acceptodds
Under review as a conference paper at ICLR 2027

Virtual Cell Models Make Systematic and Transferable Prediction Errors

Abstract

Single-cell perturbation models predict transcriptional responses to unseen genetic or chemical interventions, but their prediction errors are commonly treated as residual noise. We present the first systematic study of whether these errors recur across perturbations and transfer to unseen interventions. Across diverse predictors and genetic, combinatorial, and drug perturbation datasets, error fingerprints estimated from training perturbations recur on held-out perturbations and anticipate which genes and biological pathways will be misestimated. The fingerprints are stable across training seeds, and much of their model-specific structure reflects persistent gene-wise over- and under-estimation. We also introduce Gap Transferability Correction (GTC), a lightweight post-hoc framework that reuses historical errors to improve unseen-perturbation predictions without retraining the base model. GTC improves differentially expressed gene recovery in all three perturbation settings. Experiments across multiple cell lines show that fingerprints transfer more reliably within a cell line than across cell lines, identifying cellular context as a boundary condition for error transfer. These results establish prediction error as a structured, transferable, and reusable property of virtual-cell models.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell perturbation prediction, Virtual cell models, Model fingerprints, Prediction error analysis, Generalization to unseen perturbations

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_9gaMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper asks whether prediction errors in single-cell perturbation models contain stable structure that transfers across settings. The authors find that different virtual cell models consistently over- or underpredict specific genes. These “error fingerprints” persist across unseen genes, combinatorial perturbations, and drug perturbations, and can help predict which genes and pathways are likely to be mispredicted.

The authors use this observation to develop Gap Transferability Correction (GTC), a simple post-hoc method that uses error patterns observed during training to adjust predictions for unseen perturbations without retraining the original model. The main contribution is to characterize prediction error as a learnable, transferable property of a model and show how it can be used for correction.

Critical strengths
  1. The paper offers a novel and practically relevant perspective on model evaluation by asking whether systematic, repeatable errors can predict future failure patterns. This question matters in single-cell perturbation studies, where experimental feedback accumulates over time. If errors have stable structure, earlier experimental results can support post-hoc correction without retraining the base model. The paper formalizes this idea through transferable error fingerprints, evaluates their transferability at the gene, pathway, and perturbation-setting levels, and uses them to build a practical correction framework.
Critical weaknesses

1. Related work does not sufficiently cover causal representation learning.

The related-work discussion focuses on single-cell perturbation prediction, reliability analysis, and model fingerprinting, with limited coverage of recent causal and intervention-aware representation learning methods. Approaches such as Discrepancy-VAE [1], sVAE+[2], and PerturbedVAE[3] address closely related questions: modeling perturbation mechanisms in latent space, separating background variation, and improving generalization to unseen or combinatorial perturbations. PerturbedVAE also examines why complex models do not consistently outperform simple linear baselines and how better representations can improve perturbation prediction. Without discussing these connections, the paper’s relationship to existing representation-based perturbation modeling approaches remains unclear.

2. The baseline coverage is somewhat outdated and lacks recent strong models.

Although the paper compares 10 perturbation predictors, most baselines date from 2024 or earlier. Given progress in virtual-cell and perturbation prediction over the past two years, this provides limited support for the broader claim that transferable error fingerprints are a general property of virtual-cell models. The evaluation should include at least one recent strong baseline, such as STATE [4]. The current evidence mainly concerns earlier models, leaving it unclear whether the finding extends to the latest generation of perturbation predictors.

[1] Zhang J, Greenewald K, Squires C, et al. Identifiability guarantees for causal disentanglement from soft interventions[J]. Advances in Neural Information Processing Systems, 2023, 36: 50254-50292.

[2] Lopez R, Tagasovska N, Ra S, et al. Learning causal representations of single cells via sparse mechanism shift modeling[C]//Conference on Causal Learning and Reasoning. PMLR, 2023: 662-691.

[3] Jiang W, Liu Y, Cai Y, et al. What Makes a Representation Good for Single-Cell Perturbation Prediction?[J]. arXiv preprint arXiv:2605.19343, 2026.

[4] Adduri A K, Gautam D, Bevilacqua B, et al. Predicting cellular responses to perturbation across diverse contexts with State[J]. Cell, 2026, 189(19): 5914-5931. e20.

Questions for authors

Q1: Could the authors expand the discussion of causal representation learning methods, including Discrepancy-VAE, sVAE+, and PerturbedVAE? How does GTC differ from methods that explicitly learn perturbation-relevant or mechanism-aware representations in terms of problem setting, objective, and OOD perturbation prediction?

Q2: Could the authors include strong baselines from 2025 and 2026, including at least STATE (Adduri et al., 2026), and test whether error fingerprints persist and GTC remains effective on these newer models?

5 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/n)

We sincerely thank the reviewer for recognizing the novelty and practical relevance of our work. We particularly appreciate the insightful suggestions on causal representation learning and baseline coverage, which have helped us strengthen the manuscript.

Summary. We expanded the discussion of causal representation learning and incorporated STATE, PerturbedVAE, and Rhaister, increasing the benchmark from ten to thirteen models. The additional experiments demonstrate that systematic prediction differences persist across diverse predictors, while historical-error correction remains effective for STATE and PerturbedVAE across multiple perturbation settings.


W1 / Q1. Relationship to causal representation learning.

AW1 / AQ1. We appreciate the reviewer highlighting these closely related studies. We have expanded Section 2 to discuss Discrepancy-VAE [1], sVAE+ [2], and PerturbedVAE [3], clarifying their relationship to our work.

These approaches investigate perturbation-relevant representations: Discrepancy-VAE studies causal identifiability under interventions, while sVAE+ and PerturbedVAE model perturbation-related changes in latent representations. Their objectives concern learning representations that improve the prediction and generalization of intervention effects.

Our work addresses a complementary question: whether systematic errors remaining in a trained predictor can anticipate future failures and be reused for correction. Rather than modifying the predictor's representations or training objectives, GTC estimates recurring errors from historical predictions and experimental observations, then transfers them to unseen perturbations without retraining the original model.

Thus, representation-learning methods seek to improve OOD prediction through learned perturbation representations, whereas GTC exploits predictable structure in the remaining errors. We further demonstrate their compatibility by applying GTC to PerturbedVAE.

Revision: We expanded Section 2.1 to discuss intervention-aware representation learning and clarify its relationship to our post-hoc error analysis, with the GTC formulation presented in Section 4.1.

References:

[1] Zhang et al. Identifiability Guarantees for Causal Disentanglement from Soft Interventions. NeurIPS, 2023. https://proceedings.neurips.cc/paper_files/paper/2023/hash/9d3a4cdf6f70559e8c6fe02170fba568-Abstract-Conference.html

[2] Lopez et al. Learning Causal Representations of Single Cells via Sparse Mechanism Shift Modeling. CLeaR, 2023. https://proceedings.mlr.press/v213/lopez23a.html

[3] Jiang et al. What Makes a Representation Good for Single-Cell Perturbation Prediction? ICML, 2026. https://proceedings.mlr.press/v306/jiang26ad.html

AuthorsOpenReview ↗

Official Comment by Authors (2/n)

W2 / Q2. Recent baselines and the generality of error fingerprints.

AW2 / AQ2. We appreciate the reviewer's suggestion to include more recent perturbation predictors. We have incorporated STATE [4], PerturbedVAE [3], and Rhaister [5] into our fingerprint and correction evaluations.

Systematic prediction differences persist across these models. Across the eight settings in Figure 2(a), STATE and PerturbedVAE achieve mean real-vs-predicted AUROCs of 0.906 and 0.994, respectively. With all thirteen predictors included, model-attribution balanced accuracy reaches 0.979, substantially exceeding the chance level of 0.077. These results demonstrate that identifiable prediction patterns extend beyond the original model collection.

Historical-error correction also remains effective. Using Global+Local GTC with predicted-response neighborhoods, STATE and PerturbedVAE improve in 9 of 10 evaluation scenarios each, covering genetic, combinatorial, and chemical perturbations. Their improvements are particularly pronounced on unseen perturbations. Rhaister shows smaller correction gains, improving in 4 of 10 scenarios.

The complete results are summarized below. Values denote mean DEG@50 changes in percentage points relative to each base predictor.

Dataset and evaluation scenarioSTATEPerturbedVAERhaister
Norman2019: held-out cells+24.09+6.03−0.03
Norman2019: unseen combinations+10.47+5.81−0.05
Replogle K562: unseen gene perturbations+12.44−0.820.00
SciPlex3: unseen drugs, shared models+8.37+1.37+0.14
SciPlex3: unseen drugs, per-cell-line models+9.61+1.26−0.05
McFarland2020: unseen cell lines+2.20+1.70+0.70
Replogle–Nadig: unseen cell lines+1.48+0.720.00
Replogle–Nadig: unseen gene perturbations+19.93+15.62+0.60
Tahoe-100M: unseen cell lines−0.43+0.24−0.18
Tahoe-100M: unseen drugs+18.86+12.42+0.10
Scenarios improved / evaluated9/109/104/10

Scores are averaged across cellular contexts within each scenario. Improvements are determined using unrounded values.

Rhaister provides a complementary perspective. Its predictions are constructed from observed experimental summary statistics [5], potentially explaining their lower real-vs-predicted separability and the limited additional gains from GTC.

We also extended the effect-space real-vs-predicted and model-attribution analyses to all six datasets (Figure S2). Together, these results provide broader evidence that systematic prediction differences persist across diverse model architectures and that historical errors can improve predictions from models with different learning objectives.

Revision: We expanded the model descriptions in Section 3.2, incorporated the additional predictors into Figure 2(a) and Tables 1–3, and provided model configurations and detailed performance results in Appendices C and D, together with expanded fingerprint analyses in Appendix G (Figure S2).

References:

[3] Jiang et al. What Makes a Representation Good for Single-Cell Perturbation Prediction? ICML, 2026. https://proceedings.mlr.press/v306/jiang26ad.html

[4] Adduri et al. Predicting Cellular Responses to Perturbation across Diverse Contexts with State. Cell, 2026.

[5] Svensson et al. Back to Basics: Observed Statistics Are Sufficient to Predict Drug Responses. bioRxiv, 2026. https://doi.org/10.64898/2026.06.09.731197

Dear Reviewer 9gaM,

We appreciate your suggestions on causal representation learning and baseline coverage. We expanded Section 2.1 to discuss Discrepancy-VAE, sVAE+, and PerturbedVAE, and incorporated STATE, PerturbedVAE, and Rhaister into our benchmark (Figure 2(a), Tables 1–3). The additional results confirm that distinctive prediction patterns persist, while GTC improves STATE and PerturbedVAE in 9/10 evaluated scenarios each (Appendix D).

We hope these additions address your concerns and would appreciate your consideration of the new evidence in your final assessment. We remain happy to clarify any outstanding questions.

Reviewer_1pCVOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This work presents an empirical study of systematic errors in single-cell perturbation prediction models. The authors demonstrate that these errors exhibit predictable patterns across perturbations and can be leveraged to improve predictions on unseen perturbations through a post-hoc correction framework (GTC). The study spans genetic, combinatorial, and drug perturbations, across multiple predictive models and evaluation metrics, providing additional robustness to the findings

Critical strengths
  • The research question is relevant and novel. The proposed perspective can potentially inform both model development and evaluation, particularly given the limitations of current virtual-cell models.
  • Broad empirical evaluation. The study covers ten prediction models across multiple settings, including single-gene, combinatorial, and chemical perturbations.
  • Broad/careful characterization of error structure. The authors investigate multiple aspects of the observed fingerprints, including reproducibility across training seeds, consistency at different levels, and the contribution of gene-specific biases.
  • The CTG framework does not only further strengthen the finding, but also provides a simple and useful mechanism to improve existing models in practical applications. The investigation of cell-line dependence, although not really novel, is particularly valuable in practical applications.
Critical weaknesses
  • Limited understanding of the sources of systematic errors. While the study demonstrates well that prediction errors exhibit reproducible patterns, it provides limited insights into what drives these patterns, e.g. why particular genes or pathways are consistently mispredicted. Further analyses connecting these errors to biological properties, model characteristics, etc., would make the findings more informative.
  • Limited validation across biological contexts. Despite evaluating multiple models and settings, the study relies on three datasets, with genetic perturbation experiments restricted to a single cell line. Additional validation across more datasets, especially more diverse biological settings, would strengthen the generalizability of the findings.
Questions for authors

See Weaknesses

5 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/n)

We sincerely thank the reviewer for recognizing the novelty and practical significance of our work. The insightful suggestions on understanding systematic errors and evaluating their generalizability have helped us strengthen the manuscript.

Summary. We added a systematic analysis of gene properties and model characteristics across 13 models and three datasets, revealing consistent factors associated with recurring prediction errors. We also extended our evaluation to three additional datasets, covering genetic and chemical perturbations across diverse cellular contexts.


W1. Limited understanding of the sources of systematic errors.

AW1. We appreciate the reviewer's important question about what drives recurring prediction errors. To investigate this, we conducted a systematic analysis across 13 models and three datasets (Norman2019, Replogle, and SciPlex3), yielding three main findings.

  • Genes with stronger perturbation responses exhibit larger prediction errors. After adjusting for baseline expression, detection rate, expression variability, and differential-expression frequency, response magnitude remains positively associated with held-out prediction errors in all 65 model–context comparisons. The median partial correlations are 0.40, 0.70, and 0.47 across the three datasets, respectively.
  • Model design is associated with distinct prediction-error patterns. Models combining control-expression inputs with an expression encoder–decoder generally make fewer large predictions when true responses are weak, but more frequently miss strong responses. In SciPlex3, the group-median rates are 0.25% vs. 5.32% for large predictions on weak responses, and 73.7% vs. 46.4% for missed strong responses. A similar trade-off is observed in Replogle.
  • Pathway-level errors partly reflect the properties of their constituent genes. Elevated errors in pathways such as E2F Targets are substantially reduced after comparison with genes matched for expression level and response magnitude. This suggests that apparent pathway-specific prediction difficulties are largely associated with the strongly responding genes within these pathways.

Together, these findings connect recurring errors to biological response properties and model-dependent prediction behavior, providing insight into why certain genes and pathways are repeatedly mispredicted.

Further identifying how specific architectural choices and training objectives shape these errors would require systematic ablations across models and datasets. We consider this a promising direction for understanding model behavior and guiding the development of more reliable predictors, while the present work focuses on characterizing transferable errors and demonstrating how historical errors can be reused.

Revision: We added a summary of the findings to Section 3.4, with detailed analyses of gene properties and model-dependent prediction behavior in Appendix H (Figure S4).

AuthorsOpenReview ↗

Official Comment by Authors (2/n)

W2. Limited validation across biological contexts.

AW2. We appreciate the reviewer's emphasis on biological diversity, particularly for genetic perturbations.

Our original evaluation covers three complementary settings: combinatorial genetic perturbations (Norman2019), unseen single-gene perturbations (Replogle K562), and chemical perturbations (SciPlex3).

We further extended the evaluation to Replogle–Nadig [1], McFarland2020 [2], and Tahoe-100M [3], incorporating additional genetic and chemical perturbations across diverse cellular backgrounds. In particular, Replogle–Nadig includes K562, RPE1, HepG2, and Jurkat, extending genetic perturbation validation beyond the original K562 setting.

Using the same Global+Local GTC variant based on predicted-response neighborhoods, we obtained the following results across 13 models:

DatasetEvaluation settingMean DEG@50: Base → GTCModels improved
Replogle–NadigUnseen context0.1415 → 0.157412/13
Replogle–NadigUnseen perturbations0.2117 → 0.296713/13
McFarland2020Unseen context0.2935 → 0.322812/13
Tahoe-100MUnseen context0.2686 → 0.280210/13
Tahoe-100MUnseen perturbations0.3339 → 0.412511/13

GTC consistently improves mean performance across all five additional evaluation settings, including unseen cellular contexts in both genetic and chemical perturbation datasets.

We also extended the effect-space real-vs-predicted and model-attribution analyses to all six datasets (Figure S2), providing further evidence that systematic prediction differences persist across diverse biological contexts.

Revision: We expanded the dataset descriptions in Section 3.2, added performance results across additional biological contexts in Section 4.2 and Table 4, and provided detailed evaluations in Appendix D.2, together with expanded fingerprint analyses in Appendix G (Figure S2).

References:

[1] Nadig et al. Transcriptome-wide analysis of differential expression in perturbation atlases. Nature Genetics, 57, 1228–1237 (2025). https://doi.org/10.1038/s41588-025-02169-3

[2] McFarland et al. Multiplexed single-cell transcriptional response profiling to define cancer vulnerabilities and therapeutic mechanism of action. Nature Communications, 11, 4296 (2020). https://doi.org/10.1038/s41467-020-17440-w

[3] Zhang et al. Tahoe-100M: Mapping drug-induced molecular phenotypes at single-cell resolution. Cell, 189, 5945–5961.e8 (2026). https://doi.org/10.1016/j.cell.2026.08.035

Dear Reviewer 1pCV,

Thank you again for your encouraging assessment and thoughtful suggestions. Following your comments, we investigated the sources of systematic errors across 13 models and three datasets, finding consistent associations with gene response magnitude and model design (Section 3.4, Figure S4). We also added three datasets covering more diverse cellular contexts, including genetic perturbations across multiple cell lines, and evaluated GTC in unseen-context settings (Table 4, Appendix D).

We would be glad to hear whether these additional analyses address your concerns about error sources and biological generalizability before the discussion period ends.

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

The paper shows that virtual-cell models make systematic errors that transfer across perturbations and introduces a post-hoc method to correct them and improve predictions on unseen perturbations.

Critical strengths

The topic is interesting and novel

Critical weaknesses
  1. The analysis is limited to pseudobulk responses, and this overlooks potentially systematic errors in the predicted single-cell distributions. It would be (more) interesting to investigate whether error fingerprints also exist at the single-cell distribution level.
  2. The paper evaluates error transfer across unseen perturbations, but not across unseen cell types, which is arguably a more widely studied generalization setting in the literature.
Questions for authors
  1. Figure 1a is confusing. Why are predictions, ground truth, and errors represented as matrices rather than vectors? What do the different rows within each matrix represent? Also, a color scale is missing, making it difficult to understand what the colors mean.
  2. Why is Tahoe not included in the evaluation?
  3. Could the authors also evaluate Rhaister as a baseline?
7 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/n)

We sincerely thank the reviewer for recognizing the novelty of our work. We particularly appreciate the constructive suggestions on biological scope, model coverage, and presentation, which have helped us strengthen the manuscript.

Summary. Our primary objective is to investigate whether historical prediction errors can anticipate failures on unseen perturbations and support post-hoc correction. We have extended the evaluation to single-cell expression distributions and additional biological contexts, while maintaining this central focus. We have also incorporated Tahoe-100M and Rhaister, alongside improvements to Figure 1.


W1. The analysis is limited to pseudobulk responses.

AW1. We appreciate the reviewer's suggestion to investigate whether systematic error patterns also exist at the single-cell level.

Our primary analysis focuses on systematic gene-level errors in perturbation responses. Aggregating single-cell profiles provides a common condition-level prediction target, reduces cell-level sampling variability, and facilitates analysis of differential expression and pathway-level responses. This choice follows established aggregation-based analyses of transcriptional changes [1, 2] and perturbation modeling using gene-level expression responses [3].

To extend this investigation beyond mean responses, we conducted two complementary classification tasks directly on individual-cell expression profiles:

  • Real vs. predicted: Distinguishing observed cells from model-predicted cells.
  • Model attribution: Identifying the prediction model from its generated single-cell expression profiles.

Across 12 evaluation scenarios, the ten models with cell-level outputs achieve mean real-vs-predicted AUROCs ranging from 0.93 to 1.00 when averaged over available scenarios for each model. Moreover, ten-class model attribution achieves 0.98 mean balanced accuracy on held-out perturbations.

These results provide strong evidence that reproducible, model-specific differences also exist at the single-cell level (Figure S3). We retain pseudobulk responses as the primary setting because they directly support our central investigation of transferable gene-level errors and their biological consequences.

We added the single-cell classification results to Section 3.3, with the evaluation protocol detailed in Appendix G and complete results presented in Figure S3.

References:

[1] Crowell et al. muscat detects subpopulation-specific state transitions from multi-sample multi-condition single-cell transcriptomics data. Nature Communications, 11, 6077 (2020). https://doi.org/10.1038/s41467-020-19894-4

[2] Squair et al. Confronting false discoveries in single-cell differential expression. Nature Communications, 12, 5692 (2021). https://doi.org/10.1038/s41467-021-25960-2

[3] Xing and Yau. GPerturb: Gaussian process modelling of single-cell perturbation data. Nature Communications, 16, 5423 (2025). https://doi.org/10.1038/s41467-025-61165-7

AuthorsOpenReview ↗

Official Comment by Authors (2/n)

W2. Error transfer is not evaluated across unseen cell types.

AW2. We appreciate the reviewer's emphasis on cross-context generalization, an important complementary setting.

As stated in our Introduction, “We ask a different question: are prediction errors themselves reproducible across perturbations?” Our primary focus is therefore error transfer to unseen perturbations. The original Figure 4 already demonstrated the importance of cellular context in determining error transferability.

We have now extended our evaluation to Tahoe-100M [4], Replogle–Nadig [5], and McFarland2020 [6], covering additional genetic and chemical perturbations across diverse cellular backgrounds.

In the unseen-context settings, GTC improves mean DEG@50 from 0.269 to 0.281 on Tahoe, 0.141 to 0.158 on Replogle–Nadig, and 0.293 to 0.323 on McFarland2020, with gains of up to +1.2, +1.6, and +2.9 percentage points, respectively.

The additional unseen-perturbation evaluations also show improvements, reaching +8.3 points on Tahoe and +8.5 points on Replogle–Nadig.

We further extended the effect-space real-vs-predicted and model-attribution analyses to these datasets (Figure S2). The expanded evaluation thus provides complementary evidence of systematic prediction differences and the utility of historical-error correction across diverse biological contexts.

Revision: We expanded the dataset descriptions in Section 3.2, added cross-context performance results in Section 4.2 and Table 4, and provided detailed evaluations in Appendix D.2, together with expanded classification analyses in Appendix G (Figure S2).

References:

[4] Zhang et al. Tahoe-100M: Mapping drug-induced molecular phenotypes at single-cell resolution. Cell, 189, 5945–5961.e8 (2026). https://doi.org/10.1016/j.cell.2026.08.035

[5] Nadig et al. Transcriptome-wide analysis of differential expression in perturbation atlases. Nature Genetics, 57, 1228–1237 (2025). https://doi.org/10.1038/s41588-025-02169-3

[6] McFarland et al. Multiplexed single-cell transcriptional response profiling to define cancer vulnerabilities and therapeutic mechanism of action. Nature Communications, 11, 4296 (2020). https://doi.org/10.1038/s41467-020-17440-w

AuthorsOpenReview ↗

Official Comment by Authors (3/n)

Q1. Figure 1a is confusing: why matrices rather than vectors, what do the rows represent, and why is there no color scale?

AQ1. Thank you for pointing this out.

Each prediction, ground truth, and error is a gene-wise vector. The original schematic arranged each vector into two rows solely for visual compactness; the rows do not represent distinct cells or additional biological dimensions.

We have revised Figure 1a to explicitly identify each square as a gene, harmonize the color encoding, and introduce labeled color scales indicating expression intensity and prediction errors.

Revision: We revised Figure 1(a) and its caption to clarify the gene-wise vector representation, harmonize color encodings, and introduce labeled color scales. We have also reviewed the remaining figures and captions to improve the consistency of notation and visual explanations.


Q2. Why was Tahoe not included?

AQ2. We appreciate the reviewer highlighting Tahoe-100M [4], an important resource for large-scale perturbation studies.

Our initial evaluation prioritized diversity across perturbation modalities, datasets, and prediction architectures, rather than dataset size alone. This enabled a consistent investigation of error fingerprints across genetic, combinatorial, and chemical perturbations.

Given Tahoe's scale of approximately 100 million cells, evaluating all predictors on the complete dataset would require substantial computational resources. Following the precedent of the MAP study [7], we adopted a representative six-cell-line subset, processed using the same VCBench pipeline.

On Tahoe, GTC improves mean DEG@50 from 0.334 to 0.417 (+8.3 points) for unseen perturbations and from 0.269 to 0.281 (+1.2 points) for unseen cellular contexts.

These findings extend our conclusions to Tahoe while preserving the diversity and comparability of the overall evaluation.

Revision: We added Tahoe-100M to Section 3.2, detailed its sampling and evaluation protocols in Appendix C.1, and reported the results in Table 4 and Appendix D.2.

References:

[4] Zhang et al. Tahoe-100M: Mapping drug-induced molecular phenotypes at single-cell resolution. Cell, 189, 5945–5961.e8 (2026). https://doi.org/10.1016/j.cell.2026.08.035

[7] Feng et al. A knowledge-driven framework for predicting single-cell responses for unprofiled drugs. Nature Machine Intelligence, 8, 1478–1491 (2026). https://doi.org/10.1038/s42256-026-01286-w


Q3. Could the authors also evaluate Rhaister as a baseline?

AQ3. We appreciate the reviewer highlighting Rhaister [8], which offers a promising alternative to conventional perturbation prediction models.

Rhaister constructs predicted responses through weighted combinations of experimentally observed screen-level summary statistics [8]. Its predictions are therefore closely grounded in real experimental responses, potentially making them less distinguishable from observed data.

We have incorporated Rhaister into our evaluation, adapting its context–perturbation formulation to the unseen-perturbation setting and applying GTC to its predicted expression changes.

Across the eight settings in Figure 2(a), Rhaister exhibits substantially lower real-vs-predicted separability, with a mean AUROC of 0.798, compared with 0.955 averaged over the other twelve predictors.

Its GTC improvements are also limited:

SettingAUROCBase DEG@50Global GTCG+L–Pert.
Norman20190.8810.75300.75300.7530
Replogle K5620.6910.26120.26120.2604
SciPlex3 shared K5620.8500.42110.42110.4216

We believe these observations are likely related to Rhaister's reliance on weighted combinations of observed experimental responses, which may reduce systematic prediction differences and leave less residual bias for GTC to correct.

These findings highlight the potential of summary-statistic-based prediction as a complementary direction for future studies of prediction reliability.

Revision: We incorporated Rhaister into Section 3.2, Figure 2(a), and Tables 1–3, documented its adaptation in Appendix C.4, and provided detailed performance and classification results in Appendices D and G, respectively.

Reference:

[8] Svensson et al. Back to basics: Observed statistics are sufficient to predict drug responses. bioRxiv (2026). https://doi.org/10.64898/2026.06.09.731197

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.