acceptodds
Under review as a conference paper at ICLR 2027

Decoupling Baseline from Condition in Single-Cell Perturbation Data with SplitDGD

Abstract

Decoupling the representation of normal cellular states from condition-specific perturbation responses is a central modelling problem in single-cell biology and beyond. Achieving this decoupling without deteriorating the representation’s utility for reconstruction and downstream analyses is a non-trivial trade-off. To address this, existing methods build on variational inference with soft regularisers or structural constraints on the latent, and we show that both strategies leak condition information into the invariant component or degrade the utility of the representation. We propose splitDGD, an extension of the Deep Generative Decoder framework (Schuster & Krogh, 2023) that enforces decoupling architecturally. This is obtained by constructing two separate latent blocks—an invariant one encoding condition-independent variations (e.g., of cell types) and a condition-specific one capturing the perturbation response—and by masking the latter when the perturbation is absent. Across controlled image benchmarks and single-cell perturbation datasets, in both unsupervised and label-supervised regimes, splitDGD is the only method to mitigate the leakage-vs-utility tradeoff that other models exhibit.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Decoupling, Single cell perturbation, Leakage, Representation learning

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_5DHyOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

First of all, I thank the authors for taking their time to write the paper. The paper tackles the problem of disentanglement in a regime with binary labels (the default regime), where one is provided observations with labels, and the task is to learn shared and condition-specific representations. Here, the former represents variation in the absence of the label, and the latter represents the portion that is correlated with changes caused by the label (e.g. cell types vs perturbations).

The paper is nicely written, and tackles the literature effectively by starting off with what most methods do: use a Variational Inference (VI) framework with additional regularizations (e.g. adversarial classifier on shared representations), prior specifications (e.g. an empirical prior) etc. In contrast to previous work, the paper extends upon the DGD framework and stands out in three aspects: MAP estimation for training, explicit masking of the conditional embedding when unperturbed, and using an empirical-like prior restricted only to the unperturbed class for the shared embeddings. Importantly, this considers both the supervised disentanglement problem in general and the specific setting it is in (i.e. perturbations), where one of the classes is a biological baseline.

In this sense, I found many parts of this manuscript particularly interesting which I outline below. Overall, I found the work well-motivated, and the approach justified both methodologically and on the application side. The remaining concerns are mostly on the empirical front; which, to be clear, is highly thorough and communicates the applicability of SplitDGD. The comments will mainly focus on specifics, especially considering the precise biological setting the model is framed in.

Critical strengths
  • The hard masking mechanism + MAP approach resolves the optimization ambiguity explicitly. Since the introduction of MultiGroupVI (Weinberger et al, 2022 ; which can be traced back to Abid & Zou, 2019), most methods operating over the VI framework used a similar idea for conditional embeddings, where the prior for unperturbed samples was set to be a delta at zero (see details of referenced paper) which was enforced throughout optimization with an additional KL term. Here, there is no encoder, and the masking gives no updates to the embedding, which essentially fixes this as part of the optimization

  • The prior term (i.e. alternate prior training) is very interesting, in that it follows recent previous ideas related to empirical priors but adapts it nicely to the perturbation setting. The prior terms (which for VI is also relevant for hard masking) are choices at the end of the day, and recent work by Slavutsky et al. 2025 (DisCoVR) has shown cases where hand-crafted priors can be harder to accurately specify in certain settings. They instead proposed an empirical prior that averaged shared embeddings per class for the conditional embeddings; and a standard normal prior for the shared embeddings (as this paper also notes). SplitDGD carries a similar idea, but with the biological context we know what the baseline is, hence it departs from the context in DisCoVR in the sense that the shared embedding is not what remains when one marginalizes out the label, but it is defined relative to what the expected baseline is (which for this context we know should be the unperturbed samples). This idea drives the optimization term for the prior, which I found quite fitting.

Critical weaknesses
  • Comparisons with competing methods are clear, but ablations on the different components would add to the story. Personally, I am curious as to which of the three ingredients (MAP estimation, and the other two mentioned in strengths) add to the gain here. I think this would also be helpful for future researchers to decide on what avenues of methodology they should pursue for similar questions. Ideally, one would have an ablation on every component starting from the base DGD framework (and also at the end maybe reverting partially to VI to see if MAP / DGD give the benefits), but given time constraints, I leave it to the authors’ judgement on what would be best.

  • For specific biological problems (e.g. counterfactual prediction) the baselines might be slightly misspecified, with some additional considerations required per method. For counterfactual prediction, it might be better to compare to a commonly used method (e.g. CPA, scGen etc.) as certain baselines (specifically DisCoVR, and as far as I can tell contrastiveVI) were neither originally designed nor evaluated for counterfactual prediction, although the procedure here has been previously used (starting with scGen, which used VAE + mean shifts in a similar way). Also, for VI based methods that use external components (e.g. adversarial classifiers for DisCoVR), it is quite often the case that hyperparameters are hard to tune. I am wondering if any hyperparameter tuning was done for these methods (outside of the parameters specified by authors themselves).

Questions for authors
  • Would it be possible to have some ablations regarding critical components of the model (as specified under weaknesses)? Particularly, finding out where the maximal gain is out of the three components would be very much appreciated.

  • Is it possible to compare with additional models that are specifically targeted for counterfactuals (Biolord / CPA)?

  • Would SplitDGD using a Gaussian likelihood vs baselines modeling counts directly change any of the results substantially?

  • (MINOR) Here are some typos I noticed: "arhe" (Sec. 2), "following the in protocol" (Sec. 4.2), "proovable" (Sec. 5), "gradients acts" (Fig. 1 caption), "an held-out", and a missing space in "Table 13.On". The image dataset is called "DF-MNIST" in the text and tables, "DigitFashion-MNIST" in Appendix B, and "Fashion-MNIST" in the Fig. 2 and Fig. 3 legends.

2 comments

We thank the reviewer for the careful reading and for the positive comments on the hard mask combined with MAP estimation and on alternate prior training.

W1/Q1 - Ablations. We ran the ablation on Haber, in both w-regimes, with the splits, the 10 seeds and the per-seed hyperparameters of the paper's runs (all six metrics are in the table of our answer to Reviewer RXQW). Following the reviewer's suggestion, the table below starts from a two-latent DGD with neither mechanism and adds them one at a time.

The largest gain comes from the hard mask, which lowers linear leakage by 0.05-0.13 and raises z-MCC by 0.23-0.32. The control-only prior adds little once the mask is in place (at most 0.008 on either metric) and more without it (z-MCC +0.03 and +0.10), so it acts as a second safeguard. Fidelity is unaffected: across the four variants RMSE stays within 0.001 of splitDGD and counterfactual R2R^2 within 0.008 points.

reverting partially to VI to see if MAP / DGD give the benefits

To compare MAP estimation of free latents to amortised variational inference, this is not directly possible in the DGD framework. Thus, we contrast splitDGD to contrastiveVI, which is the closest counterpart to splitDGD as it applies the same zeroing of the condition block within a VAE (the contrastive-analysis lineage the reviewer mentions). The results are already in Table 2 of the manuscript. On Kang and Haber its z has the highest linear leakage of all methods (0.70 and 0.68, against 0.52 and 0.55 for splitDGD), and its reconstruction error and counterfactual R2R^2 are worse than splitDGD's on all three single-cell datasets, while it recovers the perturbation subtypes on Haber better (w-MCC 0.84 against 0.60).

W2 / Q2 - Baselines comparison. For every method, splitDGD included, the counterfactual of a control cell is obtained by decoding its invariant latent with the mean condition-specific latent, the mean-shift scheme that, as the reviewer notes, goes back to scGen. Among our baselines, scDisInFact was designed for perturbation prediction, and splitDGD is ahead of it on every counterfactual entry (0.960 vs 0.871 on Kang, 0.926 vs 0.824 on Haber, 0.953 vs 0.872 on sci-Plex). As suggested, we have also run biolord (setup and full table in our answer to Reviewer wHDC, W1b): its counterfactual R2R^2 on Haber is 0.64 (unsup-w) and 0.58 (sup-w), below every method in Table 2 (at least 0.82). We did not add CPA. Like biolord, it takes the perturbation label and the cell type as inputs and has no per-cell perturbation representation, so it cannot be scored on the decoupling criteria that are the subject of the paper, for which counterfactual prediction is one of four criteria and not the target task.

I am wondering if any hyperparameter tuning was done for these methods

ContrastiveVI, DisCoVR and scDisInFact were tuned per seed on the single-cell datasets with the same multi-objective search as splitDGD (200 trials on Kang and Haber, 60 on sci-Plex), over spaces built around the authors' defaults. For DisCoVR the search includes the adversarial weight (1 to 100), the reconstruction weight, the two KL weights, the latent dimension and the learning rate. The DisCoVR paragraph of the Appendix described the authors' defaults; it was out of date and we will correct it.

Q3 - Modeling counts vs. Gaussian. We ran this comparison on Haber, training splitDGD with a negative-binomial likelihood on the raw counts and keeping everything else as in the paper's runs (splits, 10 seeds, and the per-seed hyperparameters, which were not re-tuned for the new likelihood). The table gives, for each w-regime, the paper's values and those obtained with counts (RMSE and counterfactual R2R^2 in log-normalised space).

The results do not change substantially. Leakage and reconstruction are the same within seed variability, so splitDGD keeps the lowest linear leakage and the lowest RMSE of all methods in both regimes. z-MCC and counterfactual R2R^2 are slightly worse with counts; w-MCC in unsup-w is also lower on average, within its seed variability. We have added this ablation in a new section of the Appendix.

Q4 - Typos.

Here are some typos I noticed

Thank you for spotting them. We have applied those corrections in the revised version of the manuscript.

Thanks for providing the missing tables (an edit to the original response might be needed to make the tables visible for others). I remain positive about the work and have updated my soundness evaluation accordingly.

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

This paper addresses the problem of single-cell perturbations by decomposing the cellular response into two components: zz, which captures the intrinsic state of the cell (e.g., cell type), and ww, which represents the changes induced by a treatment (e.g., the response to interferon stimulation). The objective is to ensure that zz does not reveal whether a cell has been treated, while still allowing for the identification of its cell type. The proposed method, splitDGD, dispenses with an encoder; instead, it assigns a unique zz and ww vector to each cell—treated as learnable parameters—and employs a shared decoder to map these representations back to gene expression profiles. During inference, the decoder can be used by combining the zz vector of a control cell with the average ww vector derived from treated cells of the same type. The authors evaluate splitDGD against four existing methods across two image datasets and three single-cell datasets, demonstrating that splitDGD achieves the lowest leakage of treatment-related information into the zz representation.

Critical strengths
  • The problem definition is clear. The authors establish four verifiable criteria for an effective decomposition: zz should not reveal the treatment condition, zz should identify the cell type, ww should identify the treatment type, and the model should be capable of reconstructing expression profiles and making predictions.
  • The training procedure is straightforward, requiring no additional adversarial training or auxiliary classifiers.
Critical weaknesses
  • SplitDGD essentially combines the DGD framework with the zeroing-out strategy used in contrastive analysis; a similar dual-representation structure already exists in multiDGD. Furthermore, the study fails to compare against biolord, another single-cell method that operates without an encoder.
  • While the authors claim that the separation is guaranteed by the model architecture, this guarantee effectively applies only to the control group. For treated cells, the zz vector is merely constrained by a Gaussian Mixture Model (GMM), which provides a relatively loose regularization. When evaluated using more robust classifiers, treatment-related information can still be detected within zz (achieving an accuracy of approximately 0.7 on the Kang and Haber datasets). On the Sci-Plex3 dataset, the unsupervised version of splitDGD exhibits even greater information leakage than a standard CVAE that lacks any explicit separation mechanism.
  • The ablation studies are incomplete. Given that the GMM is fitted exclusively on control data, it remains unclear whether the GMM is actually necessary. Additionally, although the paper repeatedly asserts that the encoder is the primary source of information leakage, it does not investigate the impact of including an encoder in the model.
Questions for authors

See my Weaknesses.

3 comments

Part 1/2

We thank the reviewer for appreciating the clear problem definition, with its four verifiable criteria, and the simplicity of the training procedure, which needs no adversarial or auxiliary components.

W1a - Novelty. We agree that both ingredients of splitDGD have precedents, and we credit them in the paper.

From multiDGD, splitDGD takes the supervised covariate scheme, used for the supervised-w block (Sec. 4). multiDGD's covariate representation is however symmetric: every cell, in every covariate class, has one, placed under the mixture component of its label. splitDGD is instead built around the control group: control cells have no w (the decoder sees [z;c⋅w][z; c \cdot w]), so baseline variation must be explained by z alone, and the z-prior is fitted on controls only. This also allows an unsupervised-w regime that needs only the binary condition, which multiDGD's labelled scheme does not cover.

The absence of w for controls is the zeroing of the condition-specific block used in contrastive analysis. It is new to the DGD framework and, together with the control-only prior, it is what prevents the leakage: a two-latent DGD without these two mechanisms (our new ablation on Haber, see our answer to Reviewer RXQW) leaks in the unsup-w case (linear AUC 0.68, against 0.55), and in both regimes its z retains less of the invariant factor (z-MCC 0.44 and 0.41, against 0.73 and 0.74).

W1b - Comparison to biolord. We thank the reviewer for the suggestion. biolord is indeed encoder-free, but we did not consider it a like-for-like comparison for the decomposition studied here, for three reasons:

(i) biolord learns its per-cell latent only for training cells, as free parameters, and defines no way to obtain it for test cells; evaluating it on held-out cells requires an additional inference step. The same limitation led PerturBench (Wu et al., 2025) to replace biolord's per-cell latent with an encoder in its benchmark implementation.

(ii) biolord has no per-cell perturbation representation: a perturbation enters only as an embedding of its label, shared by all cells with that label, so perturbation-structure recovery (C3) does not apply to it.

(iii) in its published sci-Plex3 experiment, biolord receives the cell type as a known attribute, whereas in our setting no method receives cell-type labels, since recovering them from z is criterion C2.

Nonetheless, we evaluated biolord on Haber with the same splits, input and tuning budget as the other baselines, giving it only the perturbation label. Its latent for every cell was inferred by optimising biolord's own training objective with the trained network fixed, similarly to what splitDGD does for test cells; counterfactuals use biolord's own prediction (source latent + target perturbation embedding).

Haber, mean (std) over 10 seedsAUC linAUC NLz-MCCRMSECF R2R^2
unsup-wbiolord0.599 (.051)0.870 (.117)0.710 (.075)0.249 (.033)0.636 (.180)
splitDGD0.550 (.021)0.694 (.029)0.730 (.044)0.198 (.007)0.885 (.044)
sup-wbiolord0.632 (.149)0.884 (.147)0.573 (.217)0.257 (.017)0.575 (.106)
splitDGD (sup-w)0.532 (.026)0.687 (.032)0.740 (.023)0.196 (.005)0.926 (.009)

In both settings biolord has the highest nonlinear leakage of all methods (0.87-0.88, against at most 0.77 for every other method on Haber), the highest reconstruction error, and the lowest counterfactual R2R^2 (0.64 and 0.58, against at least 0.82), with lower z-MCC than splitDGD. We have added this analysis of biolord to the revised manuscript, in a dedicated section of the Appendix.

Part 2/2

W2 - guarantees of separation. We acknowledge that our wording can be read as a guarantee for every cell, and we will make it more precise in the Introduction and Conclusions. splitDGD enforces the separation by construction: control cells have no w, so the decoder must explain baseline variation through z alone, under a z-prior defined by controls; treated cells share both, and the condition-specific part is left to w. Like any other model, this does not guarantee a condition-free z for treated cells, but it measurably reduces leakage on all cells: removing the mask raises the linear AUC of the condition from z on Haber from 0.55 to 0.62 (0.68 without both mechanisms; see our answer to Reviewer RXQW). Across datasets, this gives the best leakage-utility trade-off of the methods we compare.

Treatment-related information can still be detected within (achieving an accuracy of approximately 0.7)

We agree that a nonlinear probe still detects the condition in z, as it does for every method we compare. These values (nonlinear-probe AUCs, not accuracies) are among the lowest of all methods: on Haber the lowest (0.694 and 0.687), on Kang second to scDisInFact. Part of the signal is also expected, since z must encode cell type and cell type alone already predicts the condition with AUC 0.53 on Kang and 0.57 on Haber, because cell-type proportions slightly differ between control and treated cells. On the image datasets, splitDGD's leakage is at chance for the linear probe (0.50-0.52) and close to it for the nonlinear one (0.54-0.62), the lowest of all methods.

On the Sci-Plex3 dataset, the unsupervised version of splitDGD exhibits even greater information leakage than a standard CVAE

On Sci-Plex3 without compound labels, splitDGD's linear leakage is lower than CVAE's (0.591 vs 0.613) and its nonlinear leakage slightly higher (0.633 vs 0.615), in a range of 0.58 to 0.64 across all methods. With 188 compounds and no labels, no unsupervised method organises compound identity in w (w-MCC about 0 for all), which we believe limits how much of the compound signal w can absorb; with compound labels, splitDGD's leakage drops to 0.555 (linear, the lowest of all methods) and 0.587 (nonlinear).

W3 - Ablations. On the z-prior, we ran the ablation on Haber (ref to table in our answer to Reviewer RXQW). With the mask in place, fitting the z-prior on all cells instead of controls only has a minor effect: linear AUC rises by 0.005 and 0.002 and z-MCC falls by 0.008 and 0.007 in the two w-regimes. It matters as a safeguard once the mask is removed, where dropping it as well lowers z-MCC further (0.467 to 0.438, unsup-w; 0.510 to 0.411, sup-w). The mask is what carries most of the decoupling. We have added this ablation analysis to the revised manuscript, in a dedicated section of the Appendix.

it does not investigate the impact of including an encoder in the model.

The closest state-of-the-art counterpart of splitDGD with an encoder is contrastiveVI, which is part of our comparison. It has the same split latent, with the condition-specific block zeroed for controls, but both blocks are produced by encoders. As reported in Table 2 of the paper, on Kang and Haber its z has the highest linear leakage of all methods (0.70 and 0.68, at the level of CVAE, which has no separation mechanism, against 0.52 and 0.55 for splitDGD), and it has a higher reconstruction error and a lower counterfactual R2R^2 than splitDGD on all three single-cell datasets; its strength is the recovery of the perturbation subtypes on Haber (w-MCC 0.84), which comes with that leakage. These results indicate a rise in leakage, which was already reported for this model family (Weinberger et al., 2022), and agree on the advantage of latent optimisation over amortised inference, similarly to Gabbay & Hoshen (2020).

Thank you for the detailed responses and additional experiments. I appreciate the authors' efforts, particularly the biolord comparison and ablation studies. While these additions address some of my concerns, I remain unconvinced by the extent of the architectural decoupling. I therefore maintain my original rating.

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

This paper proposes splitDGD, a split-latent extension of the encoder-free Deep Generative Decoder (DGD) framework. Instead of relying on loss-level regularization, splitDGD separates invariant content from perturbation responses through its architecture. Across two controlled image benchmarks and three single-cell perturbation datasets, the authors report that splitDGD is the only method that consistently mitigates the trade-off between information leakage and utility. They also report that its cost on complementary fidelity metrics is bounded and is the lowest among methods with the same disentanglement goal.

Critical strengths
  1. The paper addresses an interesting problem, similar to latent action models in world-model research: perturbation prediction also aims to capture causal dynamics. Separating invariant cell states from perturbation effects is important for virtual cell modeling, but has received less attention than perturbation prediction itself.
  2. It explores DGD, a less common generative framework that learns representations through decoding rather than an autoencoder paradigm such as the VAE. This idea is related to recent unified multimodal approaches such as MetaQuery and BLIP3-o.
  3. The evaluation goes beyond cell perturbation and includes more general image representation tasks.
Critical weaknesses
  1. The motivation is not sufficiently convincing. The paper should explain what fundamental advantages DGD has over VAE-based methods, why it addresses the limitations of previous architectural approaches for separating perturbations from baseline states, and why it does not require the priors or constraints used by those methods.
  2. The contribution appears incremental, mainly adapting multiDGD to an architectural disentanglement setting. The paper also provides limited analysis of why the method works or why combining DGD with architectural separation is particularly effective.
  3. The empirical results are not consistently competitive: splitDGD does not clearly outperform the baselines across all metrics.
Questions for authors
  1. How does DGD prevent latent collapse when latent representations are optimized directly? Similar paradigms, including JEPA-style methods, are often prone to collapse when representations are learned only through gradient descent.
  2. Why is DGD particularly suitable for this problem? What prevents other generative frameworks from achieving similar results with the same architectural separation?
  3. Since DGD require inference-time optimization, does the inference cost match other baselines? Scaling the inference-time compute is known for better performance.
3 comments

Part 1/2

We thank the reviewer for recognising the importance of the problem, the interest of the DGD as a less common generative framework, and the breadth of the evaluation beyond single-cell data.

W1 / Q2 - Advantages of splitDGD for decoupling. Most perturbation-decoupling models add a soft penalty to a VAE objective, such as an adversarial classifier or an MMD term, so that the invariant latent carries no information about the condition. Such penalties do not guarantee the separation and are known to require difficult hyperparameter tuning and to often degrade the model performance (Elazar & Goldberg, 2018; Moyer et al., 2018; Wu et al., 2025).

The contrastive-analysis models, instead, enforce the separation architecturally by splitting the latent and setting the condition-specific block to zero for controls (Abid & Zou, 2019; Weinberger et al., 2023). They still rely on an encoder, and the separation is not guaranteed for them either. Leakage between the two blocks is acknowledged in these models (Weinberger et al., 2022), and soft penalties are often added on top of the architecture (a total-correlation term in Abid & Zou, 2019; moment-matching terms in Weinberger et al., 2022; an optional Wasserstein term in contrastiveVI).

Our insight is that such an architectural constraint is more natural and effective in a decoder-only model. In a DGD, the representation of each sample is a free parameter, estimated by MAP together with the decoder (Schuster & Krogh, 2023), so the constraint is applied to the representation itself, not to the output of an encoder. A control cell has w=0w=0, and the objective needs no decoupling penalty. Removing the encoder also removes the amortisation gap of VAE-like models (Cremer et al., 2018), and latent optimisation was found superior to amortised inference for disentanglement (Gabbay & Hoshen, 2020). In addition, DGDs perform better than encoder-based models when training samples are few relative to the number of features, as is the case in many biological datasets (Schuster & Krogh, 2021; Schuster et al., 2024).

Our results support the decoder-only model as a better alternative with respect to the methods that leverage variational inference. The closest architectural analogue with an encoder is contrastiveVI, which applies the same zeroing as splitDGD within a VAE. contrastiveVI has the highest linear leakage of all methods on Kang and Haber (0.70 and 0.68, against 0.52 and 0.55 for splitDGD); its reconstruction error and counterfactual R2R^2 are also worse than splitDGD's on all single-cell datasets.

W2 - Novelty over multiDGD. splitDGD shares the DGD backbone with multiDGD, and its supervised-w variant reuses multiDGD's use of known labels in the prior, as credited in Sec. 4, but the two models differ in purpose and structure. multiDGD gives every sample a second representation that encodes a known label, such as the experiment it comes from, to keep this unwanted variation out of the main representation; all samples are treated alike. splitDGD is instead built around the control group. A control sample has no w, so z alone must explain the unperturbed state, and the z-prior is fitted on controls only; w is left to describe how a perturbed sample departs from it. This asymmetry is the core of the method and is absent in multiDGD. It also allows a variant that needs only the binary control/perturbed indicator (unsup-w), with no perturbation labels. Furthermore, in the multiDGD work it was shown that the separation resulted in a better model, but there were no analyses of leakage of condition information among the latent representations.

To analyse the impact of each component of splitDGD, we have added an ablation study on Haber that removes the two mechanisms, one at a time and together (see our answer to Reviewer RXQW, where the contribution of each is detailed). Without them, a two-latent DGD leaks the condition (linear AUC from 0.55 to 0.68, unsup-w) and its z retains much less of the invariant factor (z-MCC from 0.73 to 0.44), while reconstruction is almost untouched.

Part 2/2

W3 - Results across metrics. We agree that splitDGD is not the best method on every metric, and we do not claim that. Our claim is that it offers the best leakage-vs-utility trade-off (Sec. 5), and the results support it in three ways. (i) On the trade-off itself, no baseline is at least as good as splitDGD on both linear leakage and utility (z-MCC or counterfactual R2R^2) on any dataset. (ii) splitDGD is best or second-best in 18 of its 27 entries in the main result tables (67%), and splitDGD (sup-w) in 16 of 22 (73%), against 48% for CVAE, which has no decoupling mechanism, and 24-30% for the three decoupling baselines. (iii) It almost never performs badly. In the robustness analysis of the Appendix (Table 6), which counts the entries where a method is worse than the leader by more than a chosen threshold, splitDGD has one such entry out of 27 (w-MCC on Haber, unsup-w) and splitDGD (sup-w) none out of 22, whereas every baseline has at least four (CVAE 10/27, DisCoVR 10/27, scDisInFact 4/17, contrastiveVI 4/17).

In contrast, CVAE keeps utility but is among the leakiest models, scDisInFact reduces leakage but falls behind on fidelity, and contrastiveVI recovers the perturbation structure on Haber at the cost of the highest leakage, so each baseline is strong on one side only.

Q1 - Latent collapse. In JEPA-style models the loss is computed between learned embeddings, so it can be minimised trivially by an encoder that returns the same output for every input, and this has to be countered by dedicated mechanisms, such as explicit regularisation or an asymmetric architecture with a momentum target encoder (Assran et al., 2023).

In a DGD, representation collapse is prevented by the objective itself: the likelihood of the observed data given the representation. The same representation for different cells would give identical reconstructions and a poor likelihood, as in any model trained to reconstruct its input; no additional mechanism is needed, as long as the representation has low capacity compared with the data (Assran et al., 2023), which holds here (at most 16 dimensions per block, against 784 to 8,630 features). Moreover, including (i) learnable mixture priors and (ii) leveraging the hard masking mechanism, favor capturing “meaningful” variations both in z and w.

Q3 - Inference cost. We acknowledge that inferring representations with splitDGD is slower than an encoder pass, and we have added an analysis of the inference cost in the paper (see Section X). The table gives the seconds needed to obtain the representations of new samples on one GPU, for the test split and for 100,000 cells: the baselines run one encoder pass, and splitDGD optimises the latents with the trained network fixed, for the number of steps of the paper's runs (650 on Haber, 550 on sci-Plex).

MethodInferenceHaber test split (1,969 cells)sci-Plex test split (26,317 cells)Haber, 100,000 cellssci-Plex, 100,000 cells
CVAEencoder pass0.020.701.272.80
DisCoVRencoder pass0.081.303.885.05
contrastiveVIencoder pass0.011.150.984.17
scDisInFactencoder pass0.033.953.6014.9
splitDGD650 / 550 steps (paper)2.3425.841.595.2

The cost is incurred once, since counterfactual predictions reuse the inferred representations, and stays under two minutes for 100,000 cells. This is negligible in practice, since these models are typically used to analyse a dataset offline

Reviewer_GVNjOpenReview ↗

Thanks for your rebuttal

Thanks for your reply. It addressed my concern about the motivation and the benefits of generative architecture is convincing. I will increase my score accordingly.

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

The paper presents splitDGD, an encoder-free model for separating baseline cellular variation from perturbation responses. It combines per-sample latent optimization, control masking, and a baseline prior fitted only from control cells. Experiments on two image datasets and three single-cell datasets evaluate condition leakage, latent-factor recovery, reconstruction, and counterfactual expression. The reported results show competitive reconstruction and a useful empirical balance between several representation objectives.

Critical strengths
  • S1. Fitting the baseline prior only from controls is a simple and well-motivated way to use control cells as a reference. Combined with direct latent optimization and control masking, it provides a clear strategy for separating baseline variation from perturbation responses.
  • S2. The evaluation considers condition leakage and representation utility jointly across image and single-cell datasets. Nonlinear probes and competitive reconstruction results provide useful empirical evidence beyond qualitative latent visualizations.
Critical weaknesses
  • W1. The comparisons provide evidence for the overall model, but the contribution of fitting the baseline prior only from controls remains difficult to isolate. A matched comparison with fitting this prior from all cells would help establish how much this design contributes to the reported leakage and utility trade-off.
  • W2. The counterfactual evidence has a limited scope. On Kang and Haber, the transferred perturbation embedding is the mean of embeddings inferred from perturbed test cells of the same cell type and perturbation, as disclosed in the paper. These results therefore do not establish prediction without access to the target response population. Sci-Plex3 provides a separate evaluation on held-out combinations, but the gains vary across perturbation-sensitive metrics.
Questions for authors
  • Could you provide a matched ablation of fitting the baseline prior from controls only versus all cells, keeping the remaining splitDGD configuration unchanged, and report condition leakage, invariant-factor recovery, and reconstruction on representative datasets?
1 comment

We thank the reviewer for the positive assessment, in particular of the control-only prior as a simple and well-motivated use of control cells, and of the joint evaluation of leakage and utility.

W1/Q1- Ablation of splitDGD. We ran the matched ablation on Haber, for both unsupervised w and supervised w, retraining splitDGD with one mechanism switched off and everything else as in the paper's runs (splits, 10 seeds, and the hyperparameters selected for each seed): w/o alternate prior, the z-prior is fitted on all cells instead of controls only (the comparison requested); w/o hard mask, the decoder sees [z;w][z; w] for every cell and every w is under the w-prior; w/o both. In parentheses, the difference from splitDGD on the same seeds. We observe that hard masking positively affects the performance of splitDGD, while adding alternate priors is beneficial, but with milder improvements.

Haber (10 seeds)AUC linAUC NLz-MCCw-MCCRMSECF R2R^2
unsup-wsplitDGD0.5500.6940.7300.5990.1980.885
w/o alternate prior0.555 (+0.005)0.708 (+0.014)0.722 (-0.008)0.640 (+0.041)0.199 (+0.001)0.888 (+0.003)
w/o hard mask0.616 (+0.066)0.722 (+0.028)0.467 (-0.263)0.567 (-0.032)0.197 (-0.001)0.891 (+0.006)
w/o both0.683 (+0.133)0.737 (+0.043)0.438 (-0.292)0.552 (-0.047)0.198 (0.000)0.885 (0.000)
sup-wsplitDGD0.5320.6870.7401.0000.1960.926
w/o alternate prior0.534 (+0.002)0.681 (-0.006)0.733 (-0.007)1.000 (0.000)0.196 (0.000)0.925 (-0.001)
w/o hard mask0.594 (+0.062)0.673 (-0.014)0.510 (-0.230)0.935 (-0.065)0.196 (0.000)0.918 (-0.008)
w/o both0.580 (+0.048)0.648 (-0.039)0.411 (-0.329)0.934 (-0.066)0.196 (0.000)0.918 (-0.008)

More in detail, without the hard mask, linear leakage rises by 0.06-0.07 AUC, z-MCC falls by 0.23-0.26 and w-MCC by 0.03-0.07, in both regimes, so the mask carries most of the decoupling. The alternate prior has a minor effect on its own: with the mask in place, fitting the z-prior on all cells worsens linear AUC and z-MCC only slightly, but consistently in both regimes (linear AUC +0.005 and +0.002, z-MCC -0.008 and -0.007), since the mask already keeps the condition out of z. It also acts as a second safeguard when the mask is removed: removing it as well then lowers z-MCC further in both regimes (0.467 to 0.438, unsup-w; 0.510 to 0.411, sup-w) and, in unsup-w, raises linear leakage further (0.616 to 0.683). In sup-w, removing both mechanisms lowers the nonlinear AUC (0.687 to 0.648), but this reflects the leakage-utility trade-off: z leaks less about the condition because it retains less information overall, as its z-MCC falls from 0.740 to 0.411.

For all ablations, RMSE is slightly unaffected (at most 0.001), while counterfactual R2R^2 fluctuates by at most 0.008. This shows that the two mechanisms only in part affect performance in fidelity

W2 - Counterfactual evidence has limited scope. On Kang and Haber every (cell type, perturbation) pair is observed in training and, as the reviewer notes, the perturbation code is the mean w of the perturbed test cells of that cell type and perturbation. The counterfactual R2R^2 (Lotfollahi et al., 2019; 2023) therefore measures whether the learned decomposition recombines, that is, whether a control cell's own z combined with a perturbation code reproduces the perturbed profile. Every method is scored by the same rule on the same cells.

To check that the scores do not depend on access to the perturbed test cells, we recomputed the counterfactual R2R^2 with the perturbation code taken from the training cells instead, so that no perturbed test cell enters the prediction (paper's protocol -> training cells, mean over 10 seeds):

KangHaber unsup-wHaber sup-w
DisCoVR0.891 -> 0.8930.831 -> 0.831-
contrastiveVI0.893 -> 0.8900.819 -> 0.819-
scDisInFact0.871 -> 0.872-0.824 -> 0.826
splitDGD0.960 -> 0.9570.885 -> 0.8750.926 -> 0.926

Every method's counterfactual R2R^2 changes by at most 0.010, and also with this variant splitDGD remains ahead of these baselines in every block.

The prediction of unobserved responses is evaluated on sci-Plex3, under PerturBench's covariate-transfer split (Wu et al., 2024): every scored (cell line, compound) pair is absent from training, and the compound's w is the mean over its training cells in the other two lines. Each metric captures a different aspect of the counterfactual prediction, and we agree with the reviewer that no model leads on all of them.

Overall, splitDGD and its supervised variant never perform poorly in any metric, while other methods always fail at least at one (e.g., CVAE achieves near random Rank because it cannot distinguish distinct perturbations). We will highlight this point in the revised manuscript.

Comments on the paper

1 comment

We apologise for the mistake: the two tables were missing from our reply. The table for W1/Q1 is the ablation table of our answer to Reviewer RXQW (W1/Q1), which we also report here (Haber, mean over 10 seeds; in parentheses, the difference from splitDGD):

Haber (10 seeds)AUC linAUC NLz-MCCw-MCCRMSECF R2R^2
unsup-wsplitDGD0.5500.6940.7300.5990.1980.885
w/o alternate prior0.555 (+0.005)0.708 (+0.014)0.722 (-0.008)0.640 (+0.041)0.199 (+0.001)0.888 (+0.003)
w/o hard mask0.616 (+0.066)0.722 (+0.028)0.467 (-0.263)0.567 (-0.032)0.197 (-0.001)0.891 (+0.006)
w/o both0.683 (+0.133)0.737 (+0.043)0.438 (-0.292)0.552 (-0.047)0.198 (0.000)0.885 (0.000)
sup-wsplitDGD0.5320.6870.7401.0000.1960.926
w/o alternate prior0.534 (+0.002)0.681 (-0.006)0.733 (-0.007)1.000 (0.000)0.196 (0.000)0.925 (-0.001)
w/o hard mask0.594 (+0.062)0.673 (-0.014)0.510 (-0.230)0.935 (-0.065)0.196 (0.000)0.918 (-0.008)
w/o both0.580 (+0.048)0.648 (-0.039)0.411 (-0.329)0.934 (-0.066)0.196 (0.000)0.918 (-0.008)

Further comments and details on this table are in our answer to Reviewer RXQW.

The table for Q3 (Haber, mean (std) over 10 seeds):

AUC linAUC NLz-MCCw-MCCRMSECF R2R^2
unsup-wsplitDGD (paper)0.550 (.021)0.694 (.029)0.730 (.044)0.599 (.065)0.198 (.007)0.885 (.044)
splitDGD (counts)0.531 (.037)0.705 (.036)0.678 (.046)0.534 (.224)0.195 (.003)0.872 (.016)
sup-wsplitDGD (paper)0.532 (.026)0.687 (.032)0.740 (.023)1.000 (.001)0.196 (.005)0.926 (.009)
splitDGD (counts)0.542 (.041)0.668 (.033)0.663 (.047)1.000 (.000)0.198 (.003)0.897 (.018)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.