acceptodds
Under review as a conference paper at ICLR 2027

Beyond the Transcriptome: Chromatin-Informed Prediction of Cell-State-Dependent Perturbation Responses

Abstract

Predicting transcriptional responses to genetic perturbations is central to understanding gene function. Existing predictors primarily rely on transcriptomic measurements, although chromatin accessibility provides complementary information about the cellular context in which perturbations act. Using this information requires linking chromatin context to specific perturbations and accounting for baseline differences between independently sampled control and perturbed populations. We propose ChromaPert, a chromatin-informed framework for predicting state-dependent perturbation responses. ChromaPert combines molecular and DNA-informed locus priors with paired control RNA–ATAC features to jointly represent target identity and measured cellular context. Its Chromatin-Guided Response Router (CGRR) builds a source bank of control-derived expression offsets and retrieves relevant sources using perturbation/locus or control-ATAC similarity. These offsets account for baseline differences, while conditional transport flow learns the remaining response. Across three unseen-target and held-out-state settings on K562 CAT-ATAC and Perturb-Multiome, ChromaPert achieves the highest all-gene response correlation among evaluated methods. When transferring known perturbations to held-out states, it improves correlation by 23.2% across all genes and 46.2% for the twenty most responsive genes over the strongest respective baselines. Correct RNA–ATAC pairing improves recovery of chromatin-associated response slopes over shuffled pairing, while cross-state predictions retain state-specific responses to the same perturbation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell perturbation prediction, Chromatin accessibility, Cell-state generalization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_YsS2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors introduce a novel chromatin-informed approach for perturbation modeling, leveraging chromatin accessibility and DNA sequence features alongside gene-level embeddings to inform post-perturbation gene expression prediction. Their approach, named ChromaPert, is designed to handle predictions for unseen perturbations as well as unseen cell states by using representations of perturbations and cell states to query a source bank of training set expression offsets and aggregating a weighted set of retrieved offsets as predictions. The authors leverage paired multimodal (RNA + ATAC) single-cell data and benchmark ChromaPert against 14 foundation model baselines, reporting performance across three separate categories of metrics: mean-response shape, pointwise accuracy, and transport/direction. The authors show that ChromaPert yields strong performance across these metrics on the unseen perturbation and unseen cell state prediction tasks. While the paper presents a novel framework for incorporating chromatin-level regulatory information into the perturbation prediction task, the presentation of the results makes the impact and merits of the work difficult to assess.

Critical strengths
  • The paper proposes a framework that uniquely attempts to leverage chromatin accessibility and DNA sequence information for improved perturbation prediction. The use of such data is currently underexplored and has the potential to provide valuable insights about the utility of other data modalities for perturbation prediction tasks and virtual cell modeling.
Critical weaknesses
  • The paper’s presentation quality is poor, with many sections that are difficult to follow, making it not easy to assess the merits and contributions of the work. Factors that contribute to this poor presentation quality include the following: the inclusion of abundant non-standard technical terminology (e.g., “source addback”, “training-fold standardization”, “statebase”, “clean-flow”, “response-slope Pearson”), inconsistent terminology (e.g., “Top-20-gene Spearman” vs. “Pearson_{Delta, DE}” vs. “DE-focused correlation” for the same metric), unexplained benchmarks (e.g., GSE).
  • The paper includes limited discussion and interpretation of results, with the Results section primarily focused on reiterating quantitative results that can be found in the papers’ tables.
  • According to the Results section text, the unperturbed control is excluded from the rankings, but this baseline seems to yield the best or second-best performance on a number of the metrics included in Table 1 (e.g., all Pointwise Accuracy metrics). It’d be helpful to include a discussion of this result to contextualize ChromaPert’s overall performance.
  • The paper contains a number of formatting errors. For example, the order of cited sections of the paper do not cohere with the sections’ numbering (e.g., Appendix B.4 comes before Appendix B.1-B.2 in the main text).
  • The ablation results (Section 4.5) appears to only show a subset of the metrics that were used in the previous tables (e.g., L2, W1, discriminative cosine). The inclusion of the complete suite of metrics would more comprehensively demonstrate the conclusions drawn from the ablation analyses.
Questions for authors
  • What do the points in Figure 3 represent? The caption mentions that the box “summarizes completed folds for the corresponding setting”, but it is presently unclear what this statement implies. The inclusion of multiple points per method in these boxplots would seem to indicate that standard deviations can be included in the Results tables as well.
  • How are the DNA sequences featurized? Which GENCODE/RefTSS annotations, ENCODE cCRE annotations, and local descriptors are used?
  • Based on the text’s description, the control context variable (cc) appears to include ℓp\ell_p twice (Equations 2 and 3). xpx_p appears to already contain ℓp\ell_p, but ℓp\ell_p is also included in the definitions for cppoolc_p^{pool} and cp,sstatec_{p,s}^{state}. Can you please clarify this definition?
  • What is meant by GSE in the following sentences? “GSE also evaluates 19 known targets across seven held-out states” (Section 4.1), “On GSE held-out-target evaluation” (Section 4.4)
9 comments

We sincerely thank you for recognizing the value of incorporating chromatin accessibility and DNA sequence information into perturbation prediction. Below, we address each point and describe the corresponding revisions.


W1.1: Technical terminology.

R: To make these operations easier to understand, we have replaced their names with descriptions of the computational steps and explained their meanings at first use. The corresponding changes are as follows:

Original termRevised wordingMeaning
Source addbackAdding back the source offsetAdd back the previously subtracted control-baseline offset to restore the pooled-response reference convention (Eq. 9).
Training-fold standardizationStandardization fitted on the training splitFit standardization parameters only on the current training split and apply them to query inputs.
StatebaseState-centered response predictionAggregate state-specific response predictions relative to matched controls (Eq. 10).
Clean-flowFlow trained on pooled responsesA flow branch trained directly on training responses Δᵢ without source-offset correction.
Response-slope PearsonPearson correlation of response slopesPearson correlation between observed and predicted response slopes, used as an auxiliary diagnostic.

To help readers understand these operations defined in Secs. 3.1–3.3, we have unified the wording in the revised Methods, Results and figure captions, and retained the definitions of the corresponding computations in Appendix C.2.

W1.2: Difference between Top-20 response-slope Spearman and PearsonΔ,DE

R: Top-20 response-slope Spearman and PearsonΔ,DE evaluate different quantities and are not different names for the same metric. Revised Sec. 4.1 defines the primary evaluation metric, while Appendix A.2 provides the calculations for both slope diagnostics and the results under true and shuffled pairing:

MetricQuantity evaluatedCalculation
PearsonΔ,DEPerturbation-induced changes in condition-mean expressionPearson correlation over the 20 genes with the largest absolute observed responses.
Top-20 response-slope SpearmanChromatin-associated response slopes in the auxiliary analysisSpearman correlation of flattened slopes over eight ATAC axes and the 20 genes with the largest absolute observed mean matched responses.

To avoid confusion, we have explicitly distinguished these metrics in revised Sec. 4.1 and the Figure 2 caption and added the definitions, calculations and numerical results for both slope metrics in Appendix A.2.

W1.3: Abbreviation for the GSE dataset.

R: GSE refers to GSE274113, the Perturb-Multiome dataset. Following your suggestion, we have explicitly defined GSE as its abbreviation when first introducing the dataset in revised Sec. 4.1 and used it consistently thereafter (revised line 304).

W2: Analysis of experimental results.

R: Thank you for suggesting more interpretation of the results. Accordingly, we have clarified what the experiments show about target generalization, state transfer and component contributions in the corresponding paragraphs.

Experimental sectionQuestion addressed and conclusion
Sec. 4.2: Unseen targets in K562Tests target generalization in a fixed cellular background, and Hallmark analysis supports recovery of shared proliferation and cell-cycle programs.
Sec. 4.3: Unseen targets in GSETests prediction of unseen-target responses in a background containing multiple cell states.
Sec. 4.4: GSE state transferResponse geometry and retrieval analyses support preservation of state-dependent differences in the response to the same perturbation.
Sec. 4.6 and Appendix A.1: Component controlsSource decomposition corrects control-baseline differences. The main advantage of flow over Direct MLP is response correlation, while absolute errors do not improve uniformly.
Sec. 4.6 and Appendix A.4: Prior ablationsMolecular and regulatory priors contribute complementary information. ESM-2 contributes strongly to target-response correlations, while the effects of cCRE and Multi-TSS differ across metrics.
Sec. 4.6 and Appendix A.1: Aggregation controlsTraining-control proportions retain most response-correlation performance, whereas uniform aggregation reduces it, showing that composition weights affect population-response aggregation.

To help readers understand the model design through the experiments, we have added these explanations to revised Secs. 4.2–4.6, directly connecting the design objectives, experimental observations and conclusions (Tables 1–6).


W3: Performance of Control baseline.

R: Control is the unperturbed reference baseline. As already noted in Sec. 4.3 of the original manuscript (lines 377–378), its low MSE and near-zero response correlation show that absolute error alone does not capture response structure. When most genes change weakly, predictions close to control can achieve low global error, consistent with the signal-dilution issue discussed in previous work [1,2]. Therefore, the task requires evaluation of both pointwise errors and perturbation-induced response patterns. To make this trade-off clearer, we retain the comparisons in which Control has lower errors in some settings and have added the explanation to revised Secs. 4.1 and 4.3.


W4: Order of appendix references.

R: Appendices and the main text have different organizational structures. Specifically, our appendices are organized into supplementary experiments (A), data and evaluation protocols (B), and methods and implementation (C). The original Sec. 4.1 follows the order in which concepts arise, citing B.4 to explain composition weights before B.1–B.2 for datasets and splits.

We apologize that this arrangement made navigation difficult. To improve navigation, we have reordered revised Sec. 4.1 to introduce datasets and splits before composition aggregation, so the references now appear as B.1–B.2 followed by B.4 (revised lines 309–310).


W5: Complete metric reporting for ablations.

R: We kindly clarify that we already reported all seven metrics, including L2, W1 and discriminative cosine, in Tables 7 and 8 of the submitted appendix. Sec. 4.5 directly cites Table 7 and points to Appendix A.5, which contains Table 8, through “Further controls appear in Appendices A.1, A.5, and A.6.” (lines 454–462). To address your concern, we have clarified this point in Section 4.6 of the revised manuscript.

Q1: The meaning of Figure 3 points and the report of fold-level SDs.

R: We kindly clarify that each Figure 3 point represents a held-out evaluation condition: a target for Target-LOO or a state–target pair for Cross. K562 and GSE Target use 18 and 19 held-out-target folds, respectively, whereas GSE Cross uses seven held-out-state folds, each evaluating 19 known targets, giving 133 condition points, with each box summarizing all condition results for its setting.

Because the paired RNA–ATAC benchmarks have few targets, we use leave-one-out evaluation to cover all available targets and evaluation states. Thank you for suggesting that we report fold-level standard deviations. We provide the PearsonΔ mean ± SD below:

Table 2: PearsonΔ mean ± SD across leave-one-out folds. For Cross, the 19 targets are first averaged within each state fold, after which SD is computed across the seven state folds.

MethodK562 (18 folds)GSE Target (19 folds)GSE Cross (7 state folds)
GEARS0.0878±0.07850.0263±0.15290.1013±0.0661
scGPT0.0676±0.11140.0416±0.12600.0978±0.0610
CellFM0.0634±0.08450.0716±0.08200.0696±0.0587
scDFM0.0100±0.1435-0.0374±0.10230.0812±0.0648
CellFlow0.1437±0.15400.0698±0.19270.1484±0.1159
ChromaPert0.2210±0.12420.1735±0.16060.1829±0.0767

Compared with CellFlow, ChromaPert has higher means and smaller fold-level SDs in all three settings. To make this clearer to you, we have explained the points and boxes more clearly in the Figure 3 caption and added the standard-deviation summaries to Appendix A.13, Table 22.


Q2: DNA encoding and annotation sources.

R: DNA is encoded using sequence-composition and short-sequence-count features from TSS windows, with GENCODE and RefTSS locating the TSSs and ENCODE cCREs providing locus-level regulatory context. Short-sequence counts are a common feature type in regulatory sequence modeling [3].

Specifically, for each gene, we average descriptors from up to three strand-oriented 2,000-bp TSS sequences: four nucleotide fractions, normalized counts of CG, GC, AT, TA, CA, TG, GATA, CACGTG, CCGGA and TATA, and GC fractions in four subwindows. The local locus interface separates sequence and accessibility components, while the regulatory bank combines GRCh38 TSS annotations, ENCODE SCREEN cCRE overlaps, guide descriptors and lightweight JASPAR proxies. Deterministic coordinate selection maps its 265 or 272 raw fields to a 128-dimensional interface. We have clarified the feature sources and concatenation in revised Sec. 3.1 and added the encoding steps to Appendix C.1.


Q3: Repeated concatenation of locus features.

R: Yes, repeated concatenation gives locus features a greater share of the conditioning vector, so it can be understood as a parameterization choice using the same underlying information. Equations 2–3 and the 576-dimensional and 736-dimensional conditions in Appendix C.1 match the implementation:

xₚ = [uₚ ∥ ℓₚ], cₚᵖᵒᵒˡ = [xₚ ∥ ℓₚ], cₚ,ₛˢᵗᵃᵗᵉ = [xₚ ∥ ℓₚ ∥ hₛ].

Because the molecular block uₚ appears once and the locus block ℓₚ twice, their relative representation in the condition input changes. The retrained single-copy comparison is shown below:

Table 3: Removing the duplicate locus copy from the flow condition on GSE Target.

Condition inputPearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowMSE↓\mathrm{MSE}\downarrow
Original Full: [xₚ ∥ ℓₚ]0.17350.36970.001054
Single locus copy: xₚ0.15980.35580.00108

Because these results support this parameterization in the reported setting, we have clarified the repeated concatenation in revised Sec. 3.1 and added the ablation to Appendix A.11, Table 20 (the rationale appears at revised line 194).


Q4: Abbreviation for the GSE dataset.

R: GSE is our abbreviation for GSE274113 Perturb-Multiome. To address your concerns, We have explicitly defined it at the dataset’s first introduction in revised Sec. 4.1 (revised line 304).


We hope these clarifications and additions address your questions and provide a basis for you to reconsider your rating. Thank you again for your careful reading. If anything remains unclear, we would be happy to explain it further.

References

[1] Ahlmann-Eltze et al., Nat Methods, 2025.

[2] Miller et al., Nat Biotechnol, 2026.

[3] Ghandi et al., PLoS Comput Biol, 2014.

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

The authors propose ChromaPert, a framework for predicting cell state-dependent responses to perturbations using paired RNA-ATAC profiles from unperturbed control cells. The architecture has three main components: (1) perturbation conditioning, which encodes each target gene using frozen protein and RNA embeddings (ESM-2, CLM-X), sequence features around the gene's TSS, and regulatory annotations (e.g. ENCODE cCREs, JASPAR motifs), with a state-aware version that also incorporates control RNA-ATAC summaries, (2) the Chromatin-Guided Response Router, which computes and routes control-derived expression offsets to account for pre-existing baseline differences in expression, and (3) conditional flow matching, implemented with an MLP in PCA space, used to predict the remaining perturbation response. For unseen target prediction, a gate fit on the training data combines four candidate predictions: the pooled flow with the baseline offset added back on (source-on), the pooled flow alone (source-off), a separate flow trained on uncorrected responses (clean-flow), and the aggregated state-specific flow (statebase). For held-out state prediction, the target’s state-dependent responses are weighted by ATAC similarity. In both settings the final response is added to the matched-control baseline to form the final prediction. The authors show that ChromaPert has the highest all-gene response Pearson correlation against 14 baselines on K562 unseen targets from CAT-ATAC (Table 1), Perturb-Multiome unseen targets (Table 2), and Perturb-Multiome cross-state transfer (Table 3). Shuffling RNA-ATAC pairing in K562 reduces response slope Pearson from 0.21 to ~0. Additional ablations show that removing ESM-2 embeddings causes the largest correlation decrease among single-prior removals in Fig. 4.

Critical strengths
  1. The method shows strong performance in evaluations for both unseen targets and held-out cell states. The authors show that ChromaPert increases all-gene response Pearson correlation on K562 unseen targets (0.1437 to 0.2210) and Perturb-Multiome unseen targets (0.1030 to 0.1735), relative to the strongest evaluated baseline. For held-out state transfer evaluation, it improves all-gene correlation (0.1484 to 0.1829) and top-20 response gene correlation (0.2262 to 0.3307). Consistent increases across these evaluations supports the framework’s strength for predicting response when either the perturbation target or the cellular context is absent during training. The included baselines are extensive (14 learned methods, including graph-based methods, single-cell foundation models, and generative methods). Each is fit or adapted within each fold using its public implementation, which is a major strength of the paper.

  2. The architecture is well-motivated and its components (i.e. the response decomposition and the flow predictor) are supported by ablation experiments. On Perturb-Multiome unseen targets, removing source decomposition reduces all-gene correlation from 0.1735 to 0.1016. Replacing flow matching with direct MLP regression (while retaining the same conditioning and source decomposition) reduces all-gene correlation to 0.1534 and top-20 response gene correlation from 0.3697 to 0.3017. External-prior ablations also support combining pretrained and regulatory priors: removing ESM-2 embeddings reduces all-gene correlation (0.1735 to 0.1569) and top-20 response gene correlation (0.3697 to 0.3268). Using only pretrained or only regulatory priors gives lower all-gene correlation compared to combining them. These experiments strengthen confidence in the proposed architecture and modeling choices.

Critical weaknesses
  1. A key claim in the paper is the use of chromatin for perturbation prediction, but the included evaluations do not isolate the contribution of chromatin to benchmark performance. As the paper is currently framed, a matched RNA-only version of ChromaPert is needed across the three eval protocols to support this claim. I acknowledge the “w/o ATAC/locus match” ablation (Table 4b), but it does not separately report the effect of removing ATAC while keeping locus matching intact. For unseen target prediction, routing uses target/locus features, and the pooled flow condition does not directly incorporate ATAC (although the state-aware branch does). The external prior ablation experiments isolate ESM-2 embeddings as the largest contributor among single-prior removals, while adding ATAC to CellFlow reduces its response correlations in held-out state transfer (Table 4a). These details cast some doubt on the importance of using chromatin in ChromaPert. The RNA-ATAC shuffle experiment does provide good supporting evidence, but it measures a separate response-slope metric instead of the benchmark’s Pearson∆, which should also be reported. These gaps are important because the benchmark gains could mostly reflect target priors, source decomposition, or post-processing choices rather than the importance of including ATAC in the framework.

  2. Configuration selection, some inconsistencies in baseline runs, and limited evaluation scope reduce confidence in the reported results. ChromaPert is a relatively complicated pipeline with a number of hand-set components: a reliability gate over four candidates and fixed hyperparameters for mean blending, stabilization, two routing temperatures, and a 20-cell fallback. Although the gate is fit on training data, the paper does not clearly describe how the hyperparameters were selected independently of LOO evaluation folds. Performance appears sensitive to some of these choices (e.g. the default blend ρ = 0.18 gives 0.1735 all-gene correlation vs. ~0.158 at 0.09 or 0.36 in Table 10). Some baseline metrics also differ substantially between the primary runs and the three-seed summaries: CellFlow shows 0.1484 in Table 3 vs. 0.0881±0.0048 in Table 6c, and 0.1437 in Table 1 vs. 0.1038±0.0117 in Table 6a. In addition, the evaluation covers only two hematopoietic lineage datasets with 18-19 targets each, plus the K562 benchmark does not evaluate transfer across held-out cell states. This scope does not establish generalization to other tissues, cell types, or larger perturbation libraries. The paper would be greatly strengthened by adding 1-2 datasets from a different lineage or tissue, ideally with a larger perturbation library. Possible extensions that come to mind include the RPE-1 CRISPRi Multiome Perturb-seq dataset of Metzner et al. 2025, and the unused iPSC and dasatinib-treated K562 subsets in CAT-ATAC to assess performance across different cell types and drug treatment contexts.

Questions for authors

Clarifications:

  1. How were the mean blend, stabilizer weight, routing temperatures, and 20-cell fallback threshold selected? Was selection performed independently of the LOO evaluation, e.g. through nested CV?
  2. Why do CellFlow's primary scores differ substantially from its three-seed summaries (0.1484 in Table 3 vs. 0.0881±0.0048 in Table 6c; 0.1437 in Table 1 vs. 0.1038±0.0117 in Table 6a)? Were the model configs and eval procedures identical, and do similar differences occur for other baselines?
  3. Were scDFM, PerturbDiff, and DeSCOPE trained separately within each eval fold, and how did authors use the released checkpoints for these methods? What hyperparameter selection budget and convergence checks were applied to the baselines compared to ChromaPert (ideally matched budget)?
  4. Were output scaling and normalization verified for FM baselines with errors substantially above the no-change control, e.g. scFoundation on Perturb-Multiome unseen targets (MSE 0.1753 vs. 0.0008 in Table 2)?
  5. What are the gate's inputs, training objective, and selection rule? How frequently does it select or weight each of the four candidates in the unseen target evaluation? It would be useful to report more on this in the paper.
  6. Can the authors report metrics on an otherwise matched ChromaPert variant without ATAC inputs across all three protocols, retaining the target/locus priors, RNA context, and other architectural components?
  7. Can the authors clarify whether the additive mean-shift control (Table 8) is equivalent to predicting the mean training perturbation response? If not, can this baseline be reported (especially given ChromaPert's blending toward the training mean)? In addition, Table 8 already includes ATAC-concatenated linear and ridge controls. Extending these simple controls across the three protocols would help establish gains of the method over these simple baselines.
  8. To quantify statistical uncertainty, can the authors perform paired performance tests between ChromaPert and the strongest relevant baselines across held-out targets and states?
  9. In related work, it would be useful to discuss the recent preprint MultiFlow (Wang et al. 2026) which also uses flow matching on paired RNA-ATAC data to predict perturbation responses in unseen contexts. Also would be good to include chromatin-informed regulatory network methods that simulate perturbations, SCENIC+ and LINGER (alongside the already included CellOracle).

Suggestions for future work:

  1. Evaluate the method on additional lineages, tissues, and contexts. Metzner et al.'s RPE-1 Multiome Perturb-seq would be suitable for a non-hematopoietic lineage, plus the unused dasatinib-treated K562 group (and possibly the iPSC group if enough targets are available) in the CAT-ATAC dataset would test cross-context transfer.
  2. Extend the framework to settings without paired RNA-ATAC controls, which is a useful setting in practice. This is already identified in the paper's limitations.
9 comments

We sincerely appreciate your insightful and professional comments and your recognition of our paper’s main contributions. Below, we respond to your questions point by point and present our supplementary experiments.


W1.1: Evaluation of chromatin’s contribution.

R: Chromatin provides the model with information about target-locus accessibility and the cellular regulatory context. To address your concerns, we compare Full, w/o ATAC and ATAC Shuffle using the primary metrics in all three settings, as shown in Table 1. Specifically, w/o ATAC removes measured ATAC and its derived state summaries, matching inputs and routing inputs, while CLM-X receives only RNA to construct representations, and cross-state context is constructed from the permitted RNA controls. Other static priors, including cCREs, DNA sequence and ESM-2, are retained.

Table 1: Primary metrics for w/o ATAC and ATAC Shuffle.

SettingModel variantPearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowMSE↓\mathrm{MSE}\downarrowMAE↓\mathrm{MAE}\downarrow
K562 TargetFull0.22100.32350.001290.01882
K562 Targetw/o ATAC0.20110.30480.0013070.018914
K562 TargetATAC Shuffle0.19520.27400.0013890.019780
GSE TargetFull0.17350.36970.0010540.018632
GSE Targetw/o ATAC0.14520.29890.0011810.019319
GSE TargetATAC Shuffle0.11460.28010.0011960.019627
GSE CrossFull0.18290.33070.0011090.01725
GSE Crossw/o ATAC0.15020.32580.0012070.018194
GSE CrossATAC Shuffle0.12590.27120.0012430.018747

Removing ATAC lowers response correlations and increases errors in all three settings, while shuffling the correspondence causes a further decline, suggesting that mismatched chromatin contexts may also misguide prediction. To present the role of ATAC more clearly, we have added this ablation to revised main-text Table 6(b) and Appendix A.7, Table 13.


W1.2: Poor performance of CellFlow[1] when using RNA-ATAC data.

R: Simple concatenation decreases response correlations while improving pointwise errors, suggesting a trade-off between response-pattern recovery and absolute prediction accuracy. To better address your concerns, we further compare with MultiFlow [2], the method you mentioned that is specifically designed for bimodal perturbation prediction, as shown in Table 2. It jointly models RNA and ATAC through coupled flows, whereas ChromaPert uses locus priors and control ATAC to organize conditioning and state matching and correct differences in control baselines.

Table 2: Comparison over all 133 GSE Cross conditions.

ModelPearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowMSE↓\mathrm{MSE}\downarrowMAE↓\mathrm{MAE}\downarrow
MultiFlow0.11400.27130.0053120.038406
ChromaPert0.18290.33070.0011090.01725

ChromaPert achieves higher response correlations and lower errors in this setting. These results support the value of chromatin information and also show that the benefit depends on how it is used to model responses. To explain the differences between these uses of chromatin, we have added the CellFlow [1] concatenation control and the explanation of this comparison to revised Sec. 4.6, Table 6(a), and Appendix A.14, Table 24.


W2.1: Hyperparameter selection.

R: We use nested cross-validation for hyperparameter selection in each setting, with inner validation restricted to the outer-training data and outer held-out folds used for evaluation. To help you understand this selection procedure, we have explained it in revised Appendix C.3 and provided one default configuration for each of the three settings in Table 31.

W2.2: Differences in CellFlow’s performance between the main text and the appendix.

R: For a fair comparison, we used benchmark-adapted configurations for the main-text experiments, while the original appendix reported CellFlow results using its official default configuration. In particular, we included cell-state features in CellFlow's covariate conditioning where applicable. We apologize for not clearly explaining this difference in the original manuscript. To address your concerns, we have added three-seed experiments using the same adapted configurations as the main text:

Table 3: Adapted CellFlow versus ChromaPert, reported as three-seed mean ± sample SD.

SettingMethodPearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowMSE↓\mathrm{MSE}\downarrowMAE↓\mathrm{MAE}\downarrow
K562 TargetCellFlow0.1315 ± 0.00720.0820 ± 0.06120.002451 ± 0.0000080.023978 ± 0.000116
K562 TargetChromaPert0.2302 ± 0.00110.3258 ± 0.00670.001289 ± 0.0000010.018822 ± 0.000005
GSE TargetCellFlow0.0713 ± 0.00830.1480 ± 0.03690.001145 ± 0.0000100.019312 ± 0.000058
GSE TargetChromaPert0.1580 ± 0.01760.3341 ± 0.03930.001083 ± 0.0000390.018831 ± 0.000222
GSE CrossCellFlow0.1444 ± 0.00210.2197 ± 0.01010.019076 ± 0.0005350.043485 ± 0.000728
GSE CrossChromaPert0.1757 ± 0.00100.3251 ± 0.00160.001101 ± 0.0000010.01727 ± 0.00001

ChromaPert has higher mean correlations and lower mean errors in all three settings. To make the comparisons between tables clearer, we have updated revised Appendix A.3, Table 8, and explained the configuration and evaluation procedure for each table.


W2.3: Evaluation across lineages and treatments.

R: Thank you for suggesting broader backgrounds. Accordingly, we have added Dasatinib-treated K562 from CAT-ATAC [3] and an independent RPE-1 CRISPRi Multiome Perturb-seq dataset [4]. The CAT-ATAC iPSC subset has too few eligible targets to support reliable evaluation under the same protocol, so we selected these two backgrounds and completed Target-LOO over 18 and 13 targets, respectively.

Table 4: Dasatinib-treated K562.

MethodPearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowL2↓L2\downarrowMSE↓\mathrm{MSE}\downarrowMAE↓\mathrm{MAE}\downarrowW1↓\mathrm{W1}\downarrowDisc. Cos.↑\mathrm{Disc.\ Cos.}\uparrow
Control——1.62280.0013690.0191766.6868—
GEARS0.06360.03052.87140.0041550.0473656.72210.0441
CellFlow0.04620.06971.89810.0018560.0211986.73990.0423
DeSCOPE0.12390.21181.62260.0013670.0198996.69920.1247
ChromaPert0.16290.28811.62080.0013610.0191416.69180.1675

Table 5: RPE-1.

MethodPearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowL2↓L2\downarrowMSE↓\mathrm{MSE}\downarrowMAE↓\mathrm{MAE}\downarrowW1↓\mathrm{W1}\downarrowDisc. Cos.↑\mathrm{Disc.\ Cos.}\uparrow
Control——2.67740.0039360.0269459.8278—
GEARS0.19050.41844.17630.0089810.0647779.88030.1891
CellFlow0.10750.27583.38730.0060220.03251410.02500.1066
DeSCOPE0.27290.37892.77830.0041300.0291609.86150.2737
ChromaPert0.37050.49612.53180.0034070.0256479.79710.3704

ChromaPert has higher response correlations than the listed baselines in both backgrounds, and all four error metrics are also lower on RPE-1. To present the model’s performance in different backgrounds, we have added both comparisons to revised Sec. 4.5, Tables 4–5. As you mentioned, this contributes significantly to the improvement of our paper.


Q1: Hyperparameter selection.

R: We have clarified hyperparameter selection in revised Appendix C.3 and Table 31.

Please see our response to W2.1 for a detailed explanation of hyperparameter selection.


Q2: Differences between CellFlow’s main and three-seed results.

R: The difference comes from the main-text adaptations and the appendix’s official default configuration. To ensure a consistent basis for comparison, we have added three-seed CellFlow experiments using the main-text configurations, with results in Table 3, updated revised Appendix A.3, Table 8.

Please see our response to W2.2 for a detailed explanation of the CellFlow configuration differences.

Q3: Baseline training, selection and convergence.

R: The additional seven-fold GSE Cross fits start from scratch, using each method’s official implementation and the recorded training configuration. The benchmark uses a different gene panel from those in the public task configurations, so these supplementary fits do not use task-pretrained checkpoints. We additionally report training budgets and late-training losses across the seven GSE Cross folds:

Table 6: Results across seven GSE Cross folds, with each model trained from scratch in each fold and losses averaged equally across folds.

MethodTraining budget and checkpointLoss over final 100 steps: start→endRelative decrease
scDFM200000 steps, batch 480.003828 → 0.0038270.033%
DeSCOPE20 epoch, batch 640.109778 → 0.1097740.0036%
PerturbDiff200,000 steps, 4 groups×32 cells, native EMA0.004379 → 0.0043780.023%
ChromaPert360epoch, batch 640.005479 → 0.0054760.055%

Loss changes over the final 100 steps are below 0.1% for all four methods, indicating that optimization has become relatively stable within the given computational budgets. The limited diversity of training perturbations and cellular backgrounds may still constrain generalization. To facilitate verification of this comparison, we have added the actual configurations and checkpoint selection criteria in Appendix A.15, Table 26.


Q4: FM output scales and normalization.

R: We checked input normalization and output scales for the scFoundation [9] and CellVQ [10] adapters, and extended their previous 15-epoch budget (official configuration) to a maximum of 100 epochs with early stopping. Limited adaptation budgets and learning-rate scheduling may have contributed to the earlier high errors.

Specifically, the frozen encoders retain native RNA context, and vocabulary-matched representations enter the GEARS [8] prediction head on the common panel, while the remaining panel genes use native GEARS embeddings. scFoundation follows its official interface: library normalization to 10,000 followed by log1p, with raw library totals retained for the count and resolution tokens. Outputs are evaluated on the common benchmark expression scale.

Table 7: Updated GSE Target results covering all 19 held-out targets.

ModelPearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowMSE↓\mathrm{MSE}\downarrowMAE↓\mathrm{MAE}\downarrow
scFoundation0.11690.30960.0155510.093400
CellVQ0.10780.31390.0206820.105988
ChromaPert0.17350.36970.0010540.018632

To explain the training procedure underlying these updated results, we have updated the results in revised Sec. 4.3, Table 2, and added a new configuration to Appendix A.15.


Q5: Gate inputs, objective and weighting rules.

R: The gate softly weights candidate responses according to training-side reliability scores, without an additional gate network or independent training loss. In each fold, inputs include target and locus features, candidate predictions and observed responses for training targets, and candidate predictions for the query target. The weights are computed without accessing the observed responses of the held-out perturbation.

Cosine similarities of standardized target and locus features produce neighborhood weights through softmax with temperature 0.22. The neighborhood score NbN_b measures response correlation and directional agreement on similar training targets, whereas the memory score MbM_b measures agreement between the query prediction and the neighborhood-weighted training response. The base score sb=0.52Nb+0.48Mbs_b=0.52N_b+0.48M_b is adjusted by availability, response-consistency and magnitude rules, then converted to candidate weights by softmax with temperature 0.12.

Table 8: Mean gate weights across 19 GSE targets.

CandidateMean weight
clean-flow0.6369
source-off0.0478
source-on0.1343
State branch0.1810

To explain how candidates combine into the final prediction, we have added the gating rules and weight analysis to revised Sec. 3.3 and Appendix A.12, Table 21.

Please see our response to W2.1 for a detailed explanation of parameter selection.


Q6: Performance of ATAC-Removal controls retaining the other components

R: To address your concerns, we have added a comprehensive ablation study of ATAC. Please refer to our response to W1.1 and Table 1 for detailed results. These results have also been incorporated into the revised Section 4.6, Table 6(b), and Appendix A.7, Table 13.

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

This paper aims to show that measured chromatin accessibility improves the prediction of transcriptional responses to genetic perturbations. The proposed method, ChromaPert, has three parts. First, static target priors (protein and RNA embeddings, promoter sequence, and regulatory annotations) are used to find similar training targets. Second, a router (CGRR) subtracts control-state RNA offsets, so that the training target becomes a response relative to composition-matched controls. Third, a flow in PCA space maps the control mean to each condition mean. Measured ATAC enters as a summary shared within each state and, in the held-out-state setting, as weights for averaging the same target’s responses across training states. On CAT-ATAC (K562 cells, 18 targets) and Perturb-Multiome (19 targets), ChromaPert reports higher Pearson delta than 14 baselines, both in the held-out-target setting on both datasets and in the held-out-state setting on Perturb-Multiome.

Critical strengths

Using measured chromatin context for perturbation prediction is a well-motivated direction, and paired multiome perturbation data is a good testbed. The paper addresses the mismatch between control and perturbed populations, covers many baselines, and documents the protocols in detail.

Critical weaknesses

The gains appear to come mainly from the composition-matched response definition and static target priors, and neither needs measured ATAC. The contribution of chromatin, which is the main claim, is not tested directly. The benchmark is also small, some numbers are inconsistent across tables, and the modeling components are close to existing practice.

  1. The role of chromatin is not tested directly. In CAT-ATAC (K562), all targets share one control pool (as mentioned in Appendix B.1), so ATAC cannot distinguish targets, and no component ablation is reported for this dataset. In the held-out-target setting, routing uses only target priors by design (Eq. 5). ATAC enters only as a summary shared within each state, so it acts much like a state label. ATAC is used for routing only in the held-out-state setting, but it is not compared with RNA-based routing, uniform weights, or a nearest-state lookup, and the shuffled-pairing test uses a diagnostic rather than the main metric (Section 4.2). It would help to report main metrics with an RNA-only control context in all three settings, and to add an RNA-routing control in the held-out-state setting.

  2. The gains mainly come from composition handling and static priors, and the comparison is asymmetric. Appendix B.4 notes that composition differences between control and perturbed cells are perturbation-induced. For TF knockouts in a differentiation system, these shifts are often the main phenotype. ChromaPert receives the test cell-type counts and composition-matched controls (Eq. 14, 17), so it only needs to predict the within-state response. On Perturb-Multiome, the two largest ablation drops come from removing source decomposition (Table 4b), after which ChromaPert is close to the best baseline (Table 2), and from keeping only the pretrained priors (Table 7). Neither part requires paired ATAC. It would help to define the task as within-state prediction, give baselines the same composition information and target priors, and report the state-centered branch (statebase) alone.

  3. Simple and chromatin-aware baselines are missing. Pearson delta can be inflated by responses shared across perturbations, and training-mean or linear baselines often match deep models (Viñas Torné et al., 2025; Ahlmann-Eltze et al., 2025). CAT-ATAC shows similar signs: shared proliferation programs dominate (Fig. 2b), the target-specific residual signal is small (Fig. A1b), and ChromaPert has no advantage on the top 20 genes (Table 1). Apart from CellOracle, no baseline uses measured chromatin. Adding training-mean and linear baselines, a metric that removes the shared response, and a baseline that feeds RNA–ATAC embeddings from CLM-X or EpiAgent (Chen et al., 2025) to the same adapter would help.

  4. The benchmark is small and some results are inconsistent. Each dataset has 18 or 19 perturbations, which leads to high variance and puts many baselines outside the data regime they were designed for. The main results in Tables 1 and 3 fall outside the three-seed ranges in Table 6, and CellFlow differs especially between Table 3 and Table 6c. Most ablation gaps are within one seed standard deviation, and both the temperature and the blend peak at their defaults. Please clarify which runs produced each table and how hyperparameters were selected. Using the Dasatinib and iPSC groups excluded from CAT-ATAC, as well as other multiome datasets, would also help.

Additional feedback (not part of the decision)

  1. The flow maps a fixed point to a condition mean, so it acts as a deterministic regression (Eq. 11; Table 9), and the W1 and distributional-accuracy claims do not apply. The gate in the held-out-target setting is also not ablated.
  2. The supplementary code contains only the forward pass, which does not match the reproducibility statement.
  3. The claim about “diverse cellular backgrounds” in the conclusion goes beyond two hematopoietic datasets.
Questions for authors
  1. What are the main metrics in all three settings with an RNA-only control context? Can component ablations be added for CAT-ATAC?

  2. How does the held-out-state setting perform with RNA-based routing, uniform weights, and a nearest-state lookup?

  3. How do training-mean and linear baselines perform? How large is the gap when baselines receive the same composition information and target priors, and what are the results without the supplied test counts?

  4. Which runs produced Tables 1 and 3 versus Table 6, and how were hyperparameters selected?

  5. What exactly do “w/o ATAC/locus match” and “w/o source decomposition” in Table 4b remove? Does the response slope use post-perturbation ATAC?

7 comments

We sincerely thank you for your constructive and positive assessment of our work. To address your concerns, we have followed your suggestions and conducted extensive supplementary experiments on the contribution of chromatin, the fairness of comparisons and evaluation reliability. We respond to each point below.


W1.1: Contribution of chromatin information.

R: To isolate measured chromatin’s contribution beyond static priors, we compare Full, w/o ATAC, and ATAC Shuffle in all three settings (as shown in Table 1). We agree that this comparison is needed to support the paper’s central claim. The w/o ATAC control removes measured ATAC and its derived context, matching and routing inputs while retaining static cCRE, DNA-sequence and ESM-2 priors and providing only RNA to CLM-X.

Table 1: ATAC removal and shuffling across three settings.

SettingModel variantPearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowMSE↓\mathrm{MSE}\downarrowMAE↓\mathrm{MAE}\downarrow
K562 TargetFull0.22100.32350.001290.01882
K562 Targetw/o ATAC0.20110.30480.001310.01891
K562 TargetATAC Shuffle0.19520.27400.001390.01978
GSE TargetFull0.17350.36970.0010540.018632
GSE Targetw/o ATAC0.14520.29890.001180.01932
GSE TargetATAC Shuffle0.11460.28010.001200.01963
GSE CrossFull0.18290.33070.0011090.01725
GSE Crossw/o ATAC0.15020.32580.001210.01819
GSE CrossATAC Shuffle0.12590.27120.001240.01875

Removing ATAC lowers response correlations, and shuffling correspondence leads to a further decline, suggesting that incorrect correspondence can introduce mismatched regulatory context. We have added this experiment to revised Table 6(b) and Appendix A.7, Table 13.

W1.2: Target specificity with shared K562 controls.

R: Shared K562 controls provide a common cellular background, while each target’s own TSS/regulatory-locus descriptors distinguish targets. Specifically, these features enter conditioning and gate neighborhoods, while GSE Cross additionally uses control ATAC for state matching. To further assess component contributions in K562, we added the following ablations:

Table 2: K562 prior-component ablations.

ModelPearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowMSE↓\mathrm{MSE}\downarrowMAE↓\mathrm{MAE}\downarrow
w/o cCRE0.21940.31060.001310.01892
w/o Multi-TSS0.21800.31790.001310.01892
w/o ESM-20.21200.30450.001310.01891
w/o CLM-X0.21090.31190.001320.01891
Pretrained priors only0.20890.30430.001340.01902
Regulatory priors only0.20000.30160.001330.01901
Full0.22100.32350.001290.01882

As shown in the table, our method outperforms all six ablations, supporting the complementary roles of molecular representations and regulatory information. To present these advantages more clearly, we have explained the computational pathways of ATAC in revised Sec. 3.1 and added the experimental results and ablation definitions to main-text Table 6(e).

Please see our response to Q2 for the detailed cross-state routing controls.


W2.1: Role of cell composition.

R: We agree that perturbation-induced changes in cell-type abundance can be an important phenotype. Our current task is expression-response prediction within cell states, followed by population aggregation; it does not predict abundance changes. To assess the sensitivity of its population-level predictions to composition information, we replaced the response-aggregation weights with training-control proportions. As shown in Table 3(a), this substitution preserves most of the model's performance.

Table 3(a): Response-aggregation weights on GSE Target.

Readout/protocolPearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowMSE↓\mathrm{MSE}\downarrowMAE↓\mathrm{MAE}\downarrow
ChromaPert0.17350.36970.0010540.018632
Training-control proportions (proxy for unknown composition)0.17100.36700.001060.01874

After replacing the weights, the correlations change only slightly, indicating that response aggregation is relatively robust to this adjustment. This control has been added to revised Appendix A.8, Table 14. At the same time, we recognize that changes in abundance are meaningful perturbation phenotypes [1], although studying this question requires data and experimental designs that support evaluating composition changes. We will continue to follow this direction in future work, and thank you again for your suggestion.

W2.2: Within-state prediction and baselines.

R: Our method also outperforms the baselines on the within-state prediction task. Specifically, we separately evaluate the state branch on the same 206 target–state groups, excluding gating and population aggregation. Scoring excludes the query gene, averages within each target, and then averages across 19 targets:

Table 3(b): Within-state prediction over 206 target–state groups.

ModelWithin-state PearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowMSE↓\mathrm{MSE}\downarrow
ChromaPert0.28210.00191
OLS0.26550.00193
Ridge0.25980.00203

ChromaPert outperforms OLS and Ridge in both correlation and MSE, while simple models remain competitive, consistent with the emphasis on strong baselines in previous work [2]. To clarify which prediction task these results address, we have explained the task scope in revised Sec. 4.1 and added the comparison to Appendix A.8, Table 15.


W3: Multimodal baselines and target specificity.

R: Thank you for suggesting that we strengthen the comparisons. Accordingly, we have added comparisons of CLM-X multimodal representations using the same adapters, together with evaluation after removing the shared response. Table 4 fixes the control cells, target and locus information, and the downstream Ridge or MLP adapter across all 19 GSE Target queries.

Table 4: Raw RNA–ATAC versus frozen CLM-X [3] representations with matched adapters.

Representation + adapterPearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowMSE↓\mathrm{MSE}\downarrowMAE↓\mathrm{MAE}\downarrow
Raw input + Ridge0.05140.08680.001150.01959
Frozen CLM-X + Ridge0.05280.08320.001130.01933
Raw input + MLP0.02250.00040.001400.02368
Frozen CLM-X + MLP0.0063-0.02080.001380.02327

CLM-X with Ridge slightly improves some metrics, but gains are inconsistent for DE correlation and the MLP adapter, indicating that representation value also depends on the downstream predictor.

To address the concern that a shared response may inflate correlations [2,4], we construct the shared response using training data only and evaluate correlation and target retrieval after removing it and excluding the query gene:

Table 5: Shared-response removal and target-retrieval diagnostics.

SettingModelAfter shared-response removal: PearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowTarget retrieval: PDS↑\mathrm{PDS}\uparrow
K562 TargetChromaPert0.12230.5879
K562 TargetCellFlow0.06890.5686
GSE TargetChromaPert0.20830.5997
GSE TargetCellFlow0.17000.5760
GSE CrossChromaPert0.38810.6003
GSE CrossCellFlow0.29100.5221

ChromaPert exceeds CellFlow [5] on both diagnostics in all three settings, providing supplementary evidence that it preserves target differences beyond the shared average response. To respond more directly to your question about target specificity, we have added these controls and evaluation definitions to revised Appendix A.9, Table 17, and Appendix A.10, Table 18.


W4.1: Results and configurations across tables.

R: CellFlow’s table discrepancies arise from configuration differences: the main text uses benchmark adaptations, including cell-state covariates where applicable, whereas the original appendix uses official defaults [5]. To address your concerns, we therefore added three-seed runs with the main-text adaptations shown in Table 6(a).

Table 6(a): Three-seed mean ± sample SD using the main-text adaptations.

SettingMethodPearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowMSE↓\mathrm{MSE}\downarrowMAE↓\mathrm{MAE}\downarrow
K562 TargetCellFlow0.1315 ± 0.00720.0820 ± 0.06120.00245±8e-60.02398±0.00012
K562 TargetChromaPert0.2302 ± 0.00110.3258 ± 0.00670.00129±1e-60.01882±5e-6
GSE TargetCellFlow0.0713 ± 0.00830.1480 ± 0.03690.00115±0.000010.01931±0.00006
GSE TargetChromaPert0.1580 ± 0.01760.3341 ± 0.03930.00108±0.000040.01883±0.00022
GSE CrossCellFlow0.1444 ± 0.00210.2197 ± 0.01010.01908±0.000540.04349±0.00073
GSE CrossChromaPert0.1757 ± 0.00100.3251 ± 0.00160.00110±1e-60.01727±0.00001

ChromaPert has higher mean correlations and lower mean errors in all three settings. To demonstrate our method more clearly, we report seed variation and condition-level uncertainty separately in revised Appendix A.3, Table 8, and Appendix A.13, Table 23.

W4.2: Parameter selection and stability.

R: We use nested cross-validation to select hyperparameters, with inner validation restricted to the outer-training data and scoring conducted on the outer held-out fold. To make this procedure clearer, we have explained it in revised Appendix C.3 and listed the default configurations for K562, GSE Target and GSE Cross in Table 31.

W4.3: Evaluation on other datasets.

R: Thank you for suggesting broader evaluation. To address your concerns, we have added Dasatinib-treated K562 [6] and non-hematopoietic RPE-1 CRISPRi Multiome Perturb-seq [7], with Target-LOO over 18 and 13 targets, respectively. CAT-ATAC iPSC has few eligible targets, so we use these two backgrounds for evaluation under the same protocol:

Table 7: Dasatinib-treated K562.

MethodPearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowL2↓L2\downarrowMSE↓\mathrm{MSE}\downarrowMAE↓\mathrm{MAE}\downarrowW1↓\mathrm{W1}\downarrowDisc. Cos.↑\mathrm{Disc.\ Cos.}\uparrow
Control——1.62280.001370.019186.6868—
GEARS0.06360.03052.87140.004160.047376.72210.0441
CellFlow0.04620.06971.89810.001860.021206.73990.0423
DeSCOPE0.12390.21181.62260.001370.019906.69920.1247
ChromaPert0.16290.28811.62080.001360.019146.69180.1675

Table 8: RPE-1.

MethodPearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowL2↓L2\downarrowMSE↓\mathrm{MSE}\downarrowMAE↓\mathrm{MAE}\downarrowW1↓\mathrm{W1}\downarrowDisc. Cos.↑\mathrm{Disc.\ Cos.}\uparrow
Control——2.67740.003940.026959.8278—
GEARS0.19050.41844.17630.008980.064789.88030.1891
CellFlow0.10750.27583.38730.006020.0325110.02500.1066
DeSCOPE0.27290.37892.77830.004130.029169.86150.2737
ChromaPert0.37050.49612.53180.003410.025659.79710.3704

ChromaPert has higher response correlations than the listed baselines in both the drug-treated and other-cell-line settings, with competitive pointwise errors. We have added these experiments to revised Sec. 4.5, Tables 4–5.


Q1: RNA-only evaluation across settings and K562 ablations.

R: Thank you for your suggestion. To address your concerns, We have completed the primary-metric evaluation without ATAC in all three settings and the six K562 component ablations, with results in Tables 1 and 2, respectively. The ATAC-removal control retains the other static priors and RNA inputs. To facilitate comparison, we have added both sets of results to revised Table 6(b, e) and Appendix A.7, Table 13.

Please see our response to W1.1 for the detailed removal scope of this control.


Q2: Cross-state routing comparisons.

R: To address your concerns, we designed additional experiments to validate the effectiveness of ATAC soft routing, as shown in Table 9. Specifically, comparisons with other routing methods on the same source-response bank demonstrate the effectiveness of ATAC soft routing. Thank you for suggesting these controls. Accordingly, we compared the following six routing rules:

Table 9: GSE Cross routing controls.

Routing rulePearsonΔ↑\mathrm{Pearson}_{\Delta}\uparrowPearsonΔ,DE↑\mathrm{Pearson}_{\Delta,\mathrm{DE}}\uparrowMSE↓\mathrm{MSE}\downarrowMAE↓\mathrm{MAE}\downarrow
ATAC soft routing0.18290.33070.0011090.01725
RNA soft routing0.14360.31280.001150.01817
Uniform weights0.13240.29990.001150.01900
Training-control frequency weights0.14810.31750.001100.01819
Nearest ATAC state0.15370.28360.001220.01869
Nearest RNA state0.14400.27290.001530.02020

This comparison supports the value of using control accessibility to match training states. The differences from uniform weights and nearest-state retrieval also demonstrate the role of soft combination. To help readers understand this matching procedure more clearly, we have added the results and routing definitions to revised main-text Table 6(f).

Q3: Matched-information baselines and test-cell counts.

R: With the same composition information, target/locus priors, source construction and readout, ChromaPert exceeds the listed simple predictors on both correlations (as shown in Table 10).

Table 10: Simple models and replacements within the same framework. Each cell reports PearsonΔ\mathrm{Pearson}_{\Delta} / PearsonΔ,DE\mathrm{Pearson}_{\Delta,\mathrm{DE}}.

ModelK562 TargetGSE TargetGSE Cross
Independent training-perturbation mean0.1605 / 0.22400.1129 / 0.26230.0378 / 0.0694
Mean0.1759 / 0.26400.1412 / 0.31310.0527 / 0.1705
Ridge regression0.1807 / 0.27160.1473 / 0.29560.1309 / 0.3056
Direct MLP0.1862 / 0.28030.1534 / 0.30180.1411 / 0.3167
ChromaPert0.2210 / 0.32350.1735 / 0.36970.1829 / 0.3307

Response aggregation can also use training-control proportions, as shown in Table 3(a), with the matched baseline held fixed. We have clarified the matched-information controls, the use of composition information, and the distinction between expression-response prediction and abundance changes in revised Appendices A.8–A.9 and Tables 14–16.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary for Submission 7976

Dear Area Chairs and Reviewers,

We sincerely thank all reviewers for their careful assessment and constructive feedback on our submission. We are particularly grateful for the recognition of the well-motivated research direction and detailed evaluation protocols (Reviewer muoQ), the strong performance and well-motivated architecture (Reviewer Jcne), and the potential of chromatin accessibility and DNA sequence information for perturbation prediction (Reviewer YsS2).

Summary of Updates

All concerns raised by the reviewers have been addressed point-by-point, including those related to chromatin contributions, baseline comparisons, target specificity, generalization across treatments and cell lineages, configuration consistency, uncertainty estimation, modality imputation, and reproducibility. We have incorporated the following key updates into our revised manuscript (revised text are highlighted in blue):

  1. Chromatin contribution. We have added Full, w/o ATAC, and ATAC Shuffle comparisons under the primary metrics across all three protocols, together with K562 component ablations and cross-state routing controls. We also compared RNA routing, uniform weights, training-control frequencies, and nearest-state retrieval, providing direct evidence for the contribution of chromatin context and the importance of correctly matching chromatin information. (See Table 6(b), (e), and (f), Appendix A.7, and Table 13.)

  2. Additional baselines. We have added comparisons with MultiFlow, SCENIC+, and LINGER, and updated our adaptations of scFoundation and CellVQ. Across all 133 GSE Cross conditions, ChromaPert achieves a PearsonΔ of 0.1829, compared with 0.1140 for MultiFlow. On matched target–gene subsets, ChromaPert outperforms regulatory-network methods, achieving higher response correlations and lower errors. (See Tables 2 and 6(a), Appendices A.14–A.15, and Table 24.)

  3. Matched-information baselines and target-specific analyses. We have added training-set mean, Ridge, and Direct MLP baselines under matched-information settings, and compared frozen CLM-X representations using the same adapters. ChromaPert outperforms the listed simple predictors on both response-correlation metrics across all three settings. We further added composition-weight controls and evaluated the state branch across 206 target–state groups to clarify the role of composition information in the prediction task. Shared-response removal and target-retrieval analyses further assess target specificity beyond shared average responses. (See Appendices A.8–A.10 and Tables 14–18.)

  4. Broader coverage. We have added the Dasatinib-treated K562 group from CAT-ATAC and an independent, non-hematopoietic RPE-1 CRISPRi Multiome Perturb-seq dataset, completing Target-LOO evaluations on 18 and 13 targets, respectively, and reporting all seven metrics. ChromaPert achieves higher response correlations than the compared predictors in both backgrounds. On RPE-1, its PearsonΔ is 0.3705, compared with 0.2729 for DeSCOPE. (See Sections 4.1 and 4.5, Tables 4–5, and Appendices A.16 and B.1.)

  5. Reliability. To address concerns about the differences in the CellFlow results, we have added three-seed comparisons using the adapted configurations reported in the main text. ChromaPert achieves higher mean response correlations and lower mean MSE and MAE across all three settings. We also separately report training-seed standard deviations, fold-level standard deviations, and paired-bootstrap intervals, allowing readers to assess performance gains together with their uncertainty. (See Appendices A.3 and A.13, Table 8, and Tables 22–23.)

  6. Unpaired controls. Following the future work suggestion to consider unpaired controls, we have evaluated CLM-X and scDiffusion-X for both RNA→ATAC and ATAC→RNA imputation on K562 and GSE. These additional diagnostics assess the feasibility and limitations of modality imputation and inform future extensions to unpaired-control settings. (See Appendix A.17 and Tables 26–27.)

  7. Readability and reproducibility. We have clarified feature construction, input boundaries, source operations, gate weighting, and training configurations. We have also distinguished benchmark metrics from auxiliary slope diagnostics, and expanded the analysis of experimental results. The supplementary training code has also been updated. (See Sections 3.1–3.3 and 4.1–4.6, Figure 3, Tables 19–21 and 25, Appendices B–C, and the Reproducibility Statement.)

We sincerely appreciate the reviewers' valuable feedback, which has helped us improve the experimental validation, methodological clarity, and overall presentation of our work.

Sincerely,

Authors of Submission 7976

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.