- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
PerturbEnergy introduces a conditional energy-based prior to encourage perturbation-aware latent representations for cellular response prediction. It trains with Langevin or mismatched negative samples and selects low-energy latent candidates during inference. Evaluation on four datasets reports improved perturbation classification and the best top-DEG RMSE on three datasets.
Critical strengths
-
The evaluation covers genetic, chemical, and cytokine perturbations and includes competitive recent baselines. Improvements on response-relevant DEG metrics provide useful evidence beyond agreement on overall expression levels.
-
The additional backbone experiments support the usefulness of the energy objective beyond the proposed VAE implementation. This makes the approach potentially useful as a reusable training component.
Critical weaknesses
-
The claimed mutual-information guarantee is not established by the proof. In Appendix A.1.4, the Bayesian derivation drops normalization constants that depend on the perturbation and therefore cannot generally be cancelled. T
-
The evidence does not clearly separate label encoding from faithful biological response modeling. The perturbation encoder already receives the perturbation label, so improved label classification can reflect retaining that input rather than learning its biological effects. Moreover, population-level fidelity can worsen despite better mean-expression prediction: Table 12 reports an energy distance of 8.65 for PerturbEnergy-Lang versus 0.70 for STATE on Replogle. These findings require a clearer explanation before interpreting latent separation as improved modeling of perturbation responses.
Questions for authors
-
How are the perturbation-dependent normalization constants accounted for in Theorem 2, and what establishes that the implemented training updates increase the stated lower bound?
-
Which representations from each model were classified in Table 3? How was direct access to perturbation labels accounted for when interpreting the classification gains?
-
How should the substantially higher energy-distance values be reconciled with improved mean-expression metrics? What do existing diagnostics indicate about preservation of variation among cells under the same condition?
8 comments
Q2/W2 Response to Reviewer 82EU
Q2/W2 Response
Q2/W2-(a) Which representations are classified?
For all methods, we classify the representation fed into the decoder, i.e., the one finally used to predict the cellular response. We aim to measure how much perturbation information remains in this representation after training, since it also encodes the cell state, context, and potential heterogeneity-driven variation. The representations used for each method are summarized below:
| Method | Representation |
| scGPT | scGPT output |
| PerturbDiff | denoised latent fed to the decoder |
| STATE | transition output fed to the decoder |
| scGen | (: perturbation embedding changes) |
| CPA | (: context embedding) |
| PerturbedVAE | |
| PerturbEnergy |
Q2/W2-(b) How is label access accounted for?
Direct access to the perturbation label is not unique to our method: all baselines above except scGPT also inject into the decoder input, yet their accuracy stays below 50% on Replogle, PBMC and Tahoe100M. Since also encodes context, basal state and cell-specific patterns, the perturbation signal can be overwhelmed when the model fits cellular heterogeneity-driven variation. The gain therefore reflects how well perturbation information is retained, not whether it is accessible.
To make it more clear, we conduct a simple experiment on Replogle, evaluating the checkpoints at 5k, 10k, 20k, 40k, and 60k training steps and track both perturbation discriminability and Top-50 DEG RMSE. Perturbation classification Acc (%, ↑) / Top-50 DEG RMSE (↓) is shown as:
| Steps (Acc / DEG) | PerturbedVAE | CPA | PerturbEnergy-Mis |
| 5k | 23.6 / 0.2642 | 18.1 / 0.3785 | 42.4 / 0.1941 |
| 10k | 26.8 / 0.2290 | 22.8 / 0.3491 | 48.5 / 0.1860 |
| 20k | 28.2 / 0.2188 | 22.6 / 0.3438 | 51.0 / 0.1780 |
| 40k | 27.3 / 0.2267 | 19.7 / 0.3618 | 50.6 / 0.1824 |
| 60k | 26.7 / 0.2325 | 20.2 / 0.3605 | 49.9 / 0.1843 |
For all methods, over-training leads to degraded Top50DEG RMSE, which means model learn non-generalizale patterns like cellular heterogeneity or noise. This also leads to lower discriminability for perturbation although label access is constant throughout training. Model forgets perturbation but remember heterogeneity patterns.
W2-(c) Improved Classification Relates to Biological Response
We can view label discriminability as a signal of how well a model captures perturbation responses, and our existing results support this (our Table 2 and Table 3). Compared with PerturbedVAE, PerturbEnergy-Mis substantially improves latent accuracy on Replogle (28.6 → 52.4), PBMC (48.2 → 61.8) and Tahoe100M (45.5 → 58.5), with correspondingly lower Top-50 DEG RMSE (0.2130 → 0.1780, 0.1071 → 0.0654, 0.3984 → 0.3123). The same pattern holds among baselines: from CPA to PerturbedVAE on Replogle, accuracy rises from 22.2 to 28.6 while RMSE drops from 0.3462 to 0.2130, and similarly on PBMC and Tahoe100M.
| Dataset | CPA | PerturbedVAE | PerturbEnergy-Mis |
| Replogle | 22.2 / 0.3462 | 28.6 / 0.2130 | 52.4 / 0.1780 |
| PBMC | 43.0 / 0.1581 | 48.2 / 0.1071 | 61.8 / 0.0654 |
| Tahoe100M | 33.3 / 0.7070 | 45.5 / 0.3984 | 58.5 / 0.3123 |
Latent Acc (%, ↑) / Top-50 DEG RMSE (↓), from Tables 2 and 3.
Q3/W2 Response to Reviewer 82EU
Q3/W2 Response
Q3/W2-(a) Energy Distance and Mean-expression Metrics
1.Metric-level explanation. We agree that the two types of metrics evaluate different dimensions of prediction like VC benchmark points[1]. Mean-expression metrics (RMSE/MSE/MAE) evaluate whether the predicted average response matches the observed one, whereas EDist compares full single-cell distributions over all cell pairs. By definition, EDist aggregates distances between all pairs of cells on HVGs. Accurate mean response does not imply a low EDist.
For our perturbation prediction, we mainly focus on perturbation-induced mean responses rather than full single-cell distributions. Besides, perturbation focus more on DEGs, which are the genes most changed by the perturbation, than all 2,000 HVGs; Actually, much of the cell-to-cell variance is not caused by the perturbation: most of the 2,000 HVGs are unaffected by the perturbation, and their variation mainly comes from cell cycle and cellular heterogeneity. We report EDist for reference, following the VC Benchmark [1]. For HVG2000, in Table 12, mean expression metrics are more stable across methods than EDist, and our method outperform other baselines on them:
| Method | RMSE ↓ | MSE ↓ | MAE ↓ | EDist ↓ |
| CPA | 0.0953 | 0.0118 | 0.0612 | 4.8973 |
| PerturbDiff | 0.0817 | 0.0079 | 0.0527 | 6.3453 |
| PerturbedVAE | 0.0745 | 0.0063 | 0.0501 | 5.5135 |
| STATE | 0.0725 | 0.0060 | 0.0491 | 0.7007 |
| PerturbEnergy-Lang | 0.0691 | 0.0054 | 0.0465 | 8.6506 |
| PerturbEnergy-Mis | 0.0704 | 0.0055 | 0.0473 | 9.0347 |
We further compute EDist on the Top-20/50 DEGs on Replogle:
| Method | Top-20 EDist ↓ | Top-50 EDist ↓ | Top-50 RMSE ↓ |
| CPA | 2.173 | 2.878 | 0.3462 |
| PerturbDiff | 2.398 | 2.605 | 0.2612 |
| PerturbedVAE | 1.966 | 2.355 | 0.2130 |
| STATE | 1.591 | 2.233 | 0.2024 |
| PerturbEnergy-Lang | 1.663 | 2.412 | 0.1870 |
| PerturbEnergy-Mis | 1.556 | 2.310 | 0.1780 |
On DEGs, the influence of cell-to-cell variation is reduced and perturbation effect becomes significant. EDist becomes much more stable across methods. Our method achieves the best Top-20 EDist (1.556 vs. 1.591 for STATE) and is close to the best on Top-50 (2.310 vs. 2.233).
2.Method-level explanation. PerturbEnergy retains perturbation-specific information and selects the lowest-energy representation at inference. This mode-seeking selection reduces the diversity of generated cells, as the model focuses on stable perturbation effects rather than perturbation-irrelevant cell-to-cell variation. Moreover, our design does not explicitly model within-condition cell-to-cell variability.
Q3/W2-(b) Diagnostics for Within-condition Variation
Our existing results do not directly measure variation among cells under the same condition because it's not direct perturbation related target. We here provide supplementary experiments to illustrate this.
Dispersion ratio. It measures the ratio between the dispersion of predicted cells and that of real cells under the same condition :
To examine the effect of minimum-energy selection, we also evaluate a variant of PerturbEnergy that samples randomly from instead of selecting the lowest-energy candidate.
| Method | Dispersion Ratio (HVG) | EDist (HVG) ↓ | Top-50 DEG RMSE ↓ |
| CPA | 0.148 | 4.8973 | 0.3462 |
| STATE | 0.584 | 0.7007 | 0.2024 |
| PerturbedVAE | 0.106 | 5.5135 | 0.2130 |
| PerturbEnergy-Mis | 0.051 | 9.0347 | 0.1780 |
| PerturbEnergy-Mis (No Minimum E) | 0.115 | 5.7360 | 0.2221 |
Our method preserves less within-condition variance (dispersion ratio 0.051), as do other VAE-based methods (0.11–0.15), whereas STATE largely inherits control-cell variance (0.584). This explains our higher EDist. The reduced variance partly stems from minimum-energy selection: without it, dispersion doubles (0.115) and EDist drops to 5.74, but Top-50 DEG RMSE rises from 0.1780 to 0.2221.
[1] Mao, Xinjie, et al. "Benchmarking virtual cell models for in-the-wild perturbation response." arXiv preprint arXiv:2604.27646 (2026).
Q1/W1 Response to Reviewer 82EU
Q1 / W1 Response
We sincerely thank you for the thoughtful questions. We provide further explanations and theoretical corrections/discussions have been updated in paper. (Theorem 2, Appendix A1.4, Appendix A1.5)
W1/Q1-(a): The Partition Function.
Thank you for pointing out this oversight. The Fα(zb,a) must be included when applying Bayes' rule; we have corrected Eq. 17 and Appendix A.1.4. Importantly, the omission only affects the form of rα(a∣za,zb,c) and the inequality in Theorem2(Eq. 16) remains valid: it is the Barber–Agakov variational bound, which holds for any normalized conditional r.
W1/Q1-(b): The Implemented Updates and Mutual Information.
We have two variants (Langevin and Mismatch) for the implementation:
Case 1: Langevin Langevin directly optimizes the Energy-ELBO, and we claimed that “this optimization increases the variational mutual information lower bound”. This describes the optimal case: when rα(a∣za,zb,c) perfectly matches the posterior p(a∣za,zb,c), the bound is maximized and recovers I(A;Za∣Zb,C) of current dataset distribution. The original version only described this intuitively. Strictly, the relation is: Assuming the variational inference model qϕα accurately approximates the conditional latent landscape, this optimization reduces an upper bound on the gap between the true conditional mutual information and its variational lower bound. Proof below:
Step 1: Define the MI gap.
The mutual information gap is:
G=I(A;Za∣Zb,C)−H(A∣Zb,C)−Ep[logrα(A∣Za,Zb,C)]=Ep[DKL(p(A∣Za,Zb,C)∥rα(A∣Za,Zb,C))]≥0.
We define two conditional joint distributions:
P(a,z_a\mid z_b,c)=\pi(a\mid z_b,c)q_{\phi_\alpha}(z_a\mid z_b,c,a),$$$$Q_\alpha(a,z_a\mid z_b,c)=\pi(a\mid z_b,c)p_\alpha(z_a\mid z_b,a).
Step 2: Use KL chain rule under two factorizations.
Since the two distributions share the marginal π(a∣zb,c), the KL chain rule gives:
D(α)=EZb,C[DKL(P∥Qα)]=EZb,C,A[DKL(qϕα∥pα)].
Alternatively, factorizing over Za yields:
D(α)=EZb,C[DKL(P(Za∣Zb,C)∥Qα(Za∣Zb,C))]+EP[DKL(P(A∣Za,Zb,C)∥rα(A∣Za,Zb,C))].
Step 3: Combine the two expressions of D(α) to obtain the final relation
Since the first KL term is non-negative, we have:
D(α)=EZb,C,A[DKL(qϕα∥pα)]≥G≥0.
Therefore, optimizing DKL(qϕα∥pα) reduces the mutual information gap G upper bound.
Case 2: Mismatch Mismatch is an efficient surrogate that optimizes a lower bound on the conditional mutual information, described as Proposition 3.
Let S=(C,Xctr,Zb). In training, a∼π(⋅∣c) with positive za+∼qϕα(⋅∣zb,xctr,a) the mismatched label a−∼π(⋅∣c,a−=a) gives negative za−∼qϕα(⋅∣zb,xctr,a−). The negative shares the same (c,xctr,zb) with the positive and differs only in the perturbation. Define the energy gap:
G(α)=E[(1−π(a∣c))(Eα(za−,zb,a)−Eα(za+,zb,a))].
Since our energy is a scaled cosine similarity, ∣Eα∣≤τ.
Proposition 3: I(A;Za∣S) ≥ G(α)−τ2/2.
Proof below:
With uniform π, (1−π(a∣c))=1−1/∣Ac∣ is constant, so the bound is proportional to the Mismatch objective.
Step 1: MI as an expected KL. Since perturbations are assigned independently of the control cell within each context, A⊥(Xctr,Zb)∣C. Hence I(A;Za∣S)=Es,a[KL(Ps,a∥Qs)], where Ps,a=q(za∣s,a) and Qs=∑a′π(a′∣c)q(za∣s,a′).
Step 2: Donsker–Varadhan bound. For any function T, KL(P∥Q)≥EP[T]−logEQ[eT]. We take T=−Eα(⋅,zb,a).
Step 3: Hoeffding’s lemma. Since T∈[−τ,τ], logEQ[eT]≤EQ[T]+τ2/2.
Step 4: Excluding the true label. Write Qs=π(a∣c)Ps,a+(1−π(a∣c))Q~s,a, where Q~s,a is exactly the distribution of mismatched negatives. Then EP[T]−EQ[T]=(1−π(a∣c))(EP[T]−EQ~[T]).
Combining steps 1–4 and averaging over (s,a) yields the claim.