acceptodds
Under review as a conference paper at ICLR 2027

Forecasting Cellular and Population Dynamics via Multi-Horizon Flow Matching

Abstract

Forecasting cellular populations from time-resolved single-cell RNA-seq requires predicting how cell states evolve and how these state-level changes reshape the population over time. We propose scHOMURA, a generative method for forecasting cellular and population dynamics from unpaired temporal snapshots. scHOMURA combines deterministic temporal displacement to capture shared state progression, conditional residual dynamics to model heterogeneous cellular outcomes, and a state-dependent growth field to model changes in relative population contribution. By directly learning transitions over multiple forecast horizons, scHOMURA reduces reliance on recursive one-step propagation and supports forecasting over both short and extended temporal intervals. Across five time-resolved single-cell datasets and seven forecasting splits, scHOMURA achieves strong distribution-, gene-expression-, and composition-level performance and remains stable as predictions extend beyond the observed temporal window. Beyond global population reconstruction, scHOMURA captures rare-state expansion, reaches future cellular states absent from the training period, and recovers developmental transitions consistent with known biological relationships. Together, scHOMURA provides a multi-horizon generative framework for forecasting cell-state dynamics and population remodeling from unpaired temporal snapshots.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell RNA sequencing, Cellular state forecasting, Conditional flow matching, Population dynamics, Emergent cellular states

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_oDV1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the challenging task of extrapolating future single-cell gene-expression distributions and population composition from unpaired temporal snapshots. It proposes scHOMURA, a generative framework that directly learns transitions across multiple forecast horizons, including non-adjacent snapshots, and combines multi-horizon predictions with partial recursion at inference to reduce reliance on repeated one-step forecasting. The method first clusters cells in a shared latent space and uses prototype-level unbalanced optimal transport to construct cell-level pseudo-pairs, reducing the computational cost of establishing correspondences. It then jointly learns deterministic mean displacements, stochastic residuals through conditional flow matching, and a state-dependent growth field that adjusts the relative contributions of source cells. Distribution-matching, cross-horizon consistency, and growth-supervision losses further support prediction of future distributions and population composition. Experiments across multiple time-resolved single-cell datasets show strong forecasting performance, particularly in Wasserstein distance, alongside favorable results in extended-horizon settings compared with the evaluated baselines. These results are encouraging for the challenging problem of forecasting cellular populations beyond the observed temporal window.

Critical strengths
  1. Clear methodological design combining multi-horizon transition learning with cellular heterogeneity modeling.
    The method constructs pseudo-pairs using prototype-level UOT across multiple forecast horizons, including non-adjacent snapshots, reducing reliance on transitions between adjacent time points. It decomposes future state changes into deterministic mean displacements and stochastic residuals generated through conditional flow matching, capturing shared progression and diverse outcomes conditional on the source state, respectively. This decomposition provides a clear framework for jointly modeling population evolution and cellular heterogeneity.

  2. Comprehensive experiments examining predictive capabilities and component contributions.
    The authors evaluate the method on five single-cell datasets with seven forecasting splits, assessing overall distributions, gene-expression changes, and population composition. Additional experiments investigate extended-horizon forecasting, rare-population expansion, and unseen-state recovery. The method achieves the lowest mean Wasserstein distance across all reported splits. Ablation studies further examine how the main components contribute across different settings and metrics, helping clarify the method’s benefits and limitations.

Critical weaknesses
  1. The model predicts relative population composition but not absolute cell numbers. Although the method uses unbalanced optimal transport and a growth field, the latter operates through normalized source-contribution weights rather than explicitly modeling mechanistic proliferation and death. Consequently, the model cannot distinguish changes in absolute population size from changes in relative composition. This limits its applicability to biological questions where absolute cell counts and population expansion or depletion are important outcomes, beyond the relative proportions of different cellular states.

  2. The ablation studies do not sufficiently isolate the sources of the reported improvements. On CD-long, removing multi-horizon learning barely changes the mean Wasserstein distance (16.35 versus 16.36), while removing growth modeling improves the mean values of all seven reported metrics, including Type-JS (0.079 to 0.061). In contrast, multi-horizon learning substantially improves abundance calibration on CD-short, increasing GrowthSkill from −1.329 to +0.172. These results suggest setting-dependent benefits rather than uniformly beneficial component contributions.

Questions for authors
  1. Why use UOT rather than balanced OT? In Appendix A.1.1, both prototype mass vectors are normalized, satisfying ∑iai=∑jbj=1\sum_i a_i=\sum_j b_j=1, so balanced OT is also applicable. Could the authors explain this choice and provide an experimental comparison with balanced OT?

  2. How are different forecast horizons sampled during training? For h=1,2,3h=1,2,3, are horizons sampled uniformly, are all eligible source–target snapshot pairs sampled uniformly, or is another scheme used? Please specify the proportion of training samples allocated to each horizon.

  3. How does the training cost compare with other methods? Although the FM regression loss is simulation-free, the MMD and consistency objectives require numerical integration. Please provide comparisons of training time and peak memory usage with other algorithms under comparable settings.

  4. Would consistency across different source snapshots improve forecasting? The current loss Lcons\mathcal{L}_{\mathrm{cons}} compares a direct two-step forecast with two composed one-step forecasts. Have the authors considered additionally matching the forecast distributions for the same target time from two observed source snapshots?

    \left( \widehat{Z}^{\mathrm{direct}}_{t\rightarrow t+2}, \widehat{Z}^{\mathrm{direct}}_{t+1\rightarrow t+2} \right)$$ Since inference mixes forecasts from different source times, would this constraint offer benefits beyond the existing distribution-matching and consistency losses?
8 comments
AuthorsOpenReview ↗

Relative versus absolute population dynamics

Response to Weakness1

Thank you for this insightful comment. We agree that distinguishing absolute population expansion or depletion from changes in relative composition is biologically important. However, inferring absolute population sizes from standard unpaired scRNA-seq snapshots faces a fundamental limitation: the total number of cells in the underlying biological system is generally not observed.

In standard scRNA-seq experiments, the number of profiled cells at each time point depends on experimental sampling, dissociation efficiency, and capture efficiency, rather than directly reflecting the total biological population size. For example, a developing tissue may undergo substantial proliferation without a corresponding increase in the number of sequenced cells. Without external measurements of total cell counts or calibrated sampling fractions, absolute population growth cannot be reliably distinguished from changes in sampling scale. Such information is unavailable in the public benchmarks considered in our study.

Accordingly, scHOMURA is designed to predict relative population composition rather than absolute cell numbers. Its state-dependent growth field adjusts the relative contributions of source cellular states to future distributions, capturing differential changes in population representation without explicitly estimating mechanistic proliferation or death rates. This distinction is important: the reported rare-population expansion metrics quantify changes in relative abundance, not absolute increases in cell numbers. Nevertheless, forecasting relative composition remains biologically informative for identifying populations that become more or less prevalent during development and differentiation.

We acknowledge that extending scHOMURA to absolute population dynamics would require additional information beyond uncalibrated scRNA-seq snapshots. Incorporating independent cell-count measurements or calibrated sampling fractions, together with explicit models of proliferation and death, represents an important direction for future work.

AuthorsOpenReview ↗

Setting-dependent contributions of model components

Response to Weakness2

Thank you for this insightful comment. We agree that the original ablation results show setting-dependent rather than uniformly positive contributions from each component. To clarify these findings, we conducted additional five-seed experiments across different levels of temporal supervision and targeted rare/emerging-population benchmarks. The results reveal complementary roles: multi-horizon learning provides clearer benefits with richer temporal supervision, residual dynamics facilitate the recovery of emerging states, and growth modeling improves relative abundance calibration.

Temporal supervision and multi-horizon learning. CD-long contains only three training snapshots (E9.5–E11.5), providing two one-step transitions, one two-step pair, and no three-step pairs. This limits the multi-step supervision available to the model and may also make long-range growth extrapolation less reliable, potentially explaining why removing growth modeling improves the aggregate metrics on CD-long (Table 4). To investigate the influence of temporal supervision, we introduced CD-medium with five training snapshots and compared it with CD-short, which has seven. As shown in Table R14-1, the full model achieves better aggregate performance across all seven metrics on CD-medium and CD-short, while providing only modest improvements on CD-long.

Table R14-1. Effect of multi-horizon learning under different levels of temporal supervision.

SplitMethodsW₂ ↓DE-AP ↑DE-dir ↑Δr ↑MMD ↓Variogram ↓Type JS ↓
CD-longFull16.348 ± 0.1690.298 ± 0.0390.768 ± 0.0350.369 ± 0.0410.132 ± 0.0150.023 ± 0.0010.079 ± 0.006
w/o multi-horizon16.358 ± 0.1570.306 ± 0.0150.768 ± 0.0730.365 ± 0.0690.141 ± 0.0110.024 ± 0.0030.083 ± 0.013
CD-mediumFull17.163 ± 0.2160.297 ± 0.0560.610 ± 0.0660.130 ± 0.1310.123 ± 0.0070.024 ± 0.0010.171 ± 0.021
w/o multi-horizon17.426 ± 0.3020.258 ± 0.0290.545 ± 0.0320.049 ± 0.0570.148 ± 0.0110.027 ± 0.0040.197 ± 0.013
CD-shortFull15.091 ± 0.1570.263 ± 0.0850.575 ± 0.1180.261 ± 0.0590.044 ± 0.0080.014 ± 0.0010.013 ± 0.005
w/o multi-horizon15.482 ± 0.3440.235 ± 0.0630.478 ± 0.0840.181 ± 0.1030.048 ± 0.0090.016 ± 0.0020.022 ± 0.012

To further assess whether these improvements are consistent across random seeds, we computed paired W2W_2 differences using five matched seeds (Table R14-2). The mean improvement increases from 0.010 on CD-long to 0.262 on CD-medium and 0.391 on CD-short. Importantly, on CR-long, the full model outperforms the ablation in all five seeds, with a mean improvement of 0.276. These results provide stronger evidence for the setting-dependent benefits of multi-horizon learning, while confirming its limited effect on CD-long.

**Table R14-2. Paired-seed analysis of multi-horizon learning across forecasting splits.**Paired improvement is defined as W2(w/o multi-horizon)−W2(Full)W_2(\mathrm{w/o\ multi\text{-}horizon}) - W_2(\mathrm{Full}). Confidence intervals are nominal 95% paired tt intervals.

SplitFull W₂ ↓w/o multi-horizon W₂ ↓Paired improvement ↑95% paired CIFull better
CD-long16.348 ± 0.16916.358 ± 0.157+0.010[−0.178, +0.199]3/5
CD-medium17.163 ± 0.21617.426 ± 0.302+0.262[+0.009, +0.515]4/5
CD-short15.091 ± 0.15715.482 ± 0.344+0.391[−0.006, +0.787]4/5
CR-long17.796 ± 0.24218.072 ± 0.169+0.276[+0.054, +0.497]5/5

Rare-population abundance calibration. Multi-horizon learning also improves the calibration of rare-population expansion on CD-short (Table R14-3). Although removing multi-horizon learning improves expansion ranking and direction accuracy, the full model nearly halves abundance-change error and improves GrowthSkill from negative to positive. The immune-cell predictions provide a concrete example (Table R14-4): both variants anticipate an increasing trend, but the single-horizon model substantially overestimates future immune-cell abundance, whereas the full model produces predictions closer to the observed proportions. These results highlight the distinction between identifying expanding populations and accurately predicting their expansion magnitudes.

Table R14-3. Effect of multi-horizon learning on rare-population expansion in CD-short.

MethodsExpn. AP ↑Dir. ↑ΔMAE ↓GrowthSkill ↑
Full0.800 ± 0.2090.650 ± 0.1370.015 ± 0.0030.172 ± 0.251
w/o multi-horizon0.967 ± 0.0750.750 ± 0.0000.029 ± 0.012−0.667 ± 0.723

The remaining results of this response are provided in the next part due to the word limit.

AuthorsOpenReview ↗

Setting-dependent contributions of model components(Part 2)

Response to Weakness2(Part 2)

Table R14-4. Comparison of predicted and observed immune-cell proportions at future developmental stages in CD-short.

Target timeTrue immune-cell proportionFullw/o multi-horizon
E16.54.19%5.80%7.69%
E17.57.42%8.26%12.27%

Contributions of residual dynamics and growth modeling. To further distinguish their roles, we evaluated the ablated variants directly on rare-state expansion and natural emergence benchmarks (Table R14-5). On CD-short, removing growth modeling substantially degrades expansion direction accuracy and abundance calibration. In contrast, removing residual flow has relatively minor effects on abundance calibration but eliminates accepted emerging-state predictions on BetaDiff. Removing growth modeling also reduces accepted emerging-cell mass from 0.064 to 0.023. These findings suggest that residual dynamics are particularly important for reaching novel cellular states, whereas growth modeling regulates their relative representation in the predicted population.

Table R14-5. Contributions of residual dynamics and growth modeling to rare-state expansion and natural emergence. The first four metrics evaluate rare-state expansion on CD-short, while the last three evaluate natural emergence on BetaDiff.

MethodsExpn. AP ↑Dir. ↑ΔMAE ↓GrowthSkill ↑Accepted ↑Unsupported ↓Precision ↑
Full0.800 ± 0.2090.650 ± 0.1370.015 ± 0.0030.172 ± 0.2510.064 ± 0.0280.043 ± 0.0170.594 ± 0.047
w/o residual flow0.883 ± 0.1830.700 ± 0.1120.016 ± 0.0050.107 ± 0.2720.000 ± 0.0000.000 ± 0.000—
w/o growth modeling0.617 ± 0.1260.250 ± 0.0000.027 ± 0.002−0.511 ± 0.1870.023 ± 0.0130.014 ± 0.0100.636 ± 0.101
Persistence0.633 ± 0.2090.450 ± 0.1120.018 ± 0.0010.000 ± 0.0000.041 ± 0.0010.055 ± 0.0030.427 ± 0.010
Noise-halo0.683 ± 0.1370.450 ± 0.1120.018 ± 0.002−0.001 ± 0.0060.101 ± 0.0060.107 ± 0.0080.485 ± 0.029

Controlled masked emergence (BetaDiff). The masked-emergence experiments further illustrate the complementary contributions of these components (Tables R14-6 and R14-7). For the rare sst_hhex population, removing residual flow eliminates detection entirely, while removing growth modeling reduces detection to 2/10 and predicted abundance from 2.20% to 0.30%. For sc_alpha, which follows a more densely observed developmental trajectory, both ablated variants still permit frequent detection. However, removing growth modeling substantially reduces predicted abundance. These results indicate that residual dynamics are particularly important for recovering poorly supported future states, while growth modeling improves their representation even when the target state is already reachable.

Table R14-6. Contributions of residual dynamics and growth modeling to masked emergence recovery of the rare sst_hhex population on BetaDiff.

MethodsDetected ↑Predicted FractionConditional W₂ ↓OOS-precision ↑
Full7/102.20%16.30.036
w/o residual flow0/100.00%——
w/o growth modeling2/100.30%17.50.022
Persistence0/100.10%——
Noise-halo0/100.10%—0

Table R14-7. Contributions of residual dynamics and growth modeling to masked emergence recovery of the sc_alpha population on BetaDiff.

MethodsDetected ↑Predicted FractionConditional W₂ ↓OOS-precision ↑
Full10/1012.10%23.30.477
w/o residual flow8/105.60%22.6—
w/o growth modeling10/103.70%23.50.724
Persistence1/100.40%23.7—
Noise-halo10/102.90%27.30.071

Together, these targeted analyses clarify why the components show setting-dependent contributions rather than uniformly improving every aggregate metric. Multi-horizon learning provides clearer benefits under richer temporal supervision and improves rare-population abundance calibration. Conditional residual dynamics facilitate recovery of novel or poorly supported future states, whereas the growth field adjusts their relative population contributions, with potential limitations under sparse temporal supervision. We have incorporated these results into the new full PDF.

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

The authors propose a new flow matching framework for scRNA-seq data. Their motivation is that the existing approaches are limited in two ways: reliance on recursive one-step propagation, which accumulates error, and a limited ability to anticipate rare or emerging populations. To address these limitations, they propose scHOMURA, which learns multi-horizon transitions instead of single-step transitions, and they argue that this improves stability over extended horizons. Moreover, their method is able to learn a state-dependent growth field.

Critical strengths
  • The idea of relying on direct multi-horizon transitions rather than only a single-step rollout sounds promising.
  • Over diverse datasets, the proposed model achieves strong performance against baselines across various metrics.
  • The proposed model is able to infer state-dependent population growth.
Critical weaknesses
  • The ablation experiments do not consistently support the paper's claims. The reported effects in Table 4 and S6 are mixed and often within one standard deviation. The most important finding is that the removal of multi-horizon learning (which is one of the main claims of the paper) has basically no effect on CD-long (16.35±0.1716.35 \pm 0.17 vs 16.36±0.1616.36 \pm 0.16) and a minor effect on CR-long (17.80±0.2417.80 \pm 0.24 vs 18.07±0.1718.07 \pm 0.17) based on W2W_2 metric. The finding in Section 4.3 is therefore not clearly attributable to multi-horizon learning. The authors should discuss these inconsistencies.

  • The paper does not report training/inference time for scHOMURA and baselines. Since single-cell datasets can be large and the method involves several stages, a brief discussion about the runtime and scalability would strengthen the paper.

Questions for authors
  • The original baseline methods differ in the input space they were built for: some were developed for PCA-reduced representations, while others can operate directly on the full 2,000-gene expression space. Could the authors specify which representation each baseline was trained on in this study?

  • Additionally, was any hyperparameter tuning performed, either for the baselines or for the proposed model? If so, please describe the search space and the validation protocol used to select hyperparameters.

  • Could the authors also provide Figure 2-like comparison for the full scHOMURA and the abblation variant w/o multi-horizon?

  • The model combines the last 3 candidates with fixed mixture weights (0.6/0.3/0.1). Could the authors please specify how they choose these specific weights and provide a sensitivity analysis over alternative weightings, such as from uniform to 1/0/0?

9 comments
AuthorsOpenReview ↗

Contribution of multi-horizon learning to long-horizon forecasting(Part 1)

Response to Weakness1

Thank you for highlighting this concern. We agree that the original ablations do not establish multi-horizon learning as the primary source of the full model's long-horizon advantage. To clarify its contribution, we conducted additional experiments examining horizon-specific performance and paired-seed consistency. Our results show that multi-horizon learning provides clearer benefits with richer temporal supervision, while also enhancing rare-population abundance calibration.

Forecasting across horizons. On CD-long, the full model and its single-horizon variant perform similarly across the seven held-out stages (Figure S3). A plausible explanation is limited temporal supervision: CD-long provides just one two-step training pair and no three-step pairs. To investigate whether richer temporal supervision makes multi-horizon learning more effective, we introduced CD-medium with five training snapshots and CD-short, which has seven. As shown in Table R7-1 and Figure S3, the full model achieves lower mean W2W_2 at every held-out stage on both CD-medium and CD-short, together with better aggregate performance across all seven evaluation metrics. These results suggest that the benefits of multi-horizon learning become more evident as additional cross-time supervision becomes available.

Table R7-1. Effect of multi-horizon learning under different levels of temporal supervision.

SplitMethodsW₂ ↓DE-AP ↑DE-dir ↑Δr ↑MMD ↓Variogram ↓Type JS ↓
CD-longFull16.348 ± 0.1690.298 ± 0.0390.768 ± 0.0350.369 ± 0.0410.132 ± 0.0150.023 ± 0.0010.079 ± 0.006
w/o multi-horizon16.358 ± 0.1570.306 ± 0.0150.768 ± 0.0730.365 ± 0.0690.141 ± 0.0110.024 ± 0.0030.083 ± 0.013
CD-mediumFull17.163 ± 0.2160.297 ± 0.0560.610 ± 0.0660.130 ± 0.1310.123 ± 0.0070.024 ± 0.0010.171 ± 0.021
w/o multi-horizon17.426 ± 0.3020.258 ± 0.0290.545 ± 0.0320.049 ± 0.0570.148 ± 0.0110.027 ± 0.0040.197 ± 0.013
CD-shortFull15.091 ± 0.1570.263 ± 0.0850.575 ± 0.1180.261 ± 0.0590.044 ± 0.0080.014 ± 0.0010.013 ± 0.005
w/o multi-horizon15.482 ± 0.3440.235 ± 0.0630.478 ± 0.0840.181 ± 0.1030.048 ± 0.0090.016 ± 0.0020.022 ± 0.012

Consistency across random seeds. To further directly address the reviewer's concern that the ablation effects in Tables 4 and S6 are mixed, we computed paired W2W_2 differences between the full model and its single-horizon variant (Table R7-2). On CD-medium, the full model outperforms the ablation in four of five seeds. On CD-short, it also performs better in four of five seeds, although the confidence interval marginally includes zero. In contrast, CD-long shows no reliable improvement. On CR-long, the other extended-horizon split highlighted in the review, the full model outperforms the ablation in all five seeds, with a mean improvement of 0.276.

Table R7-2. Paired-seed analysis of multi-horizon learning across forecasting splits. Paired improvement is defined as W2(w/o multi-horizon)−W2(Full)W_2(\mathrm{w/o\ multi\text{-}horizon}) - W_2(\mathrm{Full}). Confidence intervals are nominal 95% paired tt intervals.

SplitFull W₂ ↓w/o multi-horizon W₂ ↓Paired improvement ↑95% paired CIFull better
CD-long16.348 ± 0.16916.358 ± 0.157+0.010[−0.178, +0.199]3/5
CD-medium17.163 ± 0.21617.426 ± 0.302+0.262[+0.009, +0.515]4/5
CD-short15.091 ± 0.15715.482 ± 0.344+0.391[−0.006, +0.787]4/5
CR-long17.796 ± 0.24218.072 ± 0.169+0.276[+0.054, +0.497]5/5

Isolating the training-horizon contribution. To further determine whether the observed benefits arise from multi-horizon training itself rather than the multi-source inference strategy, we conducted a controlled ablation of the maximum direct training horizon Hmax⁡∈{1,2,3,4}H_{\max}\in\{1,2,3,4\}. All models were evaluated using the same one-step recursive inference. As shown in Table R7-3, increasing Hmax⁡H_{\max} from 1 to 3 reduces mean W2W_2 on CR-long and CD-medium. These results support Hmax⁡=3H_{\max}=3 as a practical default.

Table R7-3. Effect of the maximum direct training horizon on long-range forecasting accuracy. Comparison of mean W2W_2 across different maximum direct training horizons.

SplitHmax⁡=1H_{\max}=1Hmax⁡=2H_{\max}=2Hmax⁡=3H_{\max}=3Hmax⁡=4H_{\max}=4
CD-short15.480 ± 0.32915.736 ± 0.37215.440 ± 0.15515.384 ± 0.092
CD-medium17.542 ± 0.28117.468 ± 0.22117.219 ± 0.21217.246 ± 0.125
CR-long18.065 ± 0.17017.929 ± 0.29217.767 ± 0.24517.881 ± 0.113

The remaining results of this response are provided in the next part due to the word limit.

AuthorsOpenReview ↗

Contribution of multi-horizon learning to long-horizon forecasting(Part 2) & Runtime and scalability analysis

Response to Weakness1(Part 2)

Rare-population expansion. Beyond global forecasting, multi-horizon learning also improves rare-population abundance calibration on CD-short (Table R7-4). The full model nearly halves abundance-change error and improves GrowthSkill from negative to positive. Although removing multi-horizon learning improves expansion ranking and direction accuracy, it substantially worsens the prediction of expansion magnitudes.

Table R7-4. Effect of multi-horizon learning on rare-population expansion in CD-short.

MethodsExpn. AP ↑Dir. ↑ΔMAE ↓GrowthSkill ↑
Full0.800 ± 0.2090.650 ± 0.1370.015 ± 0.003+0.172 ± 0.251
w/o multi-horizon0.967 ± 0.0750.750 ± 0.0000.029 ± 0.012−0.667 ± 0.723

The immune-cell predictions provide a concrete example (Table R7-5). Both variants correctly anticipate an increasing trend, but the single-horizon model substantially overestimates immune-cell abundance at E16.5 and E17.5. In contrast, the full model predicts proportions, respectively, much closer to the observed values. This further illustrates how multi-horizon learning improves the calibration of future population expansion.

Table R7-5. Comparison of predicted and observed immune-cell proportions at future developmental stages in CD-short.

Target timeTrue immune-cell proportionFullw/o multi-horizon
E16.54.19%5.80%7.69%
E17.57.42%8.26%12.27%

Overall, these results suggest that multi-horizon learning improves forecasting accuracy and rare-population abundance calibration, with its benefits depending on the availability of temporal supervision. In addition, multi-horizon learning improves rare-population abundance calibration. We have incorporated these additional analyses into the revised manuscript.

Response to Weakness2

Thank you for this helpful suggestion. We conducted additional runtime and memory benchmarks covering scHOMURA and all five baselines across three datasets of increasing scale. These benchmarks include the complete multi-stage training workflow, GPU memory usage, optimization trajectories, and inference latency. The results show that scHOMURA maintains practical training costs and relatively low GPU memory usage across the evaluated datasets, with sub-second inference on CD-long.

Experimental setup. We benchmarked all six methods on CD-long (4,077 training cells across 3 time points), BetaDiff (14,942 cells across 2 time points), and Zebrafish (32,713 cells across 10 time points). Within each dataset, all methods used the same training snapshots, forecasting targets, and 2,000-gene feature set. Experiments were run sequentially on a single host with 8 CPU threads and one NVIDIA A800 80 GB GPU. We followed each baseline's recommended pretraining and main-training schedules and native early-stopping criteria, recording the actual completed epochs or iterations. In accordance with their default settings, scNODE and MIOFlow were executed on CPU.

Training efficiency and memory usage. As shown in Figure S4, scHOMURA maintains relatively low end-to-end training times and GPU memory usage across all three datasets. Our measurements include method-specific preprocessing and all training stages, including autoencoder training, prototype-level transport construction, and temporal-model optimization for scHOMURA. Although training times vary across methods due to their different architectures, training schedules, and execution devices, scHOMURA remains computationally practical even on Zebrafish, the largest evaluated dataset with 32,713 training cells.

Training convergence. Figure S5 further presents the training loss trajectories of all six methods across the three datasets. The curves illustrate the optimization behavior under each method's native objective and training schedule.

Inference efficiency. We additionally measured inference time on CD-long by generating 2,000 cells per target snapshot. As shown in Table R8-1, while the CPU-based scNODE and MIOFlow achieve shorter inference times, scHOMURA remains efficient while incorporating multi-horizon dynamics and state-dependent growth modeling.

Table R8-1. Inference time comparison on CD-long. End-to-end inference time for generating 2,000 cells per target snapshot.

MethodDeviceInference time (s) ↓
scHOMURAGPU0.236 ± 0.002
LGP-OTGPU0.311 ± 0.003
scNODECPU0.027 ± 0.001
MIOFlowCPU0.141 ± 0.035
PRESCIENTGPU0.324 ± 0.013
VGFMGPU2.784 ± 0.058

Overall, these benchmarks demonstrate that scHOMURA maintains practical computational requirements across datasets of different scales. By incorporating all training stages and reporting both resource usage and inference latency, our analysis provides a complete assessment of computational efficiency.

AuthorsOpenReview ↗

Representation space & Sensitivity analysis of loss weights and prototype count(part 1)

Response to Question1

Thank you for highlighting this ambiguity. All methods use the same set of 2,000 highly variable genes (HVGs) within each split, but differ in their internal representations and dynamics modeling spaces. Specifically, LGP-OT and scNODE receive gene-expression inputs and model dynamics in learned 32- and 50-dimensional latent spaces, respectively. MIOFlow first applies 50-dimensional PCA and subsequently learns a 10-dimensional geometric autoencoder representation. PRESCIENT operates in a standardized 50-dimensional PCA space, whereas VGFM directly models dynamics in the original 2,000-dimensional expression space. scHOMURA uses a pretrained, frozen 32-dimensional autoencoder representation.As summarized in Table R9-1, all methods are evaluated in the same 2,000-gene expression space.

Table R9-1. Input representations and dynamics modeling spaces used by scHOMURA and baseline methods.

MethodInput representation and dimensionality reductionDynamics space
LGP-OTExpression of 2,000 HVGs; no additional PCA32-dimensional GP latent space
scNODEExpression of 2,000 HVGs; representation learned by its encoder50-dimensional latent space
MIOFlowExpression reduced to 50 PCs, followed by a geometric autoencoder (GAE)10-dimensional GAE latent space
PRESCIENTStandardized expression reduced to 50 PCs50-dimensional PCA space
VGFMExpression of 2,000 HVGs directly; no external PCA2,000-dimensional expression space
scHOMURAExpression of 2,000 HVGs; pretrained and frozen autoencoder32-dimensional latent space

Response to Question2(Part 1)

Thank you for raising this important question. For the baselines, we adopted established hyperparameter configurations from previous studies and official benchmark implementations. For scHOMURA, we used fixed default hyperparameters across datasets and conducted additional sensitivity analyses to evaluate their robustness. The results show that scHOMURA is broadly stable across moderate variations in loss weights, while prototype count introduces clearer trade-offs between distributional accuracy, and emerging-state recovery.

Baseline configurations. For datasets shared with previous studies, we adopted the published or recommended hyperparameter configurations. Specifically, scNODE, MIOFlow, and PRESCIENT followed the configurations provided by the public benchmark implementation, while LGP-OT followed its documented adaptations. For additional datasets without dedicated configurations, we reused the corresponding benchmark settings. VGFM was evaluated using a fixed high-dimensional single-cell configuration and training budget.

scHOMURA hyperparameters. The default loss weights (λmean=5\lambda_{\mathrm{mean}}=5, λMMD=0.05\lambda_{\mathrm{MMD}}=0.05, and λcons=0.05\lambda_{\mathrm{cons}}=0.05) and prototype count (P=64P=64) were selected to balance the numerical scales of the regression and distributional regularizers while keeping transport-matrix computation lightweight. These settings were held fixed across datasets without per-split tuning.We assessed sensitivity by varying each loss weight by factors of 0.5× and 2× and testing P∈{32,64,128}P\in\{32,64,128\} using two random seeds (11 and 23), evaluating both global forecasting and rare/emerging population benchmarks.

Global forecasting performance. As shown in Tables R10-1 and R10-2, performance remains broadly stable across the tested loss weights on CD-short and CD-long, with some metric-specific variations. Prototype count has a more pronounced effect: P=32P=32 improves Type JS on CD-long but degrades several distribution- and expression-level metrics. Overall, loss weights show relatively low sensitivity, whereas prototype count introduces trade-offs among forecasting objectives.

Table R10-1. Sensitivity of global forecasting performance to loss weights and prototype count on CD-short.

ConfigurationW₂ ↓DE-AP ↑DE-dir ↑Δ Pearson ↑MMD ↓Variogram ↓Type JS ↓
Default15.0680.3420.6720.2970.04270.015650.00859
λmean=2.5\lambda_{\mathrm{mean}}=2.515.0680.3390.6720.2950.04270.015600.00877
λmean=10\lambda_{\mathrm{mean}}=1015.0720.3430.6720.2920.04290.015630.00885
λMMD=0.025\lambda_{\mathrm{MMD}}=0.02515.0970.3400.6600.2770.04430.015560.01048
λMMD=0.1\lambda_{\mathrm{MMD}}=0.115.0590.3320.6760.3080.04150.015710.00764
λcons=0.025\lambda_{\mathrm{cons}}=0.02515.0680.3450.6760.2940.04240.015650.00857
λcons=0.1\lambda_{\mathrm{cons}}=0.115.0740.3400.6720.2950.04310.015540.00948
P=32P=3215.3840.2940.6240.3160.04540.017330.00653
P=128P=12815.2150.3570.7020.3560.04350.014750.00631

The remaining results of this response are provided in the next part due to the word limit.

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

This paper investigates the prediction of cell-state distributions and population dynamics from unpaired temporal scRNA-seq snapshots. It focuses on prediction stability over extended forecast horizons and the prediction of rare and emerging populations. The authors propose scHOMURA, which models these changes in a latent space. It uses a deterministic displacement to capture the main change in cell state and conditional residual dynamics to capture different possible outcomes. A state-dependent growth field adjusts how much each source state contributes to the future population. The model also learns transitions directly over multiple forecast horizons. The authors evaluate scHOMURA on five datasets and seven forecasting splits. Their evaluation distinguishes between identifying populations that will expand, predicting their future proportions, and recovering cell states that emerge later. Across five time-resolved single-cell datasets and seven forecasting splits, the authors evaluate future-population, extended-horizon, and rare/emerging-population forecasting. The results suggest competitive performance across multiple evaluation metrics.

Critical strengths
  1. The evaluation covers future-population, extended-horizon, and rare/emerging-population forecasting, providing a more comprehensive assessment of model performance and potential failure modes.
  2. The framework jointly models shared temporal progression, heterogeneous cellular outcomes, and state-dependent population contributions. Experiments show improvements across multiple datasets and forecasting splits.
Critical weaknesses
  1. The contribution of multi-horizon learning to long-horizon forecasting needs clearer evidence. Although the paper evaluates longer forecast horizons, Table 4 reports nearly identical mean W2 values on CD-long for the full model and the variant without multi-horizon learning: 16.35±0.17 and 16.36±0.16, respectively. Horizon-specific results would help clarify whether multi-horizon learning improves predictions at later time points.
  2. The ablation results in Table 4 and S6 do not clearly establish the contributions of conditional residual dynamics and the growth field to emerging-population recovery. The reported metrics may primarily reflect improvements in common populations rather than emerging populations. A more targeted analysis would help clarify the roles of these components.
Questions for authors
  1. Could the authors show prediction error curves across forecast horizons, with and without multi-horizon learning? For splits with sufficient training snapshots, how does the maximum horizon used in direct training affect long-range performance?
  2. Would it be possible to evaluate variants without residual dynamics and without the growth field using the emergence metrics in Tables 2 and 3, and compare them with the full model?
  3. How do the authors explain explain why removing growth modeling improves all seven reported mean metrics on CD-long, and why removing residual flow improves several metrics on CD-short (Table 4)?
  4. How were λ_mean=5, λ_MMD=0.05, and λ_cons=0.05 selected? Could the authors test a small range of loss weights and prototype counts on representative splits, with particular attention to the prediction of rare and emerging populations?
7 comments
AuthorsOpenReview ↗

Through five-seed experiments on CD-long/medium/short, we demonstrate that multi-horizon learning is generally superior when temporal supervision is sufficient, yields limited benefits on CD-long due to fewer training snapshots, and improves the prediction of rare-population expansion.

Response to Weakness 1

Thank you for raising this important point. We agree that the nearly identical W2W_2 values on CD-long show that our current ablations do not establish a standalone benefit of multi-horizon learning for long-range forecasting. Accordingly, we have revised Section 4.3 in the updated manuscript to attribute extended-horizon stability to the full scHOMURA framework rather than to multi-horizon learning alone.

At the same time, our five-seed experiments clarify two specific roles where multi-horizon learning provides clear and consistent gains:

1. Improving multi-step forecasting accuracy when sufficient training time points are observed. As shown in Table R1-1 (full curves in Figure S3), on CD-long, the full model and its single-horizon variant perform similarly across the seven held-out stages without a consistent advantage at later horizons. This is expected because CD-long contains only three training snapshots, providing just one two-step training pair and no three-step pairs. When we evaluate splits with more observed training time points, namely CD-medium (five training snapshots) and CD-short (seven training snapshots), multi-horizon learning improves performance across all seven global metrics, including W2W_2.

Table R1-1. Effect of multi-horizon learning under different levels of temporal supervision.

DatasetMethodsW2W_2 ↓DE-AP ↑DE-dir ↑Δr ↑MMD ↓Variogram ↓Type JS ↓
CD-longFull16.348 ± 0.1690.298 ± 0.0390.768 ± 0.0350.369 ± 0.0410.132 ± 0.0150.023 ± 0.0010.079 ± 0.006
w/o multi-horizon16.358 ± 0.1570.306 ± 0.0150.768 ± 0.0730.365 ± 0.0690.141 ± 0.0110.024 ± 0.0030.083 ± 0.013
CD-mediumFull17.163 ± 0.2160.297 ± 0.0560.610 ± 0.0660.130 ± 0.1310.123 ± 0.0070.024 ± 0.0010.171 ± 0.021
w/o multi-horizon17.426 ± 0.3020.258 ± 0.0290.545 ± 0.0320.049 ± 0.0570.148 ± 0.0110.027 ± 0.0040.197 ± 0.013
CD-shortFull15.091 ± 0.1570.263 ± 0.0850.575 ± 0.1180.261 ± 0.0590.044 ± 0.0080.014 ± 0.0010.013 ± 0.005
w/o multi-horizon15.482 ± 0.3440.235 ± 0.0630.478 ± 0.0840.181 ± 0.1030.048 ± 0.0090.016 ± 0.0020.022 ± 0.012

2. Calibrating expansion abundance for rare populations. On CD-short (Table R1-2), removing multi-horizon learning preserves expansion ranking and direction prediction, but substantially degrades abundance calibration. The full model reduces ΔMAE by nearly half (from 0.029 to 0.015) and improves GrowthSkill from −0.667 to +0.172, outperforming the persistence baseline. The immune-cell trajectories provide a concrete example (Table R1-3). Both variants anticipate an upward trend, but the single-horizon variant substantially overestimates immune-cell proportions at E16.5 and E17.5, whereas the full model stays much closer to the observed fractions.

Table R1-2. Effect of multi-horizon learning on rare-population expansion in CD-short.

MethodsExpn. AP ↑Dir. ↑ΔMAE ↓GrowthSkill ↑
Full0.800 ± 0.2090.650 ± 0.1370.015 ± 0.003+0.172 ± 0.251
w/o multi-horizon0.967 ± 0.0750.750 ± 0.0000.029 ± 0.012−0.667 ± 0.723

Table R1-3. Comparison of predicted and observed immune-cell proportions at future developmental stages in CD-short.

Target timeTrue immune-cell proportionFullw/o multi-horizon
E16.54.19%5.80%7.69%
E17.57.42%8.26%12.27%

In summary, while our ablations do not demonstrate a standalone long-range advantage on sparse splits like CD-long, they confirm that multi-horizon learning calibrates rare-population expansion magnitudes and improves overall forecasting accuracy when sufficient training snapshots are available. We have added these results to Appendix A.3 (Tables S2 and S3, Figure S3) and updated Sections 4.3 and 4.6 in the revised manuscript to reflect these findings accurately.

AuthorsOpenReview ↗

Ablation experiments on the targeted benchmarks show that conditional residual dynamics are primarily responsible for reaching rare or poorly supported future states, whereas the growth field mainly regulates the representation of these states in the predicted population; the two are complementary and both indispensable.

Response to Weakness2

Thank you for this helpful suggestion. We agree that aggregate metrics alone cannot fully establish how individual components contribute to rare and emerging population recovery. We therefore evaluated the ablated variants directly on the three targeted benchmarks from Tables 2 and 3 in the original paper. These analyses reveal complementary roles of conditional residual dynamics and the growth field in future-state recovery and abundance prediction.

Rare-state expansion (CD-short). As shown in Table R2-1, the growth field plays an important role in predicting changes in rare-population abundance. Removing growth modeling substantially degrades expansion identification, direction prediction, and abundance calibration. In contrast, removing residual flow has relatively minor effects on abundance calibration. These results indicate that the growth field is particularly important for forecasting rare-population expansion.

Natural emergence (BetaDiff). As shown in Table R2-1, conditional residual dynamics are critical for reaching previously unseen cellular states. Removing residual flow reduces the accepted emerging-cell fraction from 0.064 to zero, indicating that deterministic displacement alone fails to reach the evaluated emerging-state regions. Removing the growth field also reduces the accepted fraction from 0.064 to 0.023, although with slightly higher precision (0.636 vs. 0.594). This suggests that the growth field increases emerging-state coverage, with some trade-off between coverage and precision.

Table R2-1. Contributions of residual dynamics and growth modeling to rare-state expansion and natural emergence.

MethodsExpn. AP ↑Dir. ↑ΔMAE ↓GrowthSkill ↑Accepted ↑Unsupported ↓Precision ↑
Full0.800 ± 0.2090.650 ± 0.1370.015 ± 0.0030.172 ± 0.2510.064 ± 0.0280.043 ± 0.0170.594 ± 0.047
w/o residual flow0.883 ± 0.1830.700 ± 0.1120.016 ± 0.0050.107 ± 0.2720.000 ± 0.0000.000 ± 0.000—
w/o growth modeling0.617 ± 0.1260.250 ± 0.0000.027 ± 0.002−0.511 ± 0.1870.023 ± 0.0130.014 ± 0.0100.636 ± 0.101
Persistence0.633 ± 0.2090.450 ± 0.1120.018 ± 0.0010.000 ± 0.0000.041 ± 0.0010.055 ± 0.0030.427 ± 0.010
Noise-halo0.683 ± 0.1370.450 ± 0.1120.018 ± 0.002−0.001 ± 0.0060.101 ± 0.0060.107 ± 0.0080.485 ± 0.029

Controlled masked emergence (BetaDiff). We further evaluated the two components on controlled masked-emergence benchmarks, where target populations are withheld during training. For the rare sst_hhex population (Table R2-2), removing residual flow eliminates detection entirely (7/10 to 0/10), while removing growth modeling reduces detection to 2/10 and predicted abundance from 2.2% to 0.3%. These results indicate that residual dynamics are critical for reaching this rare target state, while growth modeling further improves its recovery and predicted representation.

Table R2-2. Contributions of residual dynamics and growth modeling to masked emergence recovery of the rare sst_hhex population on BetaDiff.

MethodsDetected ↑Predicted FractionConditional W₂ ↓OOS-precision ↑
Full7/102.20%16.30.036
w/o residual flow0/100.00%——
w/o growth modeling2/100.30%17.50.022
Persistence0/100.10%——
Noise-halo0/100.10%—0

For sc_alpha (Table R2-3), which lies along a more densely observed developmental trajectory, removing either component still permits frequent detection (8/10 and 10/10, respectively). However, removing the growth field substantially reduces its predicted abundance from 12.1% to 3.7%. These findings suggest that, even when a future population is relatively accessible through learned state transitions, growth modeling remains important for predicting its representation in the future population.

Table R2-3. Contributions of residual dynamics and growth modeling to masked emergence recovery of the sc_alpha population on BetaDiff.

MethodsDetected ↑Predicted FractionConditional W₂ ↓OOS-precision ↑
Full10/1012.10%23.30.477
w/o residual flow8/105.60%22.6—
w/o growth modeling10/103.70%23.50.724
Persistence1/100.40%23.7—
Noise-halo10/102.90%27.30.071

Together, these targeted ablations demonstrate a complementary division of labor: conditional residual dynamics are particularly important for reaching rare or poorly supported future states, whereas the growth field primarily regulates their representation in the predicted population. Importantly, state recovery and abundance calibration do not necessarily improve together, reinforcing the need to evaluate these forecasting capabilities separately.

AuthorsOpenReview ↗

Through five-seed experiments comparing H_max ⁡ ∈ 1,2,3,4 on CD-short, CD-medium, and CR-long, shows that H_max ⁡= 3 is overall optimal and a practical default, while increasing it further to 4 yields no consistent benefit and even introduces trade-offs.

Response to Question1

Regarding horizon-specific forecasting performance, please refer to our response to Weakness 1. Prediction error curves across forecast horizons are provided in Figure S3.

To investigate how the maximum direct training horizon Hmax⁡H_{\max} affects long-range forecasting, we compared Hmax⁡∈{1,2,3,4}H_{\max} \in \{1, 2, 3, 4\} across five random seeds on CD-short, CD-medium, and CR-long under a fixed training budget. Within each split, the training snapshots and latent representations were kept identical. Since our default multi-source inference weights (0.6/0.3/0.1) are specifically designed for Hmax⁡=3H_{\max}=3, we evaluated all variants using standard one-step recursive inference from the latest available state to ensure a fair comparison across horizon settings. The results support Hmax⁡=3H_{\max}=3 as a practical default.

Long-range forecasting accuracy. As shown in Table R3-1, Hmax⁡=3H_{\max}=3 achieves the lowest mean W2W_2 on CD-medium (17.219) and CR-long (17.767), while Hmax⁡=4H_{\max}=4 performs slightly better on CD-short (15.384 vs. 15.440). These results indicate that increasing the maximum training horizon generally improves forecasting performance up to Hmax⁡=3H_{\max}=3 on splits with sufficient temporal supervision, whereas extending it further to 4 provides no consistent additional benefit across splits.

Table R3-1. Effect of the maximum direct training horizon on long-range forecasting accuracy. Comparison of mean W2W_2 across different maximum direct training horizons on CD-short, CD-medium, and CR-long.

SplitHmax⁡=1H_{\max}=1Hmax⁡=2H_{\max}=2Hmax⁡=3H_{\max}=3Hmax⁡=4H_{\max}=4
CD-short15.480 ± 0.32915.736 ± 0.37215.440 ± 0.15515.384 ± 0.092
CD-medium17.542 ± 0.28117.468 ± 0.22117.219 ± 0.21217.246 ± 0.125
CR-long18.065 ± 0.17017.929 ± 0.29217.767 ± 0.24517.881 ± 0.113

Performance across evaluation metrics. The additional metrics in Table R3-2 further support this observation. On CD-medium, Hmax⁡=3H_{\max}=3 outperforms both Hmax⁡=1H_{\max}=1 and Hmax⁡=2H_{\max}=2 across all seven evaluation metrics (including W2W_2), and achieves the best results on five of them among all four horizon settings. On CR-long, Hmax⁡=3H_{\max}=3 similarly outperforms the shorter-horizon settings across all six available metrics and achieves the best results on five. In contrast, increasing Hmax⁡H_{\max} to 4 introduces trade-offs: although it slightly improves some metrics, it degrades differential expression accuracy and transcriptional correlation on CR-long. CD-short shows a more mixed pattern, with no single horizon consistently dominating all metrics.

Table R3-2. Effect of the maximum direct training horizon on complementary forecasting metrics. Comparison of differential expression, transcriptional change, distributional similarity, and cell-type composition metrics across four training horizons on CD-short, CD-medium, and CR-long.

SplitHmax⁡H_{\max}DE-AP ↑DE-dir ↑Δ Pearson ↑MMD ↓Variogram ×10⁻³ ↓Type JS ↓
CD-short10.237 ± 0.0590.482 ± 0.0850.181 ± 0.0940.0488 ± 0.009515.83 ± 1.760.0222 ± 0.0099
20.245 ± 0.0740.553 ± 0.0880.246 ± 0.0960.0552 ± 0.012416.49 ± 3.060.0185 ± 0.0080
30.247 ± 0.0890.557 ± 0.1150.238 ± 0.0840.0516 ± 0.010014.72 ± 1.580.0189 ± 0.0094
40.236 ± 0.0850.541 ± 0.1280.235 ± 0.1130.0512 ± 0.010414.44 ± 1.320.0162 ± 0.0130
CD-medium10.253 ± 0.0300.533 ± 0.0350.022 ± 0.0490.1525 ± 0.010726.83 ± 3.600.2024 ± 0.0165
20.275 ± 0.0290.614 ± 0.0650.073 ± 0.1340.1399 ± 0.005725.49 ± 4.120.1875 ± 0.0136
30.304 ± 0.0550.626 ± 0.0730.163 ± 0.1270.1325 ± 0.005624.02 ± 0.780.1730 ± 0.0276
40.300 ± 0.0510.630 ± 0.0580.124 ± 0.1220.1263 ± 0.004624.15 ± 1.680.1782 ± 0.0116
CR-long10.373 ± 0.0180.713 ± 0.0550.261 ± 0.0910.3250 ± 0.011834.55 ± 1.60—
20.385 ± 0.0250.747 ± 0.0770.298 ± 0.1390.3149 ± 0.013934.65 ± 1.07—
30.393 ± 0.0270.789 ± 0.0400.376 ± 0.0980.3025 ± 0.009633.98 ± 0.78—
40.379 ± 0.0170.732 ± 0.0370.293 ± 0.0720.3101 ± 0.007633.74 ± 1.52—

Overall, these results support Hmax⁡=3H_{\max}=3 as a practical default that effectively captures multi-step dynamics without requiring excessively long training jumps. A possible explanation for the diminishing benefits beyond Hmax⁡=3H_{\max}=3 is that larger temporal offsets provide fewer valid training pairs and potentially less reliable long-range UOT couplings. We have incorporated these results into the new full PDF.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Area Chair and All Reviewers: Core Strengths and Summary of Revisions

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair and all reviewers for their time and constructive feedback. Below, we highlight the core strengths of our framework recognized across the reviews and summarize how our revisions address the main questions raised during the discussion.

1. Core Strengths of scHOMURA

We are encouraged that all reviewers acknowledged the value and empirical strength of scHOMURA across the following key aspects:

  • Unified Generative Modeling: Reviewers highlighted the clarity of our framework in decomposing population evolution into deterministic shared displacement, conditional residual flow dynamics, and a state-dependent growth field trained via prototype-level unbalanced optimal transport.
  • Consistent Empirical Performance: Reviewers recognized that scHOMURA achieves strong performance across diverse single-cell datasets and forecasting splits, including the lowest mean Wasserstein distance among all evaluated baselines.
  • Comprehensive Evaluation of Rare and Emerging States: Reviewers appreciated our broad evaluation protocol that goes beyond global distribution matching to assess rare-population expansion and the recovery of previously unseen cellular states.

2. Key Questions and Manuscript Revisions

To address the reviewers' questions thoroughly, we conducted additional five-seed experiments and computational benchmarks, incorporating all resulting clarifications and analyses directly into Sections 4.3 and 4.6, Figures S3 to S5, and the revised appendix of the updated manuscript:

  • Scope of Multi-Horizon Learning: Reviewers noted nearly identical Wasserstein distances on CD-long with and without multi-horizon learning. We revised Sections 4.3 and 4.6 to attribute extended-horizon stability to the full scHOMURA framework, and added five-seed paired analyses, stage-wise error curves in Figure S3, and horizon ablations in the revised appendix showing clear gains whenever richer multi-step supervision is available, as well as improved rare-population abundance calibration.
  • Component Roles for Rare and Emerging States: To clarify how individual modules contribute beyond aggregate metrics, we added targeted ablations on rare-state expansion and masked emergence benchmarks in the revised appendix, demonstrating that conditional residual dynamics are essential for reaching unseen states while the growth field regulates their relative proportions.
  • Computational Efficiency and Baseline Setups: We documented the input and modeling spaces of all baselines and added end-to-end training time, peak GPU memory, convergence curves, and inference latency benchmarks across datasets up to 32,713 cells in Figures S4 and S5 and the revised appendix, confirming practical scalability and sub-second GPU inference.
  • Sampling Schedule and Hyperparameter Robustness: We specified our hierarchical horizon sampling strategy and added sensitivity analyses across loss weights, prototype counts, multi-source inference weights, and an auxiliary cross-source consistency loss in the revised appendix, confirming reliable performance without per-dataset tuning.

We hope these revisions and additional analyses clearly address all points raised during the review process. We thank the Area Chair and reviewers again for their valuable time and consideration.

Sincerely, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.