acceptodds
Under review as a conference paper at ICLR 2027

PerturbRx: Learning Treatment-Conditioned Latent Transitions from Single-Cell Perturbations for Patient Drug Response

Abstract

Scarce data and tumor heterogeneity limit patient-level prediction of cancer treatment response. Many existing approaches predict response from patient profiles and drug representations without explicitly modeling treatment-induced molecular changes. We propose PerturbRx, a treatment-conditioned representation learning framework that learns drug- and dose-conditioned latent transitions from context-matched control and treated single-cell populations. The learned transition operator generalizes to compounds excluded from source-stage training, predicting held-out treatment effects more accurately than baselines (mean cosine similarity 0.229 vs. 0.198). We freeze and transfer the operator to pretreatment patient profiles, using the learned treatment-induced transformation as an additional representation of how each patient state may respond to therapy, without requiring matched post-treatment measurements. The transferred representation achieves the strongest pooled out-of-fold performance among evaluated methods on TCGA-485 (AUROC/AUPRC: 0.709/0.794) and an independent PDX cohort (0.685/0.612), with promising improvements beyond strong patient+drug controls in both domains.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
drug response prediction, single-cell perturbation, treatment-conditioned representation learning, latent state transitions, cross-domain transfer, precision oncology

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

This paper proposes PerturbRx, a two-stage framework that transfers knowledge from single-cell perturbation data to patient drug-response prediction. Using frozen scFoundation and ChemBERTa representations, it learns drug- and dose-conditioned latent transitions from context-matched but unpaired control and treated cell populations in Tahoe-100M. Training combines distributional alignment with constraints on transition direction and magnitude. The learned transition model is then frozen and applied to tumor profiles, with the resulting treatment-conditioned features combined with baseline tumor and drug representations to train a response classifier without requiring matched post-treatment measurements. Experiments assess generalization to drugs excluded from source-stage training and response prediction on TCGA-485 and an independent PDX cohort. PerturbRx achieves the highest pooled out-of-fold performance among the evaluated methods, although its incremental gains over strong patient-state-plus-drug controls are modest and not statistically significant.

Critical strengths

The paper provides a useful connection between single-cell perturbation modeling and patient drug-response prediction. The evaluation combines drug-held-out tests, patient- and model-disjoint splits on TCGA and PDX, and representation and architecture controls, strengthening the empirical assessment of transferability and the added value of the learned transition features.

Critical weaknesses

The gains over controls using only patient and drug representations are small and not statistically significant, leaving the incremental predictive value of the learned transitions uncertain. Without matched post-treatment patient measurements, it also remains unclear whether the transferred features capture patient-specific biological responses or primarily provide a useful nonlinear transformation of existing representations.

Questions for authors

Are the gains over the Patient+Drug baseline consistent across training runs with different random seeds? Could you also compare with a frozen, randomly initialized transition network of the same architecture to clarify whether perturbation pretraining provides additional value beyond a nonlinear transformation of the existing representations?

4 comments
AuthorsOpenReview ↗

Response to Reviewer jyzj

We thank the reviewer for the constructive suggestions. The concern about random-seed variability motivated an additional robustness analysis, which revealed substantial initialization sensitivity and informed our revised downstream ensemble design. We also added the requested frozen-random transition control and validation against observed patient treatment-associated changes.

1. Incremental predictive value and random-seed robustness

Following the reviewer's suggestion, we evaluated 100 matched random initializations under the original single-head configuration using identical outer folds (Appendix C.4, Tables 22–23).

The original gains over Patient+Drug were not consistently reproduced: PerturbRx achieved higher AUROC in only 35/100 TCGA and 38/100 PDX runs. Across these runs, mean AUROC was 0.691 versus 0.696 on TCGA and 0.674 versus 0.679 on PDX (PerturbRx versus Patient+Drug, respectively).

This initialization sensitivity motivated a multi-architecture, multi-seed response-head ensemble. The revised predictor averages predictions from six response-head configurations across 20 random seeds, reducing reliance on any single initialization. We evaluated the revised protocol using the same grouped outer folds and corresponding representation controls.

Under this protocol, PerturbRx achieved pooled OOF AUROC/AUPRC of 0.726/0.804 on TCGA and 0.696/0.620 on PDX, compared with 0.704/0.786 and 0.692/0.589 for Patient+Drug and PDX+Drug, respectively.

The TCGA AUROC improvement (Δ = +0.0215, p = 0.0031) and PDX AUPRC improvement (Δ = +0.0303, p = 0.0027) were statistically significant. The PDX AUROC difference was not significant.

2. Frozen random transition control

Following the reviewer's recommendation, we evaluated an architecture-matched, randomly initialized transition network with frozen weights under the revised ensemble protocol.

PerturbRx achieved numerically higher AUROC/AUPRC than this control on TCGA (0.726/0.804 versus 0.715/0.785) and PDX (0.696/0.620 versus 0.688/0.597).

On PDX, the AUPRC difference was +0.0229 (95% CI: [0.0048, 0.0409], p = 0.0142), while the AUROC difference was not statistically significant.

These results suggest that perturbation pretraining may provide additional predictive value in specific settings, but they do not establish a consistent advantage attributable specifically to pretraining rather than an additional nonlinear transformation.

3. Biological validation

We additionally evaluated paired pre- and post-treatment transcriptomes from 278 patients across 27 independent cohorts.

Compared with a leave-one-patient-out Drug Mean baseline, PerturbRx improved directional agreement in 25/27 cohorts (92.6%) and 207/278 patients (74.5%), with a mean paired cosine gain of +0.0222 (95% CI: [0.0173, 0.0272]).

Improvements were observed in both bulk RNA-seq (19/20 cohorts) and scRNA-seq (6/7 cohorts). These findings provide empirical evidence of modest but broadly positive treatment-associated directional transfer, although absolute agreement remained limited and heterogeneous. Accurate patient-specific post-treatment reconstruction was not established.

We appreciate the reviewer's suggestions, which helped us identify downstream initialization sensitivity, strengthen the evaluation of learned versus untrained transitions, and clarify the biological interpretation of transferred representations (Sections 3.4, 4.3–4.5, and Appendix C.4).

Reviewer_jyzjOpenReview ↗

Response

Thank you for the additional experiments and clarifications.

The random-seed analysis changes my assessment of the original results. PerturbRx outperformed Patient+Drug in only 35/100 TCGA runs and 38/100 PDX runs, with slightly lower average AUROC on both datasets. The original gains therefore do not appear robust.

The revised ensemble improves some comparisons, but its advantage over the matched frozen-random transition network remains nonsignificant after multiple-testing correction. It is still unclear how much the perturbation pretraining contributes beyond adding a nonlinear feature transformation. The unseen-drug results also show mixed gains over simpler models.

I appreciate the new patient-level validation and the authors’ clear discussion of these limitations. However, the additional results have not convinced me that the main contribution provides a sufficiently reliable improvement in patient drug-response prediction.

AuthorsOpenReview ↗

Follow-up Response to Reviewer jyzj

We thank the reviewer for the follow-up. We acknowledge that the original single-head results were sensitive to initialization and that the specific predictive contribution of perturbation pretraining remains uncertain across benchmarks and metrics.

However, PerturbRx's broader contribution is an inductive framework that transfers treatment-conditioned representations from single-cell perturbations to unseen patient–drug pairs. Under our revised ensemble evaluation, PerturbRx significantly outperformed all four evaluated external methods (CODE-AE, DeepSADR, TransDRP, and WISER) in AUROC on both TCGA and PDX, and in AUPRC on PDX.

The revised ensemble also showed significant improvements over Patient+Drug on TCGA AUROC (Δ = +0.0215, p = 0.0031) and PDX+Drug on PDX AUPRC (Δ = +0.0303, p = 0.0027).

Beyond downstream prediction, we independently evaluated the frozen transition operator against observed treatment-associated molecular changes. We demonstrated directional transfer on sci-Plex 3 and improved agreement with observed changes in 25 of 27 independent patient cohorts.

We therefore emphasize the transferability of treatment-conditioned representations and their independent empirical validation, while acknowledging that the incremental predictive contribution of perturbation pretraining is not uniformly established.

AuthorsOpenReview ↗

Additional Analysis: Robustness to Frozen-Random Operator Initialization

We thank the reviewer for the follow-up. We conducted an additional analysis to assess whether the frozen-random comparison depended on a particular random initialization.

Additional analysis: 20 frozen-random operator initializations

Our original frozen-random control used a single operator initialization per benchmark. We therefore evaluated 20 independently initialized frozen-random transition operators on each benchmark under the same downstream ensemble protocol.

Performance across operator initializations

BenchmarkMetricPerturbRxFrozen Random (20 runs, mean ± SD)PerturbRx wins
TCGA-485AUROC0.72560.7121 ± 0.002120/20
TCGA-485AUPRC0.80400.7835 ± 0.002420/20
PDX-791AUROC0.6960.6899 ± 0.001820/20
PDX-791AUPRC0.6200.5992 ± 0.003120/20

Paired cluster-bootstrap comparisons (PerturbRx minus mean Frozen Random)

BenchmarkMetricΔ95% CIp-value
TCGA-485AUROC+0.0135[-0.0069, 0.0339]0.1933
TCGA-485AUPRC+0.0206[-0.0059, 0.0469]0.1294
PDX-791AUROC+0.0059[-0.0059, 0.0178]0.3318
PDX-791AUPRC+0.0204[0.0020, 0.0387]0.0302

Paired cluster bootstrapping was performed at the patient level for TCGA-485 and at the PDX-model level for PDX-791, accounting for repeated drug-response episodes within clusters.

PerturbRx outperformed all 20 frozen-random initializations on both metrics and benchmarks, indicating that its numerical advantage was not driven by an unusually weak random initialization. The paired bootstrap comparisons showed a statistically significant improvement for PDX AUPRC (p = 0.0302), but not for the other three benchmark-metric combinations.

These findings strengthen the evidence for initialization-robust numerical improvements and suggest an incremental contribution from perturbation pretraining on PDX AUPRC, while leaving its consistent statistical advantage across benchmarks and metrics unresolved. We have updated the manuscript accordingly.

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

PerturbRx learns treatment-conditioned latent changes from Tahoe-100M and transfers the resulting representation to patient and PDX drug-response prediction. It obtains the highest mean results in the reported target-domain evaluations, although the gains over simpler patient-and-drug and shuffled-transition controls are small.

Critical strengths

The paper tackles an important source-to-patient transfer problem and uses careful split boundaries in both the source and target domains. The Patient+Drug and shuffled-transition controls directly test whether the learned feature adds information beyond treatment identity and static tumor state. The paper also reports weak and non-significant results transparently.

Critical weaknesses

The incremental benefit over simpler controls is small. PerturbRx has the best mean results, but the reported intervals do not establish a clear advantage over Patient+Drug or shuffled-transition controls.

The learned feature is not directly validated as a patient-specific treatment transition. It may still be a useful nonlinear representation, but the mechanistic interpretation should remain cautious.

Questions for authors
  1. Which comparison best demonstrates that the learned transition adds information beyond Patient+Drug and shuffled-transition controls?

  2. Why does the globally shuffled transition remain close to PerturbRx on PDX, and what does this imply about the transition interpretation?

  3. If suitable data are available, would the learned transition agree with observed pre- and post-treatment profiles? Even a limited analysis would help clarify the mechanistic interpretation.

2 comments
AuthorsOpenReview ↗

Response to Reviewer yybn

We thank the reviewer for the constructive questions. We address them using updated downstream evaluations, corrected shuffling controls, and validation against observed patient treatment-associated changes.

1. Incremental information beyond static and shuffled controls

We updated our downstream evaluation using a multi-architecture, multi-seed response-head ensemble and performed paired cluster-bootstrap tests.

On TCGA-485, PerturbRx achieved pooled OOF AUROC/AUPRC of 0.726/0.804, compared with 0.704/0.786 for Patient+Drug. The AUROC improvement (+0.0215) was significant (p = 0.0031). PerturbRx also significantly outperformed globally shuffled transitions on AUROC (0.726 vs. 0.689, Δ = +0.0367, p = 0.0012). The AUROC improvement over within-drug shuffling (0.711 AUROC) was also significant (Δ = +0.0143, p = 0.0495), whereas the AUPRC difference was not (p = 0.4302).

On PDX, PerturbRx achieved 0.696/0.620, compared with 0.692/0.589 for PDX+Drug. The AUPRC improvement (+0.0303, 95% CI: [0.0099, 0.0490]) was significant (p = 0.0027). PerturbRx also outperformed the global-shuffle control (0.686/0.584) on AUPRC (p = 0.0107), but not significantly on AUROC (p = 0.2726).

These comparisons support incremental predictive value beyond static patient–drug features and shuffled transitions in specific settings. However, differences vary across metrics and controls, limiting attribution of downstream gains exclusively to patient-matched transitions (Sections 4.3–4.4; Tables 2–3).

2. Why does global shuffling remain competitive on PDX?

We identified and corrected an implementation issue in the original PDX global-shuffle control: some donor transitions shared the target drug identity. The corrected procedure excludes donor transitions associated with the target drug.

Under the updated ensemble evaluation, the corrected global-shuffle control achieved 0.686 ± 0.052 AUROC and 0.584 ± 0.064 AUPRC, compared with 0.696 ± 0.027 and 0.620 ± 0.028 for PerturbRx. The AUPRC improvement was significant (p = 0.0107), whereas the AUROC difference was not (p = 0.2726).

The competitive AUROC performance of global shuffling suggests that some predictive benefit may arise from additional nonlinear representations or shared transition-feature structure, rather than exclusively from correctly matched patient–drug transitions. Accordingly, we do not interpret downstream predictive gains alone as evidence that the operator recovers patient-specific biological treatment effects. We have clarified this limitation in Sections 4.4 and 5.

3. Direct validation against observed pre/post-treatment profiles

To address this question directly, we added an independent evaluation using paired pre- and post-treatment transcriptomes from 278 patients across 27 independent cohorts, including bulk RNA-seq and scRNA-seq. We compared predicted transition directions with observed treatment-associated changes in scFoundation latent space, without retraining the source-stage operator.

Relative to a leave-one-patient-out Drug Mean baseline, PerturbRx improved directional agreement in 25 of 27 cohorts (92.6%) and 207 of 278 patients (74.5%), with a mean paired cosine gain of +0.0222 (95% CI: [0.0173, 0.0272]). We observed improvements in both bulk RNA-seq (19/20 cohorts) and scRNA-seq (6/7 cohorts).

These findings provide empirical evidence of modest but broadly positive directional transfer beyond a cohort-specific average transition. However, absolute patient-level directional agreement remained limited and heterogeneous, and we did not establish accurate post-treatment reconstruction. We therefore distinguish observed directional alignment from claims of identifiable individual-cell trajectories or accurate patient-specific state reconstruction (Sections 4.5 and 5).

AuthorsOpenReview ↗

Follow-up on Rebuttal and Revised Manuscript

We would like to briefly follow up on our November 14 response and the revised manuscript. We have clarified the incremental predictive value over static and shuffled controls, corrected the PDX global-shuffle implementation, and added validation using paired pre/post-treatment profiles from 278 patients.

As the discussion period approaches its end, we would appreciate any further feedback on whether these additional analyses address your main concerns. Thank you for your consideration.

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

This paper introduces PerturbRx, which learns drug-conditional latent state transitions from single-cell perturbation data and transfers pre-trained transition operators to the task of predicting patient drug responses. The method combines scFoundation and ChemBERTa representations to learn treatment-induced latent changes through distribution matching and state transition constraints, and is validated on the TCGA and PDX datasets.

Critical strengths
  1. This paper attempts to improve patient-level drug response prediction using single-cell perturbation data; the research question has practical significance.

  2. The authors present source-domain prediction, cross-domain transfer, and ablation experiments on multiple architectures, and employ patient-level independent partitioning; the experimental design demonstrates a certain degree of rigor.

Critical weaknesses

1. The core method makes a relatively weak practical contribution and lacks statistically significant performance improvements.

The main contribution of this paper lies in providing additional information for predicting patient drug responses by learning treatment-induced latent state transitions. However, Table 2 shows that on the TCGA-485 dataset, the simple “Patient+Drug” baseline already achieved an AUROC of 0.705, while the full PerturbRx model reached only 0.709—a difference that is not statistically significant. A similar issue exists on the PDX dataset, where the AUROC improved only from 0.673 to 0.685. Therefore, the current experimental results do not sufficiently demonstrate that introducing additional single-cell perturbation pretraining and latent transition modules can yield substantial predictive benefits.

2. The ability to predict state transitions for unseen drugs remains limited.

The paper emphasizes that the model can learn drug-induced state changes with generalization capabilities. However, Figure 2D shows that, across 42 unseen drugs, PerturbRx’s average Transition Cosine is only 0.229—a limited improvement over cAE’s 0.198—and it achieved better results for only 23 out of 42 drugs. The 95% confidence interval for the performance difference between the two is [−0.020, 0.086], which includes zero. This leaves it unclear whether the model has truly learned therapeutic state transition patterns that can be reliably generalized to new drugs.

3. The learned latent state transitions lack direct biological validation.

This paper utilizes unpaired single-cell populations to learn latent state transitions through MMD distribution matching and constraints on the direction and magnitude of transitions. However, population distribution matching cannot uniquely determine the true state transitions of individual cells. Furthermore, the model directly applies this operator to patients’ bulk RNA-seq representations but lacks paired pre- and post-treatment expression data for validation. Therefore, it is currently impossible to distinguish whether the model has learned biologically meaningful treatment-induced changes or an additional nonlinear patient–drug feature transformation. This also weakens the paper’s core interpretation regarding the modeling of treatment states.

Questions for authors

See in Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer XUKL

We thank the reviewer for the constructive feedback. We address the three concerns with updated downstream evaluations, external perturbation validation, and direct comparison with observed patient treatment-associated changes.

1. Incremental predictive value

To assess the robustness of downstream performance, we updated our evaluation using a multi-architecture, multi-seed response-head ensemble and performed paired cluster-bootstrap tests.

On TCGA-485, PerturbRx achieved pooled out-of-fold (OOF) AUROC/AUPRC of 0.726/0.804, compared with 0.704/0.786 for Patient+Drug. The AUROC improvement (+0.0215) was statistically significant (p = 0.0031).

On PDX, PerturbRx achieved 0.696/0.620, compared with 0.692/0.589 for PDX+Drug. The AUPRC improvement (+0.0303, 95% CI: [0.0099, 0.0490]) was also statistically significant (p = 0.0027), whereas the AUROC difference was not.

PerturbRx achieved the highest pooled OOF AUROC and AUPRC among all evaluated external baselines and representation controls on both benchmarks. It significantly outperformed all four external baselines on AUROC in both datasets and on AUPRC in PDX. Significant improvements were also observed over Gaussian random transitions on TCGA AUROC (p = 0.0010) and PDX AUPRC (p = 0.0013), as well as globally shuffled transitions on TCGA AUROC (p = 0.0012).

We additionally evaluated 20 independently initialized Frozen Random operators on each benchmark. On TCGA, PerturbRx achieved higher AUROC and AUPRC than all 20 operators, with mean improvements of +0.0135 and +0.0206, respectively (p = 0.1933 and p = 0.1294). On PDX, PerturbRx also outperformed all 20 operators on both metrics, with mean improvements of +0.0059 AUROC (p = 0.3318) and +0.0204 AUPRC (p = 0.0302).

These findings support incremental predictive value in specific settings and consistent numerical improvements across random operator initializations, without demonstrating uniform superiority (Sections 4.3–4.4 and Tables 2–3).

2. Generalization to unseen drugs

The Tahoe drug-held-out evaluation showed that PerturbRx achieved the highest mean transition cosine among the evaluated same-objective architectural controls (0.229, compared with 0.198 for cAE, 0.167 for MLP, and 0.137 for Linear). Although PerturbRx outperformed cAE on 23 of 42 held-out drugs, the paired difference was not statistically significant (95% CI: [−0.020, 0.086]), indicating heterogeneous generalization.

We further assessed external transfer using the A549 subset of sci-Plex 3, without retraining or adaptation. Across 188 drugs and four doses, PerturbRx achieved a mean transition cosine of 0.1318, compared with 0.1103 for the predicted Drug Mean baseline (paired gain: +0.0215, 95% CI: [0.0010, 0.0421]).

An exploratory mechanism-of-action analysis motivated a focused evaluation of 15 source-unseen Aurora kinase-related drugs. PerturbRx achieved a mean transition cosine of 0.306, outperforming cAE (0.239) and Linear (0.248), but not MLP (0.368). Relative to their respective Drug Mean baselines, PerturbRx and MLP achieved nearly identical gains (+0.208 and +0.209, respectively).

These results provide evidence of directional transfer to an independent perturbation dataset, while showing that generalization varies across drugs and is not exclusive to the PerturbRx architecture (Section 4.2 and Appendix C.2).

3. Biological validation of learned transitions

To directly evaluate whether the learned transitions capture treatment-associated molecular changes in patients, we added an independent validation using paired pre- and post-treatment transcriptomes from 278 patients across 27 independent cohorts, including both bulk RNA-seq and scRNA-seq. We compared predicted transition directions with observed treatment-associated changes in scFoundation latent space, without retraining the source-stage operator.

Relative to a leave-one-patient-out Drug Mean baseline, PerturbRx achieved higher directional agreement in 25 of 27 cohorts (92.6%) and 207 of 278 patients (74.5%), with a mean paired cosine gain of +0.0222 (95% CI: [0.0173, 0.0272]). Improvements were observed in both bulk RNA-seq (19/20 cohorts) and scRNA-seq (6/7 cohorts). These findings provide evidence of modest but broadly positive directional transfer beyond a cohort-specific average transition.

We emphasize that unpaired population supervision cannot uniquely identify individual-cell trajectories, and absolute patient-level directional agreement remained modest and heterogeneous. Accurate post-treatment reconstruction was not established. Our revised manuscript therefore distinguishes empirically supported directional transfer from claims of identifiable biological trajectories or accurate patient-specific state reconstruction (Sections 4.5 and 5).

AuthorsOpenReview ↗

Follow-up on Rebuttal and Revised Manuscript

We would like to briefly follow up on our response posted on November 14 and the revised manuscript. We have added paired patient pre/post-treatment validation, external perturbation validation, and updated downstream statistical comparisons.

As the discussion period approaches its end, we would appreciate any further feedback on whether these additional results address your main concerns. Thank you for your consideration.

Thank you for your detailed response and the experiments you conducted. Your response addressed most of my concerns, but I still feel that the methodology in this paper lacks sufficient innovation; it seems to be a combination of common methods without much algorithmic innovation. Therefore, I have only raised my score from a 1 to a 2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.