- 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.
- 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.
- 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.
- 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.
- 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.
- 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.
- 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.
- 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.