acceptodds
Under review as a conference paper at ICLR 2027

scChrono: Learning a Transcriptomic Virtual Embryo through Multi-Rate Developmental Dynamics

Abstract

Virtual Embryo aims to construct a data-backed digital twin of embryogenesis that computationally describes continuous development. At the transcriptomic level, this requires recovering continuous developmental dynamics from discrete single-cell observations. However, transcriptional changes occur at different temporal rates, with sustained developmental progression coexisting with more localized state changes, making the complete developmental process difficult to capture with a single dynamical process. We introduce scChrono, a neural stochastic differential equation (SDE) for learning multi-rate developmental dynamics and constructing a transcriptomic virtual embryo. scChrono decomposes deterministic dynamics into Persistent Dynamics for sustained progression and Fast-Transition Dynamics for temporally gated corrections, while stochastic diffusion models population-level heterogeneity. scChrono reconstructs intermediate states, completes unobserved developmental intervals, forecasts unseen future transcriptomic distributions, and transfers learned developmental representations to downstream Cell Fate Prediction. Across the two primary forecasting settings on the Mouse Developmental Cell and Lineage Atlas (mdCLA), scChrono reduces normalized mean absolute error (NMAE) by 14.1-16.0% and Co-expression Structure Score (CSS) by 4.5% in highly variable gene (HVG) space, and Sinkhorn divergence by 16.1-21.9% in principal component analysis (PCA) space, relative to the strongest baseline for each metric, while achieving the best performance on all six metrics in both temporal completion settings. On lineage-resolved LARRY, transfer improves dominant top-1 fate accuracy by 6.2% over training from scratch. These results support multi-rate developmental dynamics as a framework for building continuous transcriptomic virtual embryos from cross-sectional single-cell atlases. Code is available at https://anonymous.4open.science/r/scChrono-54CC/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell transcriptomics, multi-rate developmental dynamics, neural stochastic differential equations, long-horizon transcriptomic forecasting, transcriptomic virtual embryo

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

This paper introduces scChrono, a neural stochastic differential equation (SDE) framework for learning continuous embryonic developmental dynamics from discrete single-cell transcriptomic snapshots. To address heterogeneous transcriptomic change rates across developmental stages, scChrono decomposes latent dynamics into Persistent Dynamics for sustained developmental progression and Fast-Transition Dynamics for temporally gated corrections, while incorporating stochastic diffusion to model cellular heterogeneity.

The authors evaluate scChrono on the mdCLA dataset across interpolation, long-horizon extrapolation, and temporal completion tasks, demonstrating improvements over representative OT, ODE, SDE, and diffusion-based baselines on most metrics. They further demonstrate the transferability of learned developmental representations through cell fate prediction on the LARRY dataset. The main contribution is a structured multi-rate dynamics framework designed to improve continuous whole-embryo transcriptomic forecasting, particularly over long developmental horizons.

Critical strengths

Well-motivated framework for whole-embryo long-horizon forecasting. The paper tried to addresse an important problem: transcriptomic change rates vary substantially across developmental stages, a biologically meaningful characteristic that should be considered when modeling embryonic development. The authors attempt to address this challenge through an explicit multi-rate dynamics architecture.

Comprehensive experimental evaluation. The authors evaluate the proposed method across multiple forecasting settings and metrics, with comparisons against representative baselines. They also provide component ablations, parameter-matched controls, and analyses of the gating mechanism and forecasting horizon, which help assess the contribution of the proposed architecture.

Critical weaknesses

1. Limited and task-dependent contribution of Fast-Transition Dynamics.

Fast-Transition Dynamics (FTD) is the main methodological novelty of scChrono, but its contribution varies substantially across evaluation settings. Removing FTD slightly improves all six metrics in Inter-Time Interpolation and has little effect on Fetal Completion, while its benefits are mainly observed in Strict Long-Horizon Extrapolation. More notably, removing FTD improves all six metrics in Pre-Fetal Completion (Appendix Table 11). These results suggest that the proposed component has a more limited scope of effectiveness than the paper's broader motivation implies. I suggest emphasizing strict long-horizon extrapolation as the primary contribution and clarifying the settings in which FTD is actually beneficial.

2. Insufficient evidence for the claimed separation of developmental dynamics into distinct temporal roles.

The paper motivates the decomposition into Persistent Dynamics (PD) and Fast-Transition Dynamics (FTD) by the heterogeneous rates of transcriptomic changes across developmental stages. However, the current experiments do not directly demonstrate that PD captures sustained developmental progression or that FTD captures rapid developmental transitions. Both components are jointly optimized under the same forecasting objective, and their intended roles are mainly encouraged through architectural constraints and regularization. In particular, FTD is modulated by a manually specified temporal prior, but the authors do not show whether its learned activation or contribution corresponds to periods of rapid transcriptomic change. The current ablations establish that the two-branch architecture can improve forecasting performance, but provide limited evidence for the claimed multi-rate interpretation.

3. Unclear motivation and evaluation role of the temporal completion tasks.

Fetal Completion and Pre-Fetal Completion are introduced as additional evaluation settings, but their relationship to the central objective of long-horizon forecasting is insufficiently explained. Since 20% of observations within the target developmental period remain visible, these settings differ substantially from forecasting entirely unseen future stages. The authors should better justify what capabilities these tasks are intended to evaluate and how they complement the primary forecasting experiments. If temporal completion is considered a major contribution, its main quantitative results should also be presented in the main text rather than only in the appendix.

4. Limited evidence that the proposed method accurately recovers biologically meaningful developmental changes.

Although scChrono achieves improvements in several reconstruction and distributional metrics, its performance on differential expression recovery remains limited under Strict Long-Horizon Extrapolation. In particular, the Differential Expression Score (DES) remains negative and is worse than that of PRESCIENT. This raises concerns about whether improvements in overall transcriptomic distribution matching translate into accurate predictions of developmentally relevant gene-expression changes. Additional biological analyses would help substantiate the broader claim of constructing a predictive transcriptomic virtual embryo.

Questions for authors

1. Can the authors show whether the learned Fast-Transition Dynamics actually correspond to the rapidly changing developmental periods identified in Figure 1A? For example, analyzing the temporal variation in FTD activation or contribution magnitude and comparing it with the empirical transcriptomic change rates.

2. How can the authors verify that Persistent Dynamics and Fast-Transition Dynamics capture distinct temporal roles? Since both components are jointly optimized under the same forecasting objective, it is unclear whether they genuinely learn sustained and rapid developmental changes, respectively. Could the authors provide additional analyses of the learned dynamics to support this decomposition?

3. Why does removing FTD improve performance across all six metrics in Pre-Fetal Completion (Appendix Table 11), while its contribution is substantial in Strict Long-Horizon Extrapolation? Could the authors clarify the conditions under which FTD is beneficial and explain this discrepancy?

5 comments
AuthorsOpenReview ↗

Response to Reviewer xwFt (1/3): W1 & Q3

Response to Reviewer xwFt (Part 1/3)

W1 & Q3: Task-dependent contribution of FTD

We thank the reviewer for the careful observation and constructive suggestions. FTD is effective overall and most suitable when target visibility is low and the forecast horizon is long, which is the Strict Long-Horizon Extrapolation setting. We adopt the reviewer's suggestion to present this setting as FTD's main empirical contribution, and the manuscript has been updated accordingly (Abstract, Introduction, and Section 4.2.2).

1. Inter-Time, Fetal, and Pre-Fetal: after a minor fix, Full outperforms −FTD in two tasks and is comparable in the other two. In Pre-Fetal, backward rollout used the unsigned step ∣Δt∣|\Delta t| for FTD but the signed Δt\Delta t for PD (Appendix B.2). After the fix, Full's DES rises from 0.269 to 0.463, and "removing FTD improves all six metrics" no longer holds:

Pre-FetalNMAE-PCA ↓Sink-PCA ↓NMAE-HVG ↓Sink-HVG ↓DES ↑CSS ↓
Full (fixed)0.40411.3600.286153.3700.4630.022
−FTD0.3499.3680.291155.3410.4770.022

Full versus retrained −FTD across the four tasks (five seeds; full results: Table 1, Appendix Table 22):

TaskFull vs. −FTD
StrictBetter on all six metrics (NMAE-HVG 0.360 vs. 0.444)
FetalBetter on 3 of 5 non-tied metrics; all within 1.7%
Inter-TimeComparable; five of six within one SD
Pre-FetalComparable; better on both HVG metrics, all within one SD

FTD improves Strict and Fetal and is comparable in Inter-Time and Pre-Fetal, and with FTD, scChrono outperforms all baselines on 23 of 24 task–metric comparisons, so FTD is effective overall.

2. Conditions under which FTD helps most (Q3): weaker target constraints and longer horizons.

(1) Target visibility: the less target information, the more FTD helps. We held out 20%, 50%, or 80% of target-stage cells in Fetal and Pre-Fetal Completion (single seed; full results in Appendix Table 23):

TaskHeld-outFull winsΔDES ↑ΔNMAE-HVG ↑
Fetal20%3/6 (1 tie)−0.012−0.004
Fetal80%5/6 (1 tie)+0.018+0.002
Pre-Fetal20%3/6+0.002+0.002
Pre-Fetal80%6/6+0.049+0.023

With many visible cells, PD alone performs comparably. With few, the model relies more on FTD's gated corrections.

(2) Forecast horizon: the farther the forecast, the larger the FTD gain. In Strict Extrapolation, the NMAE-HVG gap grows from about 0.01 at Δt = 0.5 to 0.21 at Δt = 5.0 (Section 4.2.2, Figure 4). Spearman correlation between horizon and the degradation after retraining without FTD (full results in Appendix Table 20):

MetricNMAE-HVGSink-HVGCSSSink-PCANMAE-PCADES
ρ1.0000.9760.9880.9880.8670.539
p5.5×10⁻⁷2.1×10⁻⁵5.5×10⁻⁶5.5×10⁻⁶0.00220.114

Errors of a single drift accumulate along the rollout, which is consistent with FTD's larger gains at longer horizons.

(3) Bidirectional constraints: with observed stages on both sides, FTD has little room to help. Although Inter-Time fully withholds the target stage, observed stages before and after it constrain the prediction, so a single drift interpolates well.

Strict Long-Horizon Extrapolation combines all three conditions, which is why FTD is most useful there. Section 3.3.2 now unifies the signed-step update.

(Continued in Part 2/3.)

AuthorsOpenReview ↗

Response to Reviewer xwFt (2/3): W2, Q1 & Q2

Response to Reviewer xwFt (Part 2/3)

W2 & Q1 & Q2: Temporal roles of PD and FTD

We thank the reviewer for this important question. Four new analyses show that the two branches play distinct temporal and functional roles.

1. Temporal division: FTD dominates early, PD later (Q1, Q2). As requested, we plotted the gate, prior, and effective FTD against the change rates (Appendix Figure 6). The effective FTD contribution falls from about ten times PD at E7.5 to about 1% at E16.5, consistent with the high early and declining later change rates in Figure 1A:

StageE7.5E8.0E12.0E15.0E16.5
Effective FTD/PD10.752.150.1260.0230.012

2. Predictive value: the gain is not explained by capacity (W2). From E7.5, E11, and E13 under Strict Extrapolation, Full beats −FTD and a parameter-matched single drift (221,334 vs. 221,569 parameters) on all six metrics at every source (full results in Appendix Table 13):

SourceFull DES−FTD DESMatched DESFull wins (vs. −FTD / matched)
E7.50.601 ± 0.0440.557 ± 0.0130.534 ± 0.0566/6 · 6/6
E110.331 ± 0.0210.185 ± 0.0460.159 ± 0.0746/6 · 6/6
E130.138 ± 0.060−0.016 ± 0.024−0.031 ± 0.0576/6 · 6/6

Removing the gate, the prior, or both also degrades Strict performance (five-seed means; SDs in Appendix Table 14):

VariantNMAE-PCA ↓Sink-PCA ↓NMAE-HVG ↓Sink-HVG ↓DES ↑CSS ↓
Full0.82531.2510.360359.108−0.0710.021
Gate only0.91334.4640.445378.015−0.1560.028
Prior only0.96436.6660.433373.690−0.2610.021
No gate0.93933.6780.457369.671−0.2530.026

3. Biological relevance: FTD prioritizes known developmental drivers (W2, Q2). As gold driver sets require literature curation, we used 24 curated epithelial driver-TF annotations as a case study (Appendix Table 17), summarizing FTD attribution by maximum (largest absolute attribution in a window) and step-mean (mean absolute attribution to the gated step zkf−zkpz_k^{\mathrm f}-z_k^{\mathrm p}, Eq. 4; same rule as PD mean). Recall:

AttributionTop 5%Top 10%Top 20%Top 30%
FTD (maximum)37.50%37.50%54.17%66.67%
FTD (step-mean)25.00%33.33%66.67%83.33%
PD (maximum)4.17%25.00%29.17%37.50%
PD (mean)12.50%25.00%41.67%50.00%

FTD (maximum) also ranks 17 of 24 above PD (p = 0.027) and stays enriched against expression-matched genes (p = 0.016; PD p = 0.290), and 11/12 step-mean permutation tests survive BH-FDR (Appendix Table 19).

4. Relation to empirical change rates (Q1). At a coarse level, yes: effective FTD peaks early, when empirical change rates are also high. Within lineages, after detrending, the prior-free gate correlates positively with the change rates of all five lineages (Appendix Figure 7); the raw series gives ρ = −0.300:

Gate vs. change rateρp
Linear detrending0.2610.018
Quadratic detrending0.2710.014
First differences0.1190.291

Effective FTD magnitude itself is not significantly correlated with the change rate (p = 0.112), which is not contradictory: the Figure 1A rate is a snapshot-derived proxy that also reflects lineage composition, not a direct measure of transition speed. FTD provides gated corrections rather than reproducing that curve.

We therefore position PD/FTD as an empirically supported inductive bias and have revised the Abstract, Introduction, and Conclusion accordingly.

(Continued in Part 3/3.)

AuthorsOpenReview ↗

Response to Reviewer xwFt (3/3): W3 & W4

Response to Reviewer xwFt (Part 3/3)

W3: Motivation and role of Temporal Completion

We thank the reviewer for noting that the motivation of the two Completion tasks and their relation to forecasting were insufficiently explained. Temporal Completion complements long-horizon forecasting for incomplete developmental time series, while Strict Long-Horizon Extrapolation remains the main contribution.

The two tasks cover two different data gaps in developmental studies. Strict Long-Horizon Extrapolation addresses future stages with no collected data, using observed stages to forecast later transcriptomic states, a key setting for predicting cell states and trajectories [1–4]. Temporal Completion extends the model to stages that were sampled but incompletely observed: owing to limited cell numbers and cost, even a sampled stage rarely represents its full state distribution, so reconstructing unsampled cells helps complete developmental atlases. Prior work evaluates such generation by holding out 20% of cells [5], and we follow this protocol: 80% of target-period cells are visible and 20% are held out (clarified in the revision).

The two tasks lie at two ends of a continuum. As target visibility decreases, Completion increasingly resembles forecasting: raising the held-out fraction from 20% to 80% increases the FTD gain (Part 1/3).

We added this complementary role of the two tasks to Section 4.1 (Task Settings) and Appendix D.3, and Section 4.2.1 now summarizes the Completion results: scChrono achieves the best mean on all six metrics in both settings (Appendix Table 11).

W4: Recovery of developmentally relevant expression changes

We thank the reviewer for raising this point. Pooled DES under Strict Long-Horizon Extrapolation remains a limitation, but scChrono has the best DES in the other three tasks and outperforms PRESCIENT in all five major lineages.

Strict extrapolation: recovering differential expression remains a shared challenge. All evaluated methods have negative pooled DES, and scChrono ranks second:

MethodPRESCIENTscChronoTrajectoryNetscDiffEqscNODE
Strict DES ↑−0.008−0.071−0.097−0.136−0.179

We acknowledge this as a challenge shared by all methods and will improve long-range differential-expression forecasting in future versions of scChrono.

Other tasks: scChrono has the best DES in Inter-Time (0.258), Fetal (0.557), and Pre-Fetal (0.463), recovering expression changes when the target is better constrained.

Lineage level: scChrono outperforms PRESCIENT in DES in all five lineages. In a new lineage-specific analysis under Strict Extrapolation (full results in Appendix D.2, Table 9):

LineagescChronoPRESCIENT
Endothelial0.016−0.066
Epithelial−0.004−0.144
Hematopoietic0.023−0.179
Mesenchymal−0.043−0.134
Neural−0.014−0.142

scChrono leads all applicable methods in DES in four lineages and has the best NMAE, Sinkhorn, and CSS in all five lineages (25/25). Pooled DES ranks genes by population-mean changes, which combine within-lineage expression changes with shifts in lineage composition. A method can therefore lead within every lineage yet trail on pooled DES (a Simpson-type effect). Section 4.2.1 now discusses these results; for gene-level evidence, see the driver-TF analysis in Part 2/3.

References

[1] Chen, Z., King, W. C., Hwang, A., Gerstein, M., & Zhang, J. (2022). DeepVelo: Single-cell transcriptomic deep velocity field learning with neural ordinary differential equations. Science Advances, 8(48), eabq3745.

[2] Zhang, J., Larschan, E., Bigness, J., & Singh, R. (2024). scNODE: Generative model for temporal single cell transcriptomic data prediction. Bioinformatics, 40(Supplement_2), ii146–ii154.

[3] Erbe, R., Stein-O'Brien, G., & Fertig, E. J. (2023). Transcriptomic forecasting with neural ordinary differential equations. Patterns, 4(8), 100793.

[4] Yeo, G. H. T., Saksena, S. D., & Gifford, D. K. (2021). Generative modeling of single-cell time series with PRESCIENT enables prediction of cell trajectories with interventions. Nature Communications, 12(1), 3222.

[5] Turgut Ögme, S. S., Aydin, N., & Kurt, Z. (2025). A mixture of attention experts-embedded flow-based generative model to create synthetic cells in single-cell RNA-Seq datasets. PLOS Computational Biology, 21(10), e1013525.

Reviewer_x6TiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

scChrono is a neural SDE with Persistent Dynamics, temporally gated Fast-Transition Dynamics, and stochastic diffusion. Weak Lineage Anchors and a Late Organ Adapter support training and readout. The paper evaluates interpolation, Strict Long-Horizon Extrapolation, and masked temporal completion on mdCLA, with transfer to LARRY. The clearest FTD benefit is in strict extrapolation.

Critical strengths

Some empirical results show the method has good performance.

Critical weaknesses
  1. The multi-rate interpretation is not established. Removing FTD improves all six mean metrics in Inter-Time Interpolation and Pre-Fetal Completion (Tables 1 and 11), although five Inter-Time differences are within one SD of Full. The authors' horizon-wise analysis attributes FTD's main role to shaping the dynamics learned during training rather than an instantaneous rollout correction. Several large peaks in Figure 1A occur before E7.5, whereas the [7.5, 16.5] prior downweights FTD there. The relationship between the learned gate and these rates is not demonstrated.

  2. Extrapolation robustness and DES comparisons need further validation. The FTD benefit, matched-single-drift control, and horizon trend all use one fixed-source split (Section C.2; Figure 4). Table 3 varies only the prior window’s endpoint in Inter-Time Interpolation with seed 1, where FTD has little effect, leaving robustness to source stage and prior boundaries unclear. DES is unavailable for scNODE and cfDiffusion in otherwise evaluated settings because outputs are unpaired, although Eqs. 10–11 rank population-mean changes without an evident need for source-cell correspondence. Thus completion claims of best performance on all six metrics cover only available scores.

Questions for authors
  1. Could you plot the learned gate, temporal prior, and effective FTD contribution over time, and explain their relationship to the change rates in Figure 1A?
  2. Do the strict-extrapolation gains of Full over −FTD and matched-single-drift persist with another source stage and perturbed prior-window boundaries?
  3. What input required by Eqs. 10–11 is unavailable for scNODE and cfDiffusion? Could their DES be computed from predicted populations using the same reference and target groups?
5 comments
AuthorsOpenReview ↗

Response to Reviewer x6Ti (1/3): W1 & Q1

Response to Reviewer x6Ti (Part 1/3)

W1 & Q1: Necessity of FTD and the multi-rate interpretation

We thank the reviewer for the constructive comments, which are very helpful. In short, FTD is effective and most suitable when target visibility is low and the horizon is long, it acts mainly through joint training, and it is most active in the early, high-change-rate stages. Below we answer the four points in the order raised.

1. Necessity of FTD: after a minor fix, Full outperforms −FTD in two tasks and is comparable in Inter-Time and Pre-Fetal. In Pre-Fetal, backward rollout used the unsigned step ∣Δt∣|\Delta t| for FTD but the signed Δt\Delta t for PD (Appendix B.2). After the fix, Full's DES rises from 0.269 to 0.463, and "removing FTD improves all six metrics" no longer holds:

Pre-FetalNMAE-PCA ↓Sink-PCA ↓NMAE-HVG ↓Sink-HVG ↓DES ↑CSS ↓
Full (fixed)0.40411.3600.286153.3700.4630.022
−FTD0.3499.3680.291155.3410.4770.022

Full versus retrained −FTD across the four tasks (five seeds; full results: Table 1, Appendix Table 22):

TaskFull vs. −FTD
StrictBetter on all six metrics (NMAE-HVG 0.360 vs. 0.444)
FetalBetter on 3 of 5 non-tied metrics; all within 1.7%
Inter-TimeComparable; five of six within one SD
Pre-FetalComparable; better on both HVG metrics, all within one SD

FTD improves Strict and Fetal and is comparable in Inter-Time and Pre-Fetal, and with FTD, scChrono outperforms all baselines on 23 of 24 task–metric comparisons, so FTD is effective overall. Although Inter-Time fully withholds the target stage, observed stages before and after the target constrain the prediction, so a single drift interpolates well and FTD has little room to help. Conversely, as target visibility decreases, the FTD gain grows (full results in Appendix Table 23):

TaskHeld-outFull winsΔDES ↑ΔNMAE-HVG ↑
Fetal20%3/6 (1 tie)−0.012−0.004
Fetal80%5/6 (1 tie)+0.018+0.002
Pre-Fetal20%3/6+0.002+0.002
Pre-Fetal80%6/6+0.049+0.023

FTD is therefore effective overall and most suitable when target visibility is low and the horizon is long, as in Strict Long-Horizon Extrapolation. We have updated the Abstract, Introduction, and Section 4.2.2 accordingly.

2. Horizon-wise analysis: we agree that FTD acts mainly through joint training. Masking FTD only at inference (Infer-off) stays close to Full in Strict, whereas retraining without FTD degrades all six metrics (full results in Appendix Table 21). This is consistent with the multi-rate interpretation: during training, FTD absorbs rapid early changes so that PD learns a cleaner sustained drift, and the retrained gap keeps widening with horizon even after the prior suppresses FTD at later stages (Appendix D.4). We revised the Conclusion accordingly.

StrictFullInfer-offRetrained −FTD
NMAE-HVG ↓0.3600.3560.432
DES ↑−0.054−0.048−0.147

3. Peaks before E7.5: the reviewer is right that the default prior is conservative early. The prior is an inductive bias, not a biological annotation of rapid stages. Moving its onset from E7.5 to E6.5 improves all six Strict means (Part 2/3). This itself supports the multi-rate interpretation: letting FTD cover earlier high-change-rate periods improves forecasting.

4. Gate and change rates (Q1): FTD acts in the early, high-change-rate stages, and the prior-free gate follows lineage change rates. As requested, we plotted the gate, prior, and effective FTD against the Figure 1A change rates (Appendix Figure 6). The effective FTD contribution falls from about ten times PD at E7.5 to about 1% at E16.5, consistent with the high early and declining later change rates in Figure 1A:

StageE7.5E8.0E12.0E15.0E16.5
Effective FTD/PD10.752.150.1260.0230.012

Within lineages (Appendix Figure 7), the detrended prior-free gate correlates positively with the change rates of all five lineages (raw series ρ = −0.300, dominated by the shared temporal trend):

Gate vs. change rateρp
Linear detrending0.2610.018
Quadratic detrending0.2710.014
First differences0.1190.291

Effective FTD magnitude itself is not significantly correlated with the change rate (p = 0.112). This is not contradictory: the Figure 1A rate is a snapshot-derived proxy that also reflects lineage composition, and FTD provides gated corrections rather than reproducing that curve.

We position the PD/FTD decomposition as a biologically motivated, empirically supported inductive bias and have revised the Abstract, Introduction, and Conclusion accordingly. Supporting this, FTD also prioritizes known developmental drivers (Part 3/3).

(Continued in Part 2/3.)

AuthorsOpenReview ↗

Response to Reviewer x6Ti (2/3): W2 (robustness) & Q2

Response to Reviewer x6Ti (Part 2/3)

W2 (robustness) & Q2: Source stage and temporal-prior window

We thank the reviewer for the comments on extrapolation robustness and prior boundaries. New experiments show that the FTD benefit holds across source stages and under shifted, narrowed, or widened prior windows; it disappears only when the fixed prior is removed entirely.

1. Source stage: the FTD benefit holds across three source stages. Under Strict Long-Horizon Extrapolation, we used E7.5, E11, and E13 as source stages and compared five-seed results of Full, −FTD, and a parameter-matched single drift (221,334 vs. 221,569 parameters). Wins count the six metrics (full results in Appendix Table 13):

SourceFull DES−FTD DESMatched DESFull wins (vs. −FTD / matched)
E7.50.601 ± 0.0440.557 ± 0.0130.534 ± 0.0566/6 · 6/6
E110.331 ± 0.0210.185 ± 0.0460.159 ± 0.0746/6 · 6/6
E130.138 ± 0.060−0.016 ± 0.024−0.031 ± 0.0576/6 · 6/6

NMAE-HVG decreases by 0.037/0.114/0.069 relative to −FTD and by 0.060/0.140/0.098 relative to the matched single drift; at E13, both controls have negative DES while Full remains positive.

2. Prior boundaries: performance does not hinge on one window, but the best boundaries are task-dependent. Unlike the original single-seed, endpoint-only sweep in Inter-Time, we now vary both boundaries, include Strict, and repeat key windows over five seeds; we also test removing the fixed prior. Strict point estimates:

Prior windowNMAE-PCA ↓Sink-PCA ↓NMAE-HVG ↓Sink-HVG ↓DES ↑CSS ↓
[4.0, 16.5]0.75932.5880.342365.4240.0210.021
[5.0, 16.5]0.74330.3460.338363.2710.0200.022
[6.5, 16.5]0.76830.9260.337358.2690.0380.020
[7.5, 16.5] (default)0.77731.3280.342359.658−0.0300.020
[8.5, 16.5]0.77331.8840.348359.763−0.0870.019
[9.5, 16.5]0.80934.0730.375369.481−0.1550.022
[7.5, 15.5]0.80731.9240.360360.290−0.1350.021
[7.5, 17.5]0.83731.3350.371359.475−0.0900.021
Fully learned0.83734.6400.397377.278−0.1890.024

In Strict Long-Horizon Extrapolation, earlier onsets match or exceed the default, whereas delaying the onset to E9.5 or removing the prior worsens several metrics. In Inter-Time Interpolation, the fully learned gate is competitive and beats the default on all six point estimates (Appendix Table 5). In repeated runs, Strict DES is −0.058 ± 0.081, −0.050 ± 0.090, and −0.070 ± 0.085 for [4.0, 16.5], [6.5, 16.5], and the default window (Appendix C.6).

3. Perturbed windows: with any fixed prior, Full outperforms −FTD. We compared each five-seed window configuration (Appendix Table 3) with retrained −FTD (Table 1: NMAE-HVG 0.444, Sink-HVG 369.5, DES −0.168, CSS 0.027):

Prior window (5 seeds)ΔNMAE-HVGΔSink-HVGΔDESΔCSSFull wins (6 metrics)
[6.5, 16.5]+0.083+11.27+0.118+0.0076/6
[7.5, 16.5] (default)+0.077+9.88+0.098+0.0066/6
[8.5, 16.5]+0.072+9.78+0.077+0.0076/6
[7.5, 15.5]+0.074+9.25+0.096+0.0066/6
[7.5, 17.5]+0.073+10.07+0.078+0.0066/6
No fixed prior (gate only)−0.001−8.47+0.012−0.0011/6

Positive Δ favors Full.

Without the fixed prior, the advantage essentially disappears (DES only). Removing the gate, the prior, or both also degrades Strict performance (five-seed means; SDs in Appendix Table 14):

VariantNMAE-PCA ↓Sink-PCA ↓NMAE-HVG ↓Sink-HVG ↓DES ↑CSS ↓
Full0.82531.2510.360359.108−0.0710.021
Gate only0.91334.4640.445378.015−0.1560.028
Prior only0.96436.6660.433373.690−0.2610.021
No gate0.93933.6780.457369.671−0.2530.026

The strict-extrapolation benefit of FTD therefore depends on a fixed temporal prior but not on its exact boundaries, and it requires the learned gate as well.

(Continued in Part 3/3.)

AuthorsOpenReview ↗

Response to Reviewer x6Ti (3/3): W1 (cont.) & W2 (DES) & Q3

Response to Reviewer x6Ti (Part 3/3)

W1 (continued): Biological relevance of FTD

FTD prioritizes known developmental drivers more than PD. Recall of 24 curated epithelial driver annotations among top-ranked TFs (Appendix Table 17 lists the annotations):

AttributionTop 5%Top 10%Top 20%Top 30%
FTD (maximum)37.50%37.50%54.17%66.67%
FTD (step-mean)25.00%33.33%66.67%83.33%
PD (maximum)4.17%25.00%29.17%37.50%
PD (mean)12.50%25.00%41.67%50.00%

FTD (maximum) also places 17 of 24 annotations at better specificity-corrected ranks than PD (p = 0.027; Appendix Table 19).

W2 (DES) & Q3: Consistent DES evaluation

We thank the reviewer for pointing out this incomplete comparison. The reviewer is right that DES can be computed for scNODE and cfDiffusion; after adding these scores, scChrono's DES is higher than both baselines in every task where they apply.

DES does not require paired cells, so the original explanation was inaccurate. Eqs. (10)–(11) use the differential-expression changes between reference and target populations and how well the predicted population recovers them; no cell-level correspondence is needed. We corrected the text (Section 4.1, Appendix C.3) and computed DES for both baselines with the same reference and target groups:

TaskscChronoscNODEcfDiffusion
Inter-Time0.258 ± 0.0580.049 ± 0.012–
Strict−0.071 ± 0.096−0.179 ± 0.039–
Fetal0.557 ± 0.016−0.126 ± 0.0500.119 ± 0.050
Pre-Fetal0.463 ± 0.020–0.208 ± 0.015

Full benchmark results in Appendix Tables 8 and 11.

The remaining gaps reflect method scope rather than pairing. cfDiffusion conditions on discrete training days and cannot predict unseen stages, and scNODE integrates only forward from the earliest time point, so it cannot perform the backward Pre-Fetal rollout. scChrono also remains best on all six metrics in both Completion settings, and we have revised the related statements in Section 4.2.1 accordingly.

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

This paper proposes scChrono, a time-conditioned neural stochastic differential equation (SDE) framework for reconstructing and forecasting continuous embryonic transcriptomic dynamics from discrete, unpaired single-cell snapshots. The key design is a decomposition of latent drift into Persistent Dynamics for sustained evolution and a state–time-gated Fast-Transition Dynamics (FTD) for temporally localized corrections, together with stochastic diffusion. Weak lineage anchors and a late organ adapter provide additional training signals and late-stage refinement. On the mouse developmental atlas mdCLA, the authors evaluate inter-time interpolation, strict long-horizon extrapolation, and two masked temporal-completion settings. scChrono achieves the best results on all six reported metrics for interpolation and both completion settings, and five of six metrics under strict extrapolation. The work argues that an explicit multi-rate inductive bias can improve long-horizon developmental forecasting.

Critical strengths

S1. Important and concrete problem. Learning continuous dynamics from destructive, cross-sectional single-cell measurements is a meaningful challenge. The whole-embryo setting and strict future extrapolation offer a stronger test of forecasting than evaluation confined to short local intervals. The reported 8.6-fold variation in lineage-aware transcriptomic change rates provides a clear empirical motivation, while not being equated with biological speed.

S2. Clear, testable architectural hypothesis. Separating a persistent drift from a gated fast-transition drift is a plausible inductive bias for heterogeneous temporal changes. The paper does more than introduce extra parameters: a closely parameter-matched single-drift comparator performs worse on all six strict-extrapolation metrics.

S3. Substantial and relatively transparent experimental evidence. The main benchmarks span interpolation, fully unseen future targets, and masked completion, assessed with six complementary metrics. The authors report both favorable and unfavorable outcomes, including weaker strict-extrapolation DES and instances where removing FTD improves performance.

Critical weaknesses

W1. The benefit of FTD is strongly task-dependent. Under strict long-horizon extrapolation, removing FTD worsens HVG NMAE from 0.360 to 0.444, consistent with the proposed motivation. Yet removing FTD improves all six interpolation metrics and all six Pre-Fetal Completion metrics; in the latter, PCA NMAE decreases from 0.541 to 0.349 and DES increases from 0.269 to 0.477. The authors acknowledge this dependence, but it remains unclear what task properties determine the sign and magnitude of the benefit. The evidence therefore favors a conditional long-horizon advantage, not a universally superior decomposition.

W2. The learned branches are not shown to have uniquely identifiable developmental roles. The observed total drift can admit different allocations between persistent and fast-transition fields. Moreover, the horizon-wise effects differ between retraining without FTD and masking it only at inference; this supports the authors' training-time interpretation but leaves open whether FTD primarily acts as an optimization-friendly parameterization rather than as a biologically distinct rapid-transition mechanism. A visualization or quantitative validation of branch activity against independently measured developmental changes would substantially strengthen the interpretation.

W3. The roles of stochastic diffusion and population-level supervision need sharper isolation. PD-only removes both FTD and diffusion, and therefore cannot by itself attribute the resulting degradation to either component. In addition, the endpoint, anchor, and RBF-MMD terms operate on coarse-lineage time-point centroids/anchor latent collections. These objectives may align average states while missing within-lineage branching or multimodality. This is a limit of what the supervision directly constrains, not proof that the final model fails to recover distributions. A factorial dynamics ablation and targeted tests on branching populations would help.

Questions for authors

Q1. Applicability of FTD Can the authors characterize which properties of a prediction task (forecast horizon, forward versus backward direction, target-stage visibility, or observed change-rate heterogeneity) govern whether FTD helps? In particular, why does removing FTD improve all six Pre-Fetal Completion metrics?

Q2. Mechanism attribution Figure 4 and Table 10 suggest the benefit of FTD is partly mediated by joint training rather than its immediate rollout contribution. Can the authors provide branch-magnitude/gate activation analyses and relate them to the lineage-aware empirical change-rate curve? How does Full training with FTD but inference without FTD compare across settings?

Q3. Diffusion and branching Can the authors report the complete 2×2 ablation of FTD and diffusion (PD only; PD+FTD; PD+diffusion; Full)? How well do the centroid-based training signals preserve within-lineage multimodality and branching, and can this be assessed with cell-type-conditioned or lineage-resolved distribution metrics?

3 comments
AuthorsOpenReview ↗

Response to Reviewer vMyZ (1/3): W1 & Q1

Response to Reviewer vMyZ (Part 1/3)

W1 & Q1: Task-dependent benefit of FTD

We thank the reviewer for the careful observation. With FTD, scChrono outperforms −FTD in two of the four tasks, and in the other two the differences are small; with FTD, scChrono also outperforms all baselines on 23 of 24 task–metric comparisons. We agree that FTD is most effective in Strict Long-Horizon Extrapolation, and we observe that its gain grows as target visibility decreases and the forecast horizon increases.

Full outperforms −FTD in two tasks and is comparable in the other two, and scChrono leads the baselines in all four. Full versus retrained −FTD (five seeds) and versus baselines (full results: Table 1; Appendix Tables 8, 11, and 22):

TaskFull vs. −FTDscChrono vs. baselines
StrictBetter on all six metrics (NMAE-HVG 0.360 vs. 0.444)Best on 5/6 (DES: PRESCIENT)
FetalBetter on 3 of 5 non-tied metrics; all within 1.7%Best on 6/6
Inter-TimeComparable; five of six within one SDBest on 6/6
Pre-FetalComparable; better on both HVG metrics, all within one SDBest on 6/6

Pre-Fetal: after a minor implementation fix, "removing FTD improves all six metrics" no longer holds. The original code updated FTD with the unsigned step ∣Δt∣|\Delta t| while PD used the signed Δt\Delta t, so the two branches moved in inconsistent directions during backward integration; FTD now uses the signed Δt\Delta t, and forward updates are unchanged (Appendix B.2). After the fix, Full's DES rises from 0.269 to 0.463, and Full outperforms −FTD on both HVG metrics (full results in Appendix Table 22):

Pre-FetalNMAE-PCA ↓Sink-PCA ↓NMAE-HVG ↓Sink-HVG ↓DES ↑CSS ↓
Full (fixed)0.40411.3600.286153.3700.4630.022
−FTD0.3499.3680.291155.3410.4770.022

Below we show what drives the size of the FTD gain.

(1) Target visibility: the less target information, the more FTD helps. We set the held-out fraction of target-stage cells in Fetal and Pre-Fetal Completion to 20%, 50%, and 80% (full results in Appendix D.4.1, Table 23):

TaskHeld-outFull winsΔDES ↑ΔNMAE-HVG ↑
Fetal20%3/6 (1 tie)−0.012−0.004
Fetal80%5/6 (1 tie)+0.018+0.002
Pre-Fetal20%3/6+0.002+0.002
Pre-Fetal80%6/6+0.049+0.023

From 20% to 80% held out, Full wins rise from 3/6 to 5/6 (Fetal) and 6/6 (Pre-Fetal), and the DES gain grows in both tasks.

(2) Forecast horizon: the farther the forecast, the larger the FTD gain. The horizon-wise analysis (Section 4.2.2, Figure 4) shows that the FTD gain on several reconstruction and distribution metrics increases with forecast horizon: in Strict Extrapolation, the NMAE-HVG gap between retrained −FTD and Full grows from about 0.01 at Δt = 0.5 to 0.21 at Δt = 5.0 (ρ = 1.000, p = 5.5×10⁻⁷). Spearman correlation between horizon and the degradation after retraining without FTD (full results in Appendix Table 20):

MetricNMAE-HVGSink-HVGCSSSink-PCANMAE-PCADES
ρ1.0000.9760.9880.9880.8670.539
p5.5×10⁻⁷2.1×10⁻⁵5.5×10⁻⁶5.5×10⁻⁶0.00220.114

Five of six metrics widen significantly with horizon; DES shows a weaker, non-significant trend.

(3) Forward versus backward direction: direction itself is not the deciding factor. After the fix, backward Pre-Fetal and forward Fetal Completion show the same visibility trend; the anomaly that appeared only in the backward task came from the implementation error, not from the direction (table in factor 1).

(4) Change-rate heterogeneity: the FTD contribution concentrates in early, high-change-rate stages. The effective FTD contribution is strongest in early stages, where empirical change rates are also high: effective FTD/PD falls from 10.75 at E7.5 to 0.012 at E16.5 (Appendix Table 15).

(5) Bidirectional constraints: when the target is observed on both sides, the extra gain of FTD is limited. This explains Inter-Time Interpolation: although it fully withholds the target stage, observed stages before and after the target constrain the prediction, leaving little room for FTD to help (task table above).

Taken together, when the target is anchored by observations or by constraints on both sides, the additional flexibility of FTD does not consistently improve accuracy; when such anchoring is absent and the horizon is long, FTD's temporally gated freedom improves long-horizon forecasting. We have unified the signed-step update in Section 3.3.2 and added a task-wise applicability analysis in Section 4.2.2.

(Continued in Part 2/3.)

AuthorsOpenReview ↗

Response to Reviewer vMyZ (2/3): W2 & Q2

Response to Reviewer vMyZ (Part 2/3)

W2 & Q2: Temporal roles of PD and FTD and training–inference contributions

We thank the reviewer for the valuable suggestions. We added five new analyses of the temporal dynamics and biological relevance of PD/FTD and of FTD's training versus inference roles. Together they show that the two branches play distinct roles and that FTD acts mainly by shaping long-horizon dynamics during training.

1. Temporal division: FTD dominates early, PD dominates later. Following the reviewer's suggestion, we newly visualized and quantified branch activity along rollouts (Appendix Figure 6). The effective FTD contribution is highest early and then decays:

StageE7.5E8.0E12.0E15.0E16.5
Effective FTD/PD10.752.150.1260.0230.012

FTD thus provides rapid early corrections, while PD carries the sustained drift later.

2. Gate–change-rate alignment: beyond the prior, the gate follows lineage-level change rates. In a new lineage-resolved analysis (Appendix Figure 7), the prior-free learned gate, after within-lineage detrending, correlates positively with the change rates of all five lineages:

Gate vs. change rateρp
Linear detrending0.2610.018
Quadratic detrending0.2710.014
First differences0.1190.291

Both detrended correlations are significant; the first-difference correlation is positive but not significant.

3. Training versus inference: the FTD benefit comes mainly from dynamics learned in training. In a new comparison, Infer-off zeroes FTD only at inference, whereas Retrained −FTD is trained without FTD:

Task / MetricFullInfer-offRetrained −FTD
Strict NMAE-HVG ↓0.3600.3560.432
Strict DES ↑−0.054−0.048−0.147
Fetal NMAE-HVG ↓0.1170.2710.115
Fetal DES ↑0.551−0.0040.553
Inter-Time NMAE-HVG ↓0.3250.2180.306
Inter-Time DES ↑0.2350.4430.278
Pre-Fetal NMAE-HVG ↓0.3100.3240.291
Pre-Fetal DES ↑0.4380.3760.477

Full results in Appendix Table 21.

In Strict, retraining without FTD degrades all six metrics, with a gap widening with horizon (Figure 4, Table 20), whereas Infer-off stays close to Full, in line with the reviewer's training-time interpretation. In Fetal, Infer-off degrades all six metrics, so FTD also acts directly in the rollout. In Inter-Time and Pre-Fetal, Infer-off is mixed (3/6 and 2/6 metrics improve), so the inference-time role of FTD is task-dependent.

4. Capacity control: the gain is not explained by parameter count. In new experiments from E7.5, E11, and E13 under Strict Extrapolation, Full has better means than −FTD and a parameter-matched single drift on all six metrics at every source (full results in Appendix Table 13):

SourceFull DES−FTD DESMatched DESFull wins (vs. −FTD / matched)
E7.50.601 ± 0.0440.557 ± 0.0130.534 ± 0.0566/6 · 6/6
E110.331 ± 0.0210.185 ± 0.0460.159 ± 0.0746/6 · 6/6
E130.138 ± 0.060−0.016 ± 0.024−0.031 ± 0.0576/6 · 6/6

5. Biological relevance: FTD prioritizes known developmental drivers more than PD. In a new attribution analysis on 24 curated epithelial driver-TF annotations (Appendix Table 17), FTD exceeds PD at all four ranking thresholds under both maximum and step-mean summaries (recall):

AttributionTop 5%Top 10%Top 20%Top 30%
FTD (maximum)37.50%37.50%54.17%66.67%
FTD (step-mean)25.00%33.33%66.67%83.33%
PD (maximum)4.17%25.00%29.17%37.50%
PD (mean)12.50%25.00%41.67%50.00%

In addition, FTD (maximum) ranks 17 of 24 above PD (p = 0.027; Table 19). FTD is therefore not merely an optimization-friendly parameterization but carries developmental-regulatory information.

We therefore position the PD/FTD decomposition as an empirically supported dynamical inductive bias and have revised the Abstract, Introduction, and Conclusion accordingly.

(Continued in Part 3/3.)

AuthorsOpenReview ↗

Response to Reviewer vMyZ (3/3): W3 & Q3

Response to Reviewer vMyZ (Part 3/3)

W3 & Q3: FTD–diffusion ablation and within-lineage structure

We thank the reviewer for this suggestion. The new analyses show that FTD and diffusion contribute separably and complementarily, and that scChrono fits lineage-level and within-lineage distributions better.

1. 2×2 ablation: FTD and diffusion contribute separately. We added −Diff (PD + FTD), five seeds in four tasks (full results: Tables 1 and 22). Strict:

DynamicsNMAE-HVG ↓Sink-HVG ↓DES ↑
PD only0.460 ± 0.062371.8 ± 3.9−0.140 ± 0.139
PD + FTD (−Diff)0.365 ± 0.042354.2 ± 25.8−0.087 ± 0.067
PD + Diff (−FTD)0.444 ± 0.032369.5 ± 8.1−0.168 ± 0.079
Full0.360 ± 0.012359.1 ± 2.6−0.071 ± 0.096

FTD improves all six means with or without diffusion, so its gain is not from diffusion. With FTD present, diffusion improves five of six means and cuts seed SD (Sink-HVG 25.8 → 2.6); without it, NMAE-HVG rises sharply elsewhere (Inter-Time 0.320 → 1.150; Fetal 0.116 → 0.469).

2. Lineage level: scChrono leads other methods and fits within-lineage subpopulations better. Against PRESCIENT, TrajectoryNet, and scDiffEq, scChrono achieves the best mean on all NMAE, Sinkhorn, and CSS comparisons in all five major lineages (25/25). HVG-space results versus the best other method (full results in Appendix D.2, Table 9):

LineageNMAE-HVG: scChronoNMAE-HVG: best otherSink-HVG: scChronoSink-HVG: best other
Endothelial0.3526.910367.817389.6
Epithelial0.5326.965366.67209.6
Hematopoietic0.53011.418309.718422.0
Mesenchymal0.3803.343383.85301.7
Neural0.3098.848342.84619.8

Within lineages:

LineageModelCoverage ↑TVD ↓Within-mode Sinkhorn ↓
NeuralscChrono10.122 ± 0.018312.2 ± 4.7
Neural−FTD10.096 ± 0.042333.2 ± 8.0
HematopoieticscChrono10.420 ± 0.003311.8 ± 4.4
Hematopoietic−FTD10.420 ± 0.004315.1 ± 4.0

Both cover all defined subpopulations, and scChrono lowers within-mode Sinkhorn in both lineages, so its predictions go beyond lineage averages.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.