 ± 0.02219.5370 ± 0.2852
HT206B1, 278 genes (Table 4)EMD7.5191 ± 0.08927.5975 ± 0.01507.98817 ± 0.014827.86642 ± 0.003367.63725 ± 0.074777.8663 ± 0.00307.5491 ± 0.0239**
Traffic LOO (Table 5)EMD34.930 ± 0.13735.116 ± 0.093**35.806 ± 0.50335.976 ± 0.48842.891 ± 4.66935.882 ± 0.14337.497 ± 0.112
Traffic LOO (Table 5)MAE2.837 ± 0.0102.907 ± 0.011**3.101 ± 0.0543.101 ± 0.0703.714 ± 0.4213.055 ± 0.0153.028 ± 0.006
Traffic LOO (Table 5)MSE34.714 ± 0.23735.019 ± 0.110**35.430 ± 0.65835.589 ± 0.56747.678 ± 8.05235.959 ± 0.21239.983 ± 0.271
Traffic Joint Missing (Table 5)EMD34.992 ± 0.03435.237 ± 0.15635.186 ± 0.553**35.713 ± 0.21839.913 ± 2.15735.994 ± 0.12037.439 ± 0.186
Traffic Joint Missing (Table 5)MAE2.844 ± 0.0092.912 ± 0.015**3.029 ± 0.0443.036 ± 0.0443.435 ± 0.1853.036 ± 0.0093.014 ± 0.018
Traffic Joint Missing (Table 5)MSE34.884 ± 0.116**35.283 ± 0.30334.572 ± 0.92635.597 ± 0.26542.668 ± 2.57936.373 ± 0.18439.968 ± 0.395

The results show that replacing the GNN-based velocity field with an MLP generally leads to a performance decline. Nevertheless, GraphFM (MLP velocity) still achieves the best performance on 3 metrics and the second-best performance on 5 metrics out of 11 evaluated metrics. These findings further support the effectiveness of incorporating graph information into the interpolation process, even when the velocity field itself does not explicitly utilize graph structure.

The implementation samples one intermediate pseudo-triplet for the synthetic and single-cell experiments. Therefore the experiments do not show how much the multi-anchor extension itself helps.

We thank the reviewer for this insightful suggestion. To evaluate the contribution of the multi-anchor extension, we conducted additional ablation experiments on single-cell trajectory inference (EB) and traffic data imputation. We excluded the Multi dataset because only three time points are available for training under the leave-one-out setting, making three-point and multi-time regularization equivalent. We considered five loss configurations: (1) SWD only, (2) SWD with three-point linear regularization, (3) SWD with multi-time linear regularization, (4) configuration (2) with additional smoothness regularization, and (5) configuration (3) with additional smoothness regularization. Three-point regularization uses three randomly selected observed time points, whereas multi-time regularization jointly incorporates all observed time points.

Table 3. Ablation study of multi-anchor regularization.

Loss ConfigurationEB EMD ↓Traffic Joint MAE ↓Traffic Joint MSE ↓
SWD2.1106 ± 1.30682.8850 ± 0.013535.6099 ± 0.8659
SWD + Three-Point Linear Regularization1.4141 ± 0.02653.1451 ± 0.054843.0393 ± 0.7838
SWD + Multi-Time Linear Regularization1.4210 ± 0.02452.6883 ± 0.007129.3444 ± 0.2108
SWD + Three-Point Linear Regularization + Smoothness1.3711 ± 0.05322.8737 ± 0.006736.7075 ± 0.2173
SWD + Multi-Time Linear Regularization + Smoothness1.3033 ± 0.01572.8441 ± 0.009134.8839 ± 0.1163

The results show that configuration (5) consistently outperforms configuration (4) across all evaluated metrics. Furthermore, configuration (3) outperforms configuration (2) on the traffic dataset, while showing slightly worse performance on EB. These findings suggest that the multi-anchor extension can provide additional benefits over three-point regularization, particularly when combined with smoothness regularization.

Reviewer_p7BhOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper presents GraphFM, a flow-matching algorithm that incorporates graphical structure of the data across multiple marginal snapshots. The algorithm first learns a graph-based interpolant between the snapshots, then distills the learnt interpolant into the velocity field via flow matching. GraphFM is benchmarked on single-cell, spatial transcriptomics, and traffic datasets.

Critical strengths
  • The paper is well-written and the algorithm and experiments are generally clearly understood.
  • Fundamental idea of incorporating graphical information in multimarginal trajectory inference appears well-motivated.
  • Synthetic data study is a good toy example demonstrating the benefits of the approach.
Critical weaknesses

I have several concerns regarding the novelty of the approach and empirical studies in the paper. I do not think they can be addressed within the rebuttal period which is why I am rating this a 1, but I will consider raising my score if they are.

Soundness

  • GraphFM makes several design choices that are not motivated or supported by ablation studies:
    • Why the particular linear + second-order smoothing interpolation?
    • Why the two-stage distillation rather than an all-in-one training?
    • Why the parameterization of Eq. 3? There is a large space of alternatives that satisfy the endpoints.
  • Relatedly, the comparisons to existing model conflates improvements from incorporating a graph network vs. the full GraphFM algorithm, which is an unfair comparison. E.g., all existing baseline trajectory inference methods [1-6], could be directly be used with a graph-based velocity model rather than the simple MLP models generally used. This would be a much stronger baseline to compare GraphFM too, which would also help justify some of the design choices. It is not clear whether using a graph network with e.g. MMFM would still be worse than GraphFM.
  • No training time benchmarking against existing trajectory inference methods [1-6]. How much overhead does the two-stage distillation entail?
  • The two-stage distillation is not explored in detail. What does the graph interpolation actually look like? Is it indeed smooth and intuitive? How well is the flow-matching velocity model able to learn the interpolation?
  • The synthetic dataset is a good toy example but do we expect any real-world dataset to obey such clean graphical dynamics (and that we can access all the variables and relationships needed to reconstruct that graph)? For example, the paper uses data mining to define a gene regulatory network as a static graph representation of the transcriptomic state for the single-cell dataset, but this is known to be noisy / incomplete. This may be a general limitation of this graph-based approach.

Informativeness

  • The algorithm overall is a simple combination of existing ideas, so the novelty is unclear. This could be compensated for, however, with more rigorous experimental analysis and ablations of each component.
  • As mentioned above, the linear + second-order smoothing interpolation is not motivated. Why this particular interpolation? Furthermore, there have been several papers exploring smooth multimarginal interpolations such as [1-6]. It is important to engage with them (discuss differences, pros and cons) and justify why this particular choice is the best with ablation studies.
  • Also mentioned above, it is important to perform a fair comparison between existing methods and break down where any gains are coming from. Currently there are too many changes bundled together, as well as some missing related work, so this aspect is unclear.
  • Prop. 1 appears to restate a basic fact of GNNs. This and the space leading up to it could be better used to address the concerns raised above.

[1] TrajectoryNet 2020, https://arxiv.org/abs/2002.04461

[2] DMSB 2023, https://arxiv.org/pdf/2303.01751

[3] WLF 2024, https://arxiv.org/abs/2310.10649

[4] Rohbeck et. al. 2025, MMFM

[5] 3MSBM 2025, https://arxiv.org/abs/2506.10168

[6] OTP-FM 2026, https://arxiv.org/abs/2606.05327

Presentation

  • The paper would benefit from a high-level diagram describing the algorithm.
Questions for authors

I list my points of confusion and weaknesses for the authors to address above.

5 comments

Soundness 1 and 4:

We thank the reviewer for these valuable comments.

Regarding linear regularization and second-order smoothing. These design choices are motivated by prior work. Specifically, the piecewise linear reference regularizer and second-order smoothing were introduced in Ali-CFM [1]. The piecewise linear regularizer is designed to provide a flexible reference for intermediate distributions, addressing the limitations of global linear interpolation when the supports of intermediate marginals differ from those of the endpoint distributions. The second-order smoothing term, inspired by MMFM [2], further encourages smoothness along the entire interpolation trajectory.

To evaluate the effectiveness of these components, we have conducted additional ablation experiments with different loss configurations, as shown in Table 1. The results demonstrate that incorporating linear regularization and smoothness constraints generally improves performance over using SWD alone, particularly on the single-cell datasets. Moreover, the proposed multi-time linear regularization achieves the best performance on the EB dataset when combined with smoothness regularization.

We would also like to clarify that these regularization techniques are not the primary methodological contributions of GraphFM, but rather serve as components of its underlying interpolation framework. Our main contributions lie in incorporating graph structural information into the interpolation process and extending the three-point linear regularization in Ali-CFM to a multi-time setting, enabling regularization across multiple temporal anchors.

Table 1. Ablation study of different loss configurations.

Loss ConfigurationEB EMD ↓Multi EMD ↓Traffic Joint MAE ↓Traffic Joint MSE ↓
SWD2.1106 ± 1.306818.5980 ± 19.89962.8850 ± 0.013535.6099 ± 0.8659
SWD + Three-Point Linear Regularization1.4141 ± 0.02658.3862 ± 0.14673.1451 ± 0.054843.0393 ± 0.7838
SWD + Multi-Time Linear Regularization1.4210 ± 0.02458.3862 ± 0.14672.6883 ± 0.007129.3444 ± 0.2108
SWD + Three-Point Linear Regularization + Smoothness1.3711 ± 0.05327.9472 ± 0.06042.8737 ± 0.006736.7075 ± 0.2173
SWD + Multi-Time Linear Regularization + Smoothness1.3033 ± 0.01577.9472 ± 0.06042.8441 ± 0.009134.8839 ± 0.1163

Regarding the two-stage training strategy and interpolation parameterization (Eq. 3). We would first like to clarify that our two-stage training strategy is NOT a two-stage distillation. In the first stage, a neural interpolator is trained to construct continuous interpolation paths and their corresponding target velocities. In the second stage, a flow-matching velocity field is trained using the paths and target velocities provided by the learned interpolator.

Both the interpolation parameterization and the two-stage training strategy follow established practices in neural interpolation and flow-matching frameworks [1, 3, 4], rather than being introduced as the main methodological contributions of GraphFM. Our primary contribution lies in incorporating graph structural information to guide the interpolation process and the subsequent learning of system dynamics.

References

[1] Kviman et al. Multi-Marginal Flow Matching with Adversarially Learnt Interpolants. ICLR, 2026.

[2] Rohbeck et al. Modeling Complex System Dynamics with Flow Matching Across Time and Conditions. ICLR, 2025.

[3] Kapuśniak et al. Metric Flow Matching for Smooth Interpolations on the Data Manifold*.* NeurIPS, 2024.

[4] Yu et al. PACE: Geometry-Aware Bridge Transport for Single-Cell Trajectory Inference. 2026.

Soundness 3:

We thank the reviewer for this constructful comment. We agree that a training time comparison would provide useful insights into the computational cost of GraphFM. We therefore conducted an additional training time benchmark on the synthetic dataset, using the same training data for all methods. The reported times include both interpolator and velocity field training where applicable. We also included OTP-FM, a flow-matching-based method among the references suggested by the reviewer, as an additional baseline. Despite the additional training cost, GraphFM achieves superior predictive performance, as shown in Table 3 below.

Table 2 Training Time Benchmark

MethodGraphFMOT-MFMOT-ALI-CFMOT-CFMOT-MMFMOTP-FM
Interpolation StrategyNeural Interpolator (Graph-based)Neural InterpolatorNeural InterpolatorLinear InterpolationSpline InterpolationDynamic OT-based
Training Time (min)22.5365.7743.0118.078.188.32

These results are generally consistent with our expectations. GraphFM, OT-MFM, and OT-ALI-CFM involve additional training of neural interpolators, whereas OT-CFM and OT-MMFM use predefined interpolation paths, and OTP-FM constructs paths through dynamic optimal transport.


Soundness 2:

We thank the reviewer for this insightful suggestion. We agree that it is important to distinguish whether the performance gains arise from incorporating graph information into the interpolator or the velocity field.

While existing baselines are not originally designed to incorporate external graph information, adapting them with GNN-based velocity fields would require additional architectural choices. Instead, we conducted a complementary experiment by replacing the GNN-based velocity field in GraphFM with a standard MLP, while keeping the graph-informed interpolator and other training settings unchanged. The results are presented below, where bold indicates the best performance and ** indicates the second-best performance.

The results show that replacing the GNN-based velocity field with an MLP generally leads to a performance decline. Nevertheless, GraphFM (MLP velocity) still achieves the best performance on 3 metrics and the second-best performance on 5 metrics out of 11 evaluated metrics. These findings suggest that the improvements cannot be attributed solely to the GNN-based velocity field, further supporting the effectiveness of incorporating graph information into the interpolation process.

Table 3

SettingMetric ↓GraphFMGraphFM (MLP velocity)OT-MFMOT-MMFMOT-ALI-CFMOT-CFMOTP-FM
Synthetic (Table 1)MMD0.04570 ± 0.00726**0.03586 ± 0.000860.09400 ± 0.090400.09341 ± 0.001800.05757 ± 0.001200.08392 ± 0.003410.12461 ± 0.00170
Synthetic (Table 1)SWD0.08199 ± 0.01110**0.06285 ± 0.001120.14101 ± 0.125000.14393 ± 0.002800.08896 ± 0.001380.12809 ± 0.005160.24327 ± 0.00316
EB (Table 3)EMD1.3711 ± 0.05321.3400 ± 0.00561.3621 ± 0.0127**1.3842 ± 0.00921.5248 ± 0.03371.3839 ± 0.00421.4971 ± 0.0312
Multi (Table 3)EMD7.9472 ± 0.06048.1799 ± 0.0688**8.5482 ± 0.23058.4844 ± 0.01108.8042 ± 0.08668.4602 ± 0.02219.5370 ± 0.2852
HT206B1, 278 genes (Table 4)EMD7.5191 ± 0.08927.5975 ± 0.01507.98817 ± 0.014827.86642 ± 0.003367.63725 ± 0.074777.8663 ± 0.00307.5491 ± 0.0239**
Traffic LOO (Table 5)EMD34.930 ± 0.13735.116 ± 0.093**35.806 ± 0.50335.976 ± 0.48842.891 ± 4.66935.882 ± 0.14337.497 ± 0.112
Traffic LOO (Table 5)MAE2.837 ± 0.0102.907 ± 0.011**3.101 ± 0.0543.101 ± 0.0703.714 ± 0.4213.055 ± 0.0153.028 ± 0.006
Traffic LOO (Table 5)MSE34.714 ± 0.23735.019 ± 0.110**35.430 ± 0.65835.589 ± 0.56747.678 ± 8.05235.959 ± 0.21239.983 ± 0.271
Traffic Joint Missing (Table 5)EMD34.992 ± 0.03435.237 ± 0.15635.186 ± 0.553**35.713 ± 0.21839.913 ± 2.15735.994 ± 0.12037.439 ± 0.186
Traffic Joint Missing (Table 5)MAE2.844 ± 0.0092.912 ± 0.015**3.029 ± 0.0443.036 ± 0.0443.435 ± 0.1853.036 ± 0.0093.014 ± 0.018
Traffic Joint Missing (Table 5)MSE34.884 ± 0.116**35.283 ± 0.30334.572 ± 0.92635.597 ± 0.26542.668 ± 2.57936.373 ± 0.18439.968 ± 0.395

**

Comments on the paper

1 comment

We thank all reviewers for their thoughtful and constructive feedback. Overall, we appreciate their recognition of the importance of reconstructing dynamics from temporal snapshots, the motivation for incorporating graph structure into flow-matching, the clarity of ourmethod, and the breadth of our evaluation. As reconstructing continuous dynamics from sparse, unpaired temporal snapshots is intrinsically underdetermined: different trajectories and vector fields can reproduce the same observed distributions, our work tries to demonstrate the effectiveness of incorporating known interactions among variables into constrain**ing both the paths used to train flow matching and the dynamics inferred from them*.*

Summary of contributions

  • We introduce GraphFM, which incorporates the graph structure of variables hidden in the data into conditional-path construction and velocity-field learning by establishinggraph neural interpolant,restricting each variable’s dependencies to its K-hop receptive field.
  • We develop a hierarchical interpolation objective combining distribution matching at observed time points, multi-anchor trajectory regularization, and local temporal smoothness for better conditiaonal-path construction.
  • We evaluate GraphFM across synthetic dynamics, single-cell trajectories, spatial transcriptomics, and traffic imputation, and examine how its performance depends on the supplied graph.

Summary of revisions addressing reviewers' concerns

  • Graph priors on real data. We extend the graph ablation studies to single-cell trajectory inference and traffic data imputation by comparing the original graph with random graph, fully connected graph, and a graph-free MLP variant. The original graph achieves the best performance in three out of four evaluated settings, providing further evidence for the benefit of informative graph priors in real-world dynamics reconstruction.
  • Velocity-field architecture. Replacing GraphFM’s GNN velocity field with an MLP, while retaining its graph-informed interpolation, yields the best result on three metrics and the second-best on five of eleven metrics, which demonstrating the effectiveness of incorporating graph into conditional path construction in flow-matching
  • Interpolation objective. Loss ablations separate the effects of distribution matching, linear regularization, smoothness, and the multi-anchor extension. The results show that linear and smoothness regularization generally improve performance over distribution matching alone, while multi-anchor regularization further outperforms three-point regularization when combined with smoothness.
  • Additional baselines. We include OT-CFM and OTP-FM as additional baselines. GraphFM continues to achieve the best predictive performance on the synthetic benchmark.
  • Computational efficiency**.** We additionally benchmark training times under the same synthetic-data setting. Although GraphFM incurs additional computational overhead due to its two-stage training procedure, it achieves superior predictive accuracy, highlighting the trade-off between computational cost and reconstruction performance.
  • Code availability. Our implementation is publicly available at https://anonymous.4open.science/r/GraphFM-80E4A/.
  • Methodological scope. We have clarified that the endpoint-preserving parameterization technique and two-stage training strategy used by existing neural-interpolation approaches. Our methodological contributions are both the use of graph structure in the learned paths and the extension to multiple temporal anchors.

These additions make the methodological contributions and benefits of GraphFM more explicit. We hope we haveaddressed the reviewers’ concerns and strengthening the evidence for graph structure as a useful prior in dynamics reconstruction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.