acceptodds
Under review as a conference paper at ICLR 2027

PROVE: Training-Free Reference-Guided Pareto-Consensus Velocity Editing for Cross-Context Single-Cell Perturbation Prediction

Abstract

Flow-matching models learn condition-dependent velocity fields from observed context-perturbation pairs. In cross-context covariate transfer, however, the target pair is unobserved, and the training objective provides no direct supervision for its response prediction. The resulting conditional extrapolation may collapse toward the control state, align with another perturbation, or drift beyond the cross-context variation supported by available data. Although responses from related contexts do not uniquely determine the missing target distribution, they provide useful reference constraints for detecting these failures. Motivated by this, we introduce PaReto-COnsensus Velocity Editing (PROVE), a training-free inference-time method for correcting reference-inconsistent predictions from frozen flow-matching models. At each sampling step, PROVE obtains a final-state estimate from the backbone velocity and diagnoses three patterns including effect collapse, perturbation confusion, and context drift. It constructs a local correction for each active pattern, combines the frozen and corrective directions using an MGDA-inspired minimum-norm consensus, and redirects the velocity without changing its instantaneous norm. Experiments across cross-cell-line and cross-state transfer settings show that PROVE improves the original samplers across multiple back-bones and strengthens the recovery of perturbation-specific differential-expression and pathway-level effects.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single cell perturbation, Flow matching, Training-free, Multi-objective optimization

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_cuf9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PROVE improves cross-context single cell perturbation predictions by modifying the sampling direction of a pre-trained model (no retraining is required at all). At each generation step it predicts the final cell and then compares that to reference statistics derived from the training data: the mean response and the variation between context-level mean responses for the same perturbation in other contexts; the responses to other perturbations in the target context. These comparisons flag three possible inconsistencies with the references: an effect that is too small (effect collapse), a prediction more aligned with another perturbation’s response direction than with the target reference direction (perturbation confusion), and a response that is very different from what we’d expect from other contexts (context drift). PROVE generates a correction direction for each of these inconsistencies and then combines those with the model’s own direction using a minimum norm (MGDA style) consensus. It then steers the velocity towards the resulting point whilst maintaining its magnitude. Experiments cover three flow models (scDFM, CellFlow, CFM-GP) on three datasets (Replogle, Frangieh21, Srivatsan20), plus additional experiments on two diffusion models, ablations, pathway analysis and comparisons against the reference-only predictors.

Critical strengths
  • Consistent improvements as a plug-in. PROVE requires no retraining. Table 1 shows gains on all five perturbation-specific metrics across all nine model/dataset combinations. Results over three seeds (with standard deviations) are in the appendix and the diffusion model and pathway analyses provide further support.

  • Transparent comparisons and limits. Tables 6-8 make a direct comparison of PROVE with both the reference transfer and the regression baselines (even in cases where the comparison is not in PROVE’s favour). The authors also split agreement with the reference patterns from actual biological accuracy (Section 4.2.2, Appendix C.2) which makes the evidence much easier to assess.

Critical weaknesses
  • The benefit of the full editing procedure over simpler alternatives is not fully established. The ablation study does compare minimum norm consensus with normalized arithmetic averaging, but only for CellFlow on Replogle. The ablation doesn’t include comparisons with endpoint only correction or with direct interpolation between the backbone predictions and the reference responses (as acknowledged on p9 lines 447–449). These would have helped establish whether editing at each sampling step was required.

  • Table 5 shows that the effect collapse check activates on the ground-truth mean responses of 25.6% of the held-out context–perturbation pairs in Srivatsan20. The paper admits that performance is not assessed separately on those cases. The stats on variation between context means are also applied to the individual generated cells. The aggregate improvements therefore don’t tell us whether the guidance is in fact harming valid responses that differ from the references or whether it’s actually reducing the meaningful variation from cell to cell.

Questions for authors
  • Does PROVE perform better than correcting only the final prediction or taking a weighted average of the model’s prediction and a reference-based prediction, when all three approaches use the same reference data and comparable parameter tuning?

  • Does PROVE improve or degrade the predictions on the target pairs whose ground truth mean responses are what cause the reference checks to be triggered (i.e. the 25.6% that trigger the effect collapse check in Srivatsan20)? What's the effect on within condition variability?

  • Do the component ablation results generalise beyond CellFlow on Replogle? What's the actual measured inference time overhead relative to the original sampler?

  • The redirection preserves the magnitude of velocity and Appendix A.7 states that the correction is zero when the consensus direction is equal to the backbone direction. How does PROVE deal with effect collapse in this aligned case where the backbone is pointing in the right direction but the effect is too weak?

  • Even though you are following the official protocol, why do the vanilla scDFM (Frangieh21, Srivatsan20) and the retrained Squidiff (Replogle) give negative or near zero correlations? Can you confirm that the preprocessing is compatible and that they do in fact give the expected baseline performance?

9 comments

We thank the reviewer for their careful assessment and constructive feedback. We address each question below using targeted experiments and diagnostic analyses.

Q1. Benefits of Full PROVE over Endpoint-only Correction and Interpolation

Appendix C.4 (Tables 6–8) shows that direct reference transfer is competitive on Replogle. We further compared Full PROVE with endpoint-only correction and backbone–reference interpolation using CellFlow on all three datasets. All methods used identical references and paired sampling seeds (0, 42, 2026). Interpolation weights λ∈0.2,0.4,0.6,0.8,1.0\lambda\in\\{0.2,0.4,0.6,0.8,1.0\\} were selected on 20 training-side pseudo-held-out pairs per dataset using a prespecified perturbation-response score. The selected weights were 0.6/0.6/0.2 for Replogle/Frangieh21/Srivatsan20; no test outcomes informed selection.

Replogle

MethodMAE↓Pearson-Δ↑DE-Spearman↑DS↑MMD-GEX↓MMD-PCA↓AUPRC-DE↑DEG Recall↑
scDFM+PROVE0.12740.29940.37590.61990.06200.16810.30050.7130
CellFlow+PROVE0.06860.29180.57070.62530.00760.02710.29750.2953
CFM-GP+PROVE0.06490.31830.41190.64110.01120.03300.31520.3430
Endpoint-only0.06220.25960.49800.58390.00790.02790.24540.0953
Interpolation0.07020.37580.54280.63850.14480.17600.30360.3540
CC-Mean0.06910.32060.59450.72530.00950.02640.31210.3243
TwoWay-Ridge0.06990.36560.63930.69990.00800.02720.30050.3203
Analog Residual0.09650.26670.54080.76780.02760.05230.26350.5287

Frangieh21

MethodMAE↓Pearson-Δ↑DE-Spearman↑DS↑MMD-GEX↓MMD-PCA↓AUPRC-DE↑DEG Recall↑
scDFM+PROVE0.10140.46370.59670.62720.08050.16430.51980.7007
CellFlow+PROVE0.02240.85830.92810.58780.00150.00910.72600.7021
CFM-GP+PROVE0.02250.80370.91090.60080.00460.01870.62870.6915
Endpoint-only0.02060.19050.76870.56370.00240.00880.54300.6136
Interpolation0.01500.23070.80100.55140.13890.16780.52600.2354
CC-Mean0.02040.18650.71110.57910.00270.00580.52810.5405
TwoWay-Ridge0.01770.09940.48620.51780.00110.00460.12170.0014
Analog Residual0.02330.14030.79000.59920.00410.00580.48860.6148

Srivatsan20

MethodMAE↓Pearson-Δ↑DE-Spearman↑DS↑MMD-GEX↓MMD-PCA↓AUPRC-DE↑DEG Recall↑
scDFM+PROVE0.05980.80590.83220.85190.12190.26930.72420.5762
CellFlow+PROVE0.01280.87720.87540.85970.01000.12510.66900.7523
CFM-GP+PROVE0.01470.84200.84950.86100.12400.06420.62060.7412
Endpoint-only0.012920.135220.265960.560930.010590.105810.232990.35252
Interpolation0.011580.155890.276640.568640.027570.132370.251400.32706
CC-Mean0.01380.09850.21300.51550.00840.04130.15240.1470
TwoWay-Ridge0.01100.15640.28470.56270.00560.03100.19960.0363
Analog Residual0.01690.26490.38970.80180.01110.03950.27280.4453

Across perturbation-specific metrics, Full PROVE exceeds Endpoint-only in 15/15 and validation-tuned Interpolation in 11/15 comparisons, supporting a benefit from stepwise editing beyond one-shot correction or simple reference interpolation.

Q2. Performance on Ground-truth Prior-violating Conditions and Preservation of Within-condition Variability

(a) Performance on ground-truth prior-violating conditions

To address the concern that reference priors may incorrectly modify genuine context-specific biological responses, we conducted a ground-truth prior-violation stratification analysis.

Specifically, we identified target conditions whose ground-truth responses violate at least one reference prior (EC, PC, or CD) and directly evaluated PROVE's predictive performance on these conditions. Ground-truth responses were used exclusively for post-hoc stratification and evaluation, and were not involved in inference or reference construction.

The results are reported below (Δ=PROVE−Vanilla\Delta=\mathrm{PROVE}-\mathrm{Vanilla}).

DatasetΔ Pearson-ΔΔ DE-SpearmanΔ DSΔ AUPRC-DE
Replogle+0.0378+0.1427+0.0682+0.0739
Frangieh21+0.0906+0.4120+0.0584+0.5058
Srivatsan20-0.0112+0.0817+0.0206+0.0202

Even when ground-truth responses violate the reference priors, PROVE achieves positive improvements in 11 out of 12 dataset–metric comparisons.

These results indicate that disagreement between genuine biological responses and reference priors does not necessarily cause PROVE to suppress valid context-specific biological effects. Even under these more challenging conditions, PROVE continues to improve most perturbation-specific prediction metrics.

We note that this analysis uses the broader any-prior violation definition, which includes the EC-violating conditions highlighted by the reviewer, rather than separately reporting results on only the 25.6% EC-triggered conditions in Srivatsan20.

(b) Preservation of within-condition variability

To evaluate whether PROVE compromises genuine within-condition variability, we assessed three gene-wise variance consistency metrics:

  • Variance Spearman: Rank correlation between predicted and ground-truth gene-wise variances.
  • Top-20% Jaccard: Overlap between the predicted and ground-truth sets of the top 20% most variable genes.
  • NDCG@20%: How accurately highly variable genes are ranked near the top.

The changes relative to Vanilla are reported below.

DatasetΔ Variance SpearmanΔ Top-20% JaccardΔ NDCG@20%
Replogle+0.0005+0.0080+0.0097
Frangieh21-0.0040+0.0080+0.0026
Srivatsan20+0.0012+0.0011-0.0009

Overall, PROVE maintains gene-wise variability structure comparable to that of the vanilla backbone across all three datasets, without introducing substantial or systematic degradation.

Q3. Generalization of Component Ablations and Measured Inference-time Overhead

(a) Component ablation

The component ablation study in the original submission was primarily conducted using CellFlow on Replogle. To examine whether these observations generalize to another dataset, we performed leave-one-component-out ablations using CellFlow on Srivatsan20.

The results are presented below.

MethodMAE↓Pearson-Δ↑DE-Spearman↑DS↑MMD-GEX↓MMD-PCA↓AUPRC-DE↑DEG Recall↑
Full PROVE (Table 1)0.01280.87720.87540.85970.01000.12510.66900.7523
w/o EC0.01350.87380.87500.84510.01200.12520.66000.7488
w/o PC0.01300.87560.85450.85700.01170.13390.65180.7086
w/o CD0.01410.87170.87230.86030.01090.13010.66540.7523
w/o Pareto0.01380.87570.87380.86060.01220.12520.66720.7520
w/o final norm0.01370.87700.87540.85950.01200.12510.66900.7523

These results support the effectiveness of Full PROVE in this setting. In particular, removing either EC or PC degrades all eight metrics, supporting the role of these reference constraints in recovering perturbation responses.

Removing the other components also fails to yield consistent improvements, consistent with their complementary roles within the complete algorithm.

This experiment extends the original ablation observations from Replogle to Srivatsan20 using CellFlow. Component ablations on the other backbones are ongoing and will be provided subsequently.

(b) Measured inference-time overhead

We measured the mean per-cell wall-clock inference time across three backbones and three datasets, covering all nine combinations. For each combination, latency was averaged over the same 1,000 source cells and three paired repetitions. The table below reports mean per-cell sampling time.

BackboneDatasetVanilla (s/cell)PROVE (s/cell)Extra time (s/cell)Ratio
scDFMReplogle3.9184.432+0.5151.13×
scDFMFrangieh213.8674.507+0.6391.17×
scDFMSrivatsan203.8794.538+0.6591.17×
CellFlowReplogle0.0330.472+0.43914.22×
CellFlowFrangieh210.0350.304+0.2688.63×
CellFlowSrivatsan200.0340.202+0.1685.95×
CFM-GPReplogle0.00280.363+0.361130.99×
CFM-GPFrangieh210.00280.321+0.319115.06×
CFM-GPSrivatsan200.00290.232+0.22980.28×

The results show that the relative inference-time overhead is 1.13–1.17× for scDFM, 5.95–14.22× for CellFlow, and 80.28–130.99× for CFM-GP, whose original sampler is extremely fast.

Despite these differences in relative overhead, PROVE adds an average of 0.168–0.659 seconds of sampling time per cell. Therefore, its inference overhead is strongly backbone-dependent and should be assessed using both per-cell latency and relative slowdown.

Reviewer_wzBEOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a way to do inference time steering of single cell perturbation generative model such as to avoid common identified pitfalls : effect collapse, perturbation confusion, and context drift. The authors can correct the velocity at inference time and show that it results in improved perturbation prediction performance across a wide variety of base models.

Critical strengths

The proposed approach is novel as it directly attempts at correcting identified failure modes of perturbation generative models. The results show the proposed mechanism improves performance over a wide variety of based models, making it a versatile approach for perturbation modeling.

Critical weaknesses

The main weakness of this paper is whether the corrections applied may eventually ignore or dampen genuine biological signal. The authors assume that a prediction is suspicious if it has a substantially smaller effect than the average reference response, or aligns more with another perturbation or deviates too far from the mean effect. But these differences in perturbation responses are precisely what the model is supposed to learn.

Additionally, the Pareto consensus procedure can prevent corrections when they are most important. Say the correction is exactly in the opposite direction of the frozen velocity vector. PROVE would then do no correction (zero vector in convex hull) while you can argue that a correction is strongly needed in that case.

Questions for authors

Could the authors stratify performance according to whether the ground-truth response satisfies the reference constraints? This would help establish whether the method improves predictions because its reference assumptions are appropriate, or whether it systematically sacrifices certain biologically valid context-specific responses.

How do the authors justify the pathological cases identified above in terms of the Pareto consensus procedure ?

5 comments

We thank the reviewer for their careful assessment and constructive feedback. To address the two questions raised, we conducted a post-hoc analysis of predictive performance when ground-truth responses violate the reference priors, as well as a mechanistic analysis of Pareto consensus under conflicting correction directions. We respond to each point below.

Q1. Performance When Ground-truth Responses Violate Reference Constraints

To assess whether PROVE suppresses genuine context-specific biological responses, we stratified target conditions according to whether their ground-truth responses violate the reference priors.

Specifically, we identified conditions whose ground-truth responses violate at least one of the EC, PC, or CD priors (any-prior violation). Using CellFlow as the backbone, we evaluated PROVE on these conditions in Replogle, Frangieh21, and Srivatsan20. Ground truth was used solely for post-hoc stratification and evaluation; it was not used to construct references or guide inference.

The table reports performance changes relative to Vanilla (Δ=PROVE−Vanilla\Delta=\mathrm{PROVE}-\mathrm{Vanilla}).

DatasetΔ Pearson-ΔΔ DE-SpearmanΔ DSΔ AUPRC-DE
Replogle+0.0378+0.1427+0.0682+0.0739
Frangieh21+0.0906+0.4120+0.0584+0.5058
Srivatsan20-0.0112+0.0817+0.0206+0.0202

Among conditions whose ground-truth responses violate at least one reference prior, PROVE improves 11 of the 12 dataset-metric comparisons. The 25.6% EC-triggered Srivatsan20 conditions highlighted in Appendix C.2 are included in this broader any-prior subgroup, rather than evaluated as an EC-only subset. The gains therefore do not require strict agreement between genuine responses and all reference priors.

We also assessed whether PROVE affects genuine within-condition variability using three gene-wise variance-consistency metrics:

  • Variance Spearman: Rank correlation between predicted and ground-truth gene-wise variances.
  • Top-20% Jaccard: Overlap between the predicted and ground-truth sets of the top 20% most variable genes.
  • NDCG@20%: Ranking accuracy for highly variable genes near the top of the list.

Changes relative to Vanilla are shown below.

DatasetΔ Variance SpearmanΔ Top-20% JaccardΔ NDCG@20%
Replogle+0.0005+0.0080+0.0097
Frangieh21-0.0040+0.0080+0.0026
Srivatsan20+0.0012+0.0011-0.0009

Across the three datasets, PROVE preserves gene-wise variability structure at a level comparable to the Vanilla backbone, without a pronounced or systematic decline in these metrics.

Together, these findings show improved perturbation-specific prediction in the any-prior-violating subgroup and comparable relative gene-wise variability structure in the evaluated settings.

Q2. Pareto Consensus under Conflicting Correction Directions

The reviewer notes that strongly conflicting candidate correction directions could yield a zero vector under a minimum-norm Pareto consensus, leaving a violated reference constraint without an effective correction.

Pareto consensus is intended to identify a direction that accommodates multiple correction objectives. When those objectives are irreconcilable, a minimum-norm consensus may act conservatively rather than force a correction that conflicts with other constraints.

To examine the practical impact of this mechanism, we analyzed all nine combinations of three backbones (CellFlow, CFM-GP, and scDFM) and three datasets (Replogle, Frangieh21, and Srivatsan20), using three random seeds (0, 42, and 2026) and all three reference constraints (EC, PC, and CD).

We report three diagnostic measures:

  1. Strong Conflict Rate: The proportion of candidate correction directions exhibiting strong directional conflict.
  2. Near-zero Consensus Rate: The proportion for which the relative norm of the final Pareto consensus direction is near zero.
  3. Severe-violation Uncorrected Rate: The proportion with a near-zero velocity correction among conditions in the highest quartile of reference-violation magnitude.

The results are as follows.

BackboneDatasetStrong Conflict (%)Near-zero Consensus (%)Severe-violation Uncorrected (%)
CellFlowReplogle0.0000.0000.000
CellFlowFrangieh210.0000.0000.000
CellFlowSrivatsan200.0000.0000.000
CFM-GPReplogle0.0000.0000.000
CFM-GPFrangieh210.0080.0000.000
CFM-GPSrivatsan200.1670.0000.000
scDFMReplogle0.0000.1090.311
scDFMFrangieh210.0000.7990.601
scDFMSrivatsan200.0000.5200.414

Strong conflicts are rare (at most 0.167%), and near-zero consensus is observed only with scDFM (0.109% to 0.799%). Even among the highest-severity violations, the uncorrected rate remains at most 0.601%, suggesting that this failure mode is uncommon in the settings studied.

For a minimum-norm convex combination, cancellation depends on the geometry of all active directions, not just the cosine of one pair. Strong pairwise conflict and near-zero consensus are therefore distinct events.

We will include these experiments and the corresponding mechanistic analysis in the revised manuscript. We thank the reviewer again for their constructive comments.

For the biological-signal preservation concern, PROVE improves 11/12 comparisons on ground-truth prior-violating conditions. Our additional Pareto diagnostics also show that strong directional conflicts and severe-violation uncorrected cases are rare across the nine backbone–dataset combinations.

Please let us know if any concerns remain.

Reviewer_WDa3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PROVE, a method for predicting single-cell perturbations across different contexts. When target combinations—such as "knockdown of gene g in the HepG2 cell line"—lack training data, models must rely on extrapolation. This often leads to three types of failure: collapsing predictions toward the control state, misidentifying the perturbation, or drifting into regions of the state space never supported by the reference data. PROVE avoids model retraining; instead, it corrects a frozen flow-matching sampler during the inference phase. At each sampling step, the method estimates the final state based on the current velocity field. This estimate is then evaluated against three types of reference statistics: 1) the average effect of gene g across other cell lines; 2) the dispersion (variability) of these effects; and 3) the directional patterns of other perturbations within the target cell line. These statistics are used to optimize the correction direction for the velocity field. Specifically, PROVE employs the Multiple Gradient Descent Algorithm (MGDA) to identify a "consensus" direction—one that maintains a positive correlation with both the model’s original predicted direction and the correction vectors derived from the reference statistics. The final state is then adjusted toward this consensus direction, with the magnitude of the shift determined by the degree of deviation and directional alignment, while the original velocity magnitude is preserved.

Critical strengths
  • PROVE frames cross-context prediction as "extrapolation under reference constraints." It posits that while reference data may not pinpoint the exact ground truth, it can effectively rule out clearly inconsistent predictions. The three identified failure modes—effect collapse, perturbation confusion, and context drift—align with phenomena frequently observed in benchmarks like PerturBench and Systema, making the diagnostic framework both intuitive and interpretable.

  • Because PROVE does not require retraining and relies solely on end-state estimation during sampling, it can be seamlessly integrated with various flow-matching and diffusion-based backbones. This offers significant practical value in a field already populated by numerous pre-trained models.

  • The three reference-based constraints define convex sets, allowing for closed-form solutions for the necessary corrections. By incorporating the model’s original prediction into the MGDA framework, the method ensures that the adjusted direction remains positively correlated with the model’s initial output. Furthermore, by preserving the velocity norm and capping the magnitude of the adjustment, the approach ensures the sampler remains stable and robust. - Perturbation-specific metrics—such as Pearson-Δ, DE-Spearman, DS, AUPRC-DE, and DEG Recall—show improvements across all nine combinations of backbones and datasets.

Critical weaknesses
  • All three rules rely on Euclidean distance and cosine similarity; however, because scDFM operates in a normalized, high-dimensional (thousands of dimensions) gene space, while CellFlow and CFM-GP utilize their own respective preprocessed spaces, the geometric interpretations are not consistent. This warrants further clarification. In a normalized gene space, a large number of low-expression, high-noise genes are weighted equally to key response-related genes, which may disproportionately influence the calculation of distances and directions.

  • The experimental evaluation lacks baselines that utilize the same reference information but in different ways (e.g., weighted sums of correction vectors, mean-shift toward the target/endpoint, or standard flow guidance).

  • Principled guidance approaches (such as posterior sampling or Sequential Monte Carlo) are grounded in well-defined target distributions, allowing for an analysis of the sources of bias. In contrast, PROVE’s heuristic velocity-editing approach lacks an associated target distribution, making it unclear what distribution the resulting cells actually follow. For a method aimed at generating a specific distribution, any observed improvements in distributional metrics—such as MMD—can only be regarded as empirical phenomena, lacking a formal theoretical basis.

  • Since the "acceptable regions" for all three rules are centered around the reference mean pgp_g, the correction process effectively shifts predictions toward the "control plus reference mean" baseline. Consequently, the final predictions fall somewhere between the vanilla model output and this baseline. The upper bound of PROVE’s performance is, therefore, heavily dependent on the quality of the cross-context mean itself.

Questions for authors

See my Weaknesses.

7 comments

We thank the reviewer for their careful assessment and constructive feedback. We address the four concerns below by clarifying the backbone-specific geometry and theoretical scope of PROVE, and by providing additional comparisons and analyses of reference-quality dependence.

W1. Geometric Consistency Across Backbones and High-dimensional Gene Noise

We thank the reviewer for raising this concern. We would like to clarify that PROVE is designed as a plug-and-play inference-time method that operates in each backbone's existing feature space, without requiring a unified representation across different models.

As described in Section 3.1, Section 3.5, and Appendix A.8, PROVE uses the mapping ϕ\phi to construct references, evaluate geometric constraints, and compute corrections in the corresponding response space. The resulting corrections are mapped back to the backbone state space through MϕM_\phi. Specifically:

  • scDFM: Reference statistics and corrections are computed in standardized gene-expression space, with inverse feature scaling used to map correction displacements back to the model state.
  • CellFlow and CFM-GP: PROVE operates directly in their respective preprocessed gene-expression spaces, using identity displacement mappings.

Thus, PROVE does not directly compare distances or directions across different backbone feature spaces. For each backbone, reference statistics and predicted responses are represented consistently within the same space.

Regarding high-dimensional noise, PROVE follows the original preprocessing and feature representations of each backbone. Single-cell preprocessing pipelines commonly employ quality control, expression normalization, and feature selection to mitigate technical noise. As a training-free inference-time module, PROVE does not introduce additional gene filtering, learnable feature weights, or a new representation space.

Furthermore, Table 1 of the original submission evaluates PROVE across three backbones with different feature representations (scDFM, CellFlow, and CFM-GP) on three datasets (Replogle, Frangieh21, and Srivatsan20). PROVE improves all five perturbation-specific metrics across all nine backbone–dataset combinations.

These results demonstrate that PROVE can consistently improve perturbation-specific prediction across different backbone representations without requiring additional feature learning or retraining. We will further clarify this backbone-specific geometric formulation and its relationship to the plug-and-play design in the revised manuscript.

W2. Comparisons with Alternatives Using the Same Reference Information

We thank the reviewer for suggesting additional reference-matched baselines.

Appendix C.4 (Tables 6–8) of the original submission already compared PROVE-guided models against three reference-only predictors: CC-Mean, TwoWay-Ridge, and Analog Residual. Direct reference transfer is competitive on Replogle, whereas PROVE-guided models achieve stronger perturbation-specific recovery on Frangieh21 and Srivatsan20.

To further distinguish the contribution of the complete velocity-editing procedure from the use of reference information alone, we compared Full PROVE with two simpler alternatives using CellFlow on all three datasets:

  1. Endpoint-only Correction: Applies reference-guided correction only to the final generated prediction.
  2. Backbone–Reference Interpolation: Directly interpolates between backbone predictions and reference-based predictions.

All three approaches use the same reference information. For Interpolation, we evaluated λ∈{0.2,0.4,0.6,0.8,1.0}\lambda\in\{0.2,0.4,0.6,0.8,1.0\} on 20 training-side pseudo-held-out condition pairs per dataset and selected the weight using a prespecified composite score based on perturbation-response metrics. The selected weights were 0.6 for Replogle, 0.6 for Frangieh21, and 0.2 for Srivatsan20. All parameters were fixed before test evaluation, without using held-out test responses for parameter selection.

The results are summarized below.

Replogle

MethodMAE↓Pearson-Δ↑DE-Spearman↑DS↑MMD-GEX↓MMD-PCA↓AUPRC-DE↑DEG Recall↑
scDFM + PROVE0.12740.29940.37590.61990.06200.16810.30050.7130
CellFlow + PROVE0.06860.29180.57070.62530.00760.02710.29750.2953
CFM-GP + PROVE0.06490.31830.41190.64110.01120.03300.31520.3430
Endpoint-only0.06220.25960.49800.58390.00790.02790.24540.0953
Interpolation0.07020.37580.54280.63850.14480.17600.30360.3540
CC-Mean0.06910.32060.59450.72530.00950.02640.31210.3243
TwoWay-Ridge0.06990.36560.63930.69990.00800.02720.30050.3203
Analog Residual0.09650.26670.54080.76780.02760.05230.26350.5287

Frangieh21

MethodMAE↓Pearson-Δ↑DE-Spearman↑DS↑MMD-GEX↓MMD-PCA↓AUPRC-DE↑DEG Recall↑
scDFM + PROVE0.10140.46370.59670.62720.08050.16430.51980.7007
CellFlow + PROVE0.02240.85830.92810.58780.00150.00910.72600.7021
CFM-GP + PROVE0.02250.80370.91090.60080.00460.01870.62870.6915
Endpoint-only0.02060.19050.76870.56370.00240.00880.54300.6136
Interpolation0.01500.23070.80100.55140.13890.16780.52600.2354
CC-Mean0.02040.18650.71110.57910.00270.00580.52810.5405
TwoWay-Ridge0.01770.09940.48620.51780.00110.00460.12170.0014
Analog Residual0.02330.14030.79000.59920.00410.00580.48860.6148

Srivatsan20

MethodMAE↓Pearson-Δ↑DE-Spearman↑DS↑MMD-GEX↓MMD-PCA↓AUPRC-DE↑DEG Recall↑
scDFM + PROVE0.05980.80590.83220.85190.12190.26930.72420.5762
CellFlow + PROVE0.01280.87720.87540.85970.01000.12510.66900.7523
CFM-GP + PROVE0.01470.84200.84950.86100.12400.06420.62060.7412
Endpoint-only0.012920.135220.265960.560930.010590.105810.232990.35252
Interpolation0.011580.155890.276640.568640.027570.132370.251400.32706
CC-Mean0.01380.09850.21300.51550.00840.04130.15240.1470
TwoWay-Ridge0.01100.15640.28470.56270.00560.03100.19960.0363
Analog Residual0.01690.26490.38970.80180.01110.03950.27280.4453

Overall findings: Full PROVE outperforms Endpoint-only in 15/15 and validation-selected Interpolation in 11/15 perturbation-specific comparisons, including all five metrics on both Frangieh21 and Srivatsan20. Although simpler reference-based methods remain competitive on Replogle (Table 6), these results suggest that stepwise velocity editing offers benefits beyond endpoint correction or interpolation for recovering perturbation-specific responses across contexts.

W3. Distributional Interpretation and Theoretical Justification

We thank the reviewer for raising this important question regarding the distributional interpretation and theoretical guarantees of PROVE.

We would first like to clarify that PROVE does not aim to perform exact posterior sampling from a prescribed reference-conditioned distribution. Instead, it provides a geometrically constrained, inference-time correction mechanism for a frozen generative model under incomplete target-context supervision.

This distinction is important when interpreting its theoretical scope.

In the flow-matching literature, exact and approximate guidance are recognized as different approaches. For example, Feng et al. [1] develop a general framework for flow-matching guidance and investigate both exact guidance and approximate methods based on local approximations.

Although PROVE is not derived as a special case of the energy-guidance formulations in [1], it follows the broader principle of approximate inference-time steering: using task-specific reference information to selectively modify the sampling dynamics of a pretrained generative model.

In cross-context perturbation prediction, post-perturbation observations for the target context–perturbation pair are unavailable. Responses from other contexts provide useful information about perturbation direction, effect magnitude, and cross-context variation, but do not uniquely determine the true target distribution.

PROVE therefore treats these references as soft constraints for detecting and correcting reference-inconsistent extrapolations, rather than as a calibrated likelihood defining an exact Bayesian posterior.

Nevertheless, the edited velocity field induces a well-defined generative distribution. Under standard regularity assumptions for the corresponding ODE, the final distribution can be expressed as

q_1^{\mathrm{PROVE}}=(T_{v_{\mathrm{PROVE}}})_\\#q_0.

where TvPROVET_{v_{\mathrm{PROVE}}} denotes the transport map induced by the edited velocity field. The corresponding probability evolution satisfies the continuity equation

∂tqt=−∇⋅(qtvPROVE)\partial_t q_t=-\nabla\cdot(q_t v_{\mathrm{PROVE}}).

This characterizes the distribution induced by PROVE's sampling dynamics, without asserting equivalence to a prescribed posterior distribution.

The theoretical analysis in Appendix A further establishes several properties of the inference-time update:

  1. Proposition 1 establishes the compatibility of the reference-defined feasible regions.
  2. Proposition 2 bounds the approximation error of the linear endpoint estimate under the stated regularity assumptions.
  3. Proposition 3 characterizes the positive alignment property of the exact minimum-norm Pareto consensus.
  4. Proposition 4 establishes bounded editing and instantaneous velocity-norm preservation.

These results provide geometric consistency and numerical control for reference-guided velocity editing. They characterize the correction mechanism itself rather than implying a direct bound on the discrepancy from the unknown target distribution.

We additionally evaluated the distributional fidelity of the generated samples using MMD-GEX and MMD-PCA. As reported in Table 1, both metrics improve in 8/9 backbone–dataset combinations, yielding improvements in 16 out of 18 distributional comparisons.

Overall, PROVE provides an approximate inference-time correction mechanism with explicit geometric constraints and controlled velocity editing, while its effectiveness in matching observed perturbation-response distributions is supported by the empirical evaluation. We do not claim exact posterior sampling or a theoretical guarantee of monotonic MMD improvement.

We will clarify this distinction between the theoretical properties of the editing procedure and the empirical accuracy of the resulting distributions in the revised manuscript.

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

PROVE edits frozen generative samplers for cross-context perturbation prediction. It constructs reference-based preferences for effect magnitude, perturbation identity, and contextual variation, combines their correction directions with the frozen velocity through minimum-norm consensus, and preserves the velocity norm. Experiments on three datasets and three flow backbones, with additional diffusion and pathway analyses, show consistent improvements in perturbation-specific recovery.

Critical strengths
  1. Useful improvements across models and tasks. All five perturbation-specific metrics improve in all nine backbone–dataset combinations (Table 1). Pathway recovery and diffusion extensions strengthen the evidence that the correction has practical value beyond one sampler.
  2. A sensible inference-time intervention. Using available biological references to constrain otherwise unsupported extrapolation is well motivated. The comparisons with reference-only predictors in Appendix C.4 also show that guided generation can retain response information that direct transfer misses.
Critical weaknesses
  1. The necessity of the proposed machinery is not established. The method combines three hand-designed priors, projections, consensus, clipping, and norm preservation, but does not compare against endpoint-only correction or simple backbone–reference interpolation. Component removal and reference-only baselines do not answer this question. Consequently, the experiments establish the value of reference guidance more convincingly than the novelty and necessity of this particular algorithm.
  2. The priors can suppress valid contextual biology. Per-cell corrections use statistics of context-level means, and the drift radius excludes within-context variability (§3.1). Moreover, the effect-collapse diagnostic activates on 25.6% of genuine Srivatsan20 target responses (Appendix C.2). Without performance on these reference-disagreeing conditions or heterogeneity-sensitive checks, aggregate improvements do not establish that the intervention preserves the context-specific responses it aims to predict.
Questions for authors
  1. Assessment-critical comparison: With identical references and validation access, how does PROVE compare with endpoint-only correction and a simple interpolation between backbone predictions and transferred reference responses?
  2. Assessment-critical analysis: On conditions whose observed responses violate the reference priors, does PROVE improve prediction or erase valid attenuation/context specificity? How does it affect within-condition variability?
3 comments

We thank the reviewer for their careful assessment and constructive feedback. To address the two central concerns, we report three targeted analyses of the complete algorithm, predictive performance when ground-truth responses violate reference priors, and preservation of within-condition variability.

Q1. Necessity of the proposed algorithm

In Appendix C.4 (Tables 6–8) of the original submission, we compared PROVE-guided models with three reference-only predictors. The results show that direct reference transfer is highly competitive on Replogle, whereas PROVE performs best on all five perturbation-specific metrics on Frangieh21 and Srivatsan20.

To examine the contribution of the complete velocity-editing procedure, we compared it with two simpler alternatives using CellFlow on all three datasets: Endpoint-only Correction and Backbone–Reference Interpolation.

The interpolation weight was selected on training-side validation conditions rather than chosen arbitrarily. For each dataset, we evaluated λ ∈ {0.2, 0.4, 0.6, 0.8, 1.0} on 20 training-side pseudo-held-out condition pairs and selected the weight using a prespecified composite score based on perturbation-response metrics. The selected λ values were 0.6 for Replogle, 0.6 for Frangieh21, and 0.2 for Srivatsan20. All parameters were frozen before test evaluation, and no held-out test outcomes were used for parameter selection. The results are summarized below:

Replogle

MethodMAE↓Pearson-Δ↑DE-Spearman↑DS↑MMD-GEX↓MMD-PCA↓AUPRC-DE↑DEG Recall↑
scDFM+PROVE0.12740.29940.37590.61990.06200.16810.30050.7130
CellFlow+PROVE0.06860.29180.57070.62530.00760.02710.29750.2953
CFM-GP+PROVE0.06490.31830.41190.64110.01120.03300.31520.3430
Endpoint-only0.06220.25960.49800.58390.00790.02790.24540.0953
Interpolation0.07020.37580.54280.63850.14480.17600.30360.3540
CC-Mean0.06910.32060.59450.72530.00950.02640.31210.3243
TwoWay-Ridge0.06990.36560.63930.69990.00800.02720.30050.3203
Analog Residual0.09650.26670.54080.76780.02760.05230.26350.5287

Frangieh21

MethodMAE↓Pearson-Δ↑DE-Spearman↑DS↑MMD-GEX↓MMD-PCA↓AUPRC-DE↑DEG Recall↑
scDFM+PROVE0.10140.46370.59670.62720.08050.16430.51980.7007
CellFlow+PROVE0.02240.85830.92810.58780.00150.00910.72600.7021
CFM-GP+PROVE0.02250.80370.91090.60080.00460.01870.62870.6915
Endpoint-only0.02060.19050.76870.56370.00240.00880.54300.6136
Interpolation0.01500.23070.80100.55140.13890.16780.52600.2354
CC-Mean0.02040.18650.71110.57910.00270.00580.52810.5405
TwoWay-Ridge0.01770.09940.48620.51780.00110.00460.12170.0014
Analog Residual0.02330.14030.79000.59920.00410.00580.48860.6148

Srivatsan20

MethodMAE↓Pearson-Δ↑DE-Spearman↑DS↑MMD-GEX↓MMD-PCA↓AUPRC-DE↑DEG Recall↑
scDFM+PROVE0.05980.80590.83220.85190.12190.26930.72420.5762
CellFlow+PROVE0.01280.87720.87540.85970.01000.12510.66900.7523
CFM-GP+PROVE0.01470.84200.84950.86100.12400.06420.62060.7412
Endpoint-only0.012920.135220.265960.560930.010590.105810.232990.35252
Interpolation0.011580.155890.276640.568640.027570.132370.251400.32706
CC-Mean0.01380.09850.21300.51550.00840.04130.15240.1470
TwoWay-Ridge0.01100.15640.28470.56270.00560.03100.19960.0363
Analog Residual0.01690.26490.38970.80180.01110.03950.27280.4453

Overall findings: Full PROVE outperforms Endpoint-only in 15/15 and validation-selected Interpolation in 11/15 perturbation-specific comparisons, including all five metrics on both Frangieh21 and Srivatsan20. Although simpler reference-based methods remain competitive on Replogle (Table 6), these results suggest that stepwise velocity editing offers benefits beyond endpoint correction or interpolation for recovering perturbation-specific responses across contexts.

Q2. Preservation of context-specific biological responses and within-condition variability

(a) Performance on ground-truth prior-violating conditions

To assess whether reference priors might alter genuine context-specific responses, we performed a ground-truth prior-violation stratification analysis.

Using CellFlow as the backbone, we identified target conditions whose ground-truth mean responses violate at least one reference prior (EC, PC, or CD) and evaluated PROVE's predictive performance on these conditions. Ground-truth responses were used only for post-hoc stratification and evaluation; they did not enter inference or reference construction. The results are reported below (Δ = PROVE − Vanilla):

DatasetΔ Pearson-ΔΔ DE-SpearmanΔ DSΔ AUPRC-DE
Replogle+0.0378+0.1427+0.0682+0.0739
Frangieh21+0.0906+0.4120+0.0584+0.5058
Srivatsan20-0.0112+0.0817+0.0206+0.0202

Among conditions whose ground-truth responses violate a reference prior, PROVE improves 11 of the 12 dataset–metric comparisons.

The 25.6% EC-triggered Srivatsan20 conditions identified in Appendix C.2 are included in this broader any-prior subgroup; the table does not isolate EC-only violations. These results show that reference-prior disagreement does not preclude improved perturbation-specific prediction in the evaluated subgroup.

(b) Preservation of within-condition variability

To assess whether PROVE compromises within-condition variability, we examined three gene-wise variance consistency metrics: Variance Spearman (rank correlation between predicted and ground-truth gene-wise variances), Top-20% Jaccard (overlap between the predicted and ground-truth sets of the top 20% most variable genes), and NDCG@20% (how accurately highly variable genes are ranked near the top).

The changes relative to Vanilla are reported below:

DatasetΔ Variance SpearmanΔ Top-20% JaccardΔ NDCG@20%
Replogle+0.0005+0.0080+0.0097
Frangieh21-0.0040+0.0080+0.0026
Srivatsan20+0.0012+0.0011-0.0009

Overall, the measured gene-wise variability structure remains comparable between PROVE and the vanilla backbone across all three datasets, with no substantial systematic degradation.

We will incorporate these results and the corresponding implementation details into the revised manuscript. And we appreciate the reviewer's thoughtful feedback.

Regarding your two assessment-critical questions, our additional experiments show that Full PROVE outperforms Endpoint-only Correction in 15/15 and validation-selected Interpolation in 11/15 comparisons. It also improves 11/12 comparisons on ground-truth prior-violating conditions while maintaining comparable gene-wise variability.

We hope these results address your main concerns.

Comments on the paper

1 comment

We sincerely thank all reviewers for their constructive feedback. We have provided detailed point-by-point responses with additional experiments and analyses. We would like to briefly summarize the main findings:

  1. Algorithmic necessity: Using identical reference information, Full PROVE outperforms Endpoint-only Correction in 15/15 and validation-tuned Interpolation in 11/15 perturbation-specific comparisons across three datasets with CellFlow, demonstrating the benefits of stepwise velocity editing. (Responses to Reviewer ZeTx Q1, Reviewer cuf9 Q1, and Reviewer WDa3 W2.)

  2. Preservation of biological signals: On conditions whose ground-truth responses violate reference priors, PROVE improves 11/12 perturbation-specific comparisons. Additional analyses also show that gene-wise variability remains comparable to the original backbone. (Responses to Reviewer ZeTx Q2, Reviewer cuf9 Q2, and Reviewer wzBE Q1.)

  3. Mechanistic robustness and practical evaluation: Diagnostics across nine backbone–dataset combinations show that strong directional conflicts and negligible corrections under severe violations are rare. We also extended component ablations, measured inference-time overhead, and investigated the anomalous baseline correlations. (Responses to Reviewer wzBE Q2 and Reviewer cuf9 Q3–Q5.)

  4. Theoretical clarification and reference dependence: We clarified the backbone-specific geometric formulation, analyzed sensitivity to reference quality, and characterized the distributional interpretation and theoretical scope of PROVE. (Responses to Reviewer WDa3 W1, W3, and W4.)

We hope these additional results and clarifications address the reviewers' main concerns. We would greatly appreciate any further feedback on whether there are remaining questions or concerns, and would be happy to provide additional clarification during the discussion period.

We thank the reviewers again for their time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.