acceptodds
Under review as a conference paper at ICLR 2027

One Drug, MoRe Views: Response-Directed Multi-Modal Conditioning for Unseen-Drug Perturbation

Abstract

Predicting transcriptional responses to unseen drugs is central to prioritising compounds for experimental profiling, yet existing predictors generalise poorly to held-out drugs. We attribute this to two coupled limitations: drugs are represented by chemical structure alone, which only partly determines their mechanism, and models are trained to reconstruct whole expression profiles, an objective dominated by background variation rather than the perturbation effect. We introduce MoRe, a model-agnostic plug-in that addresses both. A lightweight module fuses chemical and target-informed representations into a single drug latent, and an auxiliary objective supervises the predicted mean response of each condition. Both components are trained jointly with the backbone. On an autoencoder (ChemCPA) and a diffusion model (PerturbDiff), evaluated on held-out drugs in sci-Plex3 and Tahoe-100M, MoRe improves the correlation between predicted and observed responses on the top 50 differentially expressed genes by up to 113%. Analysis shows that the drug views are complementary: they lead on different gene programmes, and fusion places each drug near training drugs with similar responses, whereas chemistry alone does not. By describing drugs by what they do as well as what they are, MoRe points to a promising direction for virtual-cell models of in silico drug perturbation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single cell biology, perturbation prediction, chemical perturbation, scRNA sequencing, representation learning, multimodal learning

Reviews

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

This paper proposes MoRe, a framework for predicting transcriptional responses to unseen drugs. It combines three drug representations (chemical descriptors, target identities, and UniPert embeddings) with an auxiliary condition-mean response supervision loss. Experiments on ChemCPA and PerturbDiff across sci-Plex3 and Tahoe-100M demonstrate improvements in perturbation-response prediction. Additional ablations and latent-space analyses investigate the contributions of multimodal fusion and response supervision.

Critical strengths

The paper addresses an important challenge in perturbation modeling: generalization to unseen drugs. Integrating complementary chemical and biological representations is well motivated, and explicitly supervising perturbation-induced responses is a reasonable approach. Evaluation across two model architectures and datasets, together with extensive ablation studies, provides a useful empirical foundation.

Critical weaknesses

Weakness 1: The empirical evidence does not convincingly establish the necessity or effectiveness of the proposed framework.
a. Marginal improvements over stronger baselines. The headline 113% improvement is measured against a weak ChemCPA baseline. More meaningful comparisons reveal substantially smaller gains: on sci-Plex3, ChemCPA with UniPert and response supervision achieves 0.247 versus MoRe's 0.249, while PerturbDiff improves only from 0.280 with fusion alone to 0.286 with the complete framework. These differences are not convincingly established as statistically reliable.
b. Insufficient justification of the proposed mechanism. The theoretical analysis relies on a simplified squared-error objective that does not directly characterize the actual backbone losses. Moreover, the large observed within-condition variance does not necessarily imply gradient-level signal dilution.
The paper does not convincingly demonstrate that its two components jointly resolve the claimed fundamental limitations, substantially weakening its methodological contribution.
Weakness 2: The evaluation does not convincingly demonstrate genuine out-of-distribution generalization or preserved biological fidelity.
a. Questionable generalization setting. Held-out drugs are represented using external target annotations and pretrained embeddings, while evaluation relies on fixed random drug splits. Without scaffold-, target-, or mechanism-disjoint evaluations, improvements may primarily reflect interpolation through known biological relationships rather than genuine extrapolation.
b. Potential degradation of cellular distribution modeling. Despite improvements in DEG-based metrics, MoRe worsens PerturbDiff's all-gene energy distance from 0.537 to 0.566. This challenges the claim that response supervision preserves distributional fidelity. These limitations directly undermine the paper's central claim of improving practically meaningful unseen-drug prediction.

Questions for authors

Q1. Given the marginal improvements over strong single-view and fusion-only baselines, can the authors demonstrate statistically significant and consistent benefits of the complete framework across multiple independent drug splits?
Q2. How much of MoRe's improvement depends on external target annotations or pretrained representations? Can the authors evaluate scaffold-disjoint and target-disjoint generalization and clarify potential overlap between representation pretraining and test compounds?
Q3. Can the authors provide direct evidence for the proposed gradient-level signal-dilution mechanism and explain why improved condition-mean prediction is accompanied by degraded all-gene distribution metrics?

6 comments
AuthorsOpenReview ↗

Response to Reviewer PVVF (1/3)

We thank the reviewer for the precise and constructive critique, which led to a stronger test of each claim. We summarise the main changes first and respond to each question in detail below:

  • W1a / Q1 (margins; significance across splits). Over five random-split folds + MoRe consistently improves both backbones, with statistical significance in 8 of the 10 settings, and it matches or exceeds the two baselines you mentioned (Tables P1, P2).
  • W1b / Q3 (gradient dilution). Measured on trained models: response supervision gives the more coherent gradient (Table P3; App. E).
  • W1c ("jointly resolve"). The two components address two different limitations, each shown directly, and the ablation shows both are needed; which one matters more depends on the backbone (Tables P2, P3).
  • W2a / Q2 (interpolation; external annotations). Three disjoint splits on both datasets: the gain is as large as on random folds when shared scaffolds or shared targets are removed (Table R1; App. M).
  • W2b / Q3 (all-gene energy distance). + MoRe matches or improves the single-cell distribution fit in 19 of the 20 comparisons; the one exception arises at low doses (Table P4).

W1a / Q1. Margins over strong baselines; significance across splits

Thank you for asking this.

1. "Statistically significant and consistent benefits across multiple independent drug splits." (Q1) Yes. We repeated the experiments on five folds of four types of drug split: random, scaffold-disjoint, target-disjoint, and both. + MoRe improves the original backbone in every setting, with statistical significance in 8 of the 10 settings (Table R1 in the general response; Table P1).

Table P1. + MoRe minus the original backbone on five random-split folds: mean ± standard deviation of the paired difference over folds. ∗: paired t-test over folds, p < 0.05.

DEG50 PearsonDEG50 directionDEG20 PearsonDEG20 direction
sci-Plex3, ChemCPA+0.139 ± 0.036∗+0.047 ± 0.045+0.135 ± 0.040∗+0.066 ± 0.048
sci-Plex3, PerturbDiff+0.030 ± 0.022+0.011 ± 0.008∗+0.044 ± 0.027∗+0.023 ± 0.008∗
Tahoe-100M, ChemCPA+0.051 ± 0.036∗+0.006 ± 0.015+0.056 ± 0.033∗+0.014 ± 0.022
Tahoe-100M, PerturbDiff+0.069 ± 0.006∗+0.030 ± 0.009∗+0.075 ± 0.006∗+0.041 ± 0.005∗

2. "Marginal improvements over stronger baselines." The two baselines you mentioned are each one component of MoRe. We retrained them on five random-split folds of sci-Plex3 (Table P2).

  • Over all doses, + MoRe matches the stronger of its two components. For ChemCPA, response supervision alone reaches the level of + MoRe and fusion alone is significantly lower; for PerturbDiff, fusion alone reaches the level of + MoRe. The two parts thus address different limitations: a backbone can lack the training signal (ChemCPA), the drug information (PerturbDiff), or both (see W1c).
  • The more reliable the signal, the better + MoRe performs. At the highest dose, the setting prior work reports (Hetzel et al., 2022; Piran et al., 2024), the signal-to-noise ratio of the observed response is 11.6 against 3.1 over all doses, and + MoRe is the best on both backbones. The ratio compares, for each held-out condition, the size of its response on the DEG50 with the sampling noise of that response, estimated from two random halves of its cells (median over conditions; App. H).
  • + MoRe is significantly better than fusion alone (p < 0.001) and than response supervision alone (p = 0.024) (Wilcoxon signed-rank test over the five folds of both datasets and both backbones; Table D4 in our reply to Reviewer dRXi).

Table P2. + MoRe against each of its two components used alone, on sci-Plex3 (DEG50 Pearson; mean over random-split folds). + fusion: the three drug views, without response supervision. + supervision: response supervision with a single drug view, chemistry (C, the original view) or UniPert (U, the best single view). SNR: signal-to-noise ratio of the observed response on the DEG50 (median over held-out conditions). Bold: best of the row. ∗: significantly different from + MoRe (paired t-test over folds, p < 0.05).

DosesSNROriginal+ fusion+ supervision (C)Best single view + supervision (U)+ MoRe
PerturbDiffAll3.10.2230.2630.2420.233∗0.253
Highest11.60.213∗0.3320.306∗0.313∗0.361
ChemCPAAll3.10.117∗0.130∗0.2500.2530.256
Highest11.60.066∗0.1000.3130.3180.324

Full results. Table 2 in the main text; App. M: all five metrics. App. H: every arm under both objectives.

References

  • Hetzel et al. Predicting cellular responses to novel drug perturbations at a single-cell resolution. NeurIPS, 2022.
  • Piran et al. Disentanglement of single-cell data with Biolord. Nature Biotechnology, 2024.
AuthorsOpenReview ↗

Response to Reviewer PVVF (2/3)

Continued: W1b / Q3 (first part) and W1c.

W1b / Q3 (first part). The scope of the decomposition, and direct evidence for gradient dilution

Thank you for both parts of this point.

1. Measured directly. Following your suggestion, we measured the gradient that the drug latent receives, on trained models and with each backbone's own loss, rather than inferring dilution from the within-condition variance. For a randomly sampled training condition, we drew 16 resamples of 64 cells and took the gradient of each loss with respect to the drug latent. The ratio ‖mean gradient‖ / mean ‖gradient‖ is 1 when all resamples agree in direction and falls as sampling noise dominates. Under each backbone's native loss a substantial part of that gradient is noise; under response supervision it is largely coherent (Table P3).

Table P3. Gradient signal-to-sampling-noise ratio with respect to the drug latent (sci-Plex3; three seeds × three checkpoints; mean ± sd over the nine runs). Both losses are evaluated on the same checkpoints.

ModelNative lossResponse supervision (ours)
ChemCPA0.49 ± 0.180.91 ± 0.10
PerturbDiff0.71 ± 0.130.87 ± 0.15

The native gradient is noisier for ChemCPA (0.49) than for PerturbDiff (0.71), so response supervision changes more for ChemCPA. This is why it carries most of the gain for ChemCPA, while fusion does for PerturbDiff (W1c).

2. The decomposition. It is stated for squared error, and its role is to motivate the form of the auxiliary term, for which it is exact (Eqs. 6 and 7). The evidence that the native objectives leave the drug latent with a noisy gradient is Table P3 (Sec. 3.3; App. E).

Full results. App. E: the decomposition with its assumptions, and the gradient ratio per checkpoint.

W1c. "Jointly resolve the claimed fundamental limitations"

Thank you for raising this point. The revision now shows directly that each component addresses its limitation, and that the two are needed together:

  • The training signal. The native objective gives the drug latent a noisy gradient, and response supervision a coherent one (Table P3).
  • The drug information. With a view noisy or missing, + MoRe stays above the model that relies on that view alone (Table K5 in our reply to Reviewer k4EA).
  • The two together. In the ablation, each component alone improves the backbone, and + MoRe is significantly better than fusion alone (p < 0.001) and than response supervision alone (p = 0.024) (Table P2 in our first comment; Table D4 in our reply to Reviewer dRXi).

Which limitation matters more depends on the backbone. ChemCPA lacks the training signal: its reconstruction objective captures little of the differential response, so response supervision carries the gain. PerturbDiff on sci-Plex3 lacks the drug information: it is expressive enough to exploit richer drug views, so fusion carries the gain. Addressing both is what lets one plug-in work on both backbones. The Introduction now makes this dependence explicit.

AuthorsOpenReview ↗

Response to Reviewer PVVF (3/3)

Continued: W2a / Q2 and W2b / Q3 (second part).

W2a / Q2. Interpolation or extrapolation; disjoint splits; dependence on external annotations

Following your suggestion, we evaluated scaffold-disjoint, target-disjoint, and scaffold- and target-disjoint splits on both datasets, five folds each (Table R1 in the general response). + MoRe improves over the original backbone on every split of both datasets.

  • The gain does not come from interpolation. When the held-out drugs share no scaffold or no direct target with the training drugs, + MoRe still improves PerturbDiff by 10–21% in DEG50 Pearson, as much as on random folds, with statistical significance in five of the six disjoint settings.
  • The gain does not depend on any single external source. With the target view or the UniPert view replaced by noise for all held-out drugs, + MoRe stays above the original backbone (Table K5 in our reply to Reviewer k4EA). With chemistry alone and response supervision, without any annotation or pretrained embedding, it still improves both backbones on both datasets (Table D4 in our reply to Reviewer dRXi; Tables 11–14).
  • Overlap with UniPert's pretraining set does not explain the gain. UniPert is trained on compound–target pairs without expression data. The gain of + MoRe is at least as large for held-out drugs outside its pretraining set as for those inside it (Table D5 in our reply to Reviewer dRXi).

Full results. App. M: every split, all metrics. App. N: the overlap with the training drugs, by split.

W2b / Q3 (second part). All-gene energy distance rises from 0.537 to 0.566

Thank you for pointing this out. We repeated the single-cell evaluation on sci-Plex3 over five random-split folds, for both backbones and by dose (Table P4).

  • + MoRe matches or improves the distribution fit in 19 of the 20 comparisons: significantly better in 7, not significantly different in 12.
  • The one exception arises at low doses. The all-gene energy distance of PerturbDiff rises over all doses (+7%, p = 0.03), driven by the lower doses, where drugs have almost no effect (Hetzel et al., 2022). At the highest dose it does not rise, and four of the five metrics improve significantly.
  • Response supervision constrains the condition mean, not the cell-to-cell spread. A strong effect shifts the mean, which + MoRe predicts better; under a weak effect the distance reflects the spread, which remains with the backbone's native objective. Pseudobulk-level prediction is also often more important for clinical discovery, since no cell is observed both before and after treatment.

The revised Sec. 5.1 states how response supervision acts on the mean and on the spread.

Table P4. Single-cell distribution metrics on sci-Plex3 over all doses and at the highest dose (mean over five random-split folds; 1,000 cells per condition). E-distance: energy distance; W1: mean gene-wise Wasserstein-1 distance; DE recall: share of the observed differentially expressed genes of a condition that are recovered from the predicted cells. Bold: better with MoRe. ∗: differs from the original backbone (paired t-test over folds, p < 0.05).

DosesModelE-distance, all genes ↓W1, all genes ↓E-distance, DEG50 ↓W1, DEG50 ↓DE recall ↑
PerturbDiffAllOriginal0.4710.02660.3710.09730.124
+ MoRe0.504∗0.02670.3680.0903∗0.152∗
HighestOriginal0.6280.03200.5890.15570.192
+ MoRe0.6160.0307∗0.530∗0.1319∗0.246∗
ChemCPAAllOriginal1.9710.08201.1800.26970.007
+ MoRe1.9700.08291.1770.27080.008
HighestOriginal2.0390.08281.3600.32200.012
+ MoRe2.0380.08361.345∗0.32080.013

Full results. App. H: Tables 16 and 17.

References

  • Hetzel et al. Predicting cellular responses to novel drug perturbations at a single-cell resolution. NeurIPS, 2022.

We greatly appreciate the time and effort you’ve put into reviewing our submission. We have carefully responded to all your concerns. We would be grateful if you could consider raising your score in light of the new experiments, explanations and revision. Thank you again for your time and thoughtful suggestions!

Best regards,

The Authors

Reviewer_dRXiOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces MoRe, a model-agnostic plug-in for predicting single-cell transcriptomic responses to unseen drugs. The authors identify two limitations of existing approaches: chemical structure alone may not adequately capture drug targets or mechanisms of action, and cell-level reconstruction objectives may dilute drug-specific signals concentrated in a small number of differentially expressed genes (DEGs).

MoRe addresses these limitations through multimodal drug conditioning and condition-mean response supervision. It combines 178-dimensional RDKit chemical descriptors, explicit Target ID representations, and 256-dimensional target-informed UniPert embeddings into a shared drug latent, primarily through concatenation. It also adds an auxiliary loss that aligns predicted and observed mean control-relative responses for each drug–dose–cell-line condition, balancing errors over the full gene panel and condition-specific DEGs.

The method is evaluated with the autoencoder-based ChemCPA and diffusion-based PerturbDiff backbones on sci-Plex3 and Tahoe-100M, using held-out drug splits containing 360 and 683 test conditions, respectively. MoRe improves residual Pearson correlation, direction agreement, and retrieval rank across both backbones and datasets, with a maximum reported relative gain of 113% in DEG50 Pearson correlation. Additional analyses suggest that the chemical and target-informed views capture complementary biological signals and that their fusion organizes drugs according to transcriptional response similarity.

Critical strengths
  1. Model-agnostic design. MoRe separates multimodal drug conditioning and condition-mean response supervision from the internal architecture of the base predictor, allowing both components to be added with minimal modification. Its successful integration with two distinct model families—autoencoder-based ChemCPA and diffusion-based PerturbDiff—demonstrates its practical applicability across different perturbation prediction frameworks.
  2. Perturbation-specific evaluation. The paper evaluates control-relative transcriptional changes rather than relying primarily on raw expression-space metrics, which can remain high even for predictors that largely reproduce the baseline cell state. The use of residual Pearson correlation, direction agreement, and retrieval rank provides a more informative assessment of whether the models capture drug-specific perturbation effects.
Critical weaknesses

1. Insufficient Conceptual and Methodological Justification and Limited Novelty

  • Problem Formulation and Characterization of Prior Work. The paper’s characterization of prior methods as predominantly chemical-only and single-view overlooks recent approaches using targets, pathways, knowledge graphs, and pretrained embeddings. Consequently, MoRe’s distinction from existing multimodal and target-aware methods is insufficiently established.
  • Limited Information in the Target Representation. Target ID does not capture binding affinity, potency, direction of action, off-target effects, or polypharmacology. Although UniPert provides additional information, it remains unclear whether it distinguishes drugs sharing targets but exhibiting different pharmacological effects.
  • Lack of Downstream Effect Propagation. MoRe does not model how target effects propagate through PPI, gene-regulatory, or signaling networks. Because the transcriptional consequences of target engagement are shaped by gene–gene interactions and downstream pathway activity, target information alone may be insufficient to represent the resulting perturbation response.
  • Limited Novelty of the Fusion Mechanism. The fusion module primarily projects and concatenates existing RDKit, Target ID, and UniPert representations. While practical, this offers limited methodological novelty beyond adding external target information to existing backbones.
  • Conceptual Limitations of DEG-Weighted Supervision. DEGs are downstream outcomes rather than mechanisms generating perturbation responses. DEG weighting may emphasize responsive genes, but it remains unclear whether the resulting gains reflect mechanistic generalization or alignment with DEG-based evaluation metrics.
  • Unclear Independent Contributions of the Two Components. Response supervision provides only a limited additional gain after multimodal fusion is applied, suggesting that the two components may

2. Insufficient Empirical Evidence for Unseen-Drug Generalization

  • Limited Robustness of the Drug Split. The evaluation uses only one fixed drug split per dataset, making the results potentially split-dependent. Multiple independent drug-level splits are needed because different training seeds do not capture uncertainty arising from data partitioning.
  • Unclear Overlap with UniPert Pretraining Data. It is unclear whether held-out drugs were included as compound–target pairs during UniPert pretraining. The authors should disclose this overlap and report results excluding overlapping compounds to separate MoRe’s contribution from drug-specific prior knowledge.
  • Limited Baseline Comparisons. Comparisons with existing multimodal and target-aware methods are limited. It is therefore unclear whether the gains arise from MoRe’s design or simply from providing additional external target information.
  • Limited Scope of the Analyses. The principal biological and latent-space analyses focus on PerturbDiff and sci-Plex3, while comparable evidence is limited for ChemCPA and Tahoe-100M. The proposed explanation may therefore not generalize across backbones and datasets.
  • Limited Statistical Evidence. Results are based mainly on three random seeds, with overlapping error ranges across several methods. Confidence intervals and significance tests are needed to establish whether fusion reliably outperforms the individual representations.
  • Imbalance Between the Scope of the Claims and the Supporting Evidence. The association between latent-space proximity and response similarity is correlational and does not establish that the learned geometry causes predictive improvements. The limited backbones, datasets, and fixed splits do not fully support broad claims about unseen-drug generalization.
Questions for authors

Clarification and Additional Validation Requests

  • Were any held-out drugs included in the UniPert pretraining data? If so, what information about these drugs was used during pretraining, and how does performance change when overlapping drugs are excluded?
  • Are the reported improvements consistent across multiple independent drug-level splits, given that the current results are based on a single fixed split?
  • What are the results when the all-gene loss and DEG-weighted loss are applied separately within condition-mean supervision? This would help distinguish whether the improvement arises from condition-level supervision itself or from aligning the training objective with DEG-based evaluation metrics.
  • What is MoRe’s substantive distinction from existing multimodal, target-aware, or network-aware drug representations? In particular, is there evidence that simple target fusion is more suitable for unseen-drug generalization than PPI- or gene-network-based representations?

Suggestion for Future Work

  • Future work could explicitly model drug–target–cell interactions and downstream effects propagated through PPI networks, gene-regulatory networks, or biological pathways.
5 comments
AuthorsOpenReview ↗

Response to Reviewer dRXi (1/5)

We thank the reviewer for the detailed and constructive review, which led to a sharper positioning and a more rigorous evaluation. We summarise the main changes first and respond to each question in detail below:

  • W1.1 / W1.4 / Q4 (prior work; fusion operator). Prior methods align the modalities into one drug vector before the predictor; MoRe keeps the views distinct inside it (Table D1). The simple fusion operator is a deliberate choice for a low-data regime.
  • W1.2 / W1.3 (target view). The target view is predictive of the response on its own, and + MoRe improves on the target-only model by 16% for held-out drugs that share a target with a training drug but differ in structure.
  • W1.5 / Q3 (DEG weighting). Without the weighting, response supervision still improves the backbone on Tahoe-100M and at the highest dose on sci-Plex3 (Tables D2, D3).
  • W1.6 (two components). Each component alone improves the backbone; + MoRe is significantly better than either alone and the best at the highest dose (Table D4).
  • W2.1 / Q2 (one split). Five random-split folds and three types of disjoint splits: + MoRe consistently improves the original backbone (Table R1 in the general response).
  • W2.2 / Q1 (UniPert pretraining). The gain of + MoRe is at least as large for held-out drugs outside UniPert's pretraining set as for those inside it (Table D5).
  • W2.3 (target-aware methods). The external target information alone does not account for the gain: an external view on its own does not beat chemistry, and fusing the three is above every single view (Table D7). External methods are positioned in Table D1.
  • W2.4 (scope of the analyses). The biological analysis is repeated on both backbones and both datasets (Table D6).
  • W2.5 (statistics). A significance test for every comparison (Table D7).
  • W2.6 (claims). The latent result is stated as a co-occurrence, following existing practice for examining a learned representation; the gain is shown on two backbones, two datasets and four types of split.

W1.1 / Q4. How MoRe differs from existing multimodal and target-aware methods

Thank you for the question. We have revised Sec. 2, and App. P gives a comparison table.

We reviewed existing multimodal and target-aware methods for chemical perturbation prediction. A common pattern among UniPert, MAP and StateXDiff is that the modalities are aligned into a single drug vector. Alignment keeps the information the views share, and gives the model no reason to keep what each view holds on its own. However, such modality-unique information can matter here. For instance, we observed that chemistry predicts the generic stress and cell-death programmes well, the target-informed views predict the programmes driven by specific proteins well, and fusing them beats the mean of the single views in 20 of 22 Hallmark programmes (Fig. 4a; App. K).

MoRe therefore adds no alignment loss. The objective is the backbone's native loss plus response supervision, and the three views enter the predictor separately, so the model learns how to weight them for each dataset instead of having that fixed in advance by a separate pretraining step.

Table D1. Where each method combines its drug modalities (App. P).

MethodDrug modalitiesCombinedViews at predictor
ChemCPA, PerturbDiffstructure–1
UniPertstructure + target sequencealigned, in pretraining1
MAP (Feng et al., 2026)structure + protein sequence + mechanism text + KGaligned, in pretraining1
StateXDiff (Shi et al., 2026)structure + L1000 + biological graphaligned, in pretraining1
MoRe (ours)structure + target identity + target-informed embeddingkept distinct, in the predictor3

Q4: network and pathway representations. The two are not alternatives. A network or pathway representation is simply another view, and MoRe can fuse it alongside the other three in the same way; the revised Discussion names this as future work. We did not add one here because it needs strictly more than we do: the same target information, plus a curated interaction network and the assumption that propagation through that network reflects the drug's downstream biology in the cell line at hand. For a novel drug, that extra is the part we would trust least.

References

  • Feng et al. A knowledge-driven framework for predicting single-cell responses for unprofiled drugs. Nature Machine Intelligence, 2026.
  • Shi et al. StateXDiff: cell state-contextualized multimodal diffusion for single-cell perturbation prediction. arXiv:2605.16104, 2026.
AuthorsOpenReview ↗

Response to Reviewer dRXi (2/5)

Continued: W1.4 and W1.2 / W1.3.

W1.4. The fusion operator

We do not claim novelty in the fusion mechanism. Projection and concatenation is a deliberate choice, given how few drugs supervise the fusion: the module sees one input per training drug (150 to 300 drugs), so an operator whose weights depend on the input has to learn a weighting per drug from very few examples, and gating can route the gradient to one view early and starve the others (Wu et al., 2022; Huang et al., 2022). Concatenation gives every view a direct gradient path and leaves the weighting to the next layer. We tried self-attention and mixture-of-experts gating, and neither gives a consistent improvement over it (App. H, Tables 11–14). The contribution we claim is the plug-in with response supervision, not the operator.

Full results. App. H: every representation and fusion operator under both objectives, for both backbones and datasets.

W1.2 / W1.3. Information in the target view; downstream propagation

Thank you for this. The target views do cover polypharmacology in part: they hold high-confidence direct targets, complex dependency genes and experimentally reported off-targets. They do not encode affinity, potency or direction of action, and that is deliberate. Such values are often unavailable for a new compound, and a predicted affinity would add its own estimation error: affinity networks largely reproduce the values of their nearest training ligands, which is least reliable for exactly the compounds we hold out (Volkov et al., 2022). A strong prior that is wrong for an unseen drug can be worse than no prior, so we leave these quantities to be learned from the response data.

On its own, UniPert has limited ability to separate drugs that share a target but act on it differently, since it is aligned with the associated targets rather than with the direction of action. This is exactly why we use several views: two such drugs can still differ in structure even though they may engage the same targets. For example, fulvestrant degrades the estrogen receptor, while toremifene, which is in training, modulates the same receptor; their mean responses on the DEGs of fulvestrant correlate at only 0.30. The target-only model predicts fulvestrant at 0.30 and + MoRe at 0.40 (DEG50 Pearson). Across the 20 held-out drugs that share a target with a training drug but are structurally distinct from every training drug, + MoRe improves on the target-only model by 16%.

Downstream effect propagation. We do not model propagation explicitly, as the downstream response is what we predict. We prefer that here because the engaged targets of a chemical perturbation are uncertain and often incomplete, so the starting point of a propagation is far less well defined than for a genetic perturbation, where the perturbed gene is known. A network view can still be added as a further view, as described under Q4.

References

  • Wu et al. Characterizing and overcoming the greedy nature of learning in multi-modal deep neural networks. ICML, 2022.
  • Huang et al. Modality competition: what makes joint training of multi-modal network fail in deep learning? (Provably). ICML, 2022.
  • Volkov et al. On the frustration to predict binding affinities from protein-ligand structures with deep neural networks. Journal of Medicinal Chemistry, 2022.
AuthorsOpenReview ↗

Response to Reviewer dRXi (3/5)

Continued: W1.5 / Q3.

W1.5 / Q3. DEG weighting and alignment with DEG-based metrics

Thank you for this question. We retrained the response term with every gene weighted equally (Table D2). The gain does not come from the objective fitting the metric.

  • Condition-level supervision helps on its own. The model with every gene weighted equally still improves the backbone on Tahoe-100M at every dose and on sci-Plex3 at the highest dose, where prior work reports its results (Hetzel et al., 2022; Piran et al., 2024). It falls below only at the lower sci-Plex3 doses, where the perturbation signal is weak, and the weighting recovers it.
  • The improvement is not confined to the genes the DEG metric scores. If the gain came from fitting the objective to the metric, it should appear only in each condition's DEG50. It does not: on the remaining 1,950 genes, response supervision raises Pearson from 0.574 to 0.604 (Table D3).
  • Nothing about the test conditions enters the objective. The DEG sets and the supervision targets are computed from training conditions only (Sec. 3.3).
  • Up-weighting responsive genes is a standard remedy for signal dilution. Up-weighting differentially expressed genes during training, or evaluating performance on the DEG50, is a common remedy to emphasise the perturbation-specific signal (Roohani et al., 2024; Mejia et al., 2026). We agree that DEGs are outcomes rather than mechanisms. Our claim is predictive accuracy for unseen drugs.

Table D2. + MoRe trained without the DEG weighting (PerturbDiff; DEG50 Pearson; mean over five random-split folds). Highest dose: recomputed from the saved condition means, which gives slightly different values from Table D4. ∗: differs from the original backbone (paired t-test over folds, p < 0.05).

DosesOriginalNo DEG weighting+ MoRe
sci-Plex3All0.2230.180∗0.253
Highest0.2150.2260.353∗
Tahoe-100MAll0.6680.720∗0.737∗
Highest0.5930.653∗0.673∗

Table D3. Pearson correlation on the DEG50 of each condition, on the rest of the panel, and on all genes, without and with response supervision (Tahoe-100M, PerturbDiff, chemistry view; mean ± standard deviation over three seeds). ∗: the two differ (Welch t-test over seeds, p < 0.05).

GenesPerturbDiff+ response supervision
DEG500.678 ± 0.0040.737 ± 0.004∗
Rest (1,950)0.574 ± 0.0060.604 ± 0.003∗
All (2,000)0.606 ± 0.0060.649 ± 0.002∗

Full results. App. D: the loss, and the effect of response supervision on the three gene sets.

References

  • Hetzel et al. Predicting cellular responses to novel drug perturbations at a single-cell resolution. NeurIPS, 2022.
  • Piran et al. Disentanglement of single-cell data with Biolord. Nature Biotechnology, 2024.
  • Roohani et al. Predicting transcriptional outcomes of novel multigene perturbations with GEARS. Nature Biotechnology, 2024.
  • Mejia et al. Needles in the haystack: addressing signal dilution improves scRNA-seq perturbation response modeling and evaluation. ICML, 2026.
Reviewer_k4EAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MoRe is a plug-in for single-cell drug-perturbation models that combines chemical, target, and UniPert drug representations with an auxiliary response-level loss. Applied to ChemCPA and PerturbDiff, it improves the reported held-out-drug results on sci-Plex3 and Tahoe-100M.

Critical strengths

MoRe is simple and works with both an autoencoder and a diffusion model without replacing either backbone. This supports its use as a reusable conditioning module. The evaluation uses perturbation-specific metrics and includes a useful perturbation-mean reference, which helps reveal models that ignore drug identity.

Critical weaknesses

The unseen-drug result uses a random held-out-drug protocol, while MoRe still receives target annotations and a target-informed embedding for each test drug. Target overlap or chemical similarity would help clarify the scope of the result.

On Tahoe-100M, output averaging matches or slightly exceeds learned fusion on several endpoints. The benefit of learned fusion is therefore somewhat mixed in this setting.

Questions for authors
  1. What is the target overlap or chemical similarity between the training and test drugs?

  2. When does learned fusion improve over output averaging, and is this pattern consistent across held-out drugs?

  3. How were the target annotations and UniPert embeddings constructed? In particular, do they use perturbation-response information related to the evaluation datasets?

  4. How does MoRe handle held-out drugs with missing or incomplete target information? This would help clarify how dependent the method is on the additional drug views.

4 comments
AuthorsOpenReview ↗

Response to Reviewer k4EA (1/4)

We thank the reviewer for the careful reading and for four concrete questions; each now has a measurement in the paper. We summarise the main changes first and respond to each question in detail below:

  • W1 / Q1 (target overlap and chemical similarity). Both held-out sets are now audited and PerturbDiff is retrained on scaffold- and target-disjoint splits; the gain holds (Table R1; App. M, N).
  • W2 / Q2 (fusion and output averaging). One + MoRe model matches the output average of three single-view models over all doses and is significantly better at the highest dose (Tables K3, K4; App. J).
  • Q3 (how the views were built). Both pipelines are now documented in the revised paper; neither uses expression data or perturbation responses (App. B).
  • Q4 (missing target information). MoRe conditions on several views, so it does not depend on any single one: with the target view missing its correlation drops by about 0.02 and stays above the target-only model (Table K5).

W1 / Q1. Target overlap and chemical similarity between training and held-out drugs

Thank you for the question. Table K1 reports the overlap. On the five-fold random split, the structural similarity between the held-out and training drugs is low (median max Tanimoto 0.31 on sci-Plex3, 0.28 on Tahoe-100M). Few held-out drugs share a scaffold with training drugs (16% and 22%), but 65% and 41% of them share a direct target.

To emphasize that the performance gain of MoRe is not purely due to this overlap between training and testing, we built three harder splits that hold out drugs with an unseen Bemis–Murcko scaffold (scaffold-disjoint, S1), an unseen direct target (target-disjoint, S2), and both an unseen Bemis–Murcko scaffold and an unseen direct target (scaffold- and target-disjoint, S3); we write S0 for the random drug split. We then retrained PerturbDiff with and without MoRe on five folds of each.

Table K1. Overlap between held-out and training drugs, by split (mean over five folds). n: held-out drugs per fold; Conditions: held-out drug × cell line × dose conditions. Tanimoto: for each test drug we take the maximum ECFP4 (Morgan, radius 2, 2048 bits) Tanimoto to any training drug, then report the median over test drugs. Scaffold / Direct target: shares a Bemis–Murcko scaffold / a direct target with a training drug. S1–S3 are new.

SplitnConditionsTanimotoScaffoldDirect target
sci-Plex3
S0303580.3116%65%
S1364340.280%64%
S2192300.247%0%
S3131510.230%0%
Tahoe-100M
S0766820.2822%41%
S1736590.260%38%
S2635700.2515%0%
S3423800.250%0%

The gain holds on S1, S2 and S3 (Table K2).

  • Stricter constraints leave fewer drugs to hold out. All drugs with overlapping direct targets must stay on one side of the split, so fewer drugs are held out in S2 and S3.
  • The harder the split, the lower the performance. On sci-Plex3, PerturbDiff falls from 0.223 on S0 to 0.153 on S3, and our model (PerturbDiff + MoRe) falls from 0.253 to 0.169. On Tahoe-100M both stay relatively flat.
  • Tahoe-100M changes less than sci-Plex3 because more of its response is shared across drugs. The perturbation mean, which uses no drug information, already reaches 0.54 on Tahoe-100M against 0.26 on sci-Plex3 (Table 1), so accuracy there depends less on the similarity to the training drugs.
  • + MoRe improves the PerturbDiff backbone on every split of both datasets, by 10–21%.
  • Removing the target overlap does not shrink that improvement. On S2, where no held-out drug shares a direct target with any training drug, our model (+ MoRe) reaches 0.209 against PerturbDiff's 0.173 on sci-Plex3 (+21%) and 0.736 against 0.662 on Tahoe-100M (+11%), both significant.

Table K2. DEG50 Pearson of PerturbDiff with and without MoRe, by type of split (mean over five folds; standard deviations in Table R1 of the general response). Gain: relative gain with MoRe. ∗: + MoRe differs from PerturbDiff (paired t-test over folds, p < 0.05).

SplitPerturbDiff+ MoRe (ours)Gain
sci-Plex3
S00.2230.253+13%
S10.1930.225∗+17%
S20.1730.209∗+21%
S30.1530.169+10%
Tahoe-100M
S00.6680.737∗+10%
S10.6600.737∗+12%
S20.6620.736∗+11%
S30.6530.726∗+11%

Full results. App. M: every split, all metrics. App. N: the overlap with the training drugs for every split of both datasets.

AuthorsOpenReview ↗

Response to Reviewer k4EA (2/4)

Continued: W2 / Q2.

W2 / Q2. Learned fusion and output averaging

Thank you for this question. One clarification first: the output average in the original manuscript already carries our response supervision, so it is not plain averaging. To provide a comprehensive comparison, we show the results of averaging without response supervision and averaging with it.

Without response supervision (Table K3). On Tahoe-100M, + MoRe is above late fusion in both metrics on every split, by about 6% in DEG50 Pearson. On the disjoint splits of sci-Plex3 it improves Pearson by up to 16% and direction agreement by 2–5%; on the random split of sci-Plex3 the two are comparable in correlation and + MoRe has the higher direction agreement. Late fusion is itself above every single view in correlation, so the three views are complementary.

Table K3. Output averaging without response supervision, against + MoRe, by split (PerturbDiff; mean over five folds). S0 random, S1 scaffold-disjoint, S2 target-disjoint, S3 scaffold- and target-disjoint. Views: C chemistry, T target identity, U UniPert. Late fusion averages the predicted condition means of three independently trained single-view models, none of which uses response supervision. r: DEG50 Pearson; dir: DEG50 direction agreement. Bold: best of the column, per dataset.

r, S0r, S1r, S2r, S3dir, S0dir, S1dir, S2dir, S3
sci-Plex3
PerturbDiff (C)0.2230.1930.1730.1530.5910.5730.5510.544
PerturbDiff (T)0.2260.1880.1430.1460.5820.5730.5340.528
PerturbDiff (U)0.2330.2050.1620.1400.5840.5720.5440.537
Late fusion (C+T+U)0.2600.2210.1800.1670.5930.5760.5410.534
+ MoRe0.2530.2250.2090.1690.6020.5890.5700.560
Tahoe-100M
PerturbDiff (C)0.6680.6600.6620.6530.7240.7210.7220.715
PerturbDiff (T)0.6760.6760.6760.6670.7300.7290.7300.726
PerturbDiff (U)0.6620.6590.6540.6520.7210.7190.7130.710
Late fusion (C+T+U)0.6980.6940.6930.6880.7420.7400.7380.732
+ MoRe0.7370.7370.7360.7260.7540.7520.7510.751

With response supervision (Table K4). + MoRe is a more efficient solution than output averaging. Output averaging is a strong baseline (Lakshminarayanan et al., 2017; Ovadia et al., 2019), but it needs multiple trained models. One + MoRe model reaches a comparable correlation and direction agreement.

Where the two differ (Table K4). Q2 asks when learned fusion helps: when the response is strong and specific to the drug. At the highest dose of sci-Plex3, the setting prior work reports (Hetzel et al., 2022; Piran et al., 2024), + MoRe is significantly better than late fusion in correlation and in direction agreement, whether the averaged models are trained with or without response supervision. On Tahoe-100M, where much of the response is shared across drugs (W1), the two are on par, within about 1% in both metrics (App. J).

Table K4. Late fusion without and with response supervision, against one + MoRe model, over all doses and at the highest dose (sci-Plex3, PerturbDiff, S0; mean over five folds). Bold: best of the row. ∗: significantly different from + MoRe (paired t-test over folds, p < 0.05).

DosesLate fusion, original modelsLate fusion, with response supervision+ MoRe
DEG50 PearsonAll0.2600.2620.253
Highest0.254∗0.330∗0.361
DEG50 directionAll0.593∗0.579∗0.602
Highest0.583∗0.591∗0.658

Full results. App. J: + MoRe, the single-view models and late fusion, for both datasets.

References

  • Lakshminarayanan et al. Simple and scalable predictive uncertainty estimation using deep ensembles. NeurIPS, 2017.
  • Ovadia et al. Can you trust your model's uncertainty? Evaluating predictive uncertainty under dataset shift. NeurIPS, 2019.
  • Hetzel et al. Predicting cellular responses to novel drug perturbations at a single-cell resolution. NeurIPS, 2022.
  • Piran et al. Disentanglement of single-cell data with Biolord. Nature Biotechnology, 2024.
AuthorsOpenReview ↗

Response to Reviewer k4EA (3/4)

Continued: Q3.

Q3. How the target annotations and UniPert embeddings were built

Thank you for the question. Neither the UniPert embedding nor the target annotation uses perturbation-response data. Both are built from structure and published literature only, without using expression data from sci-Plex3 or Tahoe-100M.

1. UniPert (U). We take the released pretrained encoder, keep it frozen, and give it SMILES as input to obtain the feature embedding. It was trained to align compounds with their protein targets and uses no expression data. The gain of + MoRe is at least as large for held-out drugs outside UniPert's pretraining set as for those inside it (Table D5 in our reply to Reviewer dRXi).

2. Target identity (T). We retrieve the targets following Li et al. (2026a): for each drug, target genes are collected from the published literature with a large language model, in three evidence tiers (direct targets, complex dependencies and reported off-targets). No expression data is used. Each target is encoded as a fixed random 512-dimensional vector, and the average over a drug's targets is taken as that drug's feature embedding.

Full results. App. B: both pipelines.

References

  • Li et al. Simple predictors, strong baselines: revisiting chemical perturbation prediction. NeurIPS 2026 AI4DD Workshop, 2026a.
AuthorsOpenReview ↗

Response to Reviewer k4EA (4/4)

Continued: Q4.

Q4. Held-out drugs with missing or incomplete target information

Thank you for this question. As the chemical features (RDKit) are by construction always available when the drug's SMILES is known, we answer this question by corrupting the other two views, or completely replacing them with noise, at inference time (Table K5).

How we corrupt a view. We replace a view by a mixture of itself and noise, x′ = (1 − α)x + αε, with ε ∼ N(0, diag(σ²)), where σ is the per-dimension standard deviation of that view over the training-partition drugs, so the noise matches the scale of the feature it replaces. α = 0 leaves the view untouched; intermediate α corresponds to incomplete or noisy information; α = 1 replaces the view by noise entirely, so it carries no information about the drug, which is the case of a missing view.

Performance under corruption (Table K5). Of the three views, target annotation is the one most likely to be missing or incomplete. We report S2, where no held-out drug's primary target appears in training, so corrupting that view is not confounded by primary target overlap. We corrupt either one view or both the target and UniPert views of + MoRe, and compare against two baselines: the single-view model that relies on the corrupted view alone, and the chemistry-only model, since the RDKit features are available for any drug with a known SMILES.

  • Performance falls as the noise rises, and falls further when both views go. + MoRe declines monotonically with α, losing 7.9% of its DEG50 Pearson when the target view is replaced by noise and 6.3% when UniPert is; corrupting both costs 12.3%.
  • With a view replaced by noise, + MoRe stays above the model that relies on that view alone: 0.191 against 0.133 for the target view and 0.194 against 0.118 for UniPert (both significant, paired t-test over folds), and its own decline is not significant.
  • Even with both views destroyed, + MoRe stays above the chemistry-only baseline, 0.182 against 0.173: it still holds the chemistry view, which was trained with response supervision.

Table K5. DEG50 Pearson when a view is replaced by noise at test time, x′ = (1 − α)x + αε (sci-Plex3, split S2, target-disjoint; mean ± s.d. over five folds). α = 0 is the uncorrupted model, so that column is the original performance; α = 1 replaces the view by noise entirely. Drop: relative loss from α = 0 to α = 1. PerturbDiff (T) and PerturbDiff (U) rely on the corrupted view alone, while + MoRe also holds the other views. PerturbDiff (C), chemistry only, is listed for reference and is never corrupted, since the chemical features exist whenever the SMILES is known (value from Table K2). The other splits, every noise level and direction agreement are in App. O.

Modelα = 0.0 (original)α = 0.2α = 0.4α = 0.6α = 0.8α = 1.0Drop at α = 1
Reference, never corrupted
PerturbDiff (C)0.173 ± 0.060––––––
Corrupting the target view (T)
PerturbDiff (T)0.145 ± 0.0490.144 ± 0.0560.144 ± 0.0610.142 ± 0.0640.137 ± 0.0630.133 ± 0.0638.5%
PerturbDiff + MoRe0.207 ± 0.0510.205 ± 0.0500.203 ± 0.0490.198 ± 0.0470.193 ± 0.0450.191 ± 0.0457.9%
Corrupting the UniPert view (U)
PerturbDiff (U)0.157 ± 0.0610.152 ± 0.0630.147 ± 0.0610.139 ± 0.0550.128 ± 0.0440.118 ± 0.03224.8%
PerturbDiff + MoRe0.207 ± 0.0510.206 ± 0.0510.204 ± 0.0510.200 ± 0.0500.197 ± 0.0500.194 ± 0.0496.3%
Corrupting both views
PerturbDiff + MoRe0.207 ± 0.0510.203 ± 0.0520.198 ± 0.0500.191 ± 0.0460.186 ± 0.0450.182 ± 0.04512.3%

Full results. App. O: the other splits and every noise level. App. H (Tables 11–14): every view and view combination under both objectives.

We greatly appreciate the time and effort you’ve put into reviewing our submission. We have carefully responded to all your concerns. We would be grateful if you could consider raising your score in light of the new experiments, explanations and revision. Thank you again for your time and thoughtful suggestions!

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank all the reviewers for their valuable and insightful feedback. We are glad that the reviewers found the problem of generalising to unseen drugs important and the combination of complementary drug representations with response supervision well motivated (PVVF\color{teal}{\textbf{PVVF}}), the design simple and model-agnostic, working with both an autoencoder and a diffusion model without replacing either backbone (k4EA\color{blue}{\textbf{k4EA}}, dRXi\color{purple}{\textbf{dRXi}}), and the evaluation perturbation-specific, across two architectures and two datasets with extensive ablations (k4EA\color{blue}{\textbf{k4EA}}, dRXi\color{purple}{\textbf{dRXi}}, PVVF\color{teal}{\textbf{PVVF}}).

In short. MoRe is a plug-in that gives a predictor several views of a drug and a training signal on its response. We showcase that adding it improves two backbones (an autoencoder and a diffusion model) on two datasets (sci-Plex3 and Tahoe-100M), over five folds for each of four types of split: random, scaffold-disjoint, target-disjoint, and scaffold- and target-disjoint.

We found the constructive feedback very helpful and have prepared an updated manuscript. We have mainly updated the positioning in Section 2, the evaluation protocol in Section 4, the results in Sections 5.1 to 5.4 (Tables 1 and 2), the Discussion, and Appendices B, D, E, H, J and K, and added Appendices M (folds and disjoint splits), N (overlap with the training drugs), O (robustness to a noisy or missing view) and P (comparison with prior methods); changes are marked in red, and table and appendix numbers below refer to the revised PDF. We reply to each reviewer in more detail in the individual responses and summarise the changes here:

  • We now report five random-split folds for both backbones on both datasets; + MoRe consistently improves the original backbone (Table 2; App. M). (dRXi\color{purple}{\textbf{dRXi}}, PVVF\color{teal}{\textbf{PVVF}})
  • We have added scaffold-disjoint, target-disjoint, and scaffold- and target-disjoint splits; without shared scaffolds or shared targets the gain is as large as on random folds (Table R1). (PVVF\color{teal}{\textbf{PVVF}}, k4EA\color{blue}{\textbf{k4EA}})
  • We have added significance tests for every comparison; the gain over the original backbone is statistically significant in 8 of the 10 settings of Table R1. (dRXi\color{purple}{\textbf{dRXi}}, PVVF\color{teal}{\textbf{PVVF}})
  • We have retrained each of the two components alone on five folds; each improves the backbone, and + MoRe is significantly better than either alone (App. H). (dRXi\color{purple}{\textbf{dRXi}}, PVVF\color{teal}{\textbf{PVVF}})
  • We have tested robustness to a noisy or missing drug view; with one view noisy or missing, + MoRe stays above the model that relies on that view alone (App. O). (k4EA\color{blue}{\textbf{k4EA}}, PVVF\color{teal}{\textbf{PVVF}})
  • We have added a direct gradient measurement; for both backbones, response supervision gives a more coherent gradient with respect to the drug latent than the native loss (App. E). (PVVF\color{teal}{\textbf{PVVF}})
  • We have positioned MoRe against multimodal and target-aware methods in a new table (Sec. 2; App. P), documented how both external views were built (App. B), and repeated the biological analysis on both backbones and datasets (App. K). (dRXi\color{purple}{\textbf{dRXi}}, k4EA\color{blue}{\textbf{k4EA}})

Table R1. DEG50 Pearson of each original backbone and with MoRe, five folds per split. Mean ± standard deviation over folds; one training seed per fold. Δ: paired difference per fold. Bold: the better of the two. ∗: + MoRe differs from the original backbone (paired t-test over folds, p < 0.05). Disjoint splits hold out whole groups of drugs linked by a shared Bemis–Murcko scaffold, a shared direct target, or either.

SplitOriginal+ MoRe (ours)Δ
sci-Plex3
ChemCPA, random0.117 ± 0.0240.256 ± 0.018∗+0.139 ± 0.036
PerturbDiff, random0.223 ± 0.0200.253 ± 0.028+0.030 ± 0.022
PerturbDiff, scaffold-disjoint0.193 ± 0.0290.225 ± 0.024∗+0.032 ± 0.007
PerturbDiff, target-disjoint0.173 ± 0.0600.209 ± 0.051∗+0.037 ± 0.025
PerturbDiff, scaffold- and target-disjoint0.153 ± 0.0410.169 ± 0.030+0.016 ± 0.042
Tahoe-100M
ChemCPA, random0.438 ± 0.0220.489 ± 0.036∗+0.051 ± 0.036
PerturbDiff, random0.668 ± 0.0090.737 ± 0.014∗+0.069 ± 0.006
PerturbDiff, scaffold-disjoint0.660 ± 0.0150.737 ± 0.009∗+0.078 ± 0.009
PerturbDiff, target-disjoint0.662 ± 0.0270.736 ± 0.028∗+0.074 ± 0.007
PerturbDiff, scaffold- and target-disjoint0.653 ± 0.0290.726 ± 0.040∗+0.073 ± 0.017

We appreciate the reviewers' time and are happy to provide any further analysis or clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.