acceptodds
Under review as a conference paper at ICLR 2027

ESFM: Reconstructing Unbalanced Dynamics via Reduction and Extended-Space Flow Matching

Abstract

Biological population dynamics are often unbalanced, involving simultaneous changes in cellular states and population abundance. Existing unbalanced flow-matching methods typically rely on cone geometry-based loss formulations. While mathematically consistent, these formulations often induce biologically unfavorable state evolutions. We propose Extended Space Flow Matching (ESFM), which lifts unbalanced dynamics into conservative transport in an extended space of state and log mass. This jointly models state transitions and abundance changes as a single conservative process, enabling prescribed conditional paths to be integrated with compatible endpoint couplings and simulation-free flow matching objectives. For linear interpolation of state and log mass, ESFM admits an analytical reduction to a tractable optimal transport formulation. On synthetic and real single-cell snapshot datasets, ESFM accurately predicts both state and abundance dynamics. Extended space formulations thus offer a flexible, biologically preferred framework for unbalanced flow matching.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
flow matching, unbalanced optimal transport, extended-space, single-cell dynamics, trajectory inference

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_h7MVOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a flow matching technique for interpolating between unbalanced experimental snapshots. The method is based on the unbalanced optimal transport problem, defining the log mass of the snapshots as an additional variable over which to learn the flow. The method is benchmarked on existing synthetic and single-cell datasets.

Critical strengths
  • Simple methodology that retains the simulation-free training benefit of flow matching.
  • Addresses an interesting problem of inferring cellular trajectories.
Critical weaknesses

Unfortunately, I have significant concerns regarding the novelty and significance of the methodology as well as the presentation.

Soundness

  • Fundamentally, the problem of inferring “cell growth” is ill-defined for two reasons:
    1. The # of cells measured by scRNA-seq in a single sample is almost entirely a function of the # of samples, sample size, and sequencing setup. Thus, rather than learning actual population growth, such a method is likely only learning technical artifacts.
    2. As far as I understand, this method is inferring trajectories per individual cell state, whereas cell growth / death clearly represents multiple (or zero) cell states. To accurately model such a phenomenon implies a need to model at a population-level rather single-cell level.
  • The prescribed linear interpolation in log mass is a highly restrictive and simplistic assumption.
  • Some existing works on trajectory inference are missing in the benchmarking: [1-3] in particular all have stronger performance than ESFM on CITE 50D [3] (the only experimental setup that is directly comparable).
    • Overall, ESFM does not demonstrate any significant improvement over prior workk.
  • One key benefit of a simulation-free training objective is efficient training, but no timing benchmarking is provided.

[1] OT-MFM 2024 https://arxiv.org/abs/2405.14780
[2] WLF-UOT 2024 https://arxiv.org/abs/2310.10649
[3] OTP-FM 2026 https://arxiv.org/abs/2606.05327

Presentation and informativeness

  • Section 4 describing the method is hard to follow. It reads like a series of equations without clear motivations and explanations of the significance and intuition behind each equation.
  • Thms. 4.1 and 4.3 in particular appear to be presented without context - their motivation, significance, and intuition should be better explained.
  • Because of this, technical differences with respect to similar methods such as VGFM and WFR-FM are also not clear. VGFM in particular appears to have an identical training objective.
  • Minor comments:
    • Some sloppy notation, e.g. Unclear in what sense \rho^\hat is a probability distribution, presumably you mean \rho^\hat(\dot, y) is a valid prob. distribution of x for each y?
Questions for authors
  1. How can you disentangle technical measurements artifacts in the dataset from true cell growth?
  2. What evidence supports the prescribed linear interpolation in log mass?
  3. Can you incorporate the prior work mentioned above and others in this space?
  4. Can you provide a training time comparison with other simulation-free methods?
  5. Can you describe the methodology clearly and explain the novelty with respect to prior work?
7 comments
AuthorsOpenReview ↗

Response to Q1/W1 and W2: Growth Interpretation and Population-Level Modeling (1/4)

Q1/W1: Technical Measurement Effects versus Biological Growth

Our objective is to reconstruct population dynamics from observed single-cell snapshots, conditional on the population abundance information provided by the data. Unbalanced OT has been used to model cell differentiation and population growth [1,2,3]. In particular, TIGON [2] and stVCR [3] employ population-level transport equations with growth terms, closely related to the dynamics considered in ESFM.

We agree that observed cell abundance can be influenced by technical factors, including sampling efficiency and sequencing protocols. In such cases, the inferred growth rate should be interpreted as an effective growth rate, potentially reflecting both biological proliferation/apoptosis and technical measurement effects. ESFM does not claim to disentangle these contributions from unpaired snapshots alone, as doing so generally requires additional information about the observation process. When the underlying technical measurement mechanism is known, the ESFM framework can accommodate an explicit observation model to account for such effects in the inference of biological population dynamics.

Accordingly, our synthetic experiments with known ground-truth growth rates validate the recovery of biological growth dynamics, whereas the real-data experiments demonstrate the reconstruction of observed population distributions and abundance changes. We will clarify this distinction and the associated limitations in the revised manuscript.

W2: Population-Level Dynamics and Cell Division/Death

We clarify that ESFM primarily models population-level dynamics rather than explicitly reconstructing individual cell lineages.

(1) Population-level dynamics. As established in Theorem 4.1, ESFM describes a conservative particle distribution ρ^t(x,y)\widehat{\rho}_t(\mathbf{x},y) in the extended space, where y=ln⁡my=\ln m. The physical population density is recovered through the mass-weighted projection

ρt(x)=∫Reyρ^t(x,y),dy.\rho_t(\mathbf{x}) = \int_{\mathbb{R}} e^y\widehat{\rho}_t(\mathbf{x},y)\\,\mathrm{d}y.

Importantly, this projected density satisfies the unbalanced continuity equation

∂tρt+∇x⋅(ρtut)=ρtgt.\partial_t\rho_t + \nabla_{\mathbf{x}}\cdot (\rho_t\mathbf{u}_t) = \rho_t g_t.

Thus, ESFM describes both state transport and state-dependent population growth or decay. In particular, the total population mass Mt=∫ρt(x),dxM_t=\int\rho_t(\mathbf{x})\\,\mathrm{d}\mathbf{x} evolves according to

dMtdt=∫Rdρt(x)gt(x),dx.\frac{\mathrm{d}M_t}{\mathrm{d}t} = \int_{\mathbb{R}^d} \rho_t(\mathbf{x})g_t(\mathbf{x}) \\,\mathrm{d}\mathbf{x}.

(2) Stochastic realization of cell division and death. Furthermore, following stVCR and USB [3,5], the learned extended-space dynamics can be used to construct a stochastic birth–death process.

Each particle evolves directly under the learned extended-space fields:

\frac{\mathrm{d}\mathbf{x}_t}{\mathrm{d}t} = v_\theta(\mathbf{x}_t,y_t,t), \qquad \frac{\mathrm{d}y_t}{\mathrm{d}t} = g_\phi(\mathbf{x}_t,y_t,t). \tag{1}

The predicted relative growth rate gϕ(x,y,t)g_\phi(\mathbf{x},y,t) determines the net balance between proliferation and apoptosis. We construct nonnegative division and death rates as

βt(x,y)=[gϕ(x,y,t)]+,dt(x,y)=[−gϕ(x,y,t)]+,\beta_t(\mathbf{x},y) = [g_\phi(\mathbf{x},y,t)]_+, \qquad d_t(\mathbf{x},y) = [-g_\phi(\mathbf{x},y,t)]_+,

such that gϕ=βt−dtg_\phi=\beta_t-d_t. Over a sufficiently small time interval Δt\Delta t, the probabilities of division and apoptosis are

\begin{aligned} P(\mathrm{division}\mid \mathbf{x}_t=\mathbf{x},y_t=y) &= [g_\phi(\mathbf{x},y,t)]_+ \Delta t+o(\Delta t),\\\\ P(\mathrm{apoptosis}\mid \mathbf{x}_t=\mathbf{x},y_t=y) &= [-g_\phi(\mathbf{x},y,t)]_+ \Delta t+o(\Delta t). \end{aligned} \tag{2}

In practice, we draw U∼Uniform(0,1)U\sim\mathrm{Uniform}(0,1) at each time step. A cell divides into two if gϕ(xt,yt,t)>0g_\phi(\mathbf{x}_t,y_t,t) > 0 and U<gϕ(xt,yt,t)ΔtU < g_\phi(\mathbf{x}_t,y_t,t)\Delta t, or undergoes apoptosis if gϕ(xt,yt,t)<0g_\phi(\mathbf{x}_t,y_t,t) < 0 and U<−gϕ(xt,yt,t)ΔtU < -g_\phi(\mathbf{x}_t,y_t,t)\Delta t. Otherwise, the cell survives without division. Upon division, the parent particle is replaced by two daughter particles initialized at (xt±ξ,yt)(\mathbf{x}_t\pm\boldsymbol{\xi},y_t), where ξ\boldsymbol{\xi} is a small random perturbation. Both daughters subsequently evolve under the same learned extended-space fields.

As illustrated in Figure 7 of Appendix B.3 in the revised manuscript, this stochastic realization produces explicit division and apoptosis events on the Dyngen dataset, complementing the continuous population-level representation. It provides one possible realization of the inferred net growth dynamics, rather than uniquely identifying biological division and death rates or actual individual cell lineages.

AuthorsOpenReview ↗

Response to Q2–Q4: Growth Prior, Additional Baselines, and Efficiency (2/4)

Q2/W3: Justification for Linear Log-Mass Interpolation

The linear log-mass path has a population-growth interpretation through the Malthusian model. This model assumes approximately constant per-capita proliferation and death rates, bb and dd, over an interval, as may be reasonable when resources are sufficient and the environment is stable. The expected total proliferation and death rates are then proportional to the current population, giving

m˙(t)=(b−d)m(t),ddtln⁡m(t)=b−d.\dot m(t)=(b-d)m(t), \qquad \frac{\mathrm d}{\mathrm dt}\ln m(t)=b-d.

Thus, the relative growth rate is constant, physical mass changes exponentially, and log mass changes linearly.

The same assumption of constant conditional relative growth is also used in VGFM [4] and USB [5]. Over short time intervals, we consider monotone conditional mass evolution a more biologically plausible prior than the transient mass decrease followed by regrowth permitted by WFR geodesics. Further discussion is provided in our response to Reviewer qGy6, Q1.

Linear interpolation in log mass is one illustrative choice within ESFM. As discussed in Appendix D and our response to Reviewer qGy6, Q2, the framework also accommodates other actions and geodesics. Appendices D.2–D.3 derive the construction for paths linear in physical mass, and Appendix D.4 recovers the standard static WFR semi-coupling formulation from the microscopic WFR action.

Q3/W4: Additional Baselines: OT-MFM, WLF-UOT, and OTP-FM

Thank you for pointing out these relevant works. We have added comparisons with OT-MFM, WLF-UOT, and OTP-FM to Tables 2 and 3 of the revised manuscript. For convenience, we reproduce the results for these three baselines and ESFM below.

Results on synthetic datasets, reproduced from Table 2 of the revised manuscript.

MethodGene W1Gene RMEDyngen W1Dyngen RMEGaussian W1Gaussian RME
OT-MFM0.311---1.767---3.794---
WLF-UOT0.618---2.056---4.101---
OTP-FM0.433---1.142---2.139---
ESFM (ours)0.0170.00050.1110.0032.4280.0001

Bold values mark the lowest W1 among the four methods shown.

W1 results on real datasets, reproduced from Table 3 of the revised manuscript.

MethodEMTEBCITEMouse
OT-MFM0.31410.72636.2177.753
WLF-UOT0.31411.61636.9909.040
OTP-FM0.31510.95337.1608.286
ESFM (ours)0.29510.13437.1476.714

Bold values mark the lowest W1 among the four methods shown.

The additional comparisons show that ESFM achieves the lowest W1 among the four methods on five of the seven datasets: Gene, Dyngen, EMT, EB, and Mouse. We acknowledge that OT-MFM and WLF-UOT obtain lower errors on CITE 50D, while OTP-FM performs better on the 1000D Gaussian benchmark. Nevertheless, these results demonstrate the competitive performance of ESFM across diverse synthetic and real-world datasets, rather than uniform superiority.

We appreciate the reviewer's suggestions and have taken care to acknowledge and cite the relevant prior work more comprehensively in the revised manuscript. We will also cite these works in our future research where relevant.

Q4/W5: Training Time and Memory Comparison

To address both Q4 and W5, we reproduce the computational-efficiency benchmark on the 100-dimensional EB dataset directly in Table 3. These are the results shown in Figure 13 of the revised manuscript (Appendix B.5). All results were obtained using a single NVIDIA A100 GPU.

Computational efficiency on the 100D EB dataset.

MethodTraining time (s)Memory (GB)
ESFM (ours)420.4
USB730.7
SF2M890.7
MMFM980.6
WFR-FM1280.9
Metric FM1364.2
UOT-FM1490.6
VGFM2850.6
MIOFlow3780.7
TIGON4530.7
DeepRUOT61710.1
Var-RUOT92020.0

Bold values mark the lowest reported value in each metric.

ESFM requires 42 seconds and 0.4 GB of memory, giving the shortest reported runtime and lowest reported memory usage among the compared methods. The ESFM runtime includes coupling computation, endpoint-pair presampling, and network training.

AuthorsOpenReview ↗

Response to Q5: Methodology and Intuition Behind the Main Theorems (3/4)

Q5/W6: Methodology and Theorem Intuition

We agree that Section 4 does not explain our method clearly enough. Below, we first clarify the methodology and then explain its novelty relative to prior work.

Clarifying the methodology.

Our starting point is to incorporate assumptions about how cell states and population mass evolve into the construction of an unbalanced flow-matching model. We express these assumptions through prescribed conditional paths and construct a compatible transport action, from which the endpoint coupling and training targets are derived.

In Section 4.1, we introduce the dynamics in the extended space, together with the density, boundary conditions, and continuity equation. In the physical state space, the mass density evolves according to the unbalanced continuity equation

\partial_t\rho_t(\boldsymbol{x}) +\nabla\cdot\left(\rho_t(\boldsymbol{x})\boldsymbol{u}_t(\boldsymbol{x})\right) =\rho_t(\boldsymbol{x})g_t(\boldsymbol{x}), \tag{3.3}

which includes a source term accounting for growth and death. By introducing log mass as an additional coordinate, we represent mass changes as transport along this coordinate, leading to the conservative continuity equation in the extended space:

\partial_t\hat\rho_t +\nabla_{\boldsymbol{x}}\cdot(\hat\rho_t\hat{\boldsymbol{u}}_t) +\partial_y(\hat\rho_t\hat g_t)=0. \tag{4.2}

Theorem 4.1 establishes the consistency between these two descriptions: under a vanishing weighted-flux condition as y→±∞y\to\pm\infty, the mass-weighted projection of (4.2) recovers (3.3), with the corresponding projected velocity and growth fields. Intuitively, transport along the log-mass coordinate becomes growth or death when viewed in the physical state space. This shows that the extended-space formulation is consistent with the dynamics in the original state space, providing a mathematical justification for modeling transport and growth in the extended space.

Building on this extended-space formulation, our framework proceeds through four steps, as illustrated in the revised Figure 1. First, we prescribe linear state interpolation and exponential mass interpolation between the endpoints:

xt=(1−t)x0+tx1,mt=m01−tm1t,t∈[0,1].\boldsymbol{x}_t=(1-t)\boldsymbol{x}_0+t\boldsymbol{x}_1, \qquad m_t=m_0^{1-t}m_1^t, \qquad t\in[0,1].

Equivalently, log mass evolves linearly as yt=(1−t)y0+ty1y_t=(1-t)y_0+ty_1, where yi=ln⁡miy_i=\ln m_i. These paths have constant conditional state velocity and constant conditional relative growth rate, dln⁡mt/dt=ln⁡(m1/m0)\mathrm d\ln m_t/\mathrm dt=\ln(m_1/m_0).

Second, we construct a microscopic action for which these prescribed paths are minimizers. These two steps are presented in Section 4.2. Third, we lift the microscopic action to a population-level dynamic OT problem and derive its equivalent static formulation in the extended space (Section 4.3).

We next ask how this extended-space static OT problem can be formulated in the original state space. This motivates Theorem 4.3 in Section 4.4: with the source log mass fixed at y0=0y_0=0 and the terminal log mass bounded by y1≤My_1\le M, where M>1M > 1, the extended static problem is equivalent to an optimization over semi-couplings in the original state space. This allows the coupling to be computed without explicitly discretizing the additional mass coordinate. The objective in (4.8) is

\iint\left[\\|\boldsymbol{x}_1-\boldsymbol{x}_0\\|^2 +\delta^2\Phi_M\\!\left(R(\boldsymbol{x}_0,\boldsymbol{x}_1)\right)\right]\\,\mathrm d\gamma_0. \tag{4.8}

Here, γ0\gamma_0 and γ1\gamma_1 represent the initial and final mass assigned to each state pair, and R=dγ1/dγ0R=\mathrm d\gamma_1/\mathrm d\gamma_0 is the local growth ratio.

In this reduced formulation, state transport retains the squared Euclidean cost ∣x1−x0∣2\\|\boldsymbol{x}_1-\boldsymbol{x}_0\\|^2, while mass change is penalized through δ2ΦM(R)\delta^2\Phi_M(R), where RR is the local growth ratio. The explicit form of ΦM\Phi_M is given in Theorem 4.3.

Finally, we sample from the resulting coupling and train the velocity and growth networks by regressing against the conditional fields determined by the prescribed paths (Section 4.5). The central idea is that the coupling and the regression targets are constructed from the same underlying paths and action, ensuring consistency between the transport problem and flow-matching training.

Reviewer_qGy6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ESFM, which extends flow matching by augmenting the state space with a (log) mass state to model unbalanced dynamics from given population snapshots. Compared to prior works, which first define WFR geometry and derive conditional paths and couplings based on it, ESFM adopts a path-first design that yields a compatible action. Experiments demonstrate accurate predictions across unbalanced snapshots compared to baselines.

Critical strengths
  • The paper is well written, and the idea of augmenting the flow with (log) mass and deriving the corresponding transport problems seems novel.
  • The experiments presented throughout the paper are convincing.
Critical weaknesses
  • It is unclear why straight-line interpolation in the augmented space is preferable to a geometry-first approach. Interpolation in flow matching yields a simple objective, and ease of optimization can matter more than the underlying geometry. However, the goal here is to model cellular dynamics. Simple interpolation between snapshots may neglect these dynamics if it does not reflect the geometry in which they evolve. I think this is the main motivation behind WFR-based methods. They first specify a geometry and then derive an optimal path under it in the hope that it aligns with the true distributional dynamics. Could the authors clarify why the prescribed path is expected to better reflect cellular dynamics?

  • Do the results in the paper rely on linear interpolation in the augmented space, or can they be generalized to other interpolations? Fixing the action and optimizing the path seems more expressive, since a general action may induce a richer family of paths than linear interpolation. Could the authors clarify the limitations of fixing the path?

  • (Minor point) I think the overview in Figure 1 could be more informative. It is currently difficult to grasp the main idea of augmenting the state space with a (log) mass state and using a simple predefined path. The figure provides little insight into how this construction can reliably model cellular dynamics. Revising Figure 1 to clarify these points would help readers understand the work.

Questions for authors

See Weaknesses

2 comments
AuthorsOpenReview ↗

Response to Q1–Q3: Biological Motivation, Alternative Paths, and Revised Overview (1/1)

Q1: Biological Motivation for the Prescribed Paths

We agree that an interpolation should reflect meaningful assumptions about cellular dynamics, rather than be chosen solely for ease of optimization. WFR provides a principled geometry that jointly models state transport and relative growth. Its dynamic formulation is

inf⁡ρ,u,g∫01!∫Rd(∣ut(x)∣2+δ2∣gt(x)∣2)ρt(x),dx,dt,s.t.∂tρt+∇!⋅(ρtut)=ρtgt,ρt=0=ρ0,ρt=1=ρ1.\begin{aligned} \inf_{\rho,\boldsymbol{u},g}\quad &\int_0^1\\!\int_{\mathbb R^d} \left(\\|\boldsymbol{u}_t(\boldsymbol{x})\\|^2+\delta^2|g_t(\boldsymbol{x})|^2\right) \rho_t(\boldsymbol{x})\\,\mathrm d\boldsymbol{x}\\,\mathrm dt,\\\\ \text{s.t.}\quad &\partial_t\rho_t+\nabla\\!\cdot(\rho_t\boldsymbol{u}_t)=\rho_tg_t, \qquad \rho_{t=0}=\rho_0,\quad \rho_{t=1}=\rho_1. \end{aligned}

Owing to its mathematical simplicity, WFR has been used to model unbalanced biological dynamics. Examples include spatiotemporal reconstruction with stVCR [1] and simulation-free velocity and growth inference with WFR-FM [2].

Along the traveling-Dirac geodesic under WFR geometry, mass evolves as a quadratic function of time (WFR-FM [2], Eqs. (3.5)–(3.6)):

m(t)=At2−2Bt+m0,m(t) =At^2-2Bt+m_0,

where

\left\\{ \begin{aligned} A&=m_0+m_1-2\sqrt{\frac{m_0m_1}{1+\tau^2}}, & B&=m_0-\sqrt{\frac{m_0m_1}{1+\tau^2}},\\\\ \tau&=\tan\theta, & \theta&=\frac{\\|\boldsymbol{x}_1-\boldsymbol{x}_0\\|}{2\delta} < \frac{\pi}{2}. \end{aligned} \right.

Along such a geodesic, mass may first decrease and then increase over time.

We consider monotone conditional mass evolution over short time intervals more likely to reflect the underlying biological dynamics than a transient decrease followed by regrowth. ESFM encodes this modeling prior through

xt=(1−t)x0+tx1,yt=(1−t)ln⁡m0+tln⁡m1,mt=m01−tm1t.\boldsymbol{x}_t=(1-t)\boldsymbol{x}_0+t\boldsymbol{x}_1, \qquad y_t=(1-t)\ln m_0+t\ln m_1, \qquad m_t=m_0^{1-t}m_1^t.

Linear state motion serves as a simple local approximation, while linear log mass corresponds to a constant conditional relative growth rate.

Q2: Alternative Interpolants and Limitations of the Path Prior

The extended-space framework is not restricted to the linear state and log-mass geodesics used in the main text. Appendix D discusses its generalization.

In Appendix D.1, we derive minimizing paths, endpoint costs, dynamic and static transport formulations, and flow-matching targets from a Lagrangian action, under the stated assumptions. We agree with the reviewer that this general construction starts from the action and allows a richer family of paths.

Appendices D.2 and D.3 consider geodesics that are linear in state and physical mass, rather than state and log mass:

xt=(1−t)x0+tx1,mt=(1−t)m0+tm1.\boldsymbol{x}_t=(1-t)\boldsymbol{x}_0+t\boldsymbol{x}_1, \qquad m_t=(1-t)m_0+tm_1.

We derive the corresponding action, reduced static transport problem, entropic coupling formulation, and regression targets, showing that the same construction remains applicable.

In Appendix D.4, we choose the microscopic WFR action and show that the resulting transport problem recovers the standard static WFR semi-coupling functional.

The geodesic used in the main text is one particular choice that we use to illustrate the framework.

We acknowledge that prescribing a conditional path introduces a modeling prior, which may be restrictive when the underlying dynamics exhibit strongly nonlinear state trajectories or non-monotonic mass evolution. However, this restriction applies to the conditional paths rather than the learned marginal dynamics. After averaging over different endpoint pairs, the learned velocity and growth fields can still exhibit nonlinear and time-dependent behavior.

Q3: Revised Overview in Figure 1

We thank the reviewer for this helpful suggestion. We have revised Figure 1 to provide a clearer and more intuitive illustration of the ESFM framework.

The revised figure first illustrates how unbalanced population dynamics are lifted into conservative transport in the extended state–log-mass space w=(x,ln⁡m)\mathbf{w}=(\mathbf{x},\ln m). It then explicitly presents the four main steps of ESFM: (1) prescribing a conditional path for state transport and mass evolution; (2) constructing a microscopic action functional for which the prescribed path is minimizing; (3) deriving the corresponding static transport problem and endpoint coupling; and (4) learning the velocity and growth fields through simulation-free flow matching.

References

[1] Peng, Q., Zhou, P., and Li, T.
stVCR: spatiotemporal dynamics of single cells.
Nature Methods, 23, 542--553 (2026).

[2] Peng, Q., Wang, Z., Ying, J., Sun, Y., Nie, Q., Zhang, L., Li, T., and Zhou, P.
WFR-FM: Simulation-Free Dynamic Unbalanced Optimal Transport.
ICLR (2026).

Thank you for the detailed response. I will keep my score.

Reviewer_9wzAOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes jointly training velocity field and growth dynamics in unbalanced transport problems by constructing a state variable and regressing to conditional targets derived from end-point marginals. The paper evaluates their method on single-cell benchmarks showing that they recover growth rate dynamics on Simulated Gene Benchmark. The paper outperforms most baselines when provided a prescribed path, however it is on par at generalising to unseen marginals. My main concerns are about the novelty and the scope of the proposed method.

Critical strengths

Potential to scale: The paper reduces simulation-based training budget in VGFM by directly regression to conditional targets to learn joint state variable.

Application to real-world systems: The paper provides insight into how well the growth dynamics is recovered and tests its methods on single-cell benchmarks.

Critical weaknesses

Novelty: The method is quite close to VGFM since both regress their velocity and growth networks to the targets constructed by end-point marginals, and while I appreciate VGFM includes a simulation-based component, that doesn’t seem to be huge cost in terms of time and memory in the smaller scale experiments presented in the paper (Figure 5). There is no evidence that avoiding simulation is a proposition that performs better at scale, which is where avoiding this component would matter significantly.

Scalability: The method wins on recovering trajectories when given a prescribed path, but at a small margin. It would be great if authors could demonstrate that their proposed method unlocks new task/performance other methods cannot capture, as it otherwise seems this might be an incremental improvement. Further on held-one-out experiments, the method is slightly better on only the lowest dimension dataset and worse on higher-dimensional data.

Questions for authors
  • In algorithm 1 it is unclear how this is trained. The theoretical section formulates joint state space vector (x_t g_t) but in the training algorithm trains them completely separately? Are losses summed up or do you need to train this twice in stages?

  • In the relation to VGFM section in the appendix it is unclear to me why VGFM does not jointly train velocity and growth dynamics since their Algorithm 1 (page 7 in the paper) seems to have similar training routine to ESFM with additional simulation-based loss? Could you explain if there are any other training differences?

  • Could you compare state-dependent growth dynamics to other unbalanced OT baselines in table 3? I am referring to the plotted comparison between ESFM and WFR-FM in Figure 2.

  • Could you also plot recovered growth-field on held out marginal task?

  • Could you plot recovered trajectories on held-out marginal? I am interested to see how well the held-out marginal is matched in comparison to baselines in table 3.

3 comments
AuthorsOpenReview ↗

Response to Q1–Q2 and W1: Joint Training and Differences from VGFM (1/2)

Q1: Joint Training of Velocity and Growth

We use separate networks for velocity and growth. Both fields are defined on the extended space w=(x,y){\bf w}=({\bf x},y) and parameterized as vθ(w,t)v_\theta({\bf w},t) and gϕ(w,t)g_\phi({\bf w},t), respectively. We train these two independent networks using the summed loss

L(θ,ϕ)=Lv(θ)+λLg(ϕ),λ>0.L(\theta,\phi)=L_v(\theta)+\lambda L_g(\phi),\lambda > 0.

Both networks are updated in every iteration, with no sequential training stages. Because their parameters are disjoint, minimizing the summed loss is equivalent to optimizing the two terms separately.

W1/Q2: Differences from VGFM

Both VGFM and ESFM jointly learn velocity and growth. They also use the same form of conditional interpolation once the endpoint states and masses are fixed. These paths and their targets are

xt=(1−t)x0+tx1,mt=m01−tm1t,yt=ln⁡mt,u∗=x1−x0,g∗=ln⁡(m1/m0).\begin{aligned}{\bf x}_t&=(1-t){\bf x}_0+t{\bf x}_1,&m_t&=m_0^{1-t}m_1^t,y_t=\ln m_t,\\\\{\bf u}^*&={\bf x}_1-{\bf x}_0,&g^*&=\ln(m_1/m_0).\end{aligned}

Let z=(x0,x1,m0,m1)z=({\bf x}_0,{\bf x}_1,m_0,m_1) and t∼U[0,1]t\sim{\cal U}[0,1]. The sampling laws qVq_{{\rm V}} and qEq_{{\rm E}} are induced by the OT solutions P∗P^* and (Γ∗,R∗)(\Gamma^*,R^*), respectively, as specified below. The conditional matching losses are

LVGFM=Et,z∼qV[∣vθ(xt,t)−u∗∣2+∣gω(xt,t)−g∗∣2],\LESFM=Et,z∼qE[∣vθ(xt,yt,t)−u∗∣2+λ∣gϕ(xt,yt,t)−g∗∣2].\begin{aligned}L_{\mathrm{VGFM}}&=\Bbb E_{t,z\sim q_{{\rm V}}}[\\|v_\theta({\bf x}_t,t)-{\bf u}^*\\|^2+|g_\omega({\bf x}_t,t)-g^*|^2],\\\L_{\mathrm{ESFM}}&=\Bbb E_{t,z\sim q_{{\rm E}}}[\\|v_\theta({\bf x}_t,y_t,t)-{\bf u}^*\\|^2+\lambda|g_\phi({\bf x}_t,y_t,t)-g^*|^2].\end{aligned}

The main difference lies in the couplings underlying qVq_{{\rm V}} and qEq_{{\rm E}}, which are used to sample endpoint pairs and growth labels. This changes the distribution of conditional regression targets and can therefore lead to different learned velocity fields.

For ESFM, the prescribed linear path in (x,y)({\bf x},y), y=ln⁡my=\ln m, minimizes the following action for fixed endpoints:

Aδ(w)=∫01(∣x˙t∣2+δ2∣y˙t∣2)dt.A_\delta({\bf w})=\int_0^1(\\|\dot{{\bf x}}_t\\|^2+\delta^2|\dot y_t|^2){\rm d}t.

Its minimum is the endpoint cost c(w0,w1)=∣x1−x0∣2+δ2∣y1−y0∣2c({\bf w}_0,{\bf w}_1)=\\|{\bf x}_1-{\bf x}_0\\|^2+\delta^2|y_1-y_0|^2, which induces the static OT formulation on the extended space W=Rd×R{\cal W}=\Bbb R^d\times\Bbb R (Eq. (4.7)):

Kδ(ρ0,ρ1)=inf⁡ρ^1∈P(W)π^∈Π(ρ^0,ρ^1)∫W2c(w0,w1)dπ^(w0,w1).K_\delta(\rho_0,\rho_1)=\inf_{\substack{\hat\rho_1\in{\cal P}({\cal W})\\\\\hat\pi\in\Pi(\hat\rho_0,\hat\rho_1)}}\int_{{\cal W}^2}c({\bf w}_0,{\bf w}_1){\rm d}\hat\pi({\bf w}_0,{\bf w}_1).

For empirical weights ai,bja_i,b_j, write Cij=∣xi0−xj1∣2C_{ij}=\\|{\bf x}_i^0-{\bf x}_j^1\\|^2 and let RijR_{ij} be the local growth ratio. The discrete reduced problem is (Eq. (A.15)):

(Γ∗,R∗)∈argmin⁡Γ≥0,0<R≤eM∑ij[Cij+δ2ΦM(Rij)]Γij,s.t.∑jΓij=ai,∑iRijΓij=bj.\begin{aligned}(\Gamma^*,R^*)\in\operatorname*{arg min}_{\Gamma\ge0,0 < R\le e^M}&\sum_{ij}[C_{ij}+\delta^2\Phi_M(R_{ij})]\Gamma_{ij},\\\\\text{s.t.}&\sum_j\Gamma_{ij}=a_i,\sum_iR_{ij}\Gamma_{ij}=b_j.\end{aligned}

In contrast, VGFM obtains its coupling from a direct extension of quadratic Wasserstein OT (VGFM, Eq. (11)):

P∗∈argmin⁡P≥0∑ijCijPij+τKL(P1∣a),s.t.∑iPij=bj.\begin{aligned}P^*\in\operatorname*{arg min}_{P\ge0}&\sum_{ij}C_{ij}P_{ij}+\tau\mathrm{KL}(P{\bf1}\\|a),\\\\\text{s.t.}&\sum_iP_{ij}=b_j.\end{aligned}

VGFM computes its coupling using an extension of quadratic Wasserstein optimal transport, whereas ESFM uses the pairwise growth cost induced by its extended-space action. Although the conditional interpolations have the same form, ESFM derives its coupling and conditional paths from a common action, whereas VGFM combines a semi-relaxed OT coupling with a separately chosen constant relative-growth interpolation. VGFM does not establish that its coupling minimizes the action associated with the conditional paths used for training.

Coupling–path consistency matters because endpoint pairing and the path connecting each pair should reflect the same dynamical assumptions. The coupling determines which endpoints are paired, while the conditional path specifies how state and mass evolve between them. In ESFM, the cost of each extended endpoint pair is exactly the minimum action of the prescribed path connecting it. The coupling therefore minimizes the total action of the conditional paths used for training. If the coupling is instead chosen using a different cost, its preferred pairings and growth allocations need not minimize this action, even when the same interpolation formula is subsequently applied. Consistency therefore allows the assumed dynamics to inform the entire construction, from endpoint selection to conditional evolution. Under the assumptions of our static–dynamic equivalence theorem, the coupling and path construction together solve a fully specified OT problem.

An additional difference is the input state: ESFM parameterizes vθ(x,y,t)v_\theta({\bf x},y,t) and gϕ(x,y,t)g_\phi({\bf x},y,t), with y=ln⁡my=\ln m, whereas VGFM uses vθ(x,t)v_\theta({\bf x},t) and gω(x,t)g_\omega({\bf x},t). Including log mass as an input expands the class of dynamics ESFM can represent.

AuthorsOpenReview ↗

Response to Q3–Q5 and W2: Growth Recovery and Held-Out Generalization (2/2)

W2: Significance and Held-Out Generalization

Our goal is trajectory inference for biological populations whose state distributions and abundances change over time. The main contribution of ESFM is a framework that starts from prescribed single-particle paths and constructs compatible transport problems and flow-matching objectives. This provides a different perspective on designing models for unbalanced biological dynamics. The linear state and log-mass path is one instance of this framework; Appendix D discusses other choices.

The empirical results support this framework across multiple datasets. In Table 3, ESFM achieves the lowest held-out distribution error on EMT and the second-lowest on EB and Mouse, both of which use 50-dimensional representations. Its competitive held-out performance therefore extends beyond the lowest-dimensional dataset.

Q3: Growth-Field Comparisons with Additional Baselines

We have completed the additional growth-field comparisons, extending ESFM and WFR-FM to include VGFM, DeepRUOT and Var-RUOT. The new results and experimental details are included in Appendix B.3 (Saturating-Growth Data), Figure 4 of the revised paper.

ESFM recovers both the high growth rates of the proliferative population and the near-zero growth rates of the approximately quiescent population, achieving R2=0.922R^2=0.922, compared with 0.5740.574 for VGFM. Thus, the additional comparison shows that ESFM better distinguishes state-dependent growth between the two populations in this experiment.

Q4: Growth-Field Recovery on Held-Out Snapshots

We have added growth-field visualizations for the held-out marginal task. On Saturating-Growth Data, each method is trained on three separate hold-out tasks by masking t=4,8,16t=4,8,16 one at a time. The predicted growth rates at the masked time points are compared with the ground truth and displayed together in Appendix B.3, Figure 5 of the revised paper.

ESFM also recovers the growth contrast at the held-out times, with a pooled R2=0.952R^2=0.952, compared with −0.095-0.095 for VGFM. ESFM therefore preserves the distinction between the two populations, whereas VGFM overestimates growth in the approximately quiescent population and underestimates it in the proliferative population.

We have also added ESFM growth maps for the held-out EMT, EB, CITE-seq and Mouse snapshots in their respective dataset descriptions in Appendix B.4, Figures 9a, 10a, 11a and 12a. Growth rates at each held-out time point are predicted by the model trained without that snapshot. As growth-rate ground truth is unavailable for these real datasets, the maps visualize learned predictions rather than establish biological growth-rate accuracy.

Q5: Trajectory and Marginal Recovery on Held-Out Snapshots

We have completed the held-out trajectory and marginal visualizations, comparing ESFM with WFR-FM, VGFM, DeepRUOT and Var-RUOT. The synthetic results are in Appendix B.3, Figure 6. The real-data results are included under the EMT, EB, CITE-seq and Mouse dataset descriptions in Appendix B.4, Figures 9b, 10b, 11b and 12b of the revised paper.

On Saturating-Growth Data, ESFM's predicted cells follow the observed progression of the proliferative population while retaining a compact approximately quiescent population on the right. In contrast, VGFM's predicted right-hand population is more dispersed and displaced from the observed cluster (Figure 6). This provides a concrete example of better held-out marginal recovery by ESFM.

The Mouse experiment provides a further example (Figure 12b). ESFM's predictions cover more of the observed upper vertical branch, whereas VGFM concentrates most predicted cells near the lower junction and underrepresents the upper branch. These examples show how ESFM better matches the observed held-out distributions than VGFM in the displayed comparisons; the gray curves are model-generated paths, rather than ground-truth single-cell trajectories.

Thank you for providing the detailed response to raised concerns. While training ESFM and difference to VGFM is clearer, I am still skeptical about method's novelty and potential to scale, which seems necessary given the concern raised about improvement by a small margin in comparison to current methods. Therefore, I will be maintaining my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.