acceptodds
Under review as a conference paper at ICLR 2027

GeoPert: Perturbation Prediction as Geometry-Constrained Distribution Modeling

26 views

Abstract

Predicting transcriptomic responses to genetic perturbations is a fundamental objective of virtual cell modeling. However, because single-cell sequencing is inherently destructive, cellular states cannot be tracked longitudinally within the same physical cell. Instead, assays such as Perturb-seq yield snapshot measurements across cell populations, framing perturbation prediction as a conditional distribution generative modeling problem. Under this formulation, generation follows a two-level hierarchical generative process:: A condition-specific cell distribution is drawn from a meta-distribution () that governs biological and technical variability across replicates, including incubation fluctuations and sequencing batch effects. : Individual cellular profiles are subsequently sampled from this realized population distribution (). Because defines an abstract distribution over probability measures, typically characterized via kernel mean embeddings in a latent RKHS, most existing frameworks collapse this hierarchy. By approaching in silico generation as a single-step marginal sampling task, they fail to account for replicate-level distributional shifts and are inherently incompatible with large-scale data augmentation schemes. To address these limitations, we propose GeoPert, a hierarchical virtual cell generative framework. GeoPert indirectly characterizes a cell population distribution in the latent RKHS by its energy distances, defined by the mean geodesic distances over a statistical manifold, against a small set of prototypical anchor populations. For a specific condition, GeoPert first generate its post-perturbation population distribution signature through conditional diffusion, supported by a foundational diffusion model via Classifier-Free Guidance (CFG). Virtual cells of a post-perturbation population are then generated through a variational inference framework conditioned on the energy distance vector signature of the population. By decoupling the virtual cell generation into a two-stage hierarchical generative process, GeoPert naturally support large-scale data augmentation, by harnessing large scale synthetic cell populations in the transcriptomic space with matching energy distance vectors, which improves cell generation authenticity. Benchmarks on public Perturb-seq datasets show that GeoPert achieves state-of-the-art performance in single-cell perturbation response prediction and is capable of generalizing to unseen gene perturbations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-Cell Perturbation Prediction, Functional Diffusion, Information Geometry

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_9mrKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GeoPert addresses single-cell perturbation prediction by modeling variation across population distributions and among cells within each population. It represents each population through a finite vector of Fisher–Rao MMD distances to selected anchor populations. A conditional diffusion model generates these population signatures from pretrained gene embeddings, and a stochastic decoder maps each signature and independent Gaussian noise to individual expression profiles. This separation also enables augmentation using synthetic populations with computable signatures. The paper provides theoretical justification for the population distance and coverage-dependent guarantees for the anchor representation, and evaluates the method on unseen single-gene perturbations across four cell lines and on combinatorial perturbations in the Norman dataset.

Critical strengths

The method is very interesting. I particularly like the idea of predicting the anchor distances as a parametrization of a population distribution. The decoder, using that signature together with Gaussian noise as a way to sample from the population, is also a great way to parametrize the sampling. While the overall formulation is similar to PerturbDiff, I think the multilateration idea and doing diffusion directly in this space are substantial contributions.

The theory backing the method is well-established, and the paper is well written. With the square-root map, normalized expression profiles are embedded into a hypersphere and the Fisher–Rao distance becomes the angular distance. This is quite intuitive to me. Using medoid-initialized farthest-first traversal to select anchors also seems well motivated, especially how Proposition 1 leads into this choice.

Critical weaknesses

The experiments are a bit weak. I would like to see some ablations on the main parts of the method, particularly whether stochastic population modelling meaningfully contributes performance improvements, and how much the method depends on pretrained gene embeddings. I would also want to see performance on larger scRNA-seq datasets if compute permits, e.g. PBMC.

A minor comment: Figure 3 is too small to be legible, especially the legend in Figure 3(d).

Questions for authors
  1. Could the authors explicitly describe how the candidate collection C\mathcal C in Equations (9)–(10) is constructed? From the implementation, my understanding is that each candidate comprises all retained cells for one training perturbation, pooled across batches, and that the selected anchors remain fixed during training. Is this correct? How should the anchor count be chosen when fewer than 100 training populations are available?

  2. Which grouping strategy was used for the reported experiments? Please clarify the discrepancy between the batch-based default described in Appendix D.1 and the random grouping specified in the released configurations, and explain how the modeled population variability should be interpreted when random partitions are used.

  3. Could the authors compare stochastic signature generation with a deterministic signature predictor while retaining the same stochastic cell decoder and decoder-training setup? This would help establish whether modeling multiple possible population distributions contributes beyond modeling variation among cells within a single predicted population.

  4. Beyond comparing different pretrained embedding sources, could the authors quantify the dependence on pretraining itself—for example, using embeddings learned from scratch or one-hot conditioning as controls? Please specify how unseen perturbation genes are handled so that these comparisons can be interpreted appropriately.

6 comments

We sincerely appreciate the reviewer’s careful reading of our manuscript and the thoughtful questions raised during the review. We are grateful that the reviewer engaged deeply with the central ideas of our work and highlighted several important questions regarding the modeling assumptions, geometric representation, and empirical validation of GeoPert. We believe that such constructive discussions are valuable for strengthening the work, and we address each concern in detail below.

Q1: How are the anchor candidate populations constructed, selected, and fixed during training, and how should the anchor number be chosen when fewer than 100 candidates are available?

Answer: Our goal is to model, for each perturbation, a distribution over the RKHS embeddings of its constituent subpopulations. Accordingly, candidates are constructed at the perturbation × batch level: each candidate comprises all retained cells from one training perturbation within one batch. Farthest-first iteratively selects the candidate whose distance to its nearest existing anchor is largest, favoring broad geometric coverage. This discourages redundant selection of nearby batch populations from the same perturbation, while allowing multiple batches to be selected when their distributions differ sufficiently.

The anchors are selected once and remain fixed, including their defining cell collections, throughout all training phases. They therefore provide a consistent reference system for population representations.

When fewer than 100 candidate populations are available, all candidates are used as anchors; the small candidate pool keeps the computational cost manageable. So actually, we use M=min⁡(Mdefault,∣C∣)M=\min(M_{\mathrm{default}},|\mathcal C|), with Mdefault=100M_{\mathrm{default}}=100. We apologize for not describing these details sufficiently clearly and will make them explicit in the camera-ready version if the paper is accepted.

Q2: Which grouping strategy was actually used in the reported experiments, and how should population variability be interpreted when random rather than batch-based grouping is used?

Answer: We thank the reviewer for requesting this clarification. We used batch-defined groups for K562, RPE1, HepG2, and Jurkat, and random groups for Norman. Although the original Norman experiment processed cells across eight 10x Chromium lanes, labeled as gemgroups 1–8, the GEARS-processed file used in our experiments does not retain the corresponding batch annotations. Technical batching can introduce variability within the same perturbation, as illustrated by our batch-resolved analysis. Such variability may remain in the pooled data even when batch annotations are unavailable. In this setting, we randomly partition cells within each perturbation to construct multiple empirical populations. Each population has its own RKHS mean embedding, yielding multiple population representations in place of a single pooled representation.

This procedure provides population-level augmentation for learning responses across different empirical realizations. Our Norman ablation supports the predictive benefit of this strategy fig9: performance improves as the number of groups increases from one to three, with three groups achieving the highest Pearson Delta and lowest RMSE. These results support modeling variability among empirical populations even without batch annotations.

Q3: Does stochastic population-signature generation provide benefits beyond modeling variation among cells within a single predicted population, compared with a deterministic signature predictor?

Answer: We appreciate the reviewer raising this important question about the necessity of meta-distribution modeling. We agree that stochastic population modeling is not strictly necessary in every setting. When only one replicate is available per perturbation, deterministic and stochastic signature prediction cannot be meaningfully distinguished from the observed data alone. However, when multiple replicates exist, their variability must be addressed rather than simply discarded through aggregation. GeoPert provides a unified way to characterize both the central tendency and variability of perturbation responses. Specifically, an empirical Fréchet mean minimizes the average Energy Distance to individual replicate distributions and represents their prototypical consensus response fig2a. Beyond this central estimate, stochastic sampling from the meta-distribution captures the range of possible population realizations fig3.

This distinction is particularly relevant when experimental conditions differ across batches or sequencing platforms. A deterministic predictor collapses each perturbation into one signature, whereas GeoPert can represent multiple possible population-level outcomes. Its architecture can also be extended to incorporate platform-specific conditioning, although suitable multi-platform perturbation benchmarks are currently lacking.

We recognize that estimating a reliable meta-distribution requires sufficient replicate observations, and current public perturbation datasets remain limited in this respect. To directly isolate the contribution of hierarchical modeling, we introduced a deterministic signature predictor while retaining the same stochastic decoder and training objective. Hierarchical modeling consistently improves Pearson Delta and RMSE across all four datasets fig2c.

Furthermore, the DDX47 and RPS28 analyses fig3 reveal substantial batch-associated variation, including occasional reversals in individual batches. Averaging multiple predicted population realizations recovers the predominant response directions across batches, demonstrating how GeoPert can extract more conservative perturbation effects while retaining population-level variability.

Reviewer_2uM5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In the manuscript, the authors proposed GeoPert, which models the perturbation prediction as a two-step generative process. Given a perturbation, it first samples the perturbed data distribution from a meta-distribution, and then generates the cell by sampling from the sampled perturbation distribution. The authors claim the benefit of this design is (1). modeling the stochasticity from batch effect and other factor that could cause different in observed data across experiments of the same perturbation, and (2). help with augmentation of the data, which improves the limited perturbation training data.

Critical strengths
  1. The combination of anchor-distance population signatures and a two-stage generative architecture is an interesting methodological contribution.
  2. The manuscript provides theoretical support for key components.
  3. The manuscript is generally well organized, and the high-level framework is clearly presented.
Critical weaknesses
  1. The motivation and empirical justification for hierarchical population modeling are insufficient.
  2. The modeling of perturbations and their relationships is insufficiently studied, despite its importance for generalization to unseen perturbations.
  3. The manuscript does not establish that the available training data, particularly the number of independent populations per perturbation, are sufficient to reliably learn the meta-distribution and support the two-stage generative process.
  4. The evaluation lack cell-level measurements.
Questions for authors
  1. GeoPert claim meta-distribution is necessary in modeling biological fluctuations and technical batch effects. However, the manuscript does not show that the data contain sufficient independent replicates to learn this meta-distribution. A comparison against direct conditional cell generation and deterministic population-signature prediction is needed.
  2. GeoPert doesn't put too much effort on the modeling of perturbations and their relationships. It relies on pretrained gene embeddings to condition response generation. The manuscript doesn't provide enough detail about the construction and biological information underlying the used embeddings, and how it compares with other embeddings. It also does not specify how embeddings are constructed for double-gene perturbations, making the mechanism for combinatorial generalization unclear.
  3. The authors should provide more robustness study on the hyper-parameters, especially anchor number and group sizes.
  4. The authors should use more evaluation metric that focus on cell within the population, rather than just population level evaluation.
  5. The authors state the importance of using augmentation to improve perturbation prediction training, and that their methods provides easy way for augmentaton. However, perturbation augmentation including augmenting with new perturbation and response, or augmenting with repeated measurement of the existing perturbation. The second case helps little in model generalization ability to new perturbations. The authors should provide more details on what they support in their augmentation, and where the performance gain comes from.
  6. As for baseline, the authors should include scLAMBDA and PerturbNet
8 comments

We would like to sincerely thank the reviewer for the thoughtful and constructive review. We particularly appreciate the reviewer’s insightful questions concerning the theoretical motivation, modeling choices, and empirical evaluation of GeoPert. These comments have provided us with an excellent opportunity to clarify the scope and contributions of our framework, as well as to further validate several of its key assumptions. We hope that the following responses will address the reviewer’s concerns and facilitate a constructive scientific discussion.

Q1: GeoPert claims meta-distribution is necessary in modeling biological fluctuations and technical batch effects but it is uncertain that the data contain sufficient independent replicates for learning this meta-distribution. A comparison against direct conditional cell generation and population signature prediction is needed.

Answer: We sincerely thank the reviewer for this insightful question, which addresses a fundamental issue in population-level single-cell modeling: how to represent multiple realizations of the same perturbation. Modeling population distributions as samples from a meta-distribution is a principled design choice rather than a strict necessity. Its key advantage is that it captures both conserved perturbation responses and inter-replicate variability.

From a machine learning perspective, deterministic signature prediction and stochastic signature generation are functionally indistinguishable when each perturbation is measured only once. However, when multiple replicates are available, simply aggregating them captures the consensus response but discards inter-replicate variability.

GeoPert addresses both aspects. It can estimate an empirical Fréchet mean (geometric barycenter) that minimizes the average Energy Distance to observed replicate distributions, providing a prototypical consensus response fig2a. Meanwhile, sampling from the learned meta-distribution generates multiple population-level realizations, explicitly characterizing inter-replicate variability fig3.

A deterministic predictor necessarily produces a single population signature and therefore cannot explicitly represent this variability. Moreover, for perturbation assays conducted across heterogeneous platforms (e.g., 10x versus BD, or 10x v3 versus v4), a single deterministic mapping may struggle to accommodate platform-specific variation. GeoPert's hierarchical formulation is naturally extensible to incorporating such technical covariates. Although current public benchmarks lack sufficient multi-platform perturbation data for direct validation, this provides an important direction for future work.

We acknowledge that the statistical confidence of meta-distribution estimation depends on the number of available replicates, which remains limited in current public datasets. To directly evaluate the benefit of hierarchical modeling, we compared GeoPert against a deterministic signature predictor while keeping the stochastic cell decoder and training objective unchanged. As shown in fig2c, hierarchical modeling consistently achieves higher Pearson Delta and lower RMSE across all four datasets.

Finally, fig3 illustrates batch-associated heterogeneity in the predicted responses to DDX47 and RPS28, including variations in effect magnitude and occasional directional reversals. Averaging predictions across sampled subpopulations recovers the predominant response directions shared across most batches, yielding more conservative and reproducible perturbation effects.

Q2: How are perturbation embeddings constructed and what biological information do they encode, particularly for double-gene perturbations and combinatorial generalization?

Answer: Our perturbation representations incorporate complementary biological priors. Language-based embeddings, such as GenePT, encode textual descriptions of gene function. Expression-derived representations capture gene dependencies: scGPT learns from large single-cell atlases, while Gene2Vec learns from coexpression contexts. UCE additionally incorporates protein-sequence information through ESM2 embeddings. Network-based representations, including STRING variants and our default KG-WaveGC, encode biological relationships through network structure. These approaches provide transferable information about gene function, expression dependencies, or functional associations. GeoPert keeps the pretrained descriptors frozen and learns a linear projection that adapts them to perturbation-conditioned generation.

Our existing ablation already compares six alternative embeddings across all four datasets fig8. Although individual metrics vary and network-based representations often perform best, the overall performance profiles remain broadly comparable across different sources. These results support GeoPert’s robustness to the choice of pretrained representation and indicate that its performance is not tied to one particular embedding.

For double-gene perturbations, we sum the two pretrained gene embeddings and apply the same trainable linear projection used for single-gene perturbations. This permutation-invariant construction allows unseen combinations to be represented using their constituent genes, without requiring a separately pretrained representation for each pair. Summing the conditioning embeddings does not impose additive transcriptional responses, because the downstream conditional generative mapping is nonlinear.

Q3: How robust is GeoPert to the choices of key hyperparameters, particularly the number of anchors and group size? We have added sensitivity analyses for both anchor number and population grouping.

Answer: Anchor number. Prediction performance generally improves as the anchor set expands from very small sizes, with diminishing gains and dataset-dependent fluctuations around 100–150 anchors fig9a. The distance-preservation analysis shows a similar pattern: Spearman correlations between original RKHS distances and distances between anchor representations typically reach high values with approximately 100 anchors, followed by saturation or gradual declines as more anchors are added fig1. Together, these results support 100 anchors as a practical default that balances geometric fidelity, predictive performance, and representation size.

Group number and size. We use observed batch labels to define groups in K562, RPE1, Jurkat, and HepG2; random grouping is used only for Norman. We therefore evaluate the number of random groups on Norman, where three groups achieve the highest Pearson Delta and lowest RMSE among the tested settings fig9b. Too few groups provide limited variation among empirical population realizations, whereas too many divide the fixed cell pool into smaller groups, increasing sampling noise in distribution estimates and anchor-distance signatures. These results favor approximately three groups in this setting, balancing population-level variation with reliable distribution estimation.

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

This paper proposes GeoPert for predicting single-cell responses to unseen genetic perturbations. It first generates a condition-specific population distribution and then samples cells from it, capturing replicate-level variability. Populations are encoded as Fisher-Rao energy distances to anchor populations, which is shown to preserve pairwise distances up to the anchor covering radius. A conditional diffusion model generates these signatures from gene embeddings, and a decoder generates cells from them with a Fisher-Rao MMD loss, with synthetic cell mixtures used for augmentation. GeoPert outperforms PerturbDiff, GEARS, scGPT, STATE and linear baselines on four Perturb-seq cell lines and the Norman dataset.

Critical strengths

The problem formulation is interesting. By explicitly modeling variability across cell populations as a distribution over population distributions, the paper goes beyond matching a single pooled distribution for each perturbation. The anchor-distance representation makes this formulation concrete and enables a two-stage generative framework with population-level data augmentation.

Critical weaknesses
  1. The Fisher–Rao loss ablation shows that it improves performance within GeoPert. However, it remains unclear how much of GeoPert’s advantage over existing methods comes from this choice of loss. Where applicable, equipping a strong baseline such as PerturbDiff or STATE with the same Fisher–Rao MMD loss would help assess the additional benefit of GeoPert’s hierarchical framework.

  2. The benefit of explicitly modeling a distribution over population distributions is not directly isolated. A useful control would predict a single deterministic population signature for each perturbation while retaining the stochastic cell decoder and the same training loss. This would help establish whether population-level stochasticity provides benefits beyond the representation and cell-level generation.

Questions for authors

see weakness

3 comments

We sincerely thank the reviewer for the careful and constructive evaluation of our work. We greatly appreciate the insightful comments and thoughtful questions, which have helped us clarify several important aspects of GeoPert and identify directions for further improvement. We value this opportunity for constructive discussion and hope that our responses below can address the reviewer’s concerns and further clarify the motivation, methodology, and empirical findings of our work.

Q1: How much of GeoPert’s improvement over existing methods comes from the Fisher–Rao loss itself versus the proposed hierarchical framework?

Answer: Euclidean distance on log-normalized expression emphasizes relative changes and strongly compresses quantitative differences at higher expression levels. For a small expression increment δ≪X\delta\ll X, the transformed differences approximately satisfy $\Delta_{\log}\approx \frac{\delta}{X+\epsilon}, \qquad \Delta_{\sqrt{\cdot}}\approx \frac{\delta}{2\sqrt{X}}, $ where XX is the baseline expression and ϵ>0\epsilon>0 is the log pseudocount. The slower decay of square-root differences preserves greater relative sensitivity to graded expression changes while remaining sensitive to on/off differences. Fisher–Rao further measures geodesic separation between normalized profiles pp and qq: $d_{\mathrm{FR}}(p,q) = 2\arccos\left(\sum_g\sqrt{p_gq_g}\right). $

We have incorporated the Fisher-Rao MMD loss into both STATE and PerturbDiff. Both models generally improve across the four datasets fig7, indicating that the benefits of Fisher-Rao geometry extend beyond GeoPert. GeoPert nevertheless retains higher PR-AUC and ROC-AUC than both Fisher-Rao variants across all four datasets fig4. These results suggest that the Fisher–Rao objective contributes to predictive performance but does not fully explain GeoPert’s advantage, supporting the additional value of its hierarchical modeling framework.

Q2: Does explicitly modeling a distribution over population distributions provide benefits beyond a deterministic population-signature predictor with the same stochastic cell decoder and training objective?

Answer: We thank the reviewer for highlighting the importance of validating the proposed meta-distribution formulation. We view stochastic population-signature generation as a principled modeling choice for representing inter-replicate variability, rather than a universally necessary assumption. While deterministic prediction may suffice when only one replicate is available, aggregating multiple replicates into a single distribution inevitably loses information about their variability.

GeoPert addresses this limitation by modeling both the consensus and dispersion of population responses. The empirical Fréchet mean provides a representative consensus distribution by minimizing the average Energy Distance to individual replicates fig2a, while multiple samples from the learned meta-distribution characterize the variability among population realizations fig3.

To experimentally assess whether this hierarchical formulation provides benefits beyond cell-level stochasticity, we constructed a deterministic signature prediction baseline, keeping the stochastic decoder and its training objective unchanged. The results in fig2c show consistent improvements in both Pearson Delta and RMSE across RPE1, K562, HepG2, and Jurkat, supporting the additional value of population-level stochastic modeling.

We further examined the biological implications using DDX47 and RPS28 fig3. Their predicted responses vary across experimental batches in both magnitude and direction. Importantly, averaging predictions across sampled subpopulations recovers the predominant response directions shared across most batches, providing a more conservative estimate of perturbation effects.

We nevertheless acknowledge that reliable estimation of inter-replicate variability depends on replicate availability, which remains a limitation of current public datasets. Moreover, the hierarchical architecture offers a natural extension for incorporating platform-specific technical variation, although this capability awaits validation on sufficiently large multi-platform perturbation atlases.

I thank the authors for their rebuttals with responses my question and additional experiments. I will maintain my original positive score.

Reviewer_8PzyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes GeoPert, a two-stage hierarchical generative approach for predicting single-cell responses to unseen genetic perturbations. Its central premise is that a perturbation does not induce one fixed distribution, but a meta-distribution over population distributions. It uses Fisher–Rao geometry and anchor-based representations to characterize cell populations, followed by conditional diffusion for population generation and a decoder for cell generation. The authors report strong performance on unseen single gene perturbations and consistent improvements over several baselines on the Norman double-gene perturbation benchmark.

The paper has a correct mathematical construction and an interesting generative architecture. However, the provided evidence is not sufficient for me to recommend acceptance.

Critical strengths
  1. The paper presents an interesting approach to representing cell-population distributions using Fisher–Rao geometry and anchor-based multilateration. This provides a finite-dimensional representation of population distributions that can be used for generative modeling.
  2. The theoretical results on distance preservation are mathematically sound under the stated assumptions. In particular, Proposition 1 provides a bound on the representation distortion in terms of the anchor covering radius.
  3. The experimental results are promising, particularly for unseen double-gene perturbations. GeoPert outperforms the baselines across all three settings, demonstrating potential for predicting responses.
  4. The two-stage architecture separates population-level distribution generation and individual cell generation. This design enables synthetic population augmentation. Ablation experiments suggest that incorporating a proportion of synthetic populations can improve prediction performance.
Critical weaknesses
  1. The paper is motivated by variability across population distributions under the same experimental protocol. The guarantee in Proposition 1 is informative only when the selected anchors sufficiently cover the relevant distribution space. The authors use 100 anchors, without showing that 100 is enough to cover unseen perturbations. If population distributions vary substantially, as the paper motivates, why should such a limited anchor set preserve their geometry?
  2. The paper uses random cell groups when experimental batch labels are unavailable (Appendix D.1). These groups may capture finite-sample variability rather than genuine differences between biological replicates. Ablation studies do not clearly isolate the contribution of hierarchical modeling. Could they compare the proposed method against a deterministic population-level model while keeping the other components unchanged?
  3. The paper does not discuss optimal transport approaches such as CellOT [1], which learn mappings between unpaired cell populations despite destructive sequencing. When the perturbation mapping is recovered, learning the effect of intervention is possible under some assumptions [2]. How does the paper compare with optimal transport approaches? And under what conditions is the complexity of the 100-anchor representation necessary?
  4. The use of Fisher–Rao geometry is an interesting contribution. However, the paper does not compare it with Euclidean distance computed after square-root.
  5. Proposition 1 provides a distortion bound using the infinity norm, while the empirical evaluation reports Euclidean distances between anchor representations. Could the authors report the actual distortion using the infinity norm? It would be very interesting to provide a more direct empirical validation of Proposition 1.

[1] Bunne et al. (2023). Learning single-cell perturbation responses using neural optimal transport. Nature Methods. [2] Jafari et al. (2026). Temporal Measurement Interval Determines Computational and Model Complexity in Single-Cell Perturbation Analysis. ICML 2026.

Questions for authors

Please address the questions in the weaknesses section.

7 comments

We appreciate the constructive comments from the reviewer and we would like to thank the reviewer for his/her careful review. The reviewer is keen at grasping and summarizing our primary contributions and the reviewer raised very insightful and interesting questions, some of which are very mind probing. We believe such healthy discussions are what drives the field forward.

Here is our attempt to constructively exchange our opinion with the reviewer and hopefully together we can find a path towards better modeling of the complex single cell transcriptomic data and how to decipher the complex cellular machinary in presence of gene expression perturbation.

Q1, Is 100 anchor enough and how can it accuractely represent an unseen distribution of arbitrary geometry?

Answer: This is a question of great depth. Retaining full resolution to represent any distribution with finite dimensions while retaining their comparative relationship is not any easy matter. Just like any data-driven machine learning task, we believe fundamentally, cellular responses to gene expression perturbations are of low rank. That means a cell, while highly polymorphic, would converge to a small set of homeostatic states, at least on a macro, landscape-wide perspective, under perturbance from external or internal environments (after sufficient time lapse). This forms the philosophical basis for predicting cellular responses to unseen perturbations: that is, while the exact observation of a gene expression perturbation is unseen, we have seen the response to a perturbation of similar effect, from which predictions are generated.

From this perspective, we can formulate the problem as follows: there is a finite, likely multi-modal domain in the transcriptomic space that describes all possible homeostatic cell states, with each model representing one homeostatic cell state. Gene expression perturbations, as well as unperturbed basal state, would drive cells into some of the many finite-sized distribution patterns among these homeostatic states. This might be a bit abstract, but in short, we would assume all perturbation distributions collectively constitute the cellular homeostatic landscape and gene perturbation is essentially a conditioned random sampling from this landscape.

In this setting, the distribution of cells after a specific perturbation, each with its own distinct geometry, is thus a random sampling from a meta distribution that characterizes all possible homeostatic distributions with or without internal or external stimulation. Whether we can accurately predict the cell distribution of a specific purtebation effect, depends on whether we have sufficient training data (perturbation distributions) that help us map the entire landscape. In this sense, if the training data is biased and there is an entire area of the meta distribution completely obscured, i.e., the testing is complete out of distribution of the training data, then it would not be possible for GeoPert, or any other machine learning algorithms in this sense, to correctly recover the geometry of the out-of-distribution perturbation geometry.

When the test data is not out of distribution, i.e., we can assume that for every unseen perturbation, there is a seen perturbation that shares similar distribution in the homeostatic landscape, then by Proposition 1, we can still accurately characterize this unseen distribution with moderately high, albeit slightly eroded resolution. This is because proposition 1 guarantees that the anchor distance vectors can characterize the distance relationship between two distributions up to the maximum to-nearest-anchor radius ϵtraining\epsilon_{training}. For an unseen distribution, while its distance to the nearest anchor might be greater than ϵtraining\epsilon_{training}, due to it shares the same distribution with the training data, we have ϵtesting≈ϵtraining\epsilon_{testing} \approx \epsilon_{training}. Therefore, as long as the test unseen perturbation is not completely OOD, the multilateration scheme can retain relatively high resolution with the training-set selected anchors.

To demonstrate this empirically, we evaluated the fidelity of the anchor representation across different numbers of anchors fig1. In both the training and test perturbations, the Spearman correlation between pairwise RKHS distances and the corresponding distances between anchor-distance vectors increases rapidly with the number of anchors and then saturates or fluctuates. Across the four cell lines, approximately 100 anchors already provide high distance fidelity, with limited gains from substantially larger anchor sets.

We further directly evaluated the ℓ∞\ell_\infty distortion predicted by Proposition 1 fig6a. For the training populations, the observed distortion remains within the theoretical reference 2ϵtrain2\epsilon_{\mathrm{train}} across anchor-set sizes, providing direct empirical support for the proposed bound. Importantly, when the same training-selected anchors are applied to unseen test perturbations, most population pairs also remain within the training-derived distortion reference, although a small fraction exceed it. This indicates that the anchor representation can preserve relative population geometry beyond the populations used to construct the anchors, consistent with the assumption that unseen perturbations remain within the covered response landscape.

When the unseen perturbation is completely out of distribution, similar to any other machine learning tasks such as natural language processing and computer vision, it would be extremely challenging for GeoPert to accurately characterize the unseen distribution. However, this challenge stems from the prediction being out of distribution rather than insufficient number of anchors. Again, by proposition 1, in this scenario, in order to retain sufficient representation resolution, there would need to be a pre-compueted synthetic distribution that serves as a close-by anchor to the OOD distribution which is not grounded to any real distributions in the training data. In this case, even if all training data are selected as anchors, $\epsilon_{testing} $remains high and the prediction would remain inaccurate. To demonstrate this, we constructed test populations that are substantially different from the training perturbation distributions fig6b. Although increasing the number of anchors continuously reduces the training coverage radius, the distance distortion of these OOD populations remains substantial and exceeds the training-derived reference even with a large number of anchors. This result suggests that adding more anchors cannot compensate for a test distribution that lies outside the geometric coverage of the training perturbation landscape. Thus, the limiting factor for OOD populations is distributional coverage rather than the absolute number of anchors.

Nonetheless, out-of-distribution generalization remains a fundamental yet notoriously challenging problem in machine learning and we seek to further investigate its manifestation in perturbation prediction in our future work.

Q2:Do random cell groups capture genuine population-level variability rather than finite-sample variation, and does hierarchical modeling provide benefits beyond a deterministic population-level model?

Answer: Thank you for this deeply insightful question, which goes to the core of population-level single-cell modeling: how to represent multiple replicates of a given condition. Modeling the cell population distribution itself as a realization of a meta-distribution is indeed a principled conceptual choice rather than a strict necessity. Beyond preserving inter-replicate variability, this hierarchical formulation provides the distinct advantage of recovering cross-population prototypical responses, capturing the conserved biological effects shared across replicates, while quantifying replicate-level uncertainty. At the same time, we fully agree with the reviewer’s astute observation that our statistical confidence in characterizing this meta-distribution is inherently tied to the number of available replicate populations in the dataset.

From a purely machine learning perspective, choosing between deterministic mapping from a perturbation to a single distributional signature versus modeling signature generation as a stochastic process is strictly a technical design choice. In fact, the two paradigms are functionally indistinguishable when each perturbation is assayed only once without biological replicates. However, whenever multiple replicates are available, even in small numbers, one must deliberately decide how to handle inter-replicate variability. One standard strategy is to first aggregate profiles across replicates and model only the conserved, consensus response. This approach, however, captures only the mean of perturbation effects and ignores the variances. Specifically, to synthesize a prototypical response representative of all replicates, GeoPert can estimate an empirical Fréchet mean (geometric barycenter) that minimizes the average Energy Distance to individual replicate distributions fig2a. Simultaneously, to complement this central tendency, GeoPert characterizes inter-replicate variance by drawing multiple stochastic realizations from the learned meta-distribution, quantifying sample-level dispersion around the conserved consensus at a meta-distribution level fig3.

From a generative standpoint, predicting the outcome of an individual perturbation assay via a deterministic direct mapping is overly restrictive, as it invariably collapses predictions onto a single Fréchet mean distribution. Furthermore, when perturbation assays are profiled across heterogeneous technologies, such as 10x Genomics versus BD platforms, or 10x v3 versus v4 chemistries, a direct mapping struggles to reconcile platform-specific technical variations. In contrast, GeoPert's meta-distribution formulation naturally accommodates these technical covariates by conditioning generation on platform-specific priors. Although current public benchmarks lack the multi-platform perturbation coverage required to evaluate this capability empirically, GeoPert's architecture is inherently extensible and future-proof as cross-platform atlases emerge.

Finally, we fully agree with the reviewer's observation: while GeoPert's architecture is designed to construct a meta-distribution capturing variation across replicates, the fidelity and statistical confidence of this estimation are directly governed by the abundance of observed replicates. As the reviewer astutely noted, given the limited replicate depth in current public perturbation datasets, it remains challenging to empirically showcase the full advantages of modeling inter-replicate meta-distributions. However, as large-scale, multi-donor, and multi-platform perturbation atlases rapidly expand, we believe GeoPert will provide a principled, scalable foundation for quantifying biological and technical uncertainty, disentangling conserved perturbation responses from batch variability, and enabling reliable in silico screening at atlas scale in the future.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.