acceptodds
Under review as a conference paper at ICLR 2027

scMomenta: A Second-order Neural ODE Framework for Single-cell Temporal Prediction

Abstract

Single-cell temporal prediction aims to reconstruct cell populations at unsampled time points from destructive measurements that provide only unpaired population snapshots. Most existing neural differential equation methods model latent cellular dynamics with a first-order equation, in which position and time alone determine the velocity. However, hidden kinetic variables, such as unspliced RNA, and dimensionality reduction can make latent dynamics depend on history that position does not record. We introduce scMomenta, a second-order neural ODE framework that augments the latent position with an explicit velocity state and learns a damped acceleration field. We also design an initial-velocity module that estimates each cell's velocity by optimal transport. On five time-course datasets, scMomenta achieves strong overall performance against six competing methods, improving the accuracy of single-cell temporal prediction. Experiments on latent dynamical order, trajectory crossing and velocity-space perturbation further show that an explicit velocity state captures temporal structure that position alone does not.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Temporal prediction, Single-cell transcriptomics, Neural ODE, Optimal transport

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

The authors propose a second-order neural ordinary differential equation framework for temporal scRNA-seq data. Previous methods rely only on the current state and time to learn first-order dynamics. The authors argue that those two are not sufficient, motivated by three aspects: (i) the equations of classical splicing models result in second-order dynamics with respect to the current state, (ii) when high-dimensional data are projected into a low-dimensional setup, the order of dynamics should increase to provide sufficient information about the gene space, and finally (iii) similar cells measured at a time do not mean that their future will be similar. To address these limitations, the authors augment the dynamics model by adding velocity information that allows the model to learn history-dependent dynamics. They estimate the initial velocity from the first two snapshots by utilizing optimal transport.

Critical strengths
  • The paper raises an important point: most existing approaches infer dynamics from the current position alone. The authors motivate, from biophysical, mathematical, and empirical perspectives, why information about history beyond position may be needed.
  • The authors evaluate on diverse datasets against established baselines. The method achieves the best results in most settings (Table 1).
  • Strong experimental design supports the paper’s claims. Section 4.3 tests Proposition 1 by projecting first-order gene-space dynamics onto fewer dimensions, showing that first-order closure breaks under compression while second-order largely restores it at the model’s latent dimension. Section 4.4 tests Proposition 2 on synthetic data with known coupling, where the first-order reduction fails as the theory predicts. The ablation in Section 4.6 separates the contributions of the velocity state and the OT initialization.
Critical weaknesses
  • Some related works are not discussed. While they do not address history-dependent dynamics, they target similar interpolation and extrapolation tasks and could be strong competitors:
    • [1] von Bassewitz, J. P., Pervez, A., Fumero, M., Robinson, M., Karaletsos, T., & Locatello, F. (2026). Learning explicit single-cell dynamics using ODE representations. In International Conference on Learning Representations.
    • [2] Balik, M. Y., & Lähdesmäki, H. (2026). Modeling Temporal scRNA-seq Data with Latent Gaussian Process and Optimal Transport. In Forty-Third International Conference on Machine Learning.
  • Proposition 3 requires fθf_\theta to be uniformly Lipschitz in vv with constant Λ\Lambda and γ>Λ\gamma > \Lambda. Yet, nothing in training enforces γ\gamma to be larger than Λ\Lambda. Is this condition always guaranteed to hold? It is not clear that the proposition applies to the trained model.
  • The paper does not report the learned γ\gamma or τ=1/γ\tau = 1/\gamma for the benchmark datasets. Reporting γ\gamma and Λ\Lambda per benchmark dataset and analyzing it to show whether velocity has time to depart from the value set by position and time would strengthen the paper.
Questions for authors
  • I have several concerns regarding the baselines: In the original papers, some baselines were designed specifically for PCA space, and some can handle 2000-dimensional gene space. Could the authors please clarify how they trained the baselines (i.e., on PCA space or gene space) in this work? Also, were any hyperparameters tuned for the baselines and their model? If so, please specify the search space and validation setup.

  • Could the authors please elaborate on why the first-order ablation variant performs substantially better than scNODE on SC extrapolation (53 vs 88 W2W_2) despite the general idea and architecture being similar?

  • How does the runtime (both training and inference) compare against the baselines?

  • The temporal prediction experiment considers interpolation-only and extrapolation-only setups. It would be nice to see high-sparsity (hard tasks in the scNODE paper) experiments that contain evaluation on both interpolation and extrapolation time points together.

5 comments

We sincerely thank the reviewer for the constructive comments. In our updated manuscript, we highlight the changes in blue.\color{blue}{\text{blue.}}


Weakness 1

Some related works are not discussed. While they do not address history-dependent dynamics, they target similar interpolation and extrapolation tasks and could be strong competitors:

[1] von Bassewitz, J. P., Pervez, A., Fumero, M., Robinson, M., Karaletsos, T., & Locatello, F. (2026). Learning explicit single-cell dynamics using ODE representations. In International Conference on Learning Representations.

[2] Balik, M. Y., & Lähdesmäki, H. (2026). Modeling Temporal scRNA-seq Data with Latent Gaussian Process and Optimal Transport. In Forty-Third International Conference on Machine Learning.

We have added Cell-MNN (von Bassewitz et al., 2026) and LGP-OT (Balik & Lähdesmäki, 2026) to the Related Work and as baselines to Tables 1 and 2, for both interpolation and extrapolation. Both were run from their official implementations with the settings recommended by their authors, without tuning, on the splits and seeds and with the evaluation pipeline of Tables 1 and 2 (Table R3-1).

Table R3-1. Temporal prediction results of Cell-MNN and LGP-OT; mean ± s.d. over five seeds.

MethodZBSCDRGAMA
W2W_2 ↓Pearson ↑W2W_2 ↓Pearson ↑W2W_2 ↓Pearson ↑W2W_2 ↓Pearson ↑W2W_2 ↓Pearson ↑
Interpolation
Cell-MNN31.55 ± 0.07.9909 ± .000130.06 ± 0.07.9767 ± .000158.91 ± 0.13.9673 ± .001261.17 ± 0.59.9502 ± .001521.73 ± 0.28.9668 ± .0012
LGP-OT24.45 ± 0.50.9898 ± .000729.81 ± 0.17.9721 ± .000551.77 ± 0.38.9648 ± .001162.79 ± 1.21.9334 ± .003919.12 ± 0.35.9654 ± .0017
scMomenta23.24 ± 0.14.9913 ± .000127.48 ± 0.24.9744 ± .000650.51 ± 0.25.9626 ± .001551.43 ± 0.31.9630 ± .001917.33 ± 0.36.9734 ± .0009
Extrapolation
Cell-MNN56.78 ± 0.73.9494 ± .001672.08 ± 4.81.9410 ± .0023189.1 ± 27.1.7926 ± .013791.80 ± 11.7.9131 ± .010379.84 ± 6.58.9066 ± .0073
LGP-OT40.83 ± 1.18.9673 ± .003068.90 ± 2.03.9502 ± .003179.26 ± 1.80.8750 ± .014866.92 ± 2.13.9269 ± .003035.07 ± 1.96.9378 ± .0072
scMomenta35.79 ± 0.63.9768 ± .003340.05 ± 2.20.9718 ± .004677.82 ± 1.54.9084 ± .012964.21 ± 1.73.9332 ± .006830.19 ± 1.16.9495 ± .0053

scMomenta achieves the lowest W2W_2 of the three methods on all five datasets, in both interpolation and extrapolation. It also has the highest Pearson in eight of the ten entries; the two exceptions are interpolation on SC, where Cell-MNN is higher, and on DR, where both methods are higher. LGP-OT is a strong competitor: its W2W_2 is within 5% of that of scMomenta on DR in both tasks and on GA extrapolation. Cell-MNN was evaluated only on interpolation in its paper; we obtained its extrapolations by integrating the learned dynamics beyond the last observed time point. It predicts with the closed-form solution exp⁡(AθΔt) z\exp(A_\theta\Delta t)\,z of a linear ODE fixed at the starting state, which its authors use as a local approximation, so any error in AθA_\theta is amplified exponentially as the target time moves beyond the observed range.

Weakness 2

Proposition 3 requires f_\theta to be uniformly Lipschitz in v with constant \Lambda and \gamma>\Lambda. Yet, nothing in training enforces \gamma to be larger than \Lambda. Is this condition always guaranteed to hold? It is not clear that the proposition applies to the trained model.

The paper does not report the learned \gamma or \tau = frac{1}{\gamma} for the benchmark datasets. Reporting \gamma and \Lambda per benchmark dataset and analyzing it to show whether velocity has time to depart from the value set by position and time would strengthen the paper.

Is the condition guaranteed? No. Nothing in training enforces γ>Λ\gamma>\Lambda, and we leave it unenforced on purpose. Proposition 3 is a statement about the model class: it shows that the second-order model contains first-order dynamics as its overdamped regime. Enforcing the condition would push scMomenta toward this regime, and with a large margin γ−Λ\gamma-\Lambda it would reduce the model to a first-order one, removing the history dependence that motivates the paper. Since neither training nor prediction requires the condition, we learn γ\gamma freely and let the data decide how far the model departs from the first-order limit.

Learned γ\gamma and Λ\Lambda. Following the reviewer's suggestion, Table R3-2 reports γ\gamma, τ\tau and Λ\Lambda of the trained models. τ=1/γ\tau=1/\gamma is the time scale over which damping erases the velocity, that is, how long the history carried by the velocity persists; we report it in units of the median interval Δt\Delta t between consecutive observed time points. Since the global Lipschitz constant cannot be computed, the local Λ\Lambda is the median, over points on the simulated trajectories, of the spectral norm of ∂fθ/∂v\partial f_\theta/\partial v, and we also report the fraction of these points at which γ>∥∂fθ/∂v∥\gamma>\lVert\partial f_\theta/\partial v\rVert. We have added this table to Appendix H.4.

Table R3-2. Learned damping of scMomenta; mean ± s.d. over five seeds.

DatasetTaskγ\gammaτ/Δt\tau/\Delta tLocal Λ\Lambda (median)Points with γ>∥∂fθ/∂v∥\gamma>\lVert\partial f_\theta/\partial v\rVert
ZBInterp.0.36 ± 0.022.804.70 ± 1.100%
ZBExtrap.0.24 ± 0.014.233.62 ± 0.850%
SCInterp.0.45 ± 0.032.216.21 ± 0.190%
SCExtrap.0.50 ± 0.082.039.17 ± 0.640%
DRInterp.5.16 ± 0.090.199.74 ± 0.748%
DRExtrap.4.89 ± 0.120.2012.49 ± 2.614%
GAInterp.1.82 ± 0.160.553.02 ± 2.1133%
GAExtrap.1.60 ± 0.010.631.15 ± 0.9570%
MAInterp.0.82 ± 0.301.323.97 ± 3.500%
MAExtrap.0.96 ± 0.271.126.74 ± 3.920%

The learned damping varies widely across datasets, from γ=0.24\gamma=0.24 on ZB to 5.16 on DR. On ZB, SC and MA, τ\tau exceeds one observation interval, so the velocity is not set by the current position alone, and the dynamics depend on each cell's history. On DR, τ\tau is only a fifth of an interval and the model is nearly first order; GA lies in between. This agrees with the temporal prediction results (Table 2): on DR, a first-order method (scIMF) achieves a lower extrapolation W2W_2 than scMomenta, as does scNODE on GA.

Question 1

I have several concerns regarding the baselines: In the original papers, some baselines were designed specifically for PCA space, and some can handle 2000-dimensional gene space. Could the authors please clarify how they trained the baselines (i.e., on PCA space or gene space) in this work? Also, were any hyperparameters tuned for the baselines and their model? If so, please specify the search space and validation setup.

Representation space. Each baseline was trained in the representation used in its original paper and official implementation, for example scNODE in the latent space of its VAE, and PI-SDE and scIMF in a PCA space. Every training representation, whether a PCA or an autoencoder, was fitted on the cells of the observed time points only, and the predictions of all methods were mapped back to gene space and evaluated with the same pipeline.

Hyperparameters. Each baseline was run from its official implementation with the settings recommended by its authors, without additional tuning. scMomenta uses a single setting for all five datasets, selected using the observed time points only. Held-out time points were never used for model selection or for fitting any training representation.


Question 2

Could the authors please elaborate on why the first-order ablation variant performs substantially better than scNODE on SC extrapolation (53 vs 88 W_2) despite the general idea and architecture being similar?

Where the gap arises. The two models differ substantially only on SC extrapolation (Table R3-3a).

Table R3-3a. W2W_2 of scNODE (Tables 1 and 2) and the first-order ablation (Table 3); mean ± s.d. over five seeds.

MethodZB interp.ZB extrap.SC interp.SC extrap.
scNODE28.89 ± 1.0848.68 ± 1.0433.33 ± 0.9888.75 ± 22.5
First-order ablation28.37 ± 0.2842.55 ± 1.0131.95 ± 0.3053.32 ± 2.07

This is also the only setting in which scNODE is unstable across seeds (Table R3-3b): its coefficient of variation there is 25%, versus at most 9.2% in its other nine settings of Tables 1 and 2, whereas that of the first-order ablation is 3.9%. The gap is therefore specific to scNODE on this task, not to first-order latent dynamics or to our evaluation pipeline.

Table R3-3b. W2W_2 of scNODE on SC extrapolation for each seed.

Seed4243444546Mean ± s.d.
W2W_284.6666.49121.13100.6370.8288.75 ± 22.49

Why the gap arises there. SC extrapolation has the longest horizon in our benchmark: four held-out time points, up to day 18, each predicted by integrating from t0t_0. A deviation at the end of the training range is therefore carried into the extrapolation and grows with the horizon. How strongly it grows depends on the model, and despite the shared VAE architecture, scNODE and our first-order ablation differ in most implementation choices, including VAE training (joint vs. pretrained and frozen), the space of the loss (gene vs. latent) and the ODE solver (Euler with step size 1 vs. adaptive dopri5).

Decoder or latent dynamics? As a diagnostic, we replaced the linear decoder of scNODE with a nonlinear one (one hidden layer of width 50 and a ReLU output), keeping all other settings. This lowers the SC extrapolation W2W_2 to 62.43 ± 9.78. Its spread across seeds, however, remains large (s.d. 9.78, 4.7 times that of the first-order ablation), so the instability originates in the latent dynamics of scNODE beyond the last observed time point, which a decoder cannot correct.

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

The paper proposes scMomenta, a second-order neural ODE for single-cell temporal prediction from unpaired snapshots. It augments position zz with velocity vv and learns a damped acceleration field. The unobserved initial velocity is produced by an encoder warm-started on barycentric velocities from OT between the first two observed snapshots. Training minimizes a Sinkhorn loss plus a kinetic penalty. The method reports strong overall interpolation and extrapolation performance on several benchmarks.

Critical strengths
  1. Whether position alone is a sufficient representation state is a meaningful question.
  2. The distinction between level- and velocity-space perturbations, including the CHGA confound analysis, is interesting.
  3. The evaluation protocol is consistent across methods, with five seeds and a single hyperparameter setting for scMomenta.
Critical weaknesses
  1. First-order energy regularization can favor non-crossing couplings. Although the second-order architecture provides a structural inductive bias, the objective contains no explicit penalty on acceleration or changes in velocity. Its kinetic regularizer is the action used in the Benamou–Brenier formulation of dynamic OT, and the initial velocities are derived from a quadratic-cost OT coupling. These choices do not explicitly favor preserving momentum across observations. Because the Sinkhorn term depends only on the marginals, it cannot distinguish couplings. Among paths that match the marginals equally well, the objective's preference is set by the regularizer. When the true trajectories cross between observation times, this preference is for a non-crossing coupling, since kinetic-energy minimizers matching the marginals are piecewise W2W_2 geodesics, which do not cross between observations. The crossings recovered in Sec. 4.4 therefore appear to rely on implicit bias from initialization and optimization. The paper should analyze this (e.g., crossing accuracy as a function of λv\lambda_v) and test whether an acceleration penalty changes the selected coupling.

  2. The benefit of the explicit velocity structure is not isolated from state augmentation. The first-order ablation removes all 50 additional state dimensions. The gains could therefore partly reflect state augmentation instead of the specific constraint z˙=v\dot z = v. A matched-state augmented baseline is needed to distinguish these explanations.

  3. Part of the headline gains may not come from the second-order structure. On SC extrapolation, the first-order ablation already outperforms every external baseline: its reported W2W_2 is 53.32, compared with 61.12 for PRESCIENT and 88.75 for scNODE. The full model improves further to 40.05, which is substantial, but the strong first-order ablation suggests that differences in training setup or other components account for part of the headline gains. The particularly large gap from scNODE, whose latent space scMomenta adopts, warrants explanation.

  4. The CE crossing experiment does not establish a structural limitation of first-order models. The CE crossing in Figure 4 occurs in a two-dimensional projection of the 50-dimensional latent space. High-dimensional first-order trajectories can cross in such projections while remaining distinct in the full state space.

  5. The closest related baselines are excluded from comparison. The idea of second-order dynamics is not new. As the authors stated, DMSB (Chen et al., 2023) and 3MSBM (Theodoropoulos et al., 2025) model population dynamics in position–velocity space and infer unobserved velocities. Both include single-cell experiments with held-out intermediate marginals, directly overlapping with the interpolation task here. The stated reason for excluding them is not convincing. For extrapolation, their learned forward dynamics can be integrated past the last marginal, just as scMomenta evaluates its time-dependent field fθ(z,v,t)f_\theta(z, v, t) at times not seen in training. A comparison on both tasks would help establish the value of scMomenta relative to existing phase-space approaches. The specific methodological novelty over existing phase-space population models is unclear.

Questions for authors
  1. How sensitive is the synthetic crossing accuracy to λv\lambda_v and initialization? Does adding acceleration regularization change coupling recovery?
  2. Can the authors compare against an augmented first-order NODE with a matched 2L2L-dimensional state, learned initial augmentation, and comparable representation, loss, and training budget?
  3. What explains the large gap between the first-order ablation and scNODE on SC extrapolation?
  4. Do the two CE lineages in Figure 4 actually come close in the full 50-dimensional latent space (e.g., their minimum distance relative to typical cell–cell distances), or only in the two-dimensional projection?
  5. Can DMSB and 3MSBM be included in both the interpolation and extrapolation benchmarks, together with a simple constant-velocity continuation baseline for extrapolation?
6 comments

We sincerely thank the reviewer for the constructive comments. In our updated manuscript, we highlight the changes in blue.\color{blue}{\text{blue.}}


Weakness 1 & Question 1

W1. First-order energy regularization can favor non-crossing couplings. Although the second-order architecture provides a structural inductive bias, the objective contains no explicit penalty on acceleration or changes in velocity. Its kinetic regularizer is the action used in the Benamou–Brenier formulation of dynamic OT, and the initial velocities are derived from a quadratic-cost OT coupling. These choices do not explicitly favor preserving momentum across observations. Because the Sinkhorn term depends only on the marginals, it cannot distinguish couplings. Among paths that match the marginals equally well, the objective's preference is set by the regularizer. When the true trajectories cross between observation times, this preference is for a non-crossing coupling, since kinetic-energy minimizers matching the marginals are piecewise W_2 geodesics, which do not cross between observations. The crossings recovered in Sec. 4.4 therefore appear to rely on implicit bias from initialization and optimization. The paper should analyze this (e.g., crossing accuracy as a function of \lambda_v) and test whether an acceleration penalty changes the selected coupling.

Q1. How sensitive is the synthetic crossing accuracy to \lambda_v and initialization? Does adding acceleration regularization change coupling recovery?

The objective itself does not select the crossing: the Sinkhorn term is indifferent between couplings, and the kinetic term, which we include to keep the velocity bounded under long integration, favours the non-crossing coupling. By Proposition 4, the coupling of any method trained on marginals is decided by something other than the data-fitting term; in a first-order model, it is decided by the kinetic term, which selects the non-crossing coupling.

In scMomenta, the coupling is decided by the OT initialization of the velocity. We estimate each cell's initial velocity by OT between the first two observed snapshots and warm-start the velocity encoder on these velocities, and the velocity state carries this direction into the crossing interval. In our synthetic data, these two snapshots lie far from the crossing, where the OT coupling is unambiguous, so the preference of quadratic-cost OT for non-crossing couplings is not carried into the crossing interval.

As the reviewer suggests, we tested the effect of λv\lambda_v, the OT initialization and an acceleration penalty λa\lambda_a on the synthetic data of Appendix F.2 (Table R2-1).

Table R2-1. Sensitivity of crossing recovery on the synthetic data; mean ± s.d. over five seeds.

SettingCoupling accuracy ↑W2W_2 ↓Trajectory MSE ↓
scMomenta (default)1.00 ± 0.000.311 ± 0.0560.038 ± 0.015
λv=10−2\lambda_v=10^{-2}1.00 ± 0.000.313 ± 0.0550.043 ± 0.017
λv=10−1\lambda_v=10^{-1}1.00 ± 0.000.325 ± 0.0550.071 ± 0.013
λv=1\lambda_v=11.00 ± 0.000.493 ± 0.0400.242 ± 0.006
λa=10−4\lambda_a=10^{-4}1.00 ± 0.000.311 ± 0.0560.038 ± 0.015
λa=10−3\lambda_a=10^{-3}1.00 ± 0.000.311 ± 0.0560.038 ± 0.015
λa=10−2\lambda_a=10^{-2}1.00 ± 0.000.311 ± 0.0560.038 ± 0.014
w/o OT initialization0.00 ± 0.000.261 ± 0.0840.951 ± 0.013
First-order model (reference)0.00 ± 0.000.241 ± 0.0810.938 ± 0.006

Neither regularizer changes the outcome. As λv\lambda_v increases, every seed keeps the crossing, even when the kinetic term clearly degrades the fit, and adding the acceleration penalty leaves the results unchanged. Preserving momentum thus does not require an acceleration term in the objective: the velocity state preserves it by itself.

The outcome is decided by the OT initialization. Without it, every seed falls back to the same non-crossing solution as the first-order model, while the fit to the marginals does not get worse. The marginals alone thus cannot distinguish the two couplings, and it is the OT initialization that selects the crossing. The crossing is therefore not an incidental bias of optimization: it is robust to the regularizers and across seeds, and is determined by the OT initialization, a component designed into the method. This agrees with our ablation on real data (Table R2-2): z˙=v\dot z=v helps only together with the OT initialization. We have added this analysis to Appendix F.2.

Weakness 2 & Question 2

W2. The benefit of the explicit velocity structure is not isolated from state augmentation. The first-order ablation removes all 50 additional state dimensions. The gains could therefore partly reflect state augmentation instead of the specific constraint \dot{z}=v. A matched-state augmented baseline is needed to distinguish these explanations.

Q2. Can the authors compare against an augmented first-order NODE with a matched 2L-dimensional state, learned initial augmentation, and comparable representation, loss, and training budget?

Reducing the model to first order removes both the extra state and the constraint z˙=v\dot z=v, so our original ablation could not separate the two. Following augmented neural ODEs (Dupont et al., 2019), we have added the variant w/o z˙=v\dot z=v (2L2L ODE state). Its state (z,ξ)∈R2L(z,\xi)\in\mathbb{R}^{2L} has an extra coordinate ξ\xi of the dimension of vv but without a prescribed meaning, initialized by a learned encoder ξ(t0)=hψ(z0)\xi(t_0)=h_\psi(z_0) trained jointly with the dynamics. The representation, training procedure and other settings are those of scMomenta (Table R2-2).

Table R2-2. Extended ablation; mean ± s.d. over five seeds.

VariantZB interp.ZB extrap.SC interp.SC extrap.
W2W_2 ↓Pearson ↑W2W_2 ↓Pearson ↑W2W_2 ↓Pearson ↑W2W_2 ↓Pearson ↑
Full model23.24 ± 0.14.9913 ± .000135.79 ± 0.63.9768 ± .003327.48 ± 0.24.9744 ± .000640.05 ± 2.20.9718 ± .0046
w/o OT initialization26.96 ± 0.23.9910 ± .000239.27 ± 0.89.9631 ± .003828.44 ± 0.20.9744 ± .000546.60 ± 3.17.9683 ± .0061
w/o z˙=v\dot z=v (2L2L ODE state)26.93 ± 0.15.9904 ± .000338.31 ± 1.78.9628 ± .005928.88 ± 0.28.9739 ± .000747.34 ± 3.72.9584 ± .0062
Reduced to first order28.37 ± 0.28.9810 ± .000442.55 ± 1.01.9597 ± .002531.95 ± 0.30.9720 ± .000853.32 ± 2.07.9660 ± .0027

The constraint z˙=v\dot z=v helps only together with the OT initialization. Without it, z˙=v\dot z=v brings no consistent gain over the unconstrained state. With it, scMomenta improves on the matched-state variant in every setting. The two are complementary: z˙=v\dot z=v gives the extra coordinates the meaning of a velocity, so that OT displacements can initialize them, and the OT initialization gives z˙=v\dot z=v an informed initial direction to carry forward. An uninformed initial velocity gives z˙=v\dot z=v nothing to carry forward. The rest of the gain over the first-order model comes from the extra state itself. We have added the new variant to Table 3 and revised the conclusion of Section 4.6 accordingly.

Weakness 3 & Question 3

W3. Part of the headline gains may not come from the second-order structure. On SC extrapolation, the first-order ablation already outperforms every external baseline: its reported W_2 is 53.32, compared with 61.12 for PRESCIENT and 88.75 for scNODE. The full model improves further to 40.05, which is substantial, but the strong first-order ablation suggests that differences in training setup or other components account for part of the headline gains. The particularly large gap from scNODE, whose latent space scMomenta adopts, warrants explanation.

Q3. What explains the large gap between the first-order ablation and scNODE on SC extrapolation?

We agree that the first-order ablation is strong: on SC extrapolation, it outperforms other baselines, so part of the headline gains comes from components that all variants of scMomenta share. The ablations show, however, that the full model improves on each ablated variant (Table R2-2), so the second-order structure contributes in addition to these shared components. This is clearest for the matched-state variant added for Weakness 2, which has the same 2L2L-dimensional state as scMomenta but neither the constraint z˙=v\dot z=v nor the OT initialization that this constraint enables. We therefore do not attribute all of the headline gains to the second-order structure, but it does contribute to them. We now turn to the gap between the first-order ablation and scNODE (Q3).

Where the gap arises. The two models differ substantially only on SC extrapolation (Table R2-3a).

Table R2-3a. W2W_2 of scNODE (Tables 1 and 2) and the first-order ablation (Table 3); mean ± s.d. over five seeds.

MethodZB interp.ZB extrap.SC interp.SC extrap.
scNODE28.89 ± 1.0848.68 ± 1.0433.33 ± 0.9888.75 ± 22.5
First-order ablation28.37 ± 0.2842.55 ± 1.0131.95 ± 0.3053.32 ± 2.07

This is also the only setting in which scNODE is unstable across seeds (Table R2-3b): its coefficient of variation there is 25%, versus at most 9.2% in its other nine settings of Tables 1 and 2, whereas that of the first-order ablation is 3.9%. The gap is therefore specific to scNODE on this task, not to first-order latent dynamics or to our evaluation pipeline.

Table R2-3b. W2W_2 of scNODE on SC extrapolation for each seed.

Seed4243444546Mean ± s.d.
W2W_284.6666.49121.13100.6370.8288.75 ± 22.49

Why the gap arises there. SC extrapolation has the longest horizon in our benchmark: four held-out time points, up to day 18, each predicted by integrating from t0t_0. A deviation at the end of the training range is therefore carried into the extrapolation and grows with the horizon. How strongly it grows depends on the model, and despite the shared VAE architecture, scNODE and our first-order ablation differ in most implementation choices, including VAE training (joint vs. pretrained and frozen), the space of the loss (gene vs. latent) and the ODE solver (Euler with step size 1 vs. adaptive dopri5).

Decoder or latent dynamics? As a diagnostic, we replaced the linear decoder of scNODE with a nonlinear one (one hidden layer of width 50 and a ReLU output), keeping all other settings. This lowers the SC extrapolation W2W_2 to 62.43 ± 9.78. Its spread across seeds, however, remains large (s.d. 9.78, 4.7 times that of the first-order ablation), so the instability originates in the latent dynamics of scNODE beyond the last observed time point, which a decoder cannot correct.

Since scNODE and the first-order ablation are both first-order models, the gap between them reflects differences in implementation rather than in model order, and we do not use it as evidence for the second-order structure. The more direct evidence is the matched-state ablation (Table R2-2), which distinguishes generic state augmentation from structured position–velocity dynamics at the same state dimension.

Reviewer_fCE2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a method to reconstruct cell populations at unsampled time points given only unpaired snapshots. Current neural differential equation models model latent dynamics with first-order equations where position and time alone determine velocity. This implies that the two cells that have the same latent variable get the same direction of evolution.

The paper suggests that this assumption can fail in certain conditions: 1) RNA production involves hidden kinetic variables leading to higher order dynamics, 2) even first order gene dynamics lose closure in latent space 3) and empirically linear racing shows that highly similar transcriptional states can lead to different descendants. Hidden variables can make latent dynamics depend on history that position does not record.

scMomenta augments position with velocity and learns an acceleration field. This requires an estimate for initial velocity. An optimal transport module estimates initial velocity which is only used as an initialization which training can improve.

Critical strengths

The paper attempts to motivate its choice of second order model by plausible justifications for requiring higher order dynamics, with biophysical, mathematical (proposition 1) and empirical (Secton 4.4) justifications for second order dynamics.

The justifications for using a second order model are validated in data showing trajectory crossing with second order dynamics.

The paper shows ablations for OT initialized initial velocity, showing that the full model does make good use of the initialization while the variants underperform.

Critical weaknesses

The paper suggests that introducing velocity adds no new assumption. However, it does seem to me that in the learning setting it does add a new constraint which is that the first two time steps are observed over which optimal transport can be used to initialize the velocity. In my understanding, the first two snapshots are not always assumed to be available in competing methods and the second snapshot is one of the snapshots that are held-out one at a time during evaluation.

From the description of the setting in 3.1 too, it appears that t1 is not assumed to be observed. The only constraint on the observation set is that it includes 0 (line 141).

The method depends on a non-observable velocity for which there is no target.

Questions for authors

The loss term does not contain a regularization for the acceleration. Is there a reason for this?

How does your method compare with CellMNN (ICLR 26), OT-MFM etc. Can your method be used on the same datasets with the same evaluation protocol as theirs?

3 comments

We sincerely thank the reviewer for the constructive comments. In our updated manuscript, we highlight the changes in blue.\color{blue}{\text{blue.}}


Weakness 1

The paper suggests that introducing velocity adds no new assumption. However, it does seem to me that in the learning setting it does add a new constraint which is that the first two time steps are observed over which optimal transport can be used to initialize the velocity. In my understanding, the first two snapshots are not always assumed to be available in competing methods and the second snapshot is one of the snapshots that are held-out one at a time during evaluation.

From the description of the setting in 3.1 too, it appears that t1 is not assumed to be observed. The only constraint on the observation set is that it includes 0 (line 141).

scMomenta does not require t1t_1 to be observed. The OT initialization uses the first two observed time points: if t1t_1 is held out, the OT plan is computed between t0t_0 and the next observed time point, and held-out time points never enter the loss or the OT initialization (Section 3.4). The only requirement is therefore ∣O∣≥2|\mathcal{O}|\ge 2, which any method that learns dynamics from snapshots needs. OT-MFM (Kapuśniak et al., 2024) likewise couples consecutive observed snapshots: in its leave-one-out benchmark, the fold that holds out t1t_1 couples t0t_0 directly with t2t_2, as scMomenta does. Introducing the velocity therefore adds no requirement on the data.

The reviewer's reading of Section 3.1 is correct, and the inconsistency was in our notation: the Introduction refers to the first two observed snapshots, but Section 3.3 wrote the second one as t1t_1 (Equations 6–7). In the revision, Section 3.1 states ∣O∣≥2|\mathcal{O}|\ge 2, and Section 3.3 writes the second snapshot as tk1t_{k_1}, the first observed time point after t0t_0, with k1=min⁡(O∖{0})k_1=\min(\mathcal{O}\setminus\{0\}). The caption of Figure 1 and the Conclusion now also refer to the first two observed snapshots.


Weakness 2

The method depends on a non-observable velocity for which there is no target.

We agree. Single-cell sequencing is destructive, so no cell's velocity is observed, and the velocity of scMomenta has no ground-truth target. The same holds for first-order models, whose velocity fθ(z,t)f_\theta(z,t) is also learned from snapshots alone. The only additional unobserved quantity in scMomenta is the initial velocity v(t0)v(t_0), for which OT between the first two observed snapshots provides a prior and a warm start (Weakness 1). After t0t_0, vv evolves under the learned dynamics and is constrained, through z˙=v\dot z=v, by the distribution-matching loss at all later observed time points. We now state in the Limitations that the learned velocity is a model-based estimate that snapshot data cannot validate directly.


Question 1

The loss term does not contain a regularization for the acceleration. Is there a reason for this?

An acceleration penalty was not part of our original design: the only regularizer in our objective (Equation 8) is the kinetic term λv∫∥v∥2dt\lambda_v\int\lVert v\rVert^2dt, which keeps the velocity bounded under long integration. Following DMSB (Chen et al., 2023) and 3MSBM (Theodoropoulos et al., 2025), we tested adding a penalty on the total acceleration, λa∫∥z¨∥2dt\lambda_a\int\lVert\ddot z\rVert^2dt with z¨=fθ(z,v,t)−γv\ddot z=f_\theta(z,v,t)-\gamma v, while keeping the kinetic term (Table R1-1).

Table R1-1. scMomenta with an acceleration penalty on ZB and SC; mean ± s.d. over five seeds.

VariantZB interp.ZB extrap.SC interp.SC extrap.
W2W_2 ↓Pearson ↑W2W_2 ↓Pearson ↑W2W_2 ↓Pearson ↑W2W_2 ↓Pearson ↑
scMomenta (default, λa=0\lambda_a=0)23.24 ± 0.14.9913 ± .000135.79 ± 0.63.9768 ± .003327.48 ± 0.24.9744 ± .000640.05 ± 2.20.9718 ± .0046
λa=10−4\lambda_a=10^{-4}22.97 ± 0.27.9912 ± .000335.86 ± 1.27.9764 ± .002827.55 ± 0.17.9740 ± .000440.66 ± 2.40.9701 ± .0044
λa=10−3\lambda_a=10^{-3}23.05 ± 0.24.9911 ± .000435.53 ± 1.17.9766 ± .004927.54 ± 0.33.9741 ± .000840.25 ± 1.27.9704 ± .0012
λa=10−2\lambda_a=10^{-2}22.97 ± 0.21.9911 ± .000235.77 ± 0.78.9771 ± .003627.58 ± 0.38.9746 ± .001040.09 ± 0.93.9712 ± .0022

The penalty leaves the results largely unchanged. Apart from ZB interpolation, where W2W_2 decreases slightly (22.97–23.05 vs. 23.24), every change is smaller than the across-seed standard deviation of the default model. Since the penalty adds a hyperparameter without a consistent gain, we keep the simpler objective and report the variant in Appendix H.3.

Question 2

How does your method compare with CellMNN (ICLR 26), OT-MFM etc. Can your method be used on the same datasets with the same evaluation protocol as theirs?

We have added Cell-MNN (von Bassewitz et al., 2026) and OT-MFM (Kapuśniak et al., 2024) as baselines to Tables 1 and 2, for both interpolation and extrapolation. Both were run from their official implementations with the recommended settings, on the splits and seeds and with the evaluation pipeline of Tables 1 and 2 (Table R1-2).

Table R1-2. Temporal prediction results of Cell-MNN and OT-MFM; mean ± s.d. over five seeds.

MethodZBSCDRGAMA
W2W_2 ↓Pearson ↑W2W_2 ↓Pearson ↑W2W_2 ↓Pearson ↑W2W_2 ↓Pearson ↑W2W_2 ↓Pearson ↑
Interpolation
Cell-MNN31.55 ± 0.07.9909 ± .000130.06 ± 0.07.9767 ± .000158.91 ± 0.13.9673 ± .001261.17 ± 0.59.9502 ± .001521.73 ± 0.28.9668 ± .0012
OT-MFM29.76 ± 0.48.9811 ± .000230.03 ± 0.29.9719 ± .000553.40 ± 0.47.9694 ± .002562.31 ± 1.14.9423 ± .001320.51 ± 0.49.9624 ± .0016
scMomenta23.24 ± 0.14.9913 ± .000127.48 ± 0.24.9744 ± .000650.51 ± 0.25.9626 ± .001551.43 ± 0.31.9630 ± .001917.33 ± 0.36.9734 ± .0009
Extrapolation
Cell-MNN56.78 ± 0.73.9494 ± .001672.08 ± 4.81.9410 ± .0023189.1 ± 27.1.7926 ± .013791.80 ± 11.7.9131 ± .010379.84 ± 6.58.9066 ± .0073
OT-MFM975.5 ± 76.5.6454 ± .03061621 ± 131.6024 ± .05331856 ± 102.3269 ± .02061132 ± 187.7411 ± .0339211.0 ± 14.2.8951 ± .0181
scMomenta35.79 ± 0.63.9768 ± .003340.05 ± 2.20.9718 ± .004677.82 ± 1.54.9084 ± .012964.21 ± 1.73.9332 ± .006830.19 ± 1.16.9495 ± .0053

In interpolation, the task on which both methods were evaluated, scMomenta has the lowest W2W_2 on all five datasets (Pearson is higher for Cell-MNN on SC and for both methods on DR). Neither original paper evaluates extrapolation; we obtained their extrapolations by integrating the learned dynamics beyond the last observed time point, and these entries reflect how the two methods are built. OT-MFM is trained simulation-free: its vector field is fitted to the velocities of interpolation paths between consecutive observed snapshots and is never integrated during training, so beyond the last snapshot, where no such path exists, its errors accumulate unchecked along the integrated trajectory. Cell-MNN predicts with the closed-form solution exp⁡(AθΔt) z\exp(A_\theta\Delta t)\,z of a linear ODE fixed at the starting state, which its authors use as a local approximation, so any error in AθA_\theta is amplified exponentially as the target time moves beyond the observed range. We have also added both methods to the Related Work.

Thank you for the clarification, especially regarding the initialization and the baselines. I will maintain my positive score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.