acceptodds
Under review as a conference paper at ICLR 2027

KoopCell: Koopman-Based Generative Model for Learning Single-Cell Dynamics from Distribution Snapshots

18 views

Abstract

Learning population dynamics from temporally sparse, unpaired distribution snapshots is a fundamental challenge in developmental biology. Recent approaches based on neural differential equations and flow matching can interpolate between observed population snapshots, but may struggle to extrapolate beyond the training horizon and often lack an explicit mechanism for modeling developmental branching. We propose KoopCell, a unified generative framework based on Koopman–Mori–Zwanzig theory that jointly learns representations and predictive linear latent dynamics. Theoretically, using the weak continuity equation, we derive a closed-form least-squares estimator for the Koopman generator from distribution snapshots and establish convergence guarantees under suitable assumptions. To model branching dynamics, we further develop KoopCell-M, which incorporates non-Markovian memory into the latent Koopman dynamics through a Markovian embedding. Experiments on synthetic systems and three scRNA-seq datasets demonstrate the ability of our framework to recover Koopman spectra, model branching through memory, and scale to predicting high-dimensional gene expression distributions, achieving state-of-the-art performance among the evaluated methods.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Koopman learning, generative modeling, single-cell, Mori-Zwanzig

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_KWRrOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes KoopCell, a generative model that combines a VAE with Koopman-inspired linear latent dynamics learned from unpaired single-cell snapshots using a weak-form estimator. Its extension, KoopCell-M, adds hidden memory variables and mixture initialization to model branching, guided by biological annotations and developmental paths. Synthetic experiments examine spectral recovery and branching, while evaluations on three scRNA-seq datasets report improved average extrapolation performance over the tested baselines.

Critical strengths

The weak-form formulation connects population-level observations to latent generator estimation, providing a closed-form update and theoretical characterizations under explicit assumptions. This is a useful contribution to a challenging single-cell modeling problem.

Synthetic experiments examine spectral recovery and branching, while three scRNA-seq datasets test interpolation and extrapolation. The reported improvements in average extrapolation performance support the framework's practical promise for predicting unobserved cell populations.

Critical weaknesses

1. Insufficient ablation of the memory mechanism.

In Appendix D, KoopCell-M introduces additional hidden dimensions, mixture components, and learned conditional initialization. Table 3 therefore cannot isolate the contribution of memory feedback from increased model capacity or hidden heterogeneity. A controlled C=0C=0 ablation with the same architecture, initialization, and training protocol is needed to assess the role of the memory kernel. In addition, the experiments lack a comparison with Delay Flow Matching (Zhao et al., 2026), which the paper cites when motivating memory-based branching.

2. Additional biological supervision in KoopCell-M.

KoopCell-M uses annotated cell states, a predefined developmental DAG, supervised classification, and a path consistency loss, providing additional biological information unavailable to several baselines. It is therefore unclear whether the improvements arise from memory dynamics or this supervision. Ablations removing the developmental-path supervision and associated losses would help isolate their contributions.

3. Unclear methodological contribution of the weak-form estimator.

Existing methods such as WCR (Lu et al., 2024) already use weak formulations and regression to estimate dynamics from unpaired distribution snapshots. The authors should clarify the specific novelty of their estimator. Moreover, the experiments do not isolate its practical contribution from representation learning and population matching. A comparison with direct gradient-based optimization of the same latent linear dynamics would strengthen the evidence, particularly since Appendix D does not use the proposed closed-form updates.

4. Insufficient documentation of model selection.

Appendix E.2 describes baseline tuning, whereas Appendix E.3 only reports the final task-specific configurations of KoopCell and KoopCell-M. Please provide the validation splits, hyperparameter search spaces, selection metrics, tuning budgets, and checkpoint-selection rules. Explicitly confirm that held-out test time points were excluded from all model-selection decisions.

5. Missing simple temporal baselines.

The evaluation lacks simple baselines that would establish the benefit of learning continuous dynamics. For extrapolation, the last observed training-time distribution could be used to predict all future time points. For interpolation, OT-based displacement interpolation between neighboring observed snapshots should be included. If mean-expression accuracy is evaluated, linear interpolation of mean expression would also be useful. All baselines should use the same metrics and sampling protocol.

Questions for authors

1. What does the proposed weak-form estimator add to existing methods such as WCR (Lu et al., 2024)? Which components are inherited and which are new? How does KoopCell differ from the concurrent DisKO framework in its formulation, estimation procedure, and theoretical guarantees?

2. Neither KoopCell nor KoopCell-M explicitly models proliferation or death. How does this affect the interpretation of inferred trajectories and fate probabilities? Changes in lineage proportions may reflect differential proliferation or survival rather than differentiation. Please distinguish accurate population-distribution prediction from recovery of cell-fate dynamics.

3. Matching population distributions does not necessarily imply recovering correct cell-level transitions. Given the identifiability limitations acknowledged in Appendix A.4.2, what evidence supports the inferred developmental trajectories and fate probabilities? Could the authors validate these predictions using independent lineage-tracing information?

4. The paper emphasizes long-horizon extrapolation, but the real-data experiments cover relatively short extrapolation horizons. How stable are the learned dynamics over longer horizons, and how do the learned Koopman spectra relate to extrapolation errors across different seeds and temporal sampling densities?

9 comments
AuthorsOpenReview ↗

Reply to Reviewer KWRr (1/7)

We thank the reviewer for the careful review and constructive suggestions. We have uploaded a revised manuscript containing additional results and figures, together with further discussion and clarifications. We kindly refer the reviewer to these revisions, especially the visualizations cited in the responses below. We address the weaknesses and questions below.

1. Memory feedback and Delay Flow Matching (Weakness 1).

Ablation for the memory mechanism (C=0). We apologize that the original toy model did not sufficiently separate the contribution of memory feedback from that of initial hidden heterogeneity. To address this concern, we first performed a C=0C=0 ablation on the hard task of all three real datasets, keeping the architecture, other losses, and training protocol unchanged.

DataInterp. ΔOT\Delta\mathrm{OT}Extrap. ΔOT\Delta\mathrm{OT}
ZB+1.05±3.61+1.05 \pm 3.61+12.94±1.89+12.94 \pm 1.89
DR+2.61±2.56+2.61 \pm 2.56+8.09±1.98+8.09 \pm 1.98
SC+0.06±0.28+0.06 \pm 0.28+5.83±1.22+5.83 \pm 1.22

Table 1. ΔOT=OTC=0−OTfull\Delta\mathrm{OT}=\mathrm{OT}_{C=0}-\mathrm{OT}_{\mathrm{full}} (mean ± sample SD over three seeds); positive values indicate degradation. Extrapolation degrades consistently across datasets.

Beyond these empirical ablations on real datasets, Example B.1 (Appendix B.2.1; Figure 5 in the revised manuscript) provides a concrete case in which initial hidden heterogeneity produces transient branching even with C=0C=0. We should clarify that hidden heterogeneity can initiate population branching, while memory feedback can sustain the resulting branches. We will make this mechanism explicit in the revision.

Mathematically, consider the augmented linear model:

\dot H_t=CZ_t+DH_t+b_h.$$ If $A,D$ are strictly dissipative (a common assumption in Mori-Zwanzig formalism), with a common margin $\gamma>0$, the $C=0$ model satisfies $$\\|\Delta Z_t\\|\le e^{-\gamma t}\bigl(\\|\Delta Z_0\\|+t\\|B\\|\cdot\\|\Delta H_0\\|\bigr)\longrightarrow0,$$ where $\Delta$ denotes differences within a population. Thus, feedback can sustain nondecaying coupled modes despite stable $A,D$, while $C=0$ leaves only decaying effects of initial heterogeneity and cannot maintain nonvanishing branch separation under a shared Lipschitz decoder. The original toy model is not suited to studying the system's long-time behavior. We therefore designed a new 64D toy model with persistent terminal branches and strictly stable parameterizations of $A,D$; its construction is detailed in revised Appendix D.1. Apart from fixing $C=0$, the two memory variants use identical architecture, initialization, losses, and training protocol. | Model | Interp. SWD ↓ | Interp. branch $W_1$ ↓ | Extrap. SWD ↓ | Extrap. branch $W_1$ ↓ | |:---|---:|---:|---:|---:| | KoopCell | $0.1393 \pm 0.0025$ | $0.2612 \pm 0.0082$ | $0.4712 \pm 0.0001$ | $1.1357 \pm 0.0001$ | | KoopCell-M | $\mathbf{0.0654 \pm 0.0006}$ | $\mathbf{0.0550 \pm 0.0007}$ | $\mathbf{0.0723 \pm 0.0043}$ | $\mathbf{0.0832 \pm 0.0091}$ | | KoopCell-M ($C=0$) | $0.1285 \pm 0.0032$ | $0.1901 \pm 0.0100$ | $0.4712 \pm 0.0001$ | $1.1355 \pm 0.0001$ | *Table 2. The SWD and branch-coordinate $W_1$ metrics from manuscript Table 3, separated into interpolation and extrapolation (mean ± sample SD over three training seeds). Training ends at $t=64$; extrapolation uses $t=80,96,128,160$.* KoopCell-M preserves separated branches throughout the test horizon, while KoopCell and the $C=0$ control merge (Figure 15 in the revised manuscript). This ablation further separates feedback's contribution to sustained branching from hidden capacity and initial heterogeneity. > **Revised manuscript:** Appendices B.2.1 (Example B.1; Figure 5), D.1 (Table 8; Figure 15), and F.3 (Table 21).
AuthorsOpenReview ↗

Reply to Reviewer KWRr (2/7)

1. Memory feedback and Delay Flow Matching (Weakness 1), continued.
Comparison with Delay Flow Matching (DFM). We additionally evaluate DFM on all three datasets and tasks. As in PI-SDE, we fix the representation to a 50-dimensional PCA space. We explore delays τ∈0.5,1.0\tau\in\\{0.5,1.0\\}; constant [DFM(C)] and diverse linear [DFM(D)] initial histories; hidden widths 64,128,256,512\\{64,128,256,512\\} and depths 3,5\\{3,5\\}; SELU/Tanh activations; inclusion or omission of time inputs; and natural/clamped cubic-spline boundary conditions. Configurations are selected using training-only validation, and Table 3 reports the selected DFM(D) results as mean ± SD over three independent runs. [5]

DataTaskPredictionDFM(D)VGFMKoopCellKoopCell-M
ZBEasyInterp.427.80±1.38427.80 \pm 1.38441.71±3.75441.71 \pm 3.75411.42±2.26411.42 \pm 2.26403.38±1.52403.38 \pm 1.52
MediumExtrap.688.59±13.17688.59 \pm 13.17636.75±8.41636.75 \pm 8.41581.85±2.88581.85 \pm 2.88575.88±3.64575.88 \pm 3.64
HardInterp.495.49±1.39495.49 \pm 1.39508.28±1.78508.28 \pm 1.78489.51±0.36489.51 \pm 0.36488.77±2.92488.77 \pm 2.92
Extrap.681.14±3.84681.14 \pm 3.84678.25±4.69678.25 \pm 4.69622.36±1.14622.36 \pm 1.14614.26±0.19614.26 \pm 0.19
DREasyInterp.436.96±2.54436.96 \pm 2.54416.74±2.13416.74 \pm 2.13399.35±2.11399.35 \pm 2.11397.76±0.60397.76 \pm 0.60
MediumExtrap.853.54±24.31853.54 \pm 24.31708.96±39.53708.96 \pm 39.53553.82±9.01553.82 \pm 9.01545.04±3.08545.04 \pm 3.08
HardInterp.478.80±2.42478.80 \pm 2.42482.89±2.43482.89 \pm 2.43471.84±1.60471.84 \pm 1.60471.65±1.22471.65 \pm 1.22
Extrap.751.21±5.15751.21 \pm 5.15698.83±15.54698.83 \pm 15.54605.76±2.22605.76 \pm 2.22604.66±2.43604.66 \pm 2.43
SCEasyInterp.111.44±1.29111.44 \pm 1.29121.56±1.63121.56 \pm 1.6391.04±0.4191.04 \pm 0.4190.24±0.4290.24 \pm 0.42
MediumExtrap.183.55±18.04183.55 \pm 18.04168.74±0.10168.74 \pm 0.10110.01±1.13110.01 \pm 1.13112.44±1.38112.44 \pm 1.38
HardInterp.94.42±2.4694.42 \pm 2.4694.44±1.2294.44 \pm 1.2282.37±0.3182.37 \pm 0.3181.81±0.1681.81 \pm 0.16
Extrap.171.66±5.59171.66 \pm 5.59168.00±4.72168.00 \pm 4.72115.02±5.10115.02 \pm 5.10110.34±0.81110.34 \pm 0.81

Table 3. OT prediction error (mean ± sample SD; lower is better).

Like VGFM, DFM uses constructed conditional interpolation paths to supervise velocity learning. These flow-matching based algorithms not necessarily reflect the underlying biological dynamics. Our formulation instead learns a latent linear evolution directly from distribution snapshots. Consistent with this motivation, both KoopCell variants achieve lower mean errors than these flow-matching based methods, with particularly pronounced advantages in extrapolation.

Revised manuscript: Section 5; Appendices E.2 and E.4 (Tables 12–14).

2. Weak-form estimation and related methods (Weakness 3; Question 1).

Comparison with related works. We apologize for not making these connections and distinctions sufficiently clear in the original manuscript. We will make them explicit in the revised manuscript.

  1. WCR [1]. Both methods use test functions to express the Fokker–Planck equation in weak form, enabling efficient regression of parameters that enter the equation linearly. WCR estimates drift and diffusion coefficients in a prescribed state-space dictionary through sparse regression, following the SINDy paradigm. [2]

    In contrast, motivated by Koopman theory, we aim to learn coordinates z=gθ(x,t)z=g_\theta(x,t) that globally linearize the nonlinear dynamics. We jointly learn these observables and their linear generator. This formulation naturally integrates VAE representation learning with weak-form dynamics estimation and directly motivates our alternating algorithm: refine the VAE representation with the generator fixed, then estimate the generator in the updated coordinates. Our contribution lies in this Koopman formulation and dynamics-informed representation learning procedure, supported by theoretical guarantees for population-level estimator consistency and distribution prediction. Estimating dynamics in learned latent coordinates also supports scalability to 2,000-gene inputs; WCR only scales up to 20 dimensions. We have already covered these discussions in the revised related work section 5. [1]

AuthorsOpenReview ↗

Reply to Reviewer KWRr (3/7)

2. Weak-form estimation and related methods (Weakness 3; Question 1), continued.

  1. DisKO [3]. Both methods model population dynamics from unpaired snapshots. In KoopCell, the observables gθ(x,t)g_\theta(x,t) act on individual gene-expression states x∈RGx\in\mathbb{R}^{G}. DisKO instead treats an entire distribution μ∈P2(RD)\mu\in\mathcal{P}_2(\mathbb{R}^{D}) as the state, and learns distribution functionals hj(μ)h_j(\mu) (e.g., means and variances). The cell-wise VAE in KoopCell retains a distribution of individually encoded cells and yields a model of individual-cell dynamics. DisKO propagates one representation per population and does not specify individual-cell transitions. [3]

    At the estimation level, DisKO fits evolution between snapshot embeddings (obtained from a set encoder) through gradient-based learning or discrete regression; KoopCell estimates a continuous-time generator from weak equations over encoded cell populations. While convergence and stability of the learning procedure remain open questions in DisKO [3], our theory establishes consistency of the population weak-form estimator, convergence of feature propagators, and Wasserstein prediction bounds under suitable conditions (Appendix A).

Comparison between weak estimation and gradient updates. We replace weak-form estimation with joint gradient-based optimization of the same latent dynamics and representation on the hard task for all three datasets, keeping the architecture and remaining loss terms unchanged.

DataPredictionKoopCell: jointKoopCell: weakKoopCell-M: jointKoopCell-M: weak
ZBInterp.493.85±3.42493.85 \pm 3.42489.51±0.36\mathbf{489.51 \pm 0.36}494.61±4.06494.61 \pm 4.06488.77±2.92\mathbf{488.77 \pm 2.92}
Extrap.621.10±2.54\mathbf{621.10 \pm 2.54}622.36±1.14622.36 \pm 1.14615.45±1.06615.45 \pm 1.06614.26±0.19\mathbf{614.26 \pm 0.19}
DRInterp.469.80±1.32\mathbf{469.80 \pm 1.32}471.84±1.60471.84 \pm 1.60472.42±0.89472.42 \pm 0.89471.65±1.22\mathbf{471.65 \pm 1.22}
Extrap.609.57±6.96609.57 \pm 6.96605.76±2.22\mathbf{605.76 \pm 2.22}604.78±2.12604.78 \pm 2.12604.66±2.43\mathbf{604.66 \pm 2.43}
SCInterp.82.75±0.6182.75 \pm 0.6182.37±0.31\mathbf{82.37 \pm 0.31}82.19±0.6782.19 \pm 0.6781.81±0.16\mathbf{81.81 \pm 0.16}
Extrap.123.08±6.90123.08 \pm 6.90115.02±5.10\mathbf{115.02 \pm 5.10}115.24±4.73115.24 \pm 4.73110.34±0.81\mathbf{110.34 \pm 0.81}

Table 4. Hard-task OT error (lower is better), reported as mean ± SD.

KoopCell-M retains lower mean extrapolation error than KoopCell under joint training on all three datasets, suggesting that this advantage does not depend on the specific optimization method. Relative to these joint controls, weak-form estimation gives better empirical performance in most settings, with lower mean errors in 10 of 12 comparisons. Conceptually, it provides EDMD-style regression for learning the Koopman operator from unpaired distribution snapshots. [4]

We further use two spectral-recovery benchmarks to distinguish the roles of the weak loss and our alternating closed-form updates. With known observables, a single closed-form solve accurately recovers the spectrum; gradient updates require repeated epochs to reach comparable accuracy (Figure 12 in the revised manuscript). With unknown observables, we learn the representation and linear generator simultaneously, comparing alternating estimation against joint gradient updates with and without the weak loss (Figure 13 in the revised manuscript). Alternating estimation recovers the reference spectra faster and gives lower mean extrapolation errors at a matched epoch budget (500 epochs). Joint updates with the weak loss approach the reference spectral structure more slowly, whereas removing it fails to recover the Duffing oscillatory pair even after 2,000 epochs. These results show that alternating closed-form updates accelerate spectral recovery, while weak constraints support reliable recovery across both benchmarks.

Revised manuscript: Section 5 (WCR discussion); Appendices C.4 (Figures 12–13) and F.2 (Table 20).

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

This paper introduces KoopCell and KoopCell-M, a generative framework that adapts Koopman operator theory to learn single-cell population dynamics from unpaired temporal transcriptomic snapshots. By combining a variational autoencoder with the weak form of the continuity equation evaluated on random Fourier test functions, the authors derive a closed-form weighted ridge regression solver for the latent linear drift generator, enabling one-step arbitrary-horizon rollout via matrix exponentials without numerical ODE integration. To further capture cell-fate bifurcations beyond autonomous linear flows, KoopCell-M incorporates a Mori–Zwanzig-inspired auxiliary memory state together with a path-alignment loss over coarse developmental graphs, achieving strong interpolation and extrapolation performance across synthetic and real single-cell datasets.

Critical strengths
  1. Extending classical trajectory-level Koopman operator learning to unpaired population measures via the weak continuity equation is conceptually elegant and practically effective, as the closed-form generator update and matrix-exponential propagation completely eliminate the expensive numerical integration and adjoint backpropagation required by latent neural ODE methods.

  2. Constraining latent evolution to a linear semigroup provides strong global dynamical regularization that prevents velocity fields from diverging in unseen regions of state space, yielding consistent improvements on multi-step temporal extrapolation benchmarks across 2,000-dimensional real single-cell developmental datasets compared to unconstrained neural ODE and SDE baselines.

Critical weaknesses
  1. Because an autonomous linear flow in latent space preserves Gaussian unimodality and cannot spontaneously split a connected progenitor distribution into disjoint fate branches, KoopCell-M relies on a pretrained cell-state classifier and a manually specified directed acyclic graph of developmental lineages via the path-alignment loss (LpathL_{\text{path}}) to guide branching (Weinreb et al., 2020). This introduces a dependence on prior lineage annotations and cell-type labels, leaving open how effectively the memory variables alone could discover previously uncharacterized bifurcations in an unsupervised manner.

  2. In the alternating optimization scheme, the time integrals in the weak continuity equation are approximated between adjacent snapshots using a two-point trapezoidal rule. While accurate when temporal sampling is dense, this linear endpoint interpolation may introduce systematic discretization bias into the estimated generator when experimental snapshots are sparse and separated by rapid nonlinear transitions.

  3. Although the visible drift matrix and affine bias (Θ=[A,bz]\Theta = [A, b_z]) admit closed-form updates in Step (a), the auxiliary memory coupling matrices (B,C,ΛB, C, \Lambda) and branch initializations in KoopCell-M still require gradient-based optimization through all-pairs Sinkhorn divergences and dynamic-programming path alignment in Step (b). Reporting wall-clock training and inference times alongside neural ODE and SDE baselines would better quantify the practical computational trade-off between KoopCell and KoopCell-M.

  4. The theoretical guarantees in Section 3.1 rely on asymptotic test-function enrichment, whereas the practical implementation uses a fixed bank of m=2,048m=2,048 random Fourier features and an exponential moving average (α=0.99\alpha=0.99) to stabilize the alternating updates. Including an ablation on the number of random Fourier test functions and the EMA smoothing coefficient would clarify the empirical sensitivity of the closed-form solver.

Questions for authors

see weaknesses

5 comments
AuthorsOpenReview ↗

Reply to Reviewer tnBf (1/3)

We thank the reviewer for the careful review, positive assessment, and constructive suggestions. We have uploaded a revised manuscript containing additional results and figures, together with further discussion and clarifications. We kindly refer the reviewer to these revisions, especially the visualizations cited in the responses below. We address the four concerns below.

1. Branching without biological supervision. In the original branching toy experiment (Section 4.2 and Appendix D), KoopCell-M is trained only with the VAE negative ELBO and population matching losses. Without branch labels, a cell-state classifier, and DAG, it recovers the bifurcation from unpaired snapshots (Figure 2 in the revised manuscript). Furthermore, on the real datasets, we compare removing CTC path supervision alone with removing the broader biological guidance. The former retains cell-state supervision and the DAG; the latter removes these together and uses anonymous mixture components with the same fixed route count.

DataWithout CTC: ΔOT\Delta\mathrm{OT}Without biological guidance: ΔOT\Delta\mathrm{OT}
ZB+2.87±3.39+2.87 \pm 3.39+6.36±1.90+6.36 \pm 1.90
DR−2.46±1.82-2.46 \pm 1.82+2.33±4.29+2.33 \pm 4.29
SC+2.94±1.70+2.94 \pm 1.70+8.19±1.91+8.19 \pm 1.91

Table 1. Biological-supervision ablation.

Removing CTC alone has a modest, dataset-dependent effect. Removing the broader guidance increases mean extrapolation error, most clearly on ZB and SC; the change on DR is small.

Following Reviewer KWRr's suggestion, we also added lineage-based validation on LARRY [1], including a no-DAG variant. We kindly refer the reviewer to these additional results in the revised manuscript. DAG-based biological guidance does affect model performance, and reliable discovery of unknown lineage structure in complex data remains an important direction for future work.

Revised manuscript: Appendices E.6 (Table 16) and F.4 (Table 22); Section 6 (biological supervision).

2. Temporal sampling and quadrature bias. We agree that temporally sparse snapshots and rapid transitions can bias the trapezoidal weak-form estimate. To assess this sensitivity empirically, we repeat Toy I spectral recovery experiment with unknown observables and vary the number of training snapshots over the fixed horizon [0,1.2][0,1.2], using N∈2,3,5,7,9N\in\\{2,3,5,7,9\\} nested time points. The N=2N=2 setting uses only t=0t=0 and t=1.2t=1.2.

Number of SnapshotsSpectral MAE ↓Relative SWD at t=2.4t=2.4 (%) ↓Relative SWD at t=6t=6 (%) ↓
220.1991±0.12610.1991 \pm 0.12615.95±0.915.95 \pm 0.9119.87±9.2719.87 \pm 9.27
330.0223±0.00910.0223 \pm 0.00911.35±0.401.35 \pm 0.404.01±1.234.01 \pm 1.23
550.0109±0.00270.0109 \pm 0.00271.31±0.131.31 \pm 0.133.90±0.783.90 \pm 0.78
770.0097±0.00270.0097 \pm 0.00271.34±0.541.34 \pm 0.543.60±1.313.60 \pm 1.31
990.0071±0.00150.0071 \pm 0.00151.19±0.331.19 \pm 0.333.13±0.793.13 \pm 0.79

Table 2. Temporal sampling density in Toy I.

With learned observables, the 3–9-snapshot settings give comparable relative SWD at t=2.4t=2.4 (1.19–1.35%), while using only the endpoints gives 5.95%. At t=6t=6, the mean error decreases from 19.87% with two snapshots to 3.13% with nine. These results support good prediction on this simple, smoothly evolving system while confirming that sparse sampling increases error.

A related dependence on temporal discretization arises in JKO-based methods. JKOnet* [2] fits discrete JKO optimality conditions using OT couplings between successive populations. Its discretization differs from our quadrature, but it likewise approximates continuous dynamics using finite-time relations between snapshots. Such finite-time approximations can also introduce systematic discretization bias when snapshots are sparsely spaced, particularly across rapid transitions.

One possible improvement is to predict midpoint statistics with the current generator and use Simpson quadrature. Another is to learn a neural probability path constrained by the observed snapshots and weak dynamics, and evaluate the weak regression integrals along it. Path optimization and generator regression could alternate. Both are promising future directions for reducing discretization bias.

Revised manuscript: Appendix C.5 (Table 6; Figure 14); Section 6 (quadrature bias).

AuthorsOpenReview ↗

Reply to Reviewer tnBf (2/3)

3. Computational efficiency. We thank the reviewer for suggesting a runtime comparison. In KoopCell and KoopCell-M, linear latent dynamics allow direct propagation via matrix exponentials, so propagations across all forward time pairs can be evaluated in parallel using batched tensor operations, avoiding sequential numerical integration in neural ODE and SDE solvers. Encoding, decoding, and OT computations can also be parallelized. Based on these observations, we implemented parallel acceleration for both models and compared their training and inference times with scNODE and PI-SDE on the hard task of all three datasets.

DataMethodUpdatesTraining (min)Time per update (s)Inference (ms)
ZBKoopCell4,0004{,}0005.37\mathbf{5.37}0.0806\mathbf{0.0806}14.79\mathbf{14.79}
KoopCell-M4,0004{,}00019.9219.920.29890.298920.09‾\underline{20.09}
scNODE4,0004{,}0007.79‾\underline{7.79}0.1168‾\underline{0.1168}53.1053.10
PI-SDE3,0003{,}00074.5374.531.49071.49071274.161274.16
DRKoopCell4,0004{,}0004.27\mathbf{4.27}0.0640\mathbf{0.0640}15.55\mathbf{15.55}
KoopCell-M4,0004{,}00013.3013.300.19960.199620.53‾\underline{20.53}
scNODE4,0004{,}0006.45‾\underline{6.45}0.0968‾\underline{0.0968}55.8455.84
PI-SDE3,0003{,}00048.4148.410.96820.96821199.141199.14
SCKoopCell4,0004{,}0007.26\mathbf{7.26}0.1090\mathbf{0.1090}18.02\mathbf{18.02}
KoopCell-M4,0004{,}00022.2422.240.33350.333524.89‾\underline{24.89}
scNODE4,0004{,}00013.55‾\underline{13.55}0.2032‾\underline{0.2032}71.9271.92
PI-SDE3,0003{,}000236.82236.824.73644.73642359.102359.10

Table 3. Wall-clock runtime on NVIDIA Tesla V100 (32 GB). Inference generates 2,000 cells per held-out time point. Best and second-best times are bold and underlined, respectively.

KoopCell achieves the shortest training and inference times. Although KoopCell-M requires longer training, both variants provide faster inference than scNODE and are over 50×50\times faster than PI-SDE across all three datasets.

Revised manuscript: Section 4.3; Appendix E.7 (Table 17).

AuthorsOpenReview ↗

Reply to Reviewer tnBf (3/3)

4. Sensitivity to the EMA coefficient and the number of Fourier test functions. We evaluated KoopCell on the hard task for all three datasets, varying one factor at a time: α∈0,0.9,0.95,0.99\alpha\in\\{0,0.9,0.95,0.99\\} with m=2048m=2048, and m∈512,1024,2048,4096m\in\\{512,1024,2048,4096\\} with α=0.99\alpha=0.99. Each of the seven distinct settings was trained in three independent runs, with 500 pretraining epochs followed by 4,000 alternating-training epochs.

αZB Interp.ZB Extrap.DR Interp.DR Extrap.SC Interp.SC Extrap.
00521.02±3.03521.02 \pm 3.03664.03±22.49664.03 \pm 22.49472.60±4.29472.60 \pm 4.29627.05±7.78627.05 \pm 7.7886.59±2.4686.59 \pm 2.46132.76±7.31132.76 \pm 7.31
0.90.9487.27±0.68487.27 \pm 0.68622.39±0.27622.39 \pm 0.27473.59±0.64473.59 \pm 0.64620.14±13.02620.14 \pm 13.0282.05±0.6182.05 \pm 0.61113.25±0.82113.25 \pm 0.82
0.950.95489.47±2.47489.47 \pm 2.47620.46±5.10620.46 \pm 5.10471.26±0.82471.26 \pm 0.82615.21±11.68615.21 \pm 11.6882.30±0.4782.30 \pm 0.47113.40±2.66113.40 \pm 2.66
0.990.99488.92±1.98488.92 \pm 1.98623.63±3.40623.63 \pm 3.40472.62±1.64472.62 \pm 1.64604.04±2.19604.04 \pm 2.1982.23±0.5782.23 \pm 0.57113.77±3.83113.77 \pm 3.83

Table 4. EMA ablation with m=2048m=2048 fixed. Hard-task OT errors are reported separately for interpolation and extrapolation as mean ± sample standard deviation over three training runs. We first average the corresponding held-out time points within each run.

mZB Interp.ZB Extrap.DR Interp.DR Extrap.SC Interp.SC Extrap.
512512487.10±0.40487.10 \pm 0.40620.80±1.92620.80 \pm 1.92471.41±2.10471.41 \pm 2.10614.04±14.95614.04 \pm 14.9582.66±0.4582.66 \pm 0.45118.78±2.20118.78 \pm 2.20
10241024489.52±1.69489.52 \pm 1.69626.38±6.40626.38 \pm 6.40472.29±3.07472.29 \pm 3.07614.36±11.43614.36 \pm 11.4382.42±0.7382.42 \pm 0.73113.62±2.57113.62 \pm 2.57
20482048488.92±1.98488.92 \pm 1.98623.63±3.40623.63 \pm 3.40472.62±1.64472.62 \pm 1.64604.04±2.19604.04 \pm 2.1982.23±0.5782.23 \pm 0.57113.77±3.83113.77 \pm 3.83
40964096490.07±2.35490.07 \pm 2.35625.39±2.41625.39 \pm 2.41471.58±0.88471.58 \pm 0.88605.20±2.95605.20 \pm 2.9582.88±1.3082.88 \pm 1.30113.42±2.86113.42 \pm 2.86

Table 5. Fourier test-function ablation with α=0.99\alpha=0.99 fixed, using the same aggregation as Table 4. The shared default setting appears in both tables for comparison.

As shown in Table 4 and the EMA sensitivity plot in the revised manuscript (Figure 26), EMA smoothing generally improves prediction both within and beyond the training horizon and reduces variability across training seeds in most settings relative to α=0\alpha=0. These results highlight EMA as an important mechanism for stabilizing learning. At the same time, final predictive performance is relatively insensitive to the precise smoothing coefficient, with α∈0.9,0.95,0.99\alpha\in\\{0.9,0.95,0.99\\} yielding comparable results. We have also added training loss curves to the revised manuscript (Figure 28), showing that EMA markedly stabilizes training dynamics by reducing loss fluctuations and suppressing large spikes. Similarly, Table 5 and the Fourier sensitivity plot in the revised manuscript (Figure 27) show broadly consistent predictive performance across m∈512,1024,2048,4096m\in\\{512,1024,2048,4096\\}, indicating that the method is relatively insensitive to the number of Fourier test functions over the tested range.

Revised manuscript: Appendix F.1 (Tables 18–19; Figures 26–28).

References

[1] Weinreb et al. Lineage tracing on transcriptional landscapes links state to fate during differentiation. Science, 2020. Paper.

[2] Terpin et al. Learning diffusion at lightspeed. NeurIPS, 2024. Paper.

We remain open to further discussion and suggestions to improve the clarity and rigor of this work.

Reviewer_g483OpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper deals with learning dynamics of gene-expressions at the population level, where we have snapshots at different times and no way of knowing which cell pairs between timesteps (only population level information is available). The authors learn a latent space by a classic VAE reconstruction task of the gene vector + timestep (~2000 dimensional vector), and then assume a linear time evolution z˙=Az+b\dot{z} = Az + b which is fit via probing functions (Fourier features). Training alternates between fitting the VAE (w/ a Sinkhorn loss) and (A,b)(A,b) via regression. Authors also introduce extra latent variables from domain knowledge (annotations and labels), where the probabilistic graphical model is informed by their knowledge of biology.

Disclaimer: This paper is very out of distribution for me, and is very applied to biology. I have no prior experience with these datasets or models. I did not check the math extensively. I do not know if the baselines they compare to are proper/SOTA. I also don't know if the dowstream metrics are the best ones.

Critical strengths
  • Authors propose a solution with an unpaired population loss, without having to construct pairs for learning.
  • Rolling out the linear dynamics model is very cheap (as far as I understand), and learning is cheap as well (regression).
  • Extra hidden variables (i.e., KoopCell-M) is domain-specific and biologically informed.
Critical weaknesses
  • KoopCell-M requires supervised data.
  • Method does not end up giving cell trajectories and instead learns population statistics (essentially models the marginal distribution?)
Questions for authors
  • Can the authors comment on how hard it is to gather supervised data for the KoopCell memory?
  • Can the authors motivate better why time marginals are enough for applications?
4 comments
AuthorsOpenReview ↗

Reply to Reviewer g483 (1/2)

We thank the reviewer for the careful review, positive assessment, and constructive suggestions. We have uploaded a revised manuscript containing additional results and figures, together with further discussion and clarifications. We kindly refer the reviewer to these revisions, especially the visualizations cited in the responses below. We address the questions and concerns below.

1. Concerns about baselines, evaluation metrics and the stated performance. We refer the reviewer to scNODE [1] for the benchmark setting, evaluation metrics, and comparative results. Our study uses the same three datasets, interpolation/extrapolation tasks, and OT metric. We also include LGP-OT [2] (ICML 2026), one of the latest methods for this task, which reports state-of-the-art performance and serves as a strong baseline. Our models achieve superior mean extrapolation performance over the evaluated baselines (manuscript Table 1). While OT assesses population prediction; the independent lineage-based evaluation below additionally tests cell fate prediction beyond population prediction.

2. Computational efficiency. Thank you for highlighting this advantage. Linear latent dynamics allow matrix-exponential propagation and thus parallel computation across time pairs. We additionally compare training and inference times with neural ODE and SDE baselines to strengthen the evidence for this efficiency advantage:

DataKoopCellKoopCell-MscNODE (ODE)PI-SDE (SDE)
ZB5.37/14.79\mathbf{5.37} / \mathbf{14.79}19.92/20.09‾19.92 / \underline{20.09}7.79‾/53.10\underline{7.79} / 53.1074.53/1274.1674.53 / 1274.16
DR4.27/15.55\mathbf{4.27} / \mathbf{15.55}13.30/20.53‾13.30 / \underline{20.53}6.45‾/55.84\underline{6.45} / 55.8448.41/1199.1448.41 / 1199.14
SC7.26/18.02\mathbf{7.26} / \mathbf{18.02}22.24/24.89‾22.24 / \underline{24.89}13.55‾/71.92\underline{13.55} / 71.92236.82/2359.10236.82 / 2359.10

Table 1. Wall clock training time (minutes) / inference time (milliseconds) on one Tesla V100 (32 GB), hard tasks. Training updates: 4,000 for KoopCell, KoopCell-M, and scNODE, and 3,000 for PI-SDE. Inference generates 2,000 cells per held-out time. Best and second-best times are bold and underlined.

KoopCell inference is 3.6–4.0× faster than scNODE and 77–131× faster than PI-SDE; KoopCell-M achieves 2.6–2.9× and 58–95× speedups, respectively. KoopCell also reduces wall-clock training time; KoopCell-M requires slightly longer training than scNODE for its additional memory mechanism.

Revised manuscript: Section 4.3; Appendix E.7 (Table 17).

3. Biological relevance and validation beyond population prediction. KoopCell and KoopCell-M model continuous-time cell dynamics and can generate trajectories from individual initial cells. Since scRNA-seq is destructive, time-course datasets usually contain unpaired expression snapshots; the original prediction tables therefore evaluate the resulting time-marginal distributions using OT distance at held-out time points. In computational biology, reconstructing continuous dynamics from these snapshots supports downstream analyses such as inferring fate probabilities, identifying lineage-specific marker genes, and studying how gene regulation changes during differentiation. For example, our original ZB analysis identifies known mesodermal and neural markers and characterizes gene-expression programs associated with muscle and nervous-system development (Appendix E.5).

However, the three original datasets provide population snapshots without lineage-tracing information, allowing us to evaluate distribution prediction but not directly validate fate probabilities. We therefore added LARRY [3] to evaluate fate inference beyond marginal prediction, also addressing Reviewer KWRr's related concern. LARRY combines RNA sequencing with inherited barcodes during mouse blood progenitor differentiation. Cells sharing a barcode belong to the same clone, so we can compare fate predictions from early cells with the later fate composition of their clones.

AuthorsOpenReview ↗

Reply to Reviewer g483 (2/2)

3. Biological relevance and validation beyond population prediction, continued. On LARRY, we train on day 2/well 0 and days 4/6/well 1, withholding barcode relationships. For the 460 test clones (693 day-2 cells in total), we roll out 2,000 trajectories per day-2 cell to day 6. We estimate each cell's fate probabilities from the endpoint states using the same classifier fitted on training data for all methods. We then average these probabilities within each clone, normalize over mature fates, and compare them with the observed mature-fate proportions of that clone in well 2 (days 4/6).

MethodBias rr ↑Neu/Mono AUROC ↑Macro AUROC ↑Accuracy (%) ↑−CE ↑Brier ↓
scNODE0.275±0.1130.275 \pm 0.1130.653±0.0700.653 \pm 0.0700.519±0.0120.519 \pm 0.01243.12±3.2643.12 \pm 3.26−3.279±0.271-3.279 \pm 0.2710.903±0.0670.903 \pm 0.067
PRESCIENT (no growth)0.368±0.0100.368 \pm 0.0100.709±0.0050.709 \pm 0.0050.617±0.0060.617 \pm 0.00647.54±0.3347.54 \pm 0.33−3.087±0.151-3.087 \pm 0.1510.845±0.0260.845 \pm 0.026
PRESCIENT + KEGG0.465±0.011‾\underline{0.465 \pm 0.011}0.774±0.007‾\underline{0.774 \pm 0.007}0.641±0.0160.641 \pm 0.01656.01±0.98‾\underline{56.01 \pm 0.98}−2.685±0.044-2.685 \pm 0.0440.732±0.0110.732 \pm 0.011
KoopCell0.346±0.0780.346 \pm 0.0780.684±0.0400.684 \pm 0.0400.564±0.0390.564 \pm 0.03938.33±3.3738.33 \pm 3.37−2.485±0.210-2.485 \pm 0.2100.822±0.0430.822 \pm 0.043
KoopCell-M0.451±0.0200.451 \pm 0.0200.766±0.0100.766 \pm 0.0100.699±0.019\mathbf{0.699 \pm 0.019}57.75±2.54\mathbf{57.75 \pm 2.54}−1.512±0.040\mathbf{-1.512 \pm 0.040}0.601±0.014\mathbf{0.601 \pm 0.014}
KoopCell-M (no DAG)0.476±0.017\mathbf{0.476 \pm 0.017}0.776±0.007\mathbf{0.776 \pm 0.007}0.670±0.003‾\underline{0.670 \pm 0.003}51.96±3.3751.96 \pm 3.37−1.675±0.072‾\underline{-1.675 \pm 0.072}0.633±0.019‾\underline{0.633 \pm 0.019}

Table 2. LARRY mature clonal-fate prediction, mean ± sample SD over three training runs. Best and second-best means are bold and underlined. KEGG supplies an additional growth prior. The no-DAG model is trained without biological state labels, a developmental DAG, or barcode relationships.

The no-DAG variant improves all six mean scores over scNODE, no-growth PRESCIENT, and KoopCell, reaching a bias correlation of 0.476 and a Neu/Mono AUROC of 0.776. These lineage-based results support fate inference beyond population prediction, even without additional DAG supervision.

To show the continuous cell trajectories underlying these fate predictions, we have added Figure 24 to the revised manuscript. For each selected clone, we initialize KoopCell-M, scNODE, and PRESCIENT from the same day-2 cell and plot eight trajectories per model through day 6 in a shared UMAP, alongside the observed day-4/day-6 cells from that clone. These results further validate our model beyond population prediction.

Revised manuscript: Section 4.3 (Figure 4); Appendix E.6 (Table 16; Figure 24 and Figure 25).

4. Cost of obtaining memory supervision. The additional experimental burden is low for these benchmarks: supervision consists of coarse cell-state annotations and a small developmental DAG (7–11 states), without requiring measured cell trajectories or lineage-tracing experiments. We reuse published marker genes for ZB, atlas annotations for DR, and published reference cell sets for SC. A classifier fitted on training snapshots supplies state labels; established lineage knowledge and training-snapshot transport construct/refine the DAG. No additional sequencing is required. Reliable annotation and topology still require biological expertise, while the no-DAG LARRY results above show that informative fate prediction is possible without these biological supervision inputs.

Revised manuscript: Appendix B.2.1 (biological state annotation and candidate paths; Figure 6); Appendix E.6 (no-DAG results in Table 16).

References

[1] Zhang et al. scNODE: generative model for temporal single cell transcriptomic data prediction. Bioinformatics, 2024. Paper.

[2] Balik and Lähdesmäki. Modeling Temporal scRNA-seq Data with Latent Gaussian Process and Optimal Transport. ICML, 2026. Paper.

[3] Weinreb et al. Lineage tracing on transcriptional landscapes links state to fate during differentiation. Science, 2020. Paper.

We remain open to further discussion and suggestions to improve the clarity and rigor of this work.

Thank you for the clarifications to my questions. I understand now that having only the population-level statistic is a fundamental limitation of the field itself.

I've raised my score but I've kept my confidence low to reflect my lack of knowledge

Thank you again for your careful review and constructive discussion. We are pleased that our clarifications addressed your questions and sincerely appreciate your time and thoughtful reassessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.