acceptodds
Under review as a conference paper at ICLR 2027

CHARM-Flow: Cellular Hierarchy-Aware Response Modeling for Single-Cell Perturbations

Abstract

Predicting single-cell perturbation responses from unpaired observations requires capturing both population-level expression changes and variation among responding cells. We introduce CHARM-Flow, a cellular hierarchy-aware generative framework that models a population-level response together with heterogeneous cellular variation around that response. A shared control-derived population–state–cell hierarchy coordinates response prediction, conditional source composition, and cell-level residual generation. Condition-dependent weights organize empirical control cells into a state-structured source, while each sampled source cell remains available as a persistent reference during generation. A control-derived gene partition parameterizes the source and evolving residual representations, and conditional flow matching provides the residual update mechanism. Across Replogle, PBMC, and Tahoe, CHARM-Flow captures perturbation-level population responses, achieving PDCorr values of 0.477, 0.870, and 0.959, respectively. Controlled analyses on Replogle show that state structured empirical sources and persistent source access improve agreement with observed cellular distributions within perturbations. The shared hierarchy connects population-level response prediction, empirical source construction, and heterogeneous cell-level generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell perturbation prediction, cellular heterogeneity, hierarchical generative modeling, conditional flow matching, empirical source modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fbaMOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

CHARM-Flow predicts post-perturbation gene expression distributions using a population–state–cell hierarchy. It combines mean response prediction, conditional flow matching, and persistent access to source-cell references to model perturbation responses and cell-to-cell variation. Experiments on Replogle, PBMC, and Tahoe cover genetic perturbations, cytokine stimulation, and drug treatments, with the clearest improvements over baselines observed on Replogle.

Critical strengths
  1. The population–state–cell hierarchy provides a reasonable way to combine population-level response patterns with within-population cellular variation. Retaining source-cell references during generation is a useful design choice for modeling heterogeneous responses.

  2. Experiments on Replogle show improvements over multiple baselines in mean-response and differential-expression prediction. Ablations also provide evidence that the state hierarchy and persistent source-cell access contribute to performance.

Critical weaknesses
  1. The biological significance and robustness of the inferred states remain unclear. The state partition affects both state summaries and source-cell sampling, but the paper provides limited evidence that these states correspond to meaningful biological differences. The sensitivity of the results to the number of states or the clustering method is also unclear.

  2. The evidence for improved modeling of cellular heterogeneity is limited. Much of the main evaluation focuses on mean-response metrics, while distributional and dispersion analyses are concentrated on Replogle and model ablations. Comparisons with existing generative baselines across PBMC and Tahoe are needed to establish whether CHARM-Flow consistently captures cell-to-cell variation more accurately.

Questions for authors
  1. Could the authors characterize representative states using marker genes, pathway activity, or available cell-state annotations? How sensitive are the results to the number of states (K) and the choice of clustering method?

  2. Could the authors extend distributional and dispersion evaluations to PBMC or Tahoe and compare against a representative generative baseline under the same protocol? This would help establish whether the model's advantage extends beyond mean-response prediction.

3 comments
AuthorsOpenReview ↗

Response to Reviewer fbaM (Q1–Q2; W1-W2): State Interpretation, Clustering Sensitivity, and Distributional Evaluation

Q1 and W1: Biological interpretation of states and sensitivity to K and clustering

Our design is motivated by making fuller use of the initial expression state of individual cells. When evaluation primarily relies on population means, heterogeneous perturbation outcomes are averaged together, and the diversity of initial cellular expression is difficult to assess. Through the population–state–cell hierarchy, we jointly use population context, initial transcriptional states, and individual-cell expression to preserve response differences associated with distinct initial cellular backgrounds during generation.

To characterize these initial states, we identified marker genes through state-versus-rest differential-expression analysis within each biological background. Candidate markers were discovered in one half of the control cells and validated in the other half, requiring a consistent direction of expression difference and an adjusted q<0.05q<0.05. The methods and results are provided in Appendix E.7 and Figure 3. Representative states exhibit interpretable biological features: ASPM, CCNB1, CDC20, and PLK1 mark proliferation-associated programs in HepG2; IFIT2, IFIT3, GBP1, and HLA-B characterize interferon-associated programs in PBMC Donor1 CD4-memory cells; and distinct states in Tahoe NCI-H460 exhibit stress-associated and replication/proliferation-associated features.

On Replogle, under the minimum-support constraint of 20 control cells per state, we compared K=64K=64, 128, and 192, together with flat k-means at K=128K=128. Their PDCorr values are 0.478, 0.477, 0.477, and 0.478, respectively, and the ranges of all three centered distribution metrics differ by less than 1%. Actual control members and within-state variation remain available throughout the model. K therefore controls the granularity with which initial-state information is organized, and further subdivision has little effect on overall prediction once sufficient resolution is reached.

The value of covering multiple source states is also illustrated by an inference comparison using the same trained model. For PPP6C perturbation in HepG2, sampling from multiple states according to the learned weights reduces NCED, CMMD, and NCSW1 by 66.2%, 41.9%, and 40.7%, respectively, compared with restricting sampling to a single state, while keeping mean prediction unchanged. Further subdivision provides finer biological interpretation. For example, under ZDHHC7 perturbation, a state at K=64K=64 splits into two subgroups at K=128K=128, with predicted G2M-program means of 0.171 and 0.334. We use 128 as a balance between state resolution and sample support per state.

The current data do not provide paired observations of the same cell before and after perturbation. Our evidence therefore supports improved prediction at the population-distribution level; it does not establish recovery of each individual cell's true perturbation outcome. An important next step is to combine the model with wet-lab measurements linking initial cellular states to post-perturbation outcomes, enabling more direct evaluation of cell-level response prediction.

Q2 and W2: Distributional and dispersion comparisons on PBMC and Tahoe

We have extended the distributional evaluation to PBMC and Tahoe and added comparisons with PerturbDiff across all three datasets using the same evaluation protocol within each dataset (Appendix E.8, Table 22). CHARM-Flow achieves lower centered distribution errors in NCED, CMMD, and NCSW1 on all three datasets, with NCED reduced by 45.5%, 90.9%, and 83.0% on Replogle, PBMC, and Tahoe, respectively. These comparisons support more accurate recovery of within-population cellular variation beyond the population mean response across the three datasets.

I thank the authors for their work. The rebuttal adds helpful new results and addresses some of my concerns. Still, open questions persist. The biological validity and robustness of these inferred states are not fully settled, and the sensitivity analysis remains limited. Though the new experiments strengthen the empirical support, my core doubts about state construction and interpretation are unresolved.

AuthorsOpenReview ↗

Follow-up to Reviewer fbaM (Q1/W1): Biological Interpretation of Control States and Sensitivity to K and Clustering Method

Thank you for clarifying your concern. We examined two related questions: what each state represents biologically within a given partition, and how its interpretation changes with the resolution of the partition.

We interpret each state based on the expression features that distinguish it from other control cells in the same biological background. We identify candidate markers in one half of the control cells and retain those that also show significantly higher expression in the same state relative to the other controls in the validation half. We interpret these markers alongside established gene-program scores, as illustrated in Appendix E.7 and Figure 3.

Replogle. In HepG2, state 3 shows elevated expression of MCM10, CDC6, and GINS2, together with higher S-phase and E2F scores, consistent with DNA replication. State 0 shows elevated expression of CCNB1, CDC20, and PLK1, along with a higher G2/M score, consistent with mitosis. Both states show proliferation-associated expression, with state 3 emphasizing DNA replication and state 0 emphasizing mitosis.

PBMC. Among CD4 memory T cells from Donor 1, state 61 shows elevated expression of IFIT2, IFIT3, and GBP1, together with higher interferon-α and interferon-γ response scores. State 0 from the same donor and cell type exhibits prominent mTORC1- and IL-2/STAT5-associated expression. These states capture distinct basal immune expression programs within the same annotated cell type.

Tahoe. Among DMSO-treated NCI-H460 control cells, state 0 shows elevated expression of GADD45A, MDM2, and CXCL8, together with prominent p53 and TNF/NF-κB signatures, consistent with a stress-associated profile. State 2 shows elevated expression of EXO1 and NEIL3, along with prominent E2F, S-phase, and G2/M signatures, consistent with a replication-associated profile. Both patterns remain detectable when using only genes excluded from the clustering panel.

To quantify marker support beyond these examples, we counted testable states with at least one gene outside the clustering panel whose expression was higher than in the remaining controls from the same background (adjusted q < 0.05 in both the discovery and validation halves). This criterion was satisfied by 468 of 512 states in Replogle (91.4%), 1,946 of 2,585 states across the four PBMC evaluation donors (75.3%), and 5,947 of 6,204 states in Tahoe (95.9%).

Effect of K on biological interpretation. To examine how clustering resolution affects the interpretation of cell states, we tracked the same HepG2 cells across different resolutions. With a maximum of 64 states, state 20 contained 93 control cells. When the maximum was increased to 128 states, these same cells were divided into two groups of 67 and 26 cells. Both groups retained elevated TOP2A expression and prominent E2F and G2/M signatures, preserving the proliferation-associated interpretation of the original state. However, the finer partition revealed differences in the strength of these shared expression signatures: the two groups had mean G2/M scores of 0.547 and 0.650, and mean E2F scores of 0.471 and 0.567, respectively.

We further assessed whether the biological interpretations remained consistent across different partitions. Using only genes excluded from clustering, we estimated each state's mean gene-program score from one half of the HepG2 control cells and evaluated its agreement with the measured scores of the same 2,515 validation cells. Across state limits of 64, 128, and 192, Pearson correlations ranged from 0.810 to 0.816 for E2F and from 0.674 to 0.679 for G2/M. At K = 128, recursive and flat k-means produced substantially different cell assignments (adjusted Rand index = 0.211), yet both achieved a correlation of 0.816 with measured E2F scores, while their correlations for G2/M were 0.675 and 0.687, respectively.

Together, these results indicate that the proliferation-associated expression patterns in HepG2 remain detectable across the clustering resolutions and methods evaluated, even when the exact cell assignments differ. The 93-cell example shows what finer resolution adds: both subgroups retain the proliferation-associated profile of the parent state, while their measured E2F and G2/M expression scores reveal a difference in the strength of that shared signature.

Reviewer_wqk3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents CHARM-Flow, which samples control cells from learned state mixtures and transforms them using residual flow. Each sampled cell remains available as a reference during generation. The model predicts the population mean separately and constrains the generated population to have that mean.

Critical strengths
  • I find the method well motivated.
  • The strongest result is the state-structured source on Replogle: it reduces centered distribution errors by about 24–30%, with improvements for roughly 90% of perturbations (Tables 4 and 16). The source-only comparison also shows that residual flow helps, particularly on DE precision.
  • The detailed appendices and three independent Replogle runs are useful.
Critical weaknesses
  • The evidence that the method improves the modeling of cellular variation relative to existing methods is limited. The NCED, CMMD, and NCSW1 comparisons on Replogle are currently restricted to internal variants. Table 15 shows that most main metrics are unchanged when generated cells are replaced by their group means, so those metrics do not tell us how well the model generates cellular variation. Reporting centered distribution metrics for representative generative baselines would strengthen this claim.

  • The results outside Replogle are mixed. On PBMC, OCOO-T has higher PDCorr and DE scores, while CHARM-Flow has higher PDS scores and lower reported MSE and MAE. On Tahoe, CHARM-Flow has higher PDCorr, but its MSE is about 6.4 times OCOO-T’s, and its DE precision is below several simple and learned baselines. These differences need more discussion.

  • The comparability of the imported baseline results also needs clarification, particularly whether populations, controls, preprocessing, and scoring settings match. This is especially important for Tahoe’s subsampled evaluation. If matching settings cannot be established, a local comparison with at least one strong baseline would help. Repeated runs would also help: the three-run Replogle DEPrec average is 0.2395 ± 0.0049, numerically close to OCOO-T’s reported 0.240, so multi-seed main results and ablations would clarify which differences are reliable.

  • The contribution of the DE-related training objectives is not sufficiently isolated. The soft rank-effect objective is closely related to the statistic used in DE evaluation, and the Replogle support objectives use training DE labels. These are reasonable choices, but an ablation would help separate their contribution from that of the architecture.

Questions for authors
  • How consistent is the benefit of residual flow? Table 14 shows an NCED improvement from 0.0242 to 0.0224 and a DEPrec improvement from 0.1941 to 0.2430 over the prepared source. I would like to see this comparison across the existing Replogle runs and on PBMC and Tahoe. Comparing independent samples of real treated cells would also help put the distribution errors in context.

  • Could the prior-only predictor be included for all three datasets? On Replogle, it reaches PDCorr of 0.3956, compared with 0.4773 for the full model. This is a meaningful improvement, but it would help to know how much the learned model adds on the other datasets. Removing state-level inputs from the mean predictor, while retaining the prior and response losses, would also clarify the hierarchy’s contribution.

  • In Table 12, removing persistent source access during retraining also increases mean-prediction MSE from approximately 0.0058 to 0.0134. Beyond the inference-time intervention in Table 18, could the authors compare retrained residual generators using a common frozen mean predictor? This would isolate the benefit of persistent source access for residual generation.

  • I would also appreciate a brief explanation of how the loss weights and dataset-specific settings were selected. Results broken down by biological context would be useful as a secondary analysis, especially for interpreting the pooled PBMC and Tahoe scores.

2 comments
AuthorsOpenReview ↗

Response to Reviewer wqk3 (Q1–Q3): Distributional Comparisons and Component Ablations

W1 and W3: Distributional comparisons and evaluation comparability

We have added comparisons with PerturbDiff across all three datasets using distributional and dispersion metrics, as reported in Appendix E.8, Table 22. In these matched distributional comparisons, both models use common treated and control populations, the same gene panel, and the same scoring settings within each dataset. CHARM-Flow achieves lower NCED, CMMD, NCSW1, NStdL1, and NIQRL1. Specifically, NCED is reduced by 45.5%, 90.9%, and 83.0% on Replogle, PBMC, and Tahoe, respectively. These results demonstrate more accurate recovery of post-perturbation cellular distributions and gene-wise variability. The Tahoe comparison uses the same evaluation cells and drug–dose aggregation units. Table 13 also reports three independent Replogle runs, with PDCorr of 0.475±0.0020.475\pm0.002 and DEPrec of 0.239±0.0050.239\pm0.005.

Q1: Consistency of the improvement over source-only

We have extended the source-only comparison to PBMC and Tahoe, as reported in Appendix E.8, Table 23. Source-only preserves variation from sampled control cells around a predicted perturbed population mean. CHARM-Flow achieves lower NCED, CMMD, and NCSW1 on all three datasets. For example, NCSW1 decreases from 0.1741341 to 0.1662931 on Replogle, from 0.0542373 to 0.0484798 on PBMC, and from 0.1503978 to 0.1430160 on Tahoe. DEPrec also increases from 0.194, 0.540, and 0.298 to 0.243, 0.569, and 0.326, respectively.

The gene-wise dispersion results vary across datasets: both errors decrease on Replogle; Tahoe shows lower NIQRL1 with nearly unchanged NStdL1; and source-only has lower dispersion errors on PBMC. Overall, CHARM-Flow improves recovery of the joint cellular distribution across all three datasets, while the improvement in individual-gene variability depends on the dataset.

Q2: Contributions of the transferred prior, learned prediction, and state information

We have added prior-only comparisons for all three datasets in Appendix E.9, Table 24. Relative to prior-only, CHARM-Flow increases PDCorr from 0.396 to 0.477 on Replogle, from 0.850 to 0.870 on PBMC, and from 0.944 to 0.959 on Tahoe. Population-mean MSE decreases by 22.5%, 17.2%, and 33.7%, respectively, and DEPrec also improves on all three datasets. These results show that learned prediction further improves the responses provided by the transferred prior.

To examine the contribution of state information, we retrained the model after removing state-level inputs from the mean predictor while retaining the transferred prior and response objectives. State organization remains available to the cellular generator. The results are reported in Appendix E.9, Table 25. The effect on mean prediction is not pronounced on Replogle or Tahoe. On PBMC, however, removing state inputs increases MSE by 25.2% and decreases PDCorr from 0.870 to 0.851.

PBMC states are constructed within the same donor and cell type. Existing annotations identify control states characterized by elevated expression of interferon-associated genes such as IFIT2, IFIT3, and GBP1, providing fine-grained context for cytokine-response prediction. Population means compress these internal differences, while the cross-donor prior primarily provides a calibrated average response. The hierarchy additionally retains state expression profiles and relative abundances, allowing prediction to account for the internal composition of the target population. Retaining state inputs lowers population-mean MSE for 51 of the 62 cytokines, demonstrating the additional value of this fine-grained information for PBMC response prediction.

Q3: Retrained residual generators with a common frozen mean predictor

To isolate residual generation, we fixed a common pretrained mean predictor and retrained two generators with persistent or initialization-only source access. Both share the frozen upstream representations and paired source draws. This differs from the end-to-end retraining in Table 18 and inference-time intervention in Table 19. Appendix E.10, Table 26 reports these two generators alongside the original CHARM-Flow reference.

Both generators achieve PDCorr of 0.477 and population-mean MSE of 5.831×10−35.831\times10^{-3}. With mean prediction held fixed, persistent source access reduces NCED, CMMD, and NCSW1 by 15.53%, 11.51%, and 14.78%, respectively, relative to initialization-only access.

Persistent source access provides a stable basal-cell reference throughout generation, allowing the model to continue using the initial cell's expression information as the residual evolves. These results show that persistent source references improve recovery of the post-perturbation joint cellular distribution when differences in mean prediction are controlled.

AuthorsOpenReview ↗

Response to Reviewer wqk3 (Q4; W2, W4): Parameter Selection, Biological Context, and DE Supervision

Q4 and W2: Parameter selection and interpretation across biological contexts

The loss coefficients were determined through a combination of empirical settings and small-scale pilot experiments, considering training stability, population-response prediction, and differential-expression recovery. For example, an early Replogle pilot compared population-mean MSE weights of 1 and 2, using 1,000 training updates and evaluation on 60 validation perturbations. In this pilot, increasing the MSE loss weight from 1 to 2 reduced validation MSE by 5.3%, while PDS-L2 and DEOver decreased by 1.6% and 14.5%, respectively. We therefore retained weight 1 to balance the evaluation objectives. We also conducted small-scale comparisons of assignment mass regularization to examine trade-offs between objectives. The remaining coefficients were set according to objective scales and training stability.

Dataset-specific state capacities and sampling settings account for the available control-cell support within each biological background. Controls are organized by cell line for Replogle and Tahoe, and by donor and cell type for PBMC. Consequently, the available control counts differ across backgrounds, and the settings accommodate this support within the population–state–cell structure.

The evaluation metrics capture different aspects of response prediction. On PBMC, OCOO-T has higher PDCorr and some DE scores, whereas CHARM-Flow has higher PDS scores and lower MSE and MAE. On Tahoe, CHARM-Flow has higher PDCorr, but its reported MSE is approximately 6.4 times OCOO-T’s, and some DE scores remain lower. Correlation of response patterns, absolute response errors, and DE recovery therefore need to be examined separately. We additionally connect the existing control-state annotations with observed and predicted transcriptional-program responses within the same biological backgrounds, as reported in Appendix E.12, Table 28.

In Replogle HepG2 cells, CHARM-Flow closely recovers the G2M-program decrease and cellular variation under RPL37A perturbation. It also recovers the response direction under PLK4 perturbation, although the predicted response magnitude and dispersion are lower. In PBMC Donor1 CD4-memory cells, both the observed and predicted interferon-program changes under IL-4 are −0.116. Under IFN-ω, the model recovers the increase but predicts a change of 0.068, compared with the observed 0.105. In Tahoe Hs 766T cells, the TNF/NF-κB-related response to 5 μM Homoharringtonine is closely recovered, with observed and predicted changes of 0.433 and 0.399. For 5 μM Fusidic acid, the corresponding changes are 0.100 and 0.055, indicating recovery of the direction with an underestimated response magnitude.

W4: Contribution of DE-specific training objectives

We have added a retraining ablation on Replogle that disables the relevant DE-specific supervision while retaining the architecture, remaining training objectives, and training budget, as reported in Appendix E.11, Table 27. Specifically, we disable the soft rank-effect objective and the associated rank and call objectives, together with support supervision based on training DE labels and the DE-count constraint.

Removing these objectives decreases PDCorr from 0.477 to 0.474 and changes population-mean MSE from 5.831×10−35.831\times10^{-3} to 5.804×10−35.804\times10^{-3}, showing little effect on mean prediction. DEOver and DEPrec decrease from 0.254 and 0.243 to 0.250 and 0.236, respectively. Retaining these objectives also reduces NCED, CMMD, and NCSW1 by 38.9%, 31.3%, and 22.5%, respectively, relative to the ablated model. This comparison quantifies the additional contribution of DE-specific supervision under the same architecture, reflected primarily in improved DEOver, DEPrec, and recovery of post-perturbation cellular distributions.

Reviewer_Daw1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces CHARM-Flow, a generative model that combines population-level mean expression prediction with flow matching to capture cellular heterogeneity in response to perturbations.

Critical strengths

The paper addresses the important problem of modeling cellular heterogeneity in perturbation responses

Critical weaknesses
  1. The paper does not establish how much the learned flow improves over simply shifting sampled control cells to the predicted perturbed mean while preserving their original variation.
  2. The model defines transcriptional states exclusively from control cells, even though perturbations can induce new states that are absent from the control population. While the flow may in principle generate such states, the paper does not explicitly model or evaluate this capability.
Questions for authors
  1. Related to the second weakness, can you evaluate perturbations that induce novel states or substantially alter cell-state composition? That would test whether the proposed hierarchy captures perturbation-induced heterogeneity rather than simply preserving the pre-existing variation.

  2. Why do you not report metrics such as Spearman correlation of LFC on differentially expressed genes (present in Cell-Eval)?

  3. How much of the performance comes from the transferred perturbation-response prior rather than the learned mean predictor? Could the authors provide an ablation of these components?

  4. Could the authors compare the full model against simply shifting sampled control cells to the predicted perturbed mean, without applying flow matching? This would help isolate the contribution of the learned flow.

4 comments
AuthorsOpenReview ↗

Response to Reviewer Daw1 (Q1–Q2): Perturbation-Induced Heterogeneity and Evaluation Metrics

Q1 and W2: Control-derived states and heterogeneous perturbation outcomes

Control-derived states describe initial cellular contexts rather than restricting the possible response endpoints. Control and treated cells are unpaired, and population means alone cannot reveal whether predicted cells recover the diversity of treated responses. We therefore use NCED, CMMD, and NCSW1 to assess distributional agreement between predicted and observed cells under the same perturbation. After independently centering both populations, these metrics quantify differences in cell-to-cell expression variation and overall distributional structure, thereby assessing recovery of post-perturbation cellular heterogeneity.

Across all 380 held-out Replogle perturbations, we compared CHARM-Flow with a model retrained without state-structured source organization. After independently centering the predicted and observed populations, NCED decreased from 0.032277 to 0.022435, CMMD from 0.142751 to 0.108315, and NCSW1 from 0.217532 to 0.166293 (Table 4). The corresponding reductions were 30.49%, 24.12%, and 23.55%. These results support the hierarchy's contribution to recovering variation among responding cells beyond the population mean response.

Q2: LFC Spearman and additional evaluation metrics

The original tables primarily presented representative metrics of population-response prediction, perturbation discrimination, and differential-expression recovery. We have now added LFCSpear to the main tables for all three datasets to assess rank agreement between predicted and observed responses among genes that are differentially expressed in the real data. We have also added ES, AUROC, and AUPRC in Appendix F, Table 29 to provide a more comprehensive evaluation.

AuthorsOpenReview ↗

Response to Reviewer Daw1 (Q3–Q4): Prior Contributions and the Source-Only Comparison

Q3: Contributions of the transferred response prior and learned mean prediction

We have added a three-way comparison on Replogle: prior-only, a model independently retrained without the transferred prior, and CHARM-Flow. Prior-only combines the target-context control mean with the calibrated response prior without neural mean correction. The two trained models use the same training budget and otherwise identical settings. Complete results are provided in Appendix E.2, Table 16.

The three configurations achieve PDCorr of 0.396, 0.436, and 0.477, and MSE of 7.521×10−37.521\times10^{-3}, 6.320×10−36.320\times10^{-3}, and 5.831×10−35.831\times10^{-3}, respectively. Compared with prior-only, CHARM-Flow reduces MSE by 22.47%, with DEPrec and AUPRC reaching 2.78 and 4.05 times the prior-only values. Removing the prior and retraining increases MSE by 8.39% relative to CHARM-Flow. These results show that learned prediction improves upon the transferred response estimate and that the transferred prior also improves the learned model.

Q4 and W1: Comparison with source-only

We compared CHARM-Flow with source-only and extended the original Replogle comparison to PBMC and Tahoe. Source-only preserves variation from sampled control cells around its predicted perturbed mean. Complete results are reported in Appendix E.8, Table 23.

On Replogle, compared with source-only, CHARM-Flow reduces NCED from 0.0241900 to 0.0224353 (7.25%), CMMD from 0.1140708 to 0.1083148 (5.05%), and NCSW1 from 0.1741341 to 0.1662931 (4.50%). On PBMC, NCED decreases from 0.0017139 to 0.0014029 (18.15%), CMMD from 0.0305167 to 0.0276995 (9.23%), and NCSW1 from 0.0542373 to 0.0484798 (10.62%). On Tahoe, NCED decreases from 0.0163695 to 0.0154450 (5.65%), CMMD from 0.0875167 to 0.0839558 (4.07%), and NCSW1 from 0.1503978 to 0.1430160 (4.91%).

Improvements across all three centered distribution metrics and all three datasets demonstrate more accurate recovery of post-perturbation cell-to-cell variation.

Thank you for the response.

Regarding Q1/W2, thank you for the additional experiment. I think it demonstrates that using control-derived states improves the agreement between predicted and observed perturbed distributions. However, my concern was about whether the model can capture novel transcriptional states induced by perturbations that are absent from the control population. This could be evaluated by identifying perturbations that induce such states and testing whether the generated cells reproduce them, including their expression signatures and relative abundances.

Regarding Q4/W1, the improvements are relatively small, particularly on Replogle and Tahoe (around 4-7%). The modest improvement over a baseline without flow raises questions about how much additional value the flow component provides relative to its complexity.

AuthorsOpenReview ↗

Follow-up to Reviewer Daw1 (Q1/W2 and Q4/W1): Recovery of Perturbation-Enriched States and the Benefit of Residual Flow

Thank you for clarifying the question. We used the CHARM-Flow predictions to test whether generated cells recover both the abundance and the expression signature of states that become prominent after perturbation.

Q1/W2. Recovery of perturbation-enriched states

We evaluated 223 held-out Replogle perturbation conditions in HepG2. In each of the two folds, we defined states using measured cells alone. A candidate state had to contain at least 20 treated cells, account for at least 10% of treated cells, and increase by at least 10 percentage points relative to controls. We required at least 100 batch-matched controls, a state frequency of at most 1% in those controls, and an upper bound of at most 5% for the 95% Wilson confidence interval. Four perturbations showed states that met these enrichment criteria. The response states for EIF2B5 and EIF2S1 qualified in both folds, while those for PHB and EIF2B1 qualified in the first and second folds, respectively. The held-out cells confirmed enrichment in all six comparisons.

State abundance. In all six comparisons, the generated state fraction was closer to the observed treated fraction than the matched-control fraction was. The mean absolute abundance error was 12.9 percentage points, with equal weight assigned to each of the four perturbations. EIF2S1 illustrates the magnitude of the shift. Across the two folds, the state accounted for only 0.8% and 0.4% of controls, then rose to 82.9% and 91.2% of treated cells. CHARM-Flow predicted 80.0% and 91.2%, respectively. For EIF2B5, the state was undetected in one held-out control sample (0/112) and rare in the other (2/715). It accounted for more than half of both observed treated and generated cells in both folds.

Expression signature. Each state was also evaluated with 20 markers selected using real discovery cells from among genes excluded from clustering. Relative to matched controls, the observed and generated states agreed on the direction of change for 19–20 of the 20 markers in every comparison. Across the six comparisons, the generated cells reproduced 58–96% of the observed increase in the mean signed marker score. For PHB, for example, this score rose from 0.02 in matched controls to 3.53 in observed state cells and 3.37 in generated state cells. These results show that CHARM-Flow captures both the enrichment of these states that were rare in controls and the accompanying shift in expression.

Q4/W1. Contribution of residual flow

We agree that the added complexity of flow should be weighed against its modest average gains on Replogle and Tahoe, and that source-only is a reasonable choice when inference cost is the priority. Source-only already places measured control-cell variation around the predicted perturbed mean. Nevertheless, residual flow still provides measurable improvements in distributional fidelity, supporting its value in further refining the predicted perturbed-cell distribution. We also tested a learned one-step alternative on 380 held-out Replogle perturbations. With the same training budget, auxiliary supervision, and terminal reconstruction, a one-step MLP attained a centered distribution error (NCED) of 0.0266, compared with 0.0242 for source-only and 0.0224 for CHARM-Flow. In this comparison, the one-step transformation did not reproduce the distributional gain of residual flow.

We further isolated the contribution of residual flow on Replogle by switching it on and off in the same trained model, with identical sampled control cells, predicted means, and terminal reconstruction. For response states recovered in both settings, we compared the distributions of single-cell marker scores with those in real treated cells. Across the 217 perturbations with comparable states, flow reduced the mean Wasserstein distance from 0.489 to 0.439 (10.1%). Flow also improved the prediction of state abundances on average. Across the 107 perturbations identified from real data as having substantial composition changes, the mean total variation distance between predicted and observed state proportions decreased from 0.3166 to 0.2984 (a paired reduction of 0.0182). With the mean response already fixed, the additional transformation therefore improves agreement with treated cells in both the distribution of marker scores within states and the relative abundance of those states.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.