acceptodds
Under review as a conference paper at ICLR 2027

MeCoGen: Mechanistically Conditioned Generative Forecasting of Single-Cell Dynamics

Abstract

Forecasting single-cell populations beyond the observed time horizon is challenging because cellular states and regulatory programs can transition into unseen regimes, particularly during iPSC differentiation, while observations are limited to heterogeneous, unlabeled population snapshots. We identify a fundamental conditioning bottleneck: although modern generative models can represent complex transcriptional distributions, they require reliable future biological conditions for accurate long-horizon forecasting. To address this challenge, we introduce MeCoGen (Mechanistically Conditioned Generation), a mechanistic-data-driven hybrid framework for single-cell forecasting. MeCoGen first uses a stochastic mechanistic model, informed by a subset of core regulatory genes and interactions, to predict a biologically structured future state capturing regulatory dynamics and developmental progression. This predicted state then conditions a generative model that synthesizes future high-dimensional transcriptomic profiles. To capture the sparse and stochastic nature of single-cell expression data, the generator separately learns gene occurrence and expression magnitude using discrete diffusion and conditional flow matching, respectively. We evaluate MeCoGen on three temporal single-cell systems spanning differentiation and cellular reprogramming. Across all datasets, MeCoGen consistently outperforms recent temporal forecasting baselines, while controlled conditioning experiments demonstrate that mechanistically predicted future states substantially improve long-horizon predictions over purely data-driven extrapolation. Ablation studies further validate the benefits of separately modeling gene occurrence and expression magnitude.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell forecasting, temporal single-cell modeling, out-of-time prediction, hybrid mechanistic-data-driven approach, gene regulatory networks, developmental dynamics, conditional generation, discrete diffusion, flow matching, single-cell RNA sequencing (scRNA-seq).

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ynAtOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces MeCoGen, a hybrid mechanistic–generative framework for out-of-time forecasting of single-cell populations. The method is evaluated on human iPSC-to-cardiomyocyte differentiation, mouse fibroblast reprogramming, and human pancreatic endocrine differentiation.

Critical strengths

The proposed framework attempts to integrate biological regulatory knowledge with generative modeling, and the authors evaluate it on three temporal single-cell datasets with comparisons to existing forecasting methods.

Critical weaknesses

1. Unclear methodological presentation and insufficient specification of key components.

The writing and methodological organization are highly confusing, making it difficult to reconstruct the actual model from the main text alone. For example, Section 4.1 describes how a predefined signed GRN is incorporated into stochastic dynamics through Hill-type regulatory functions, but does not adequately explain how the GRN itself is constructed. It remains unclear what data or biological evidence are used to select the core genes and regulatory interactions, and whether the complete GRN specifications are available for reproduction. Furthermore, although Section 4.1 briefly mentions post-simulation calibration, the main text does not clearly explain the complete forecasting pipeline. Appendix E introduces two important post-mechanistic calibration components, M3 and D3, which directly modify the structured condition supplied to the downstream generator. These components constitute a substantial part of the actual method but are largely relegated to the appendix. Consequently, several key methodological details only become apparent after reading the supplementary material, making the proposed framework difficult to understand and evaluate.

2. Heavy reliance on manually specified biological priors and insufficient evidence of generalizability.

The reported results are not sufficiently convincing to establish the generalizability of the proposed framework. MeCoGen relies on prior gene-regulatory knowledge and manually curated developmental-state transition structures, which are specified separately for each dataset. This substantially limits the applicability of the approach, as applying MeCoGen to a new biological system would require reconstructing the corresponding regulatory network, developmental states, and transition topology. The paper does not demonstrate how these priors could be obtained systematically or how sensitive the model is to their quality. Moreover, even if the method is intended for a specific biological setting, such as iPSC differentiation, the experiments do not evaluate generalization across multiple independent datasets from the same setting. As a result, the current results primarily demonstrate performance on individually configured datasets, providing limited evidence that the proposed framework can generalize to new data or biological contexts.

3. Missing ablations for the post-mechanistic calibration components and insufficient isolation of the mechanistic contribution.

The post-mechanistic calibration pipeline contains two substantial components, M3 and D3, but their individual contributions are not evaluated. Since the final conditioning state is obtained only after these calibration and refinement stages, it is unclear how much of the reported improvement originates from the underlying mechanistic dynamics and how much is attributable to the subsequent data-driven corrections. In particular, M3 explicitly calibrates gene occurrence probabilities and expression magnitudes using empirical training distributions, while D3 employs an additional discrete diffusion model to refine multigene occurrence dependencies. These operations can substantially alter the expression distributions produced by the original mechanistic simulator. The authors should compare the raw mechanistic simulator (M0), M0+M3, and M0+M3+D3 under identical evaluation settings. Additional ablations removing or perturbing the GRN interactions would also be necessary to establish whether the proposed mechanistic dynamics provide meaningful predictive advantages over simpler state-transition models.

4. Potentially unfair baseline comparisons and confounding effects of gene-expression sparsity.

The comparison with existing methods raises concerns about whether the reported improvements genuinely reflect better temporal forecasting. MeCoGen explicitly separates gene occurrence and expression magnitude, whereas the compared methods primarily generate continuous expression values. As shown in Figure 2 and Appendix I.4, the generated gene occurrence rates of scIMF and scNODE are nearly 1.0, compared with approximately 0.5 in the real data. This discrepancy may arise partly from the inability of continuous generators to produce exact zeros, rather than from a failure to learn the underlying temporal dynamics. Consequently, a substantial portion of MeCoGen's advantage in zero-fraction error, marginal distribution metrics, and gene co-occurrence statistics may be attributable to its explicit sparsity modeling. The paper does not adequately disentangle improvements in temporal forecasting from improvements in expression-distribution modeling. A more convincing evaluation would include baselines with comparable zero-expression handling or training-only calibrated postprocessing, together with comparisons of positive-expression distributions and other metrics less directly affected by exact-zero occurrence.

5. Limited transcriptomic coverage and insufficient validation of the claimed high-dimensional forecasting capability.

Although the paper repeatedly emphasizes high-dimensional transcriptomic generation, the experiments only generate 200 downstream genes in addition to approximately 33–45 mechanistically modeled core genes. This represents a small subset of the transcriptome, and the downstream genes are selected from highly variable genes. Therefore, the current experiments do not establish whether the proposed framework can scale to realistically sized single-cell transcriptomic profiles containing thousands of genes. The relatively small output dimension also makes it difficult to assess the necessity and scalability of the proposed generative architecture. Experiments with substantially larger gene panels, together with analyses of computational scalability and prediction accuracy across genes with different expression characteristics, would be needed to support the broader claims.

6. Limited methodological novelty and insufficient justification for the complexity of the proposed framework.

The proposed framework largely combines established methodological components, including stochastic GRN modeling, predefined developmental-state transitions, discrete diffusion, conditional flow matching, and distributional calibration. While their integration is potentially useful, the paper does not sufficiently demonstrate why this particular combination is necessary or what fundamental methodological advance it provides over simpler alternatives. For example, the experiments do not establish whether a simpler model using the same developmental-state priors and empirical state-conditioned gene-expression distributions could achieve comparable forecasting performance. Similarly, the generator ablations demonstrate the benefit of joint occurrence and magnitude modeling, but do not establish the necessity of using discrete diffusion and flow matching specifically. Without stronger comparisons against simpler, prior-matched alternatives, the methodological contribution remains insufficiently substantiated for ICLR.

Questions for authors
  1. Construction of biological priors: How are the core regulatory genes, signed GRN interactions, and developmental-state transition topologies determined for each dataset? What specific data sources or biological evidence support these choices? Can the authors provide the complete GRN specifications and a reproducible procedure for constructing these priors?

  2. Generalizability: How would MeCoGen be applied to a previously unseen dataset or biological system for which a curated GRN and developmental-state topology are unavailable? Have the authors evaluated the model across independent datasets from the same biological setting, or investigated its sensitivity to incomplete or incorrect biological priors?

  3. Contribution of individual mechanistic components: Can the authors provide ablations comparing M0, M0+M3, and M0+M3+D3? How much of the performance gain can be attributed to the stochastic GRN dynamics, rather than the predefined developmental topology and subsequent empirical calibration? Would a simpler state-transition model with the same calibration pipeline achieve comparable results?

  4. Fairness of baseline comparisons: Given that the baseline methods produce almost entirely nonzero expression profiles, how much of MeCoGen's improvement is attributable to its explicit modeling of gene-expression sparsity? Can the authors provide comparisons using consistent zero-expression handling or appropriately calibrated continuous-generation baselines?

  5. Scalability and architectural necessity: Why are only 200 downstream genes considered? How does the framework scale to thousands of genes? Additionally, can the authors justify the necessity of combining discrete diffusion and conditional flow matching by comparing against simpler conditional generative models under the same mechanistic conditioning?

Reviewer_JuX8OpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents MeCoGen, a hybrid framework for forecasting single-cell population distributions beyond observed time points. A stochastic mechanistic model extrapolates core regulatory gene expression and developmental states, which subsequently condition a transcriptome generator. The generator separately models gene occurrence using discrete diffusion and positive expression magnitudes using conditional flow matching. Experiments on three temporal single-cell datasets report improvements over existing forecasting baselines, alongside conditioning and generator ablations.

Critical strengths
  • The paper addresses an important problem and presents a scientifically motivated framework that combines mechanistic extrapolation with conditional generation.
  • The reported improvements span three biological systems and multiple distributional, marginal, and sparsity metrics, providing encouraging evidence for the framework’s empirical utility.
  • The controlled conditioning experiment holds the downstream generator fixed, providing useful evidence that the quality of future conditioning states materially affects forecasting performance.
Critical weaknesses

W1. The provenance of biological priors and state supervision is insufficiently transparent.
The model prespecifies developmental topology, terminal fates, and regulatory interactions, and relies on cell-state labels for fitting and calibration. However, the paper does not fully document state coverage within each temporal split or the provenance of the annotations and prior configurations. For states absent or sparsely represented during training, it remains unclear whether the corresponding dynamics and calibration statistics can be reliably determined. Several topology and model choices are also dataset-specific, with insufficient justification for their selection. Consequently, the benefits of mechanistic extrapolation are difficult to distinguish from those attributable to additional biological information.

W2. The independent contribution of mechanistic dynamics after post-processing has not been sufficiently established.
M3/D3 substantially reconstruct core-gene occurrence patterns and introduce empirical training expression distributions and quantile calibration. Although some mechanistically predicted magnitudes are retained, the existing ablations do not adequately establish whether this dynamical information is necessary for the final predictions. It therefore remains unclear how much of the improvement arises from gene-regulatory mechanisms, rather than state-composition forecasting and state-conditioned expression reconstruction. This directly affects the evidence supporting the paper’s central claim that mechanistic conditioning improves extrapolation.

W3. The baseline comparisons do not sufficiently isolate the benefits of mechanistic structure from those of the generative model.
MeCoGen uses curated biological priors, state supervision, and explicit sparsity modeling, whereas the external baselines differ in both available information and output distribution structure. The nearly complete absence of zeros in baseline predictions suggests that sparsity mismatch may substantially affect several evaluation metrics. The current results therefore support the empirical advantage of the complete system but do not clearly attribute that advantage to mechanistic extrapolation. Although Table 2 fixes the downstream generator, differences in state priors and conditioning calibration remain insufficiently controlled. The adequacy of baseline training and the suitability of their configurations are also insufficiently validated.

W4. The interpretation of evaluation metrics, statistical uncertainty, and consistency of reported results raise concerns.
Closeness to the Real–Split Sinkhorn value does not directly establish greater predictive accuracy. In Table 3, Joint-Occ. substantially outperforms the complete model in �0��0� while performing worse on other metrics, and this ranking discrepancy is not adequately explained. The use of reductions in raw �1��1� to summarize forecasting improvements also requires reconciliation with the discussion of this metric’s limitations. Furthermore, evaluation sample sizes and the source of uncertainty estimates in the main table are unclear, while the reported repetitions over generation seeds cannot capture variability across independent fits. Cross-table discrepancies also remain unexplained: for example, the four-split average of scIMF’s Zero error in Table 7 is 0.642, compared with 0.550 in Table 1.

W5. The extrapolation assumptions of the conditional generator lack empirical support.
The model assumes that temporal and experimental effects on the remaining genes are fully mediated by core expression and discrete state, and uses a conditional generator without explicit time dependence. However, the paper does not establish that this state representation is sufficiently informative for forecasting or adequately characterize generation when future conditions depart from the training support. The existing analyses do not clearly distinguish errors in forecasting the conditioning state from failures of the conditional generation relationship. Consequently, the claims of robust prediction under unseen regulatory regimes and distribution shifts remain insufficiently supported.

W6. The experimental scale and system-specific design limit the generality of the method.
The experiments cover only approximately 233–245 genes, which is insufficient to establish transcriptome-scale predictive quality and computational scalability. Each dataset also requires distinct regulatory networks, state topologies, and simulation configurations, while sensitivity to misspecified priors is insufficiently evaluated. The results therefore primarily demonstrate an effective framework tailored to particular biological systems, and the transferable, reusable methodological contribution remains unclear. Incomplete prior configurations and implementation details further limit independent verification of these conclusions.

Questions for authors

Q1. Provenance and coverage of priors. How were the developmental topology, regulatory priors, state annotations, and dataset-specific configurations determined? Were these choices independent of the held-out populations, and how are states absent or sparsely represented during training handled in fitting and calibration?

Q2 . Contribution of mechanistic dynamics. What evidence establishes that mechanistically predicted expression contributes materially after M3/D3 calibration, beyond the benefits of forecasting state proportions and reconstructing state-conditioned expression distributions?

Q3. Comparability of baselines. How are differences in biological priors, state supervision, sparsity modeling, and conditioning calibration accounted for when attributing improvements to mechanistic extrapolation? What evidence supports the adequacy of baseline training and configurations?

Q4. Evaluation and uncertainty. What justifies interpreting closeness to the Real–Split Sinkhorn value as improved fidelity, particularly given the conflicting rankings in Table 3? What sample sizes and sources of randomness underlie the reported metrics and error bars? How is the scIMF Zero-error discrepancy between Tables 1 and 7 explained?

Q5. Validity of conditional extrapolation. What evidence supports the assumption that core expression and discrete state sufficiently mediate temporal and experimental effects on the remaining genes? How are conditioning-state prediction errors distinguished from failures of the conditional generator when future conditions depart from the training support?

Q6. Generality and scalability. To what extent do the results extend beyond the evaluated gene panels and dataset-specific configurations? How sensitive is performance to misspecified biological priors, and which components can be transferred across biological systems without substantial manual redesign?

1 comment

Thank you for the detailed responses and the substantial additional experiments. The rebuttal meaningfully strengthens the empirical support for the framework and addresses several of my original concerns.

In particular, applying the same M3 calibration to both conditioning predictors while holding the downstream generator fixed provides stronger evidence that the forecasting advantage is not explained by calibration alone. The same-topology Markov-chain comparison and regulatory perturbations further support the value of the proposed state dynamics beyond the supplied developmental graph. The sparsity-calibrated baseline comparisons and positive-expression evaluations also substantially alleviate my concern that the improvements primarily reflect differences in zero handling.

I also appreciate the expansion to 2,000 downstream genes, the independent training repetitions, and the explanations and corrections concerning cross-table discrepancies. Removing D3 simplifies the framework and makes the contribution of the remaining components clearer.

Some reservations remain. Application to a new biological system still requires system-specific regulatory and developmental priors, and the proposed curation workflow does not yet establish transferability across independent datasets. Moreover, the conditional generator’s reliability for states absent from training remains uncertain, as the authors acknowledge. The evidence is therefore stronger for forecasting within a specified biological setting than for general extrapolation to previously unseen cellular states. The methodological contribution also remains primarily an integration of established components, although the additional controlled experiments now provide a more convincing justification for that integration.

Overall, the additional evidence has sufficiently alleviated my main concerns about contribution attribution and empirical validation. I therefore raise my rating from 2 (Weak Rejection) to 3 (Weak Acceptance), while retaining the above reservations regarding generalizability and the scope of the extrapolation claims.

Reviewer_53nJOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

MeCoGen predicts cell-population distributions beyond the observed time range from cell snapshots collected at a limited number of time points. It first extrapolates core biological states using a mechanistic model of gene regulation, state transitions, and population changes. Discrete diffusion then generates which genes have nonzero expression, while conditional flow matching generates expression magnitudes. Experiments on cardiomyocyte differentiation, pancreatic endocrine differentiation, and cellular reprogramming predict core genes and 200 additional genes, reporting improvements over the compared baselines on several distributional and sparsity metrics. Experiments that hold the generator fixed while changing the future conditioning states further support the importance of conditioning quality for prediction.

Critical strengths
  1. The division of labor between mechanistic extrapolation and high-dimensional generation is well motivated. Combining a mechanistic predictor of future states with separate modeling of expression sparsity and magnitude offers a distinctive approach to prediction beyond the observed time range.

  2. Results across several biological systems and evaluation metrics are complemented by experiments that hold the generator fixed. Together, these provide substantive support for the overall approach and for the importance of future conditioning states as a bottleneck in prediction.

Critical weaknesses
  1. The experiments do not fully isolate the benefit of mechanistic dynamics. Mechanistic and data-driven conditioning differ in their state priors and calibration procedures, so the entire improvement cannot be attributed to the mechanistic component. This leaves uncertainty about which part of the conditioning pipeline accounts for the gain.

  2. The relationship between two key experimental configurations is unclear. For iPSC-CM, Table 1 reports W₂ = 57.40, whereas the average over the two time points in Table 2 is approximately 46.97. Differences in the generator, calibration, or evaluation could explain these values, but they need to be specified before the two tables can be interpreted as connected evidence.

Questions for authors
  1. Attribution to mechanistic dynamics: Are the mechanistic and data-driven conditioning methods compared using matched state priors and calibration procedures? If not, what controlled evidence separates the contribution of mechanistic dynamics from those other differences?

  2. Clarification of Tables 1 and 2: What accounts for W₂ = 57.40 in Table 1 versus an average of approximately 46.97 across the two time points in Table 2 for iPSC-CM? Please specify any differences in the generator, calibration, evaluation time points, or aggregation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.