- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper investigates when learned representations for single-cell perturbation prediction support OOD generalization and proposes scCausalDiff, a three-stage framework: (1) projecting control-cell expression into a basal latent space via a variational encoder and predicting an additive shift with a perturbation encoder; (2) performing causal propagation on the combined representations using a learned linear SCM; and (3) feeding the propagated representation into a conditional diffusion generator. In particular, the SCM is trained under sparsity and acyclicity regularization.
The paper also provides a theoretical analysis establishing latent-distribution identifiability under a fixed conditional kernel and an injectivity assumption. It further discusses structural identifiability under additional constraints, proves oracle Bayes optimality across domains, and derives a generation-error bound controlled by representation error and generator approximation error.
The empirical evaluation spans held-out genetic perturbations on Replogle, unseen gene combinations on Norman, and held-out compound-dose conditions on sci-Plex3. Based on the reported figures, scCausalDiff is competitive in terms of mean-response accuracy, while its most substantial advantage lies in distributional metrics on sci-Plex3 and Norman.
Critical strengths
The strengths lie primarily in the empirical results, demonstrating practical value for distributional prediction. scCausalDiff achieves superior performance in terms of MMD and energy distance on sci-Plex3, and Appendix D.1 reports consistent improvements across all five folds on Norman. The evaluation of both mean responses and cell-population distributions is appreciated.
Critical weaknesses
- The theoretical analysis does not establish identifiability for the jointly learned model or OOD generalization from source-domain training.
Theorem 4.4 treats the conditional kernel as fixed and assumes that the observation operator is injective, whereas the proposed method jointly learns the latent SCM and the conditional operator. For example, jointly scaling the latent representation and reparameterizing the conditional kernel can leave the observed distribution unchanged. This does not contradict Theorem 4.4, but shows that its fixed-kernel guarantee leaves the joint-learning ambiguity unresolved. Also, Theorem 4.10 relies on Assumption B.1(c–d) to bound target-domain representation and generator errors, but it does not show how source-domain training actually controls these terms. Neither result explains why the proposed causal propagation improves OOD generalization.
- The existing ablation study also can not support the benefits of the learned DAG.
Figure 5 replaces the propagation map only at inference time, while keeping all other weights fixed. The drop in performance shows that the trained model depends on this map, but it does not show that training with a DAG is better than training with other architectures. For example, a comparison would be models trained from scratch with and with an unconstrained learned linear map, using the same embeddings, denoiser, and training budget. Also, sci-Plex3 is split by compound-dose conditions, but it does not establish compound-disjoint evaluation. These results therefore do not establish generalization to completely unseen compounds.
Questions for authors
-
Theorem 4.4 fixes , while Assumption B.1(c–d) already bounds target-domain errors. What do these results guarantee for the actual model, which learns and jointly from source-domain data?
-
Does the DAG outperform models trained from scratch with or an unconstrained learned linear map, using the same embeddings, denoiser, and training budget?
-
How do you position your novelty compared to DiscrepancyVAE (Zhang et al., 2023), SAMS-VAE (Bereket and Karaletsos, 2023), and PerturbedVAE (Jiang et al., 2026) in its modeling assumptions, identifiability guarantees, and prediction of unseen perturbations? Where the prediction settings overlap, are direct empirical comparisons with these methods available?
References:
Zhang J, Greenewald K, Squires C, et al. Identifiability guarantees for causal disentanglement from soft interventions. NeurIPS 2023.
Bereket M, Karaletsos T. Modelling cellular perturbations with the sparse additive mechanism shift variational autoencoder. NeurIPS 2023.
Jiang W, Liu Y, Cai Y, et al. What Makes a Representation Good for Single-Cell Perturbation Prediction? ICML 2026.
3 comments
Thank you for the detailed response and additional experiments. The retrained identity ablation, new baseline comparisons, and clarification of the compound-disjoint splits address important aspects of my empirical concerns. Three points remain:
- Connection between the theory and the jointly learned model. I agree that Theorem 4.10 can apply to jointly trained representations and generators. However, Assumption B.1(c–d) already requires control of their errors in the target domains. My concern is how source-domain training provides this control; the response does not establish that connection. Similarly, Theorem 4.4 identifies latent distributions under a fixed kernel. The observation that unique recovery of is unnecessary for prediction does not resolve the ambiguity in latent distributions when the kernel is learned jointly.
- The specific benefit of the DAG constraint. The retrained identity comparison supports the value of learned propagation. However, the reported similarity between DAG and unrestricted propagation does not establish that acyclicity improves prediction. Comparable performance with fewer parameters could be useful, but please provide the full comparison and clarify the parameterization. The response describes a ''triangular DAG'', whereas Section 3 specifies NOTEARS without a prescribed topological order. Were the new comparisons conducted using the model described in the manuscript?
- Incorporation of the rebuttal evidence. I could not find these additions in the supplied manuscript. Appendix C.1 still describes sci-Plex3 splits as formed over conditions, Figure 5 still reports inference-time replacements, and the new baseline comparisons are absent. Please incorporate the corrected split description, retrained ablations, and baseline results, with sufficient implementation detail to assess their comparability.
I appreciate the stronger empirical evidence. My remaining concerns focus on what the theory establishes for the trained model and what the experiments show specifically about the DAG constraint.
We thank the reviewer for acknowledging the additional evidence and clarifications. To maintain a consistent manuscript version for all reviewers during the ongoing discussion, we have presented the additional results and clarifications in our responses. We plan to incorporate these updates together in the final manuscript, should the paper be accepted.
We provide further clarification on the two remaining points below.
On joint learning and transfer to unseen perturbations
To connect source-domain fitting with OOD prediction under joint learning, we consider the observable cell-population expression distributions. Let and denote the true and predicted distributions for perturbation , with finite first moments, and define .
Following our structural formulation, write the true and learned propagated states as and , respectively. Within each model, the basal-state distribution is shared across perturbations. Suppose and are Lipschitz in the perturbation embedding, with constants , and the corresponding generation kernels are -Lipschitz in their latent inputs, with constants . Let and .
By coupling the same basal state across perturbations within each model, we have
\qquad W_1(\hat P_p,\hat P_q)\leq \hat M\|r_p-r_q\|.$$ Applying the triangle inequality through any source perturbation $q$ gives $$\boxed{ D_p\leq \min_{q\in\mathcal P_{\mathrm{src}}} \left[ D_q+(M+\hat M)\|r_p-r_q\| \right]. }$$ Hence, source-domain prediction accuracy can extend to unseen perturbations when the response mappings vary smoothly and the target perturbations are sufficiently covered in the embedding space. The bound applies to the jointly fitted predictor and depends on observable source-domain errors rather than requiring alignment between the true and learned latent coordinates. The same reasoning applies to two independently fitted models, even if their latent representations and generators differ. If their observable response mappings have Lipschitz constants $M_1,M_2$, then $$\boxed{ \begin{aligned} &W_1(\hat P_p^{(1)},\hat P_p^{(2)})\\ &\leq \min_{q\in\mathcal P_{\mathrm{src}}} \left[ D_q^{(1)}+D_q^{(2)} +(M_1+M_2)\|r_p-r_q\| \right], \end{aligned} }$$ where $D_q^{(i)}=W_1(P_q,\hat P_q^{(i)})$. Thus, different latent parameterizations need not lead to substantially different OOD predictions: their observable disagreement is controlled by source-domain fitting accuracy and response variation for nearby unseen perturbations. This analysis complements Theorem 4.4, which concerns latent identifiability under a fixed shared kernel, and Theorem 4.10, which bounds generation error through representation and generator approximation errors. It further relates the jointly learned model's source-domain predictive accuracy to its behavior on unseen perturbations under the stated regularity and coverage conditions. ___ ### On the DAG constraint and unrestricted propagation The "triangular DAG" in our earlier response refers to the structural property that a DAG admits a topological ordering under which its adjacency matrix is strictly triangular, with at most $d(d-1)/2$ possible edges. In our implementation, this property is enforced through the NOTEARS acyclicity constraint. When the constraint is satisfied exactly, the learned graph likewise has at most $d(d-1)/2$ edges. This describes the DAG structure, not the number of optimized parameters. The fully unrestricted variant starts from a randomly initialized dense matrix and jointly optimizes its entries without acyclicity or L1 constraints, allowing cyclic dependencies and self-loops. **DAG vs. unrestricted propagation.** Signed mean differences over five folds are computed as DAG minus free for ↑ metrics and free minus DAG for ↓ metrics. Positive values favor the DAG. | Dataset | Δ-P ↑ | Δ-S ↑ | PDS-cos ↑ | PDS-L1 ↑ | MMD ↓ | E-dist ↓ | |---|---:|---:|---:|---:|---:|---:| | Replogle | −0.0055 | −0.0047 | +0.0055 | +0.0038 | −0.0005 | −0.0148 | | sci-Plex3 | −0.0048 | −0.0070 | +0.0011 | +0.0006 | −0.0022 | −0.0531 | The results show broadly comparable predictive performance. The value of the DAG constraint lies in its structural inductive bias: perturbation effects propagate along directed paths without returning through feedback cycles, consistent with our latent SCM formulation, rather than through arbitrary cyclic feature mixing. Together with the end-to-end retrained $A=0$ ablation, these findings support two complementary observations: learning propagation improves prediction over identity propagation, and these improvements can largely be retained under DAG-based structural constraints.
We thank the reviewer for the careful reading and the precise questions. We address them in turn below.
On what the theory guarantees for the jointly learned model (Weakness 1, Question 1)
The fixed-kernel condition in Theorem 4.4 concerns population-level identifiability, rather than a restriction on joint optimization. Given a shared conditional kernel satisfying injectivity, the theorem establishes identifiability of perturbation-induced latent distributions and their relative shifts. While jointly learned latent representations may admit equivalent parameterizations, our objective is accurate perturbation-response prediction rather than unique recovery of the structural matrix A. Such structural identifiability is not required for our OOD results.
Crucially, Theorem 4.10 provides a generation-error guarantee for the jointly learned model. It explicitly considers both the learned representation U^p and the learned generator Qgen, without requiring separate optimization, independent estimation errors, or a uniquely identifiable A. Under Assumption B.1, the target-domain generation error is bounded by ω(δU)+δQ. Conditions B.1(c–d) specify the required target-domain approximation accuracy; the theorem establishes how these two errors jointly control the observable prediction error, rather than assuming accurate predictions themselves. Thus, the guarantee applies to jointly learned models satisfying these conditions, without requiring the generation kernel to remain fixed during training.
On whether the learned DAG improves over alternative architectures (Weakness 2, Question 2).
We trained the learned DAG and A=0 variants from scratch on the same five folds of Replogle and sci-Plex3, using the same perturbation embeddings and denoiser architecture within each dataset. The table reports mean differences in the original metric units. Positive values favor the DAG and negative values favor A=0. We subtract A=0 from the DAG for Δ-P, Δ-S, and PDS, and reverse the subtraction for MMD and E-dist.
The DAG improved Δ-P, Δ-S, MMD, and E-dist in all five folds on both datasets, supporting the use of a learned propagation operator.
In a separate five-fold comparison, the learned DAG and unrestricted propagation achieved similar overall predictive performance. The triangular DAG used approximately half as many free structural coefficients (d(d−1)/2 versus d2).
On compound-disjoint evaluation for sci-Plex3 (Weakness 2)
Our description in Appendix C.1 ("formed over conditions") was imprecise. The sci-Plex3 folds are compound-disjoint: the 187 compounds are partitioned across the five folds, and all four doses of a held-out compound go to the test set as a block. We re-verified on the released split files that the train–test compound overlap is exactly zero on every fold, with no cell of a test compound in training. Table 1 therefore measures generalization to completely unseen compounds.
On DiscrepancyVAE, SAMS-VAE, and PerturbedVAE (Question 3)
Modeling. DiscrepancyVAE uses latent soft interventions, SAMS-VAE sparse offsets, and PerturbedVAE label-dependent SCM edges. Ours predicts sp=f(rp) from perturbation representations, uses shared B=(I−A)−1, and adopts conditional diffusion for unseen perturbations.
Identifiability. DiscrepancyVAE identifies DAGs/targets up to CD-equivalence under intervention coverage; PerturbedVAE identifies responsive factors under invertibility and sufficient environments; SAMS-VAE has no formal identifiability theorem. Our Theorem 4.4 instead identifies latent distributions under Assumption 4.3, without requiring identifiable A or per-node intervention coverage.
Experiments. We retrained all three baselines on our five-fold Norman and Replogle splits.
Norman
Replogle
Ours leads on all six Replogle metrics and five of six Norman metrics.
We thank the reviewer for the constructive feedback and hope our responses address their concerns.