acceptodds
Withdrawn from ICLR 2027

Hitting the Right Targets: Simple Pseudobulk Predictors for Chemical Perturbation Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Predicting transcriptional responses to chemical perturbations can accelerate drug discovery. Single-cell perturbation models are increasingly complex, yet they are typically evaluated on aggregated pseudobulk profiles, where simple baselines remain competitive. We revisit chemical perturbation prediction at pseudobulk resolution with a lightweight MLP on frozen compound and cell-line embeddings, and benchmark twelve compound representations spanning molecular structure, morphology-aligned phenotype, and drug targets. On sci-Plex3 and Tahoe-100M, the predictor outperforms learning-free baselines and four single-cell models (ChemCPA, BioLord, Doloris, and PerturbDiff) on most metrics while using up to 55 fewer parameters. To understand why, we examine the two ingredients of the predictor: its prediction target and its compound representation. Aggregation yields a less noisy target, and retraining a representative single-cell model (ChemCPA) on pseudobulk profiles narrows, but does not close its gap to our model. On the representation side, more sophisticated molecular embeddings bring little consistent gain; instead, performance tracks how closely the geometry of a representation aligns with that of the transcriptional responses. These results suggest that predictive performance depends not only on model complexity but also on the data, and position pseudobulk predictors as strong baselines and potential priors for single-cell models.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single cell perturbation prediction, chemical perturbation, representation learning, benchmark

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_vsE4OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

the authors proposed a lightwieght mlp method to predict chemical perturbation effect on pseudobulk level based on fozen compound/cell-line embeddings. they compared to non-learning baseline and other four single-cell models on 12 compound representations in sciplex3 and tahoe-100m. results demonstrates that this simple method can get better performance though with smaller parameter size. they showed that pseudobulk aggregation can decrease the noise in target prediction.

Critical strengths
  1. they proposed a simple baseline but gained better performance empirically suggesting that to choose a matched eval resolution for the prediction objectives matters more than scaling.
  2. the paper contains multiple pseudobulk level experiments and thorough alignment analyses on compound representation/transcriptional response geometry to study the reason behind simple model performas better. most importantly, revealing the limitation of distributional metrics on evaluating population heterogeneity induced by perturbation.
Critical weaknesses
  1. current evaluation cannot differentiate drug-specific signals from shared drug response. e.g. the best mlp gains on \delta-pearson seems not statistically significant compared to cell line perturbation mean, whereas degnerates on \delta-direction.
  2. missing explanation on drug target annotation coverage. the coverage seems only refer to that part of drugs have at least one target in db, but not explaining the target set precision. performance gain from drug-target knowledge is not clear.
Questions for authors
  1. can u remove shared pertubration component in evaluation then update table2? e.g. subtract cell line specific training perturbation mean from prediction as well as from ground truth.
  2. can u compare to simple regression based method (e.g. ridge regression) or other regularized linear predictors, to resolve concerns whether performance gain is from direct psuedobulk supervision or nonlinear architecture.
  3. please also address weakness #2
Reviewer_tzsVOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper shows that a complex single cell model and a molecular representation does not necessarily result in a strong performance. They instead introduce a lightweight MLP that predicts the mean transcriptional effect. They evaluate their model on two chemical perturbation datasets: Tahoe 100M and sciplex and show that such a simple model could produce competitive results compared to single cell based models. They also argue that prediction at the pseudobulk level could be a more reproducable prediction target and even improves performance. Finally, this paper shows that the performance depends more on how well the representation geometry aligns with transcriptional response.

Critical strengths
  • The paper asks and answer an interesting question.
  • The analysis of why pseudobulk has advantage over single cell is interesting.
Critical weaknesses
  • My main concern is the interpretation and generality of the target-based representation results. The strongest evidence for Target ID comes from sciPlex3, especially on the DEG50 evaluation. On Tahoe the differences among Target ID, RDKit, MolFormer, CellCLIP, and UniPert are very small. I do not think the experiments establish a general conclusion that target-based representations are preferable for unseen-drug prediction.
  • The definition of "novel drug" in this paper requires additional discussion. Target-based prediction assumes that the targets of the held-out molecule are already known. In this work, these targets are retrieved from existing literature using an LLM-based pipeline. I would like to see the target-based analysis stratified by target availability and novelty. For example, how does the method perform when (1) the test drug has known targets that also occur among training compounds, (2) its targets are known but absent from the training set, and (3) no target information is available?
  • Two datasets provide different conclusions regarding representation choice. On sciPlex3, representation choice matters considerably, while on Tahoe most representations perform nearly identically and even random neighbors have relatively similar responses.
Questions for authors
  • How should the Target ID approach be applied to a genuinely novel compound for which no reliable target annotations are available?
  • Why do target-based representations provide a clearer advantage on sci-Plex3 but essentially no advantage over structure-based representations on Tahoe
Reviewer_yjvLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors investigate chemical perturbation prediction at pseudobulk resolution, arguing that lightweight models trained to predict mean transcriptional responses can outperform substantially more complex single-cell generative models when evaluated on aggregate-response metrics. They propose a simple MLP conditioned on frozen drug and cell-line embeddings that predicts the transcriptional shift relative to an observed control profile.

The authors systematically compare twelve compound representations, including molecular descriptors, chemical language-model embeddings, morphology-aligned representations, and protein-target information. On sci-Plex3 and Tahoe-100M, their predictor outperforms four representative single-cell models and learning-free baselines on most pseudobulk metrics, while using substantially fewer trainable parameters.

The paper further investigates the role of pseudobulk aggregation, demonstrating that aggregation improves target reproducibility and that training ChemCPA on pseudobulk profiles improves its performance. Finally, the authors study the relationship between representation geometry and transcriptional-response similarity. Their most interesting observation is that simple representations encoding protein-target identity can outperform considerably more sophisticated molecular representations, particularly when held-out compounds share targets with training compounds.

Overall, the paper provides useful empirical observations concerning model complexity, prediction resolution, and inductive biases in chemical perturbation prediction. However, several important experiments are missing to establish the proposed explanations and distinguish the contribution of the learned predictor from simpler statistical models.

Critical strengths
  1. A useful reassessment of prediction complexity. The authors correctly distinguish between predicting the mean transcriptional response and predicting the full distribution of single-cell responses. Their results reinforce the importance of evaluating sophisticated generative models against strong, task-appropriate baselines. This is a worthwhile contribution in a field increasingly dominated by complex architectures.

  2. A systematic comparison of compound representations. Evaluating molecular structure, chemical foundation-model embeddings, morphology-related features, and biological targets under a common prediction architecture makes the comparison informative. Particularly interesting is the finding that target-identity representations, containing no explicit biological similarity between different proteins, can outperform substantially richer molecular embeddings.

  3. An interesting investigation of representation geometry. The authors analyze the correspondence between similarity in compound-representation space and similarity of transcriptional responses, using local neighborhood statistics and centered kernel alignment. The observation that target-based representations benefit from shared mechanisms between training and test compounds offers an intuitive explanation of their predictive performance.

  4. Extensive empirical comparisons and ablations. The experiments include two substantial perturbation datasets, multiple drug representations, four single-cell models, learning-free baselines, loss-function and cell-line encoding ablations, and analyses of pseudobulk reproducibility. Retraining ChemCPA at pseudobulk resolution and evaluating shifted and unshifted control-cell distributions are particularly useful controls.

Critical weaknesses
  1. Missing strong learned baselines for target-based representations

The most interesting result is the strong performance of the Target ID representation. However, its effectiveness may have a considerably simpler explanation than the authors investigate. Each gene target is assigned a fixed random vector, and a drug is represented by averaging the vectors associated with its targets. In sufficiently high dimensions, these vectors are approximately orthogonal, so the resulting representation essentially encodes the drug–target incidence matrix, with inner products reflecting target-set overlap. The natural baselines are therefore regularized linear regression on drug–target indicators, low-rank target-response factorization, and kernel ridge regression with a target-overlap kernel. None is evaluated. The authors instead compare against a learning-free lookup baseline that averages the transcriptional responses of drugs sharing at least one target. This is a much weaker comparison than a learned regression model that can estimate the contributions of different targets. Without these baselines, it is unclear whether the neural network provides any substantial benefit over conventional statistical learning supplied with the same target annotations.

  1. The proposed explanation of the pseudobulk advantage is incomplete

The authors attribute part of their predictor's success to the reduced noise obtained by aggregating individual-cell responses into pseudobulk profiles. Although aggregation certainly improves the reproducibility of estimated condition means, this observation alone does not establish an advantage for training a conditional-mean predictor. In fact, for a predictor that produces the same output for all cells under a given condition, minimizing squared error over individual cells is mathematically equivalent, up to a constant and appropriate weighting, to minimizing squared error against their average: N|f_\theta(z)-\bar{x}|^2+ \sum_i|x_i-\bar{x}|^2.

Therefore, aggregation need not improve the statistical estimation of a conditional mean under an MSE objective. Differences may instead arise from architecture, optimization, weighting of conditions, or the learning objective. The experiment retraining ChemCPA on pseudobulks is useful, but it does not resolve this issue because ChemCPA differs from the proposed MLP in several respects beyond training resolution.

An experiment training the same predictor on individual-cell and pseudobulk observations, with matched objectives and condition weighting, would provide a substantially more convincing explanation.

  1. Generalization is primarily demonstrated within a shared mechanism space

The authors show that the target-based representation performs better when held-out compounds share protein targets with training compounds. Indeed, they explicitly acknowledge that this corresponds to generalization within a partially shared mechanism space rather than extrapolation to unseen mechanisms. This substantially limits the interpretation of the results. The Bemis–Murcko scaffold split on sci-Plex3 ensures structural novelty but does not prevent test compounds from engaging previously observed targets. The random drug splits used for Tahoe-100M provide an even weaker test of mechanistic generalization. A more informative experiment would evaluate target-disjoint or mechanism-of-action-disjoint splits and compare against structure-based representations under the same conditions.

In particular, I would like to understand whether the target-based representation retains any advantage when the test compounds act on proteins not represented in the training data. This would help distinguish generalization through known biological mechanisms from the more challenging problem of predicting genuinely novel perturbations.

  1. Target annotations introduce substantial external information and potential confounding

The strongest representations depend on protein-target annotations retrieved from scientific literature using GPT-5.4. This provides the model with extensive drug-specific biological knowledge that is not available to methods operating only on molecular structure. The authors validate these annotations by reporting that at least one reference target was recovered for 219 of 264 annotated Tahoe compounds. However, this measure establishes coverage rather than precision. Given that the retrieved target lists often contain numerous genes, recovering at least one annotated target does not guarantee that the complete representation is reliable. Furthermore, the annotation procedure combines direct binding targets, complex-dependency genes, and experimentally reported off-targets. These are mechanistically distinct categories, and their relative contributions to predictive performance are not investigated.

The authors should report annotation precision on a manually curated subset and evaluate these evidence categories separately. Comparison against a standard curated target database would also be useful. This is not necessarily a problem of information leakage, but rather one of understanding how much of the performance comes from external biological knowledge and how reliably that knowledge can be obtained for genuinely new compounds.

  1. The evaluation does not always demonstrate a substantial improvement in perturbation-specific prediction

The absolute expression metrics are strongly influenced by the basal transcriptional state. For example, on sci-Plex3, the cell-line perturbation-mean baseline already attains an R² of 0.840, compared with 0.886 for the target-based predictor. On the full gene panel, the target-based model improves residual Pearson correlation from 0.336 to 0.356, whereas the simpler RDKit representation reaches 0.360. The gains are more substantial on DEG50, but the authors acknowledge in Appendix A that the top-50 differentially expressed genes are selected from the ground-truth response of each held-out condition. This makes DEG50 an oracle-defined evaluation subset, not a directly realizable prospective prediction setting. Moreover, on Tahoe-100M, the cell-line perturbation-mean baseline remains superior on DEG50 direction accuracy. It would be useful to report performance on strongly responding genes without using the held-out condition to select them, and to quantify how accurately the models predict the magnitudes of perturbation-specific transcriptional shifts.

  1. The representation-geometry explanation is largely retrospective

The observation that better representations tend to preserve transcriptional-response geometry is intuitively appealing. However, the analysis computes alignment using experimentally measured response profiles, including those of held-out compounds. Consequently, the proposed alignment measures are primarily explanatory diagnostics rather than predictors of generalization performance. Furthermore, the correspondence is not entirely consistent: GenePT representations exhibit favorable local geometry on sci-Plex3 but perform poorly on several full-panel metrics. On Tahoe-100M, representation differences and alignment differences are both small, with substantial transcriptional similarity even among randomly selected compounds. The authors could strengthen this analysis by computing alignment exclusively on training compounds and testing whether it predicts downstream performance on held-out drugs across representations and splits.

  1. The comparison with single-cell generative models is not entirely symmetric

The proposed model is optimized directly for predicting population means, whereas the compared single-cell approaches attempt to represent more complex response distributions. Its advantage on pseudobulk metrics is therefore not necessarily surprising. The authors also evaluate distributional reconstruction by shifting observed control cells according to the predicted mean response. Interestingly, even unshifted controls perform competitively on EMD and Wasserstein distance. This suggests that the selected distributional metrics may be dominated by the control-cell distribution and insufficiently sensitive to perturbation-induced changes in cellular heterogeneity. A stronger comparison would explicitly measure changes in gene-expression variance, covariance, subpopulation composition, or state-dependent responses. These are precisely the properties that a global mean-shift model cannot reproduce.

The authors are justified in arguing that sophisticated single-cell models should be compared against strong pseudobulk baselines. Their results do not, however, establish that explicitly modeling heterogeneity provides no predictive benefit.

  1. Miscellaneous
  • The effective number of distinct training compounds is relatively small compared with the millions of trainable parameters in the predictor. Learning curves and comparisons with appropriately regularized low-capacity models would be informative.
  • The significance tests appear to aggregate results over cell lines and repeated random splits. Since observations across splits are not independent, a compound-level or hierarchical statistical analysis would provide stronger evidence for the reported improvements.
  • The authors address plate confounding in Tahoe-100M by training and evaluating on individual plates. While sensible, this does not establish robustness to unseen experimental batches or cross-plate generalization.
  • The reported parameter-efficiency advantage concerns the trainable predictor, not the computational or informational cost of obtaining pretrained embeddings and literature-derived target annotations. This distinction should be explicit.
  • The behavior of some representations changes considerably between the highest-dose experiment and the all-dose experiment. In particular, GenePT and Qwen-derived representations perform poorly on several main sci-Plex3 metrics but become among the strongest representations when all doses are included. This dependence on training conditions deserves further investigation.
Questions for authors
  1. How does the Target ID predictor compare against ridge regression on drug-target indicators, low-rank target-response models, and kernel ridge regression using target-overlap similarity? Is nonlinear learning necessary?

  2. Have the authors compared the same MLP trained directly on individual-cell observations versus aggregated pseudobulk observations, using matched squared-error objectives and condition weighting? How do they reconcile the claimed denoising advantage with the mathematical equivalence of these objectives for conditional-mean prediction?

  3. How does performance change when held-out drugs have no target overlap with training compounds? Can the authors evaluate a target-disjoint or mechanism-of-action-disjoint split?

  4. What is the precision of the GPT-derived target annotations? What happens when the target representation includes only validated direct binding targets rather than pooling all three evidence categories?

  5. How accurately does the model recover strongly responding genes without selecting them using the ground-truth test response? Can the authors report errors in the magnitudes of transcriptional changes?

  6. Can the authors demonstrate that representation–response geometry alignment computed using only training compounds predicts held-out performance across different drug representations?

  7. Can the authors evaluate single-cell distributional predictions using metrics sensitive to perturbation-induced changes in variance, covariance, and cell-state composition, rather than relying exclusively on EMD and Wasserstein distance?

  8. How sensitive are the results to the number of available training compounds, the cell-line representation, and the statistical dependence between repeated data splits?

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

Related papers

Discussion (0)

Sign in to comment.