acceptodds
Under review as a conference paper at ICLR 2027

Rotation Operators for Single-Cell Perturbation Response Prediction

10 views

Abstract

Predicting how cells respond to genetic and chemical perturbations is becoming central to mechanistic biology and drug discovery. However, the space of possible combinations grows quadratically, and identifying which pairs interact non-additively before measuring them remains an open problem. To address this, we introduce OpPert, a new deep learning framework that models each perturbation as a norm-preserving rotation of the cell's latent representation. To model combinations of two perturbations, their rotations compose into a pair prediction via Baker-Campbell-Hausdorff, and the resulting Lie bracket gives their interaction strength. This interaction score is computed from single-perturbation operators alone, without supervision from the candidate pair, and we show it correlates with measured non-additivity. To infer effects of unseen perturbations, a flow-matching model generates their rotations, with sampling variance serving as a confidence estimate. We validate OpPert computationally on known benchmarks and experimentally in the wet lab. Across held-out evaluations on three genetic screens (Norman, K562, RPE1), OpPert achieves 81-88% directional accuracy, above existing approaches. We further evaluate OpPert on a small-molecule screen and validate predicted Belinostat targets by RT-qPCR in MCF7 cells, elucidating apoptotic regulators never previously associated with this drug.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Machine Learning, Single Cell, Computational Biology

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_DM9HOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method based on norm-preserving rotation to learn cell representations and predict perturbation effects. The results look promising, and the analysis is very interesting. The authors also demonstrate their prediction effect based on wet-lab validation with a control group.

Critical strengths

This paper proposes a method based on norm-preserving rotation to learn cell representations and predict perturbation effects. The results look promising, and the analysis is very interesting. The authors also demonstrate their prediction effect based on wet-lab validation with a control group.

Critical weaknesses

I have some questions to check:

  1. The motivation is not clear to me. Choosing from matching does not seem like a very strong method novelty in this area, and I also cannot see the strength directly. I am also not quite confident about the rotational operator, and it seems like cherry-picked geometric modelling, lacking solid theoretical support.
  2. How does preserving latent norm ensure that rotated cells remain within the decoder’s training distribution, especially if that distribution is anisotropic or contains distinct cell-state clusters?
  3. Because rotations generally satisfy RpRq≠RqRpR_pR_q \neq R_qR_p, how does OpPert choose an ordering for simultaneous perturbations, and how sensitive are predictions to reversing that order?
  4. Since training includes measured combinations through the composition loss, how well does the Lie-bracket interaction score perform when the entire model is trained exclusively on single-perturbation data?
  5. How are the top-20 differentially expressed genes selected for uncertainty estimation on unseen perturbations? Does the confidence score remain useful when those genes must be selected without access to measured test responses?
  6. What additional experiments would establish whether the Belinostat-associated expression changes in PDE4D, PRKCA, and PRKCB contribute causally to apoptosis, rather than reflect downstream stress responses?
  7. More ablation studies and baselines should be included to defend the strength of this method; e.g., other methods might also produce wet-lab-strong results.
Questions for authors

See the weakness above

4 comments
AuthorsOpenReview ↗

Comment (1/4): Motivation and Theoretical Support for the Rotation

Dear Reviewer DM9H,

Thank you for the positive and constructive review. We address each point below, beginning with the motivation and the theory behind the rotation. Line references [L] point to the submitted paper and [R] to the revised paper.

Motivation and novelty

…The motivation is not clear to me. Choosing from matching does not seem like a very strong method novelty in this area, and I also cannot see the strength directly.

  • The main novelty lies in the rotation operator and the interaction score it provides, whereas the flow-matching head serves to generate rotations for perturbations unseen in training. The motivation is to predict unmeasured combinations, since the number of possible pairs grows quadratically, and we therefore need to find the pairs most likely to interact before measuring them [L012-014, L036-038].
  • Our first contribution is the perturbation operator, which represents each perturbation as a norm-preserving rotation and lets two perturbations compose algebraically, and the second is the bracket score, an interaction score computed from the single perturbations alone [L078-085]. The flow-matching head is the third contribution, which maps biological descriptors to rotation generators for unseen perturbations and provides a confidence signal through its sampling variance, and the fourth tests whether predictions transfer to an independent experimental modality through RT-qPCR in Belinostat-treated MCF7 cells [L086-092].

Why the rotation, and its theoretical support

…I am also not quite confident about the rotational operator, and it seems like cherry-picked geometric modelling, lacking solid theoretical support.

  • This is a fair concern, since one could choose a geometric operator to fit the data. However, this cannot happen in our case, as the rotation follows from three requirements on a perturbation operator, namely that it should i) act on each cell's own latent state, ii) leave the norm of that state unchanged and iii) vary continuously from the identity as the effect of a perturbation grows. The linear maps that preserve the norm of every input are exactly the orthogonal maps, and those reachable continuously from the identity are exactly the rotations, which therefore form the only linear family that meets all three requirements.
  • For a generator AA acting on a latent state zz, with symmetric part S=(A+A⊤)/2S=(A+A^\top)/2, we have ddt∣etAz∣2=2(etAz)⊤S(etAz)\frac{d}{dt}\\|e^{tA}z\\|^2=2(e^{tA}z)^\top S(e^{tA}z), and the norm of every state is therefore conserved exactly when AA is skew-symmetric. The generator therefore lies in the Lie algebra so(n)\mathfrak{so}(n), a vector space that a flow can predict and in which two generators compose by the BCH formula with the bracket as its first correction [L207-208].
  • In addition, a rotation does not amplify prediction errors, since two skew-symmetric generators AA and A′A' satisfy ∣eAz−eA′z∣≤∣A−A′∣∣z∣\\|e^Az-e^{A'}z\\|\leq\\|A-A'\\|\\|z\\|. An error in a predicted generator therefore reaches the decoder input at most scaled by the latent norm, whereas a general linear generator can amplify it further by up to emax⁡(μ(A),μ(A′),0)e^{\max(\mu(A),\mu(A'),0)}, with μ\mu the largest eigenvalue of its symmetric part.
  • The paper establishes the remaining properties formally. Lemma 3 fixes the norm of the decoder input [L966-968], Theorem 5 gives convergence of the BCH series for small generators [L980-984], and Proposition 9 bounds the interaction score, with zero exactly when the two generators commute [L1026-1027]. In practice, the learned rotations are small, with a median eigenangle of 0.114 rad [L205-207].
  • Therefore, OpPert offers a new way to model a perturbation. Existing methods decompose the effect of a perturbation into a shift that they add alike to the latent state of every cell, which gives them no inductive bias for the state of the cell [L041-044]. In contrast, OpPert models each perturbation as an operator that acts on the state of each cell. Rotations already serve as such operators in other areas of machine learning, since RotatE models each relation of a knowledge graph as a rotation, which lets relations compose [1], and rotary position embeddings encode relative position in transformers through rotations [2]. In recurrent networks, orthogonal and unitary weights avoid vanishing and exploding gradients [3], and the exponential map of a skew-symmetric matrix, which OpPert applies to each block, parametrises them without constraints [4].

[1] Sun et al. ICLR (2019).

[2] Su et al. Neurocomputing (2024).

[3] Arjovsky et al. ICML (2016).

[4] Lezcano-Casado et al. ICML (2019).

AuthorsOpenReview ↗

Comment (2/4): Alternatives Tested and the Norm Constraint

Continuing from our first comment, we address the alternatives we tested and the norm constraint.

Alternatives we tested

…More ablation studies and baselines should be included to defend the strength of this method; e.g., other methods might also produce wet-lab-strong results.

  • To test the rotation against the operators one could choose instead, we trained a general linear control, GL(4)32\mathrm{GL}(4)^{32}, which has 512 parameters per perturbation against 192 for the rotation. Each block uses a full matrix generator and its matrix exponential, and to hold everything else fixed, we kept the encoder, decoder, descriptor encoder, losses, optimiser, schedule and split, and retrained the rotation backbone with the same code. For each operator, we trained three backbones differing in data order and dropout and selected each on the 13 validation perturbations. We report the mean ± population SD over these runs, each averaged over three flow seeds.

    Norman top-20 DA in percent under the same training procedure

    OperatorParameters26 singles79 pairs
    General linear51276.6 ± 4.591.3 ± 1.0
    OpPert rotation19282.4 ± 2.093.0 ± 0.3
  • As the table above shows, the general linear control scores below the rotation in each of the three training runs despite 2.7 times the parameters, and a two-sided signed-rank test over the 105 perturbations gives p=4×10−6p = 4\times10^{-6}. Its accuracy also varies more between runs, whereas the rotation reaches nearly the same accuracy in every run. The gap is largest on the held-out singles, where the flow predicts every generator for an unseen gene, which agrees with the error bound of our first comment.

  • Moreover, Section 4.6 replaces the rotation with the vector-additive latent shift of scGen, CPA and GEARS under the same encoder, decoder, flow and descriptor encoder, and Table 1 reports a rotation gain of 13 to 16 points in directional accuracy [L338-339, L450-457]. Within the rotation family, Table 11 finds the highest pair accuracy at r=4r=4, since r=2r=2 makes every bracket vanish and larger blocks add parameters that the 70 training singles cannot supervise [L1594-1602].

  • In addition, we train and evaluate every learned method in Table 1 on the GEARS-standard splits under one pipeline, and Appendix F reports the sensitivity to nine further hyperparameters [R288-289, L462-466]. The RT-qPCR tests whether OpPert's predictions transfer to an independent experimental modality, rather than comparing methods, which Table 1 does on held-out data [L419-421].

Prompted by this comment, in the revised manuscript we have now added the general linear control to Appendix N.3 and Table 28 [R2299-2318]. The code is in the supplementary, where the rotation_general_linear setting in oppert/model.py switches every block to GL(4), and scripts/score_gl4_control.py reproduces the table from results/norman_gl4_control_per_perturbation.csv.

Norm and the decoder's training distribution

…How does preserving latent norm ensure that rotated cells remain within the decoder's training distribution, especially if that distribution is anisotropic or contains distinct cell-state clusters?

  • We also considered this case when we chose the operator, and in our setting the learned rotations are small, with nearly all eigenangles below π/4\pi/4, which limits how far a rotation moves a latent state from its starting point [L319-321]. In addition, a rotation keeps the norm of every latent state zz it acts on, ∣Rpz∣=∣z∣\\|R_pz\\|=\\|z\\| as Lemma 3 shows [L966-968]. However, in general, norm preservation does not constrain the direction of a latent state, as the anisotropic and clustered cases in your question show.

In the revised manuscript, we have now stated the guarantee in terms of the norm of the decoder input, in the introduction, Section 3.2, the contribution statement, the Figure 2 caption and Appendix B.5, to be more precise and avoid confusion [R049, R079, R148-149, R178-179, R1026-1030].

AuthorsOpenReview ↗

Comment (3/4): Pair Order and Training on Single Perturbations

Continuing from our second comment, we address the order of the two genes and training on single perturbations.

Order of simultaneous perturbations

…Because rotations generally satisfy (R_pR_q \neq R_qR_p), how does OpPert choose an ordering for simultaneous perturbations, and how sensitive are predictions to reversing that order?

  • This is an interesting point for discussion, since simultaneous perturbations have no intrinsic order and exchanging the two genes reverses the sign of their Lie bracket [θp,θq][\theta_p,\theta_q], as [θq,θp]=−[θp,θq][\theta_q,\theta_p]=-[\theta_p,\theta_q]. However, this cannot change the interaction score in our case, as Eq. 4 uses only the norm of the bracket, which is the same in either order [L213-220]. For the prediction itself, we use alphabetical gene order during training and evaluation, which defines one prediction per pair.

  • To measure the sensitivity you ask about, we exchanged the genes in each held-out Norman pair with the checkpoint and flow seeds unchanged, and to isolate the bracket contribution, we restored alphabetical order and set s=0s=0 at inference [R1766-1770, R2283-2289]. The table reports directional accuracy, DA, on the top differentially expressed genes, together with the pp-value of a two-sided Wilcoxon signed-rank test over the 79 pairs, each averaged over the three flow seeds and excluding zero differences.

    Norman pair DA in percent from one trained model

    DE genesOpPertExchanged orders=0s=0OpPert minus exchangedpOpPert minus s=0s=0p
    Top-2092.6292.4592.81+0.170.56−0.190.48
    Top-5092.4892.2292.42+0.260.50+0.060.92
  • As the table above shows, neither change makes a significant difference to pair DA at either cutoff within this trained model, which suggests that the ordering convention and the bracket scale have no significant effect on the accuracy that OpPert reports, consistent with Section 3.3 [R213-214]. We expect this, since both changes modify only the second-order term of the generator, which is small for most learned blocks [L205-212, L319-321]. Moreover, setting s=0s=0 leaves the interaction score of Eq. 4 unchanged, because the score contains no ss [L1722-1723].

Prompted by this comment, in the revised manuscript we have now i) stated the alphabetical order in Section 3.3, ii) reported the paired test of s=0s=0 in the caption of Table 17, and iii) added both paired tests to Appendix N.2 and Table 27 [R214-216, R1766-1770, R2283-2298].

Training on singles only

…Since training includes measured combinations through the composition loss, how well does the Lie-bracket interaction score perform when the entire model is trained exclusively on single-perturbation data?

  • In OpPert, training uses the 49 measured combinations of the GEARS training split, which calibrate ss through the training pair generator of Appendix N.6, whereas we hold out the 79 evaluated pairs, which never contribute to this gradient [L1374-1375, R254-261, R2353-2363]. We therefore compute each score without supervision from the candidate pair, as the abstract states [L018-019].

  • As suggested, we also retrained the backbone with every training combination cell removed, which leaves the 70 training singles and the control cells, with the rest of the training procedure unchanged. The bracket scale then stays at its initial value of 1, since only combination cells train it.

    Norman top-20 DA in percent with and without training pairs

    Training data26 singles79 pairs
    Singles only81.0 ± 0.592.1 ± 0.3
    Singles and the 49 training pairs82.4 ± 2.093.0 ± 0.3
  • As the table above shows, the model trained without any pair reaches nearly the same accuracy on the 79 held-out pairs, and a two-sided signed-rank test over pairs finds no significant difference (p=0.52p = 0.52). Therefore, OpPert predicts held-out pairs without any pair in training. In this model, BCH composes every held-out pair from its two single rotations with the analytic bracket scale s=1s = 1, which shows that the composition rule itself, rather than pair supervision, produces the pair predictions. The bracket score of Eq. 4 needs only the two single generators, which the model trained on singles also provides [L213-220].

In the revised manuscript, we have now added this training on single perturbations to Appendix N.4 and Table 29 [R2319-2332].

AuthorsOpenReview ↗

Comment (4/4): Confidence Genes, Causal Role and Ethics

Continuing from our third comment, we address the genes of the confidence score, the causal role of the genes and the wet-lab ethics.

Genes for the uncertainty score

…How are the top-20 differentially expressed genes selected for uncertainty estimation on unseen perturbations? Does the confidence score remain useful when those genes must be selected without access to measured test responses?

  • In OpPert, we obtain the confidence signal from the flow itself, since the spread of the decoded expression shifts across independent flow draws measures how uncertain the flow is about an unseen perturbation, and Eq. 7 takes their coefficient of variation [L239-242, L376-378, L402-405]. A perturbation with a larger coefficient of variation therefore receives a lower confidence.

  • In Figure 5, we evaluate this score on the top-50 DEGs of each perturbation, which come from the measured response and match the genes on which we compute the directional accuracy shown on its y-axis [R397-400, R414-415]. However, we do not use that response to predict any unseen perturbation, and only the choice of genes for scoring depends on it.

  • To compute the score without any measured held-out expression, as suggested, we recomputed Eq. 7 on all genes for K562 and Norman, and on the 50 genes with the largest predicted change for Norman, since neither set requires measured held-out expression [R2058-2062]. The table reports the Spearman correlation of each score with prediction error, SPCC, the DA of the most confident 10% and 50% of targets, and the rank agreement of each score with the Figure 5 score.

    Confidence score in percent for 270 K562 and 105 Norman targets

    DatasetGene setSPCCDA at 10% coverageDA at 50% coverageRank agreement
    K562Figure 5 top-50 DEGs47.585.383.2100
    K562All genes39.385.081.998.1
    NormanFigure 5 top-50 DEGs30.698.494.5100
    NormanAll genes25.998.293.797.1
    NormanTop-50 predicted25.298.093.597.5
  • As the table above shows, the score retains a positive association with prediction error on every gene set, and the most confident predictions reach the highest accuracy on both datasets, also when we compute the score on gene sets that use no measured held-out expression. In addition, these scores rank the targets almost exactly as the Figure 5 score does.

Prompted by this comment, in the revised manuscript we have now stated the gene set of the score in the Eq. 7 text and the Figure 5 caption, and added this analysis to Appendix N.5 and Table 30 [R397-400, R414-415, R2334-2350], with the per-target scores in the supplementary files uncertainty_per_pert.json and norman_uncertainty_top50_predicted_degs.csv.

Causal role of the expression changes

…What additional experiments would establish whether the Belinostat-associated expression changes in PDE4D, PRKCA, and PRKCB contribute causally to apoptosis, rather than reflect downstream stress responses?

  • That is a good point for discussion, since PRKCA and PDE4D regulate apoptosis in MCF7 cells, where knocking down PKCα induces apoptosis [5] and inhibiting PDE4D triggers apoptosis through cAMP and ER-stress signalling [6]. Their significant change under Belinostat is consistent with an adaptive survival response that accompanies the apoptotic phenotype [L426-430]. Loss-of-function or overexpression of these genes under Belinostat, with an apoptosis readout, would test their contribution directly, whereas our RT-qPCR tests whether OpPert's predictions transfer to an independent experimental modality [L419-421].

Ethics of the wet-lab stage

…Since they have wet lab results, I think it will be helpful to check the wet lab stage and if it fits the ethics rules.

  • The study used the commercial human cell line MCF7, ATCC-HTB-22, authenticated by STR analysis, and it involved no human participants, no patient material and no animals [L2157-2158].

In summary, your detailed comments led to additional analyses, which show that i) a general linear control with more parameters scores below the rotation in every training run, ii) the order of the two genes has no significant effect on pair accuracy, iii) OpPert predicts held-out pairs without any pair in training, and iv) the confidence score ranks predictions without measured held-out expression. We are grateful for these suggestions, which have led us to i) state the norm guarantee and the gene order precisely, ii) add Appendix N and iii) extend the supplementary.

We hope these clarifications and revisions address your concerns, and we would be glad to answer any further questions.

[5] Cui et al. Fundam. Clin. Pharmacol. (2019).

[6] Mishra et al. Clin. Cancer Res. (2018).

Reviewer_ASVuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes that a perturbation should rotate a cell's internal representation rather than shift it. Most current models encode a cell as a point in a learned space and represent a perturbation as a fixed step in some direction. OpPert instead represents each perturbation as a rotation of that point. The authors argue rotations have two advantages. They keep predictions from drifting into territory the model never saw. And two rotations can be combined with a well-known formula from Lie theory, whose correction term (the "bracket") should flag gene pairs that interact rather than simply add up.

To handle perturbations the model has never seen, a flow-matching model generates the rotation from a description of the gene or drug. Running it several times gives a spread that serves as a confidence estimate. The authors test on three CRISPR screens and one drug screen, compare against eleven methods, and add a small qPCR experiment on Belinostat-treated MCF7 cells. The headline results are that OpPert gets the direction of change right for 81 to 88% of the top responding genes, which beats every baseline. Swapping rotations for ordinary shifts costs 13 to 16 points. The bracket score correlates moderately with how non-additive real gene pairs are (Rho = 0.49).

Critical strengths
  • Treating perturbations as rotations is an elegant idea. It gives a natural way to combine two perturbations and a pair score you can compute before running the experiment. The most convincing evidence is the authors' own comparison: with everything else held fixed, replacing rotations with ordinary shifts drops accuracy by 13 to 16 points on all three CRISPR screens and makes the predicted cell populations worse. If that holds up, it is a simple idea other groups could adopt.

  • Four datasets, eleven baselines retrained under one pipeline, five random seeds, population-level metrics, released code, and a wet-lab experiment is more than most papers in this area attempt. The built-in confidence estimate also beats a method designed specifically for uncertainty (PRESCRIBE) on K562. That breadth would make the comparison useful once the problems below are fixed.

Critical weaknesses

Main criticism: the results presented to not match the claims;

  • The title and abstract center on the bracket, the term that is supposed to capture how two perturbations interact. But when the authors switch it off, accuracy on gene pairs barely changes: 92.4% without it, 92.9% with it.
  • The rotation still helps a lot on the two screens that have no gene pairs at all, so the benefit can't come from combining perturbations. Keeping the cell's representation from drifting is not the reason either, because ordinary shifts only drift by about 2% anyway. The more likely explanation is simpler: a rotation lets the effect depend on the starting cell, and it has more parameters than a fixed shift. The paper never tests that.
  • Mathematically, the bracket measures whether the order of two perturbations matters. In these screens both genes are hit at the same time, so order is meaningless. Two perturbations can interact strongly even when the bracket is zero.
  • The Rho = 0.49 correlation is never compared to simple alternatives, such as how strong each single effect is or how similar the two genes' annotations are. Either could plausibly do just as well.
  • The abstract says the qPCR identified "apoptotic regulators never previously associated with this drug." What it shows is that 3 of 4 hand-picked genes changed in the predicted direction. None of the four is a classic apoptosis gene. There are no negative controls and no replicate counts. Belinostat's effect on MCF7 cells was already measured in the training dataset's source screen, and similar HDAC inhibitors are in training.
  1. The paper reports different values for the same quantity in different places:
  • For pairs where neither gene was seen in training, the main text says 68% while Table 20 says 93.4%.
  • It is defined on the top 20 genes in some sections and the top 200 in others.
  • SciPlex appears 79.3, 76, and 70.0
Questions for authors
  • Please compare rotations against a general (non-rotation) linear transformation and against a shift with the same number of parameters, all trained identically. That would show whether it's the rotation or just the extra flexibility.
  • Please reconcile the 68% vs 93.4% result, say which accuracy definition each table uses, and report every headline number one consistent way.
  • How were the settings chosen, and why do some baselines score below chance? Was there a separate validation set? How are predicted changes centered relative to the controls? Please also add simple linear baselines: summing the two single-gene effects for pairs, and a ridge regression from the same gene descriptions.
  • Does the confidence score peek at the answer? It is computed on each perturbation's top responding genes. Are those chosen without looking at the held-out data?
  • How does the interaction score compare to simple alternatives? For example, single-effect strength, annotation overlap, or other models' predicted interactions. Does it still work if no gene pairs are used in training? And why 79 pairs rather than 84?
  • Would the average response to the other HDAC inhibitors in training have predicted the same genes?
4 comments
AuthorsOpenReview ↗

Comment (1/4): Why the Rotation Helps and the General Linear Control

Dear Reviewer ASVu,

Thank you for the detailed and constructive review. We address each point below, beginning with why the rotation helps. Line references [L] point to the submitted paper and [R] to the revised paper.

Why the rotation helps

…The rotation still helps a lot on the two screens that have no gene pairs at all…

  • As you note, the benefit on these screens does not come from composition, since a single prediction applies one rotation to each cell [L137-139, L160-161]. In OpPert, this rotation lets the effect depend on the starting cell and conserves its norm [L966-968], and it does not amplify an error in a predicted generator, since two skew-symmetric generators AA and A′A' satisfy ∣eAz−eA′z∣≤∣A−A′∣∣z∣\\|e^Az-e^{A'}z\\|\leq\\|A-A'\\|\\|z\\| for every latent state zz, whereas a general linear generator can amplify it.

Rotation against a general linear map and a shift

…The more likely explanation is simpler: a rotation lets the effect depend on the starting cell, and it has more parameters than a fixed shift. The paper never tests that.

…Please compare rotations against a general (non-rotation) linear transformation and against a shift with the same number of parameters, all trained identically.

  • Following your suggestion, we trained a general linear control, block GL(4)32\mathrm{GL}(4)^{32}, whose effect also depends on the starting cell and which has 512 parameters per perturbation against 192 for the rotation. Each block uses a full matrix generator and its matrix exponential, and to hold everything else fixed, we kept the encoder, decoder, descriptor encoder, losses, optimiser, schedule and split, and retrained the rotation backbone with the same code. For each operator, we trained three backbones differing in data order and dropout and selected each on the 13 validation perturbations. We report the mean ± population SD over these runs, each averaged over three flow seeds.

    Norman top-20 DA in percent under the same training procedure

    OperatorParameters26 singles79 pairs
    General linear51276.6 ± 4.591.3 ± 1.0
    OpPert rotation19282.4 ± 2.093.0 ± 0.3
  • As the table above shows, the general linear control scores below the rotation in each of the three training runs despite 2.7 times the parameters, and a two-sided signed-rank test over the 105 perturbations gives p=4×10−6p = 4\times10^{-6}. Its accuracy also varies more between runs, whereas the rotation reaches nearly the same accuracy in every run. The gap is largest on the held-out singles, where the flow predicts every generator for an unseen gene, which agrees with the error bound above.

  • Moreover, Section 4.6 replaces the rotation with the vector-additive latent shift, which predicts a pair by summing the two single shifts in latent space, under the same encoder, decoder, flow and descriptor encoder, and Table 1 reports a rotation gain of 13 to 16 points in directional accuracy [L338-339, L450-457]. This shift carries at most 128 parameters per perturbation and the general linear control 512. Therefore, neither the dependence on the starting cell nor the parameter count explains the gain of the rotation, since the general linear control has both and still scores lower.

Prompted by this comment, in the revised manuscript we have now added the general linear control to Appendix N.3 and Table 28 [R2299-2318]. The code is in the supplementary, where the rotation_general_linear setting in oppert/model.py switches every block to GL(4), and scripts/score_gl4_control.py reproduces the table from results/norman_gl4_control_per_perturbation.csv.

Settings, validation, below-chance baselines and centring

…How were the settings chosen, and why do some baselines score below chance? Was there a separate validation set? How are predicted changes centered relative to the controls?

  • We list the settings in Appendix C and Table 5, and Appendix F reports the sensitivity to nine further settings [L261, L462-466, L1320-1322]. In addition, we use a separate validation set, since the GEARS split holds 13 validation perturbations apart from the training and held-out sets, and we select the Table 20 model on these 13 alone [R1836-1839].
  • We train and evaluate every learned method in Table 1 on the GEARS-standard splits under one pipeline [R288-289]. Values below chance also appear in published evaluations, since PRESCRIBE reports 32.33 for GEARS on RPE1 [1], close to the 31.7 of our GEARS run [L335], which shows that our pipeline reproduces them. Predicting such unseen perturbations is what OpPert addresses, since its flow predicts a generator for every unseen perturbation from its descriptor [L239-242].
  • We centre each predicted change on the decoded control, gω(Rz)−gω(z)g_\omega(Rz)-g_\omega(z) [L407], where zz is the mean encoded control state, and each measured change on the control mean.

[1] Cheng et al. NeurIPS (2025).

AuthorsOpenReview ↗

Comment (2/4): Bracket, Interaction Score and Training without Pairs

Continuing from our first comment, we address the bracket, the interaction score and training without pairs.

The bracket, pair order and pair accuracy

…The title and abstract center on the bracket, the term that is supposed to capture how two perturbations interact. But when the authors switch it off, accuracy on gene pairs barely changes: 92.4% without it, 92.9% with it.

…Mathematically, the bracket measures whether the order of two perturbations matters. In these screens both genes are hit at the same time, so order is meaningless…

  • We would like to clarify the role of the bracket, since the title names the rotation operators, and the abstract assigns the pair prediction to their composition and the interaction strength to the bracket [L001-002, L016-020].

  • This is an interesting point for discussion, since simultaneous perturbations have no intrinsic order and exchanging the two genes reverses the sign of their Lie bracket [θp,θq][\theta_p,\theta_q], as [θq,θp]=−[θp,θq][\theta_q,\theta_p]=-[\theta_p,\theta_q]. However, this cannot change the interaction score in our case, as Eq. 4 uses only the norm of the bracket, which is the same in either order [L213-220]. For the prediction itself, we use alphabetical gene order during training and evaluation, which defines one prediction per pair.

  • To measure both effects directly, we exchanged the genes in each held-out Norman pair with the checkpoint and flow seeds unchanged, and to isolate the bracket contribution, we restored alphabetical order and set s=0s=0 at inference [R1766-1770, R2283-2289]. The table reports directional accuracy, DA, on the top differentially expressed genes, together with the pp-value of a two-sided Wilcoxon signed-rank test over the 79 pairs, each averaged over the three flow seeds and excluding zero differences.

    Norman pair DA in percent from one trained model

    DE genesOpPertExchanged orders=0s=0OpPert minus exchangedpOpPert minus s=0s=0p
    Top-2092.6292.4592.81+0.170.56−0.190.48
    Top-5092.4892.2292.42+0.260.50+0.060.92
  • As the table above shows, neither change makes a significant difference to pair DA at either cutoff within this trained model, which suggests that the ordering convention and the bracket scale have no significant effect on the accuracy that OpPert reports, consistent with Section 3.3 [R213-214]. We expect this, since both changes modify only the second-order term of the generator, which is small for most learned blocks [L205-212, L319-321]. Moreover, setting s=0s=0 leaves the interaction score of Eq. 4 unchanged, because the score contains no ss [L1722-1723].

In the revised manuscript, we have now i) stated the alphabetical order in Section 3.3, ii) reported the paired test of s=0s=0 in the caption of Table 17, and iii) added both paired tests to Appendix N.2 and Table 27 [R214-216, R1766-1770, R2283-2298].

The interaction score

…Two perturbations can interact strongly even when the bracket is zero.

…How does the interaction score compare to simple alternatives?…

  • That is a good point for discussion, since the bracket tests whether the two latent generators commute, and a zero bracket can therefore coexist with a non-additive measured response. For this reason, the discussion describes the score as non-commuting latent action rather than mechanistic epistasis, and the paper reports its correlation with measured non-additivity within this scope [L081-085, L2117-2119].

Training without pairs

…Does it still work if no gene pairs are used in training?…

  • As suggested, we retrained the backbone with every training combination cell removed, which leaves the 70 training singles and the control cells, with the rest of the training procedure unchanged. The bracket scale then stays at its initial value of 1, since only combination cells train it.

    Norman top-20 DA in percent with and without training pairs

    Training data26 singles79 pairs
    Singles only81.0 ± 0.592.1 ± 0.3
    Singles and the 49 training pairs82.4 ± 2.093.0 ± 0.3
  • As the table above shows, the model trained without any pair reaches nearly the same accuracy on the 79 held-out pairs, and a two-sided signed-rank test over pairs finds no significant difference (p=0.52p = 0.52). Therefore, OpPert predicts held-out pairs without any pair in training.

Prompted by this comment, in the revised manuscript we have now added this training on single perturbations to Appendix N.4 and Table 29 [R2319-2332].

AuthorsOpenReview ↗

Comment (3/4): Confidence Score and gene cut offs

Continuing from our second comment, we address the confidence score and reconcile Figure 4b with Table 20.

Confidence without held-out data

…Does the confidence score peek at the answer? It is computed on each perturbation's top responding genes. Are those chosen without looking at the held-out data?

  • In OpPert, we obtain the confidence signal from the flow itself, since the spread of the decoded expression shifts across independent flow draws measures how uncertain the flow is about an unseen perturbation, and Eq. 7 takes their coefficient of variation [L239-242, L376-378, L402-405].

  • In Figure 5, we evaluate this score on the top-50 DEGs of each perturbation, which come from the measured response and match the genes on which we compute the directional accuracy shown on its y-axis [R397-400, R414-415]. However, we do not use that response to predict any unseen perturbation, and only the choice of genes for scoring depends on it.

  • To compute the score without any measured held-out expression, as suggested, we recomputed Eq. 7 on all genes for K562 and Norman, and on the 50 genes with the largest predicted change for Norman [R2058-2062]. The table reports the Spearman correlation of each score with prediction error, SPCC, and the rank agreement of each score with the Figure 5 score across targets.

    Confidence score in percent for 270 K562 and 105 Norman targets

    DatasetGene setSPCCRank agreement
    K562Figure 5 top-50 DEGs47.5100
    K562All genes39.398.1
    NormanFigure 5 top-50 DEGs30.6100
    NormanAll genes25.997.1
    NormanTop-50 predicted25.297.5
  • As the table above shows, the score retains a positive association with prediction error on every gene set, also when we compute it on gene sets that use no measured held-out expression, and these scores rank the targets almost exactly as the Figure 5 score does.

In the revised manuscript, we have now stated the gene set of the score in the Eq. 7 text and the Figure 5 caption, and added this analysis to Appendix N.5 and Table 30 [R397-400, R414-415, R2334-2350], with the per-target scores in the supplementary files uncertainty_per_pert.json and norman_uncertainty_top50_predicted_degs.csv.

Figure 4b and Table 20

…For pairs where neither gene was seen in training, the main text says 68% while Table 20 says 93.4%.

…Please reconcile the 68% vs 93.4% result, say which accuracy definition each table uses, and report every headline number one consistent way.

  • Thank you for spotting this, and we would like to clarify where the 68% in Figure 4b and the 93.4% in Table 20 come from. As the revised Figure 4b and Section 4.2 show, the 68% bar belongs to the pairs with both genes observed, whereas the pairs with neither gene observed score higher on the same genes [R306-311, R363-365]. In addition, Figure 4b scores only the genes that OpPert and PerturBench share, whereas Table 20 scores the full Norman gene panel, and the table below lists the pairs, genes and cutoff behind each value.

    Evaluation behind each reported value

    SourceCategorynGenesCutoffDA
    Figure 4bUnseen 052,382 sharedTop-2068
    Figure 4b pairsUnseen 05Full panelTop-2098.0
    Figure 4bUnseen 272,382 sharedTop-2092
    Submitted Table 20Unseen 217Full panelTop-20093.4 ± 1.7
    Revised Table 20Unseen 217Full panelTop-20091.6 ± 0.2
    Revised Table 20Unseen 217Full panelTop-2089.9 ± 0.1
  • As the table above shows, the 68% and the 93.4% describe different categories, and the pairs with neither gene observed score similarly under every evaluation, including the top-200 cutoff that Appendix G.8 states for the submitted Table 20 [L1897, R1841-1847]. In addition, the five pairs behind the 68% reach 98.0% on the full gene panel, which shows that this value reflects the shared gene set rather than the pairs.

  • Both values use directional accuracy, DA, as Section 4.1 defines it, and we use the top-20 DE genes for the main comparison [L280-282], since not all genes show regulation in response to a perturbation [2]. The supplementary file norman_per_perturbation_scores.csv lists the score, DE-gene cutoff and setting of every target and reproduces Table 20 and the tests of our second comment.

Prompted by this comment, in the revised manuscript we have now i) updated the group labels of Figure 4b and Section 4.2, ii) reported each category of Table 20 at the top-20, top-50 and top-200 cutoffs, and iii) stated the DE-gene cutoff of each result [R306-311, R363-365, R1465-1466, R1836-1846].

[2] Nadig et al. Nat. Genet. (2025).

AuthorsOpenReview ↗

Comment (4/4): In-vitro validation and biological insights

Continuing from our third comment, we address the remaining values and the RT-qPCR experiment.

Norman counts

…And why 79 pairs rather than 84?

  • We score every held-out perturbation whose genes all have a gene descriptor, which leaves 79 of the 82 held-out pairs of the GEARS split, since 1+IGDCC3+TGFBR2, 1+ZBTB10 and 2+ZBTB10 each contain a numeric token rather than a gene [R1442].

In the revised manuscript, we have now updated Table 6 to the 108 held-out perturbations of the GEARS split and stated the 105 scored perturbations in the dataset description [R275, R1477-1480].

SciPlex3 values

…SciPlex appears 79.3, 76, and 70.0

  • The 79.3 and the 76 come from two evaluations with different purposes. Table 2 reports the benchmark protocol of Table 1, directional accuracy on the top-20 DE genes [L280-282, L378], whereas Figure 6B reports the 15 held-out drugs at 10 µM, the dose of our RT-qPCR experiment [L425, L443-446].

Prompted by this comment, in the revised manuscript we have now removed the bootstrap diagnostics of the submitted Figure 10, which gave the 70.0, to report each SciPlex3 value under one of these two protocols [R378, R445-447].

RT-qPCR design and controls

…What it shows is that 3 of 4 hand-picked genes changed in the predicted direction. None of the four is a classic apoptosis gene. There are no negative controls and no replicate counts.

  • In our RT-qPCR, three of the four selected genes changed significantly in the predicted direction, whereas ACSBG1 did not reach significance [L426-429, Figure 6A]. We selected the four candidates by their predicted differential expression under Belinostat in MCF7 cells, and the RT-qPCR tests whether these predictions transfer to an independent experimental modality [L419-425].
  • PRKCA and PDE4D regulate apoptosis in MCF7 cells, since knocking down PKCα induces apoptosis [3] and inhibiting PDE4D triggers apoptosis through cAMP and ER-stress signalling [4]. Their significant change under Belinostat is therefore consistent with an adaptive survival response that accompanies the apoptotic phenotype, which is why the abstract describes them as regulators of apoptosis rather than as core apoptotic genes [L027, L426-430].
  • In addition, every measurement compares Belinostat with the DMSO vehicle control, and the Methods specify the cell authentication, treatment, normalisation and statistical test [L425, L2157-2158, L2182, L2203-2204, L2210, L2212-2215]. We measured each gene in three replicates per condition (n = 3).

In the revised manuscript, we have now added this count to the Methods [R2254-2255].

Belinostat and the training drugs

…Belinostat's effect on MCF7 cells was already measured in the training dataset's source screen, and similar HDAC inhibitors are in training.

…Would the average response to the other HDAC inhibitors in training have predicted the same genes?

  • As you note, SciPlex3 measures Belinostat in MCF7 cells. However, this response cannot inform our prediction in our case, as Belinostat is one of the 15 held-out drugs of Figure 6B and the model that selected the four genes never sees its cells in training [L443-446]. Figure 6B scores its prediction against the measured SciPlex3 response at 92% at 10 µM, above the mean of 76% across held-out drugs, and the RT-qPCR measures the four candidates in a separate culture under an independent modality [L419-421, L425, L445-446].
  • This is an interesting point for discussion, since OpPert uses no drug-class label and its flow predicts the generator of each unseen drug from that drug's own chemical descriptor [L1447-1449]. Predicting a drug that was never measured from its structure alone is exactly the purpose of OpPert, whereas a class mean requires a class label and the measured responses of related drugs. In addition, the rotation lets the response depend on each cell, whereas current methods mostly reproduce the response shared across perturbations [5] and do not yet outperform simple linear baselines [6].

In summary, your detailed comments led to additional analyses, which show that i) a general linear control with more parameters scores below the rotation in every training run, ii) exchanging the two genes leaves the interaction score unchanged and does not change pair accuracy significantly, iii) OpPert predicts held-out pairs without any pair in training, and iv) the confidence score ranks predictions without measured held-out expression. We are grateful for these suggestions, which have led us to i) revise Figure 4b, Table 6 and Table 20, ii) add Appendix N and iii) extend the supplementary.

We hope these clarifications address your concerns, and we would be glad to answer any further questions.

[3] Cui et al. Fundam. Clin. Pharmacol. (2019).

[4] Mishra et al. Clin. Cancer Res. (2018).

[5] Viñas Torné et al. Nat. Biotechnol. (2025).

[6] Ahlmann-Eltze et al. Nat. Methods (2025).

Reviewer_tD53OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

OpPert encodes a cell in a 128-dimensional latent space and represents each perturbation by block-diagonal SO(4) rotations. A descriptor-conditioned flow predicts generators for unseen interventions. The paper uses operator composition and a second-order BCH expression for pairs, proposes a normalized Lie bracket as an interaction score, and derives confidence from variation across flow samples. Experiments cover three genetic screens and SciPlex3, with architectural ablations and an MCF7 Belinostat RT-qPCR study. The central claim is that rotation geometry supports accurate unseen-perturbation prediction and exposes non-additive interactions.

Critical strengths
  1. The rotation parameterization gives a tractable structural constraint

The direct product of small SO(4) blocks avoids a dense 128-by-128 matrix exponential for every perturbation while preserving latent norm exactly. It also provides an explicit algebra in which single-perturbation generators can be compared and combined. That is a concrete modeling choice with testable consequences. The appendix explores block size and coordinate locking, which are more closely connected to the proposed geometry than a generic width sweep. The reported bracket/non-additivity association is also a useful empirical question even if the bracket is ultimately interpreted as a diagnostic rather than a biological mechanism.

  1. The paper includes several forms of empirical checking

The genetic evaluations span Norman, K562 and RPE1, and the chemical task uses fifteen held-out drugs. Stratifying Norman pairs by the number of unseen constituent genes is the right way to expose an easy-versus-hard combination distinction. Table 17 includes an s=0 control, which is important for isolating the bracket. The Belinostat assay adds evidence in another measurement modality: three of four selected expression changes agree significantly with the prediction. These checks give the paper substance, although the transcriptional assay validates selected expression directions and cannot identify direct drug binding targets or establish the proposed interaction algebra.

Critical weaknesses
  1. The written objective cannot learn the bracket coefficient, and the pair action is ambiguous

Equation 3 on p. 4 introduces a learned bracket coefficient s. Appendix B.6 then states that s converges to approximately 2.42 through the combination-cell loss. But Equation 24 on p. 23 trains g(R_p R_q z_control), and Equation 23 defines each R from its single-perturbation generator. Neither expression contains s. The reconstruction, DEG and adversarial terms in the stated Stage 1 objective do not contain it either. Consequently the written data loss has derivative zero with respect to s. Adding s to AdamW does not supply a response-dependent learning signal. The authors need to show which actual forward computation produced the reported fitted value.

This ambiguity propagates to inference. Exact R_p R_q multiplication and exp(theta_p + theta_q + s[theta_p,theta_q]/2) are different operations when the BCH series is truncated and its bracket coefficient is changed. For simultaneous gene perturbations there is a further issue: the intervention {p,q} has no intrinsic order, but reversing the operands changes the sign of the bracket. A canonical ordering, an averaging rule, or a specified sequential-intervention task is needed to define the prediction. None is apparent in the described procedure. Showing predictions after swapping pair labels would be a direct check.

The ablation does not establish a large predictive contribution from the bracket. Table 17 reports combination DA of 92.4 +/- 1.4 for s=0, 92.6 +/- 1.5 for s=1 and 92.9 +/- 1.5 for learned s. The 0.5-point difference requires a paired analysis; overlapping standard deviations are not a significance test, but neither do these values explain the much larger advantage over a different additive predictor. That comparison also changes the representation and decoder.

Some supporting mathematical claims are overstated. Figure 2 says norm preservation bounds decoded predictions to the training distribution. A rotation can move a point from a thin training ellipse into a direction never observed while keeping its norm unchanged. Lemma 7 is false globally: in so(4), set Omega=2piJ12 and Phi=0.3*J23, where Jij is the skew-symmetric generator for coordinate plane (i,j). Then exp(Omega)=I commutes with exp(Phi), although the generator bracket is nonzero. The lemma could be replaced by a suitably local statement. Orthogonality remains valid; the stronger distribution-preservation and global interaction arguments do not follow from it.

  1. The accuracy record is internally unresolved and the confidence score may require the unknown response

The evaluation discrepancy is concrete. Section 4.1 specifies directional accuracy on the top-20 DEGs, whereas Appendix D.5 and the later ablation/analysis sections use top-200. Section 4.2 and Figure 4 report about 68% DA when neither constituent gene was observed; Table 20 reports 93.4 +/- 1.7 for seventeen pairs in that category. The main figures label the Norman test set as 105 perturbations, while Table 6 lists 110, comprising 26 singles and 84 pairs. Eligibility filtering or different scoring cutoffs may explain some differences. The manuscript does not identify which target lists and configurations belong to which number. The hardest generalization category cannot be assessed from these incompatible summaries.

The confidence result has a different problem. Equation 7 on p. 8 restricts the coefficient of variation to each perturbation's top-20 DE genes. Appendix D.5 defines DEGs from measured treated-versus-control expression. If that measured held-out gene set is used in Equation 7, the score cannot be computed before observing the experiment whose reliability it is supposed to assess. The issue concerns the confidence calculation; it is not evidence that the expression predictor itself was given the held-out response.

Appendix I.3 instead refers to variance of the sampled generators. That statistic is available without the response, but it is not Equation 7. The paper needs to identify which one generated Figure 5 and rerun the prospective analysis with a fixed training-only or prediction-derived gene set if necessary. A strong risk-coverage curve is not sufficient when defining the ranking itself may require outcome information. Together with the cohort and cutoff differences, this prevents me from relying on the headline accuracy and confidence claims.

Questions for authors
  1. Please write the implemented pair forward pass and combination loss, including the exact occurrence of s, and show its response-loss gradient or equivalent implementation. Are pairs evaluated as exact products or truncated BCH predictions? What happens if p and q are exchanged, and what is the paired DA difference between learned s and s=0 with the rest of the model fixed?

  2. Please reconcile the seventeen neither-observed pairs at 68% in the main text versus 93.4% in Table 20. A single per-target score file should identify the DEG cutoff, target eligibility, seeds and model setting for every result, including why the Norman count is 105 in the figures and 110 in Table 6. The baseline scores need to use the same declared protocol.

  3. Which uncertainty statistic generated Figure 5: decoded-shift variation on top-20 DEGs, or generator variance? Can it be computed without measured held-out expression? Please provide the confidence-error association and risk-coverage curves using an outcome-independent gene set, with the point predictor and evaluated targets unchanged.

4 comments
AuthorsOpenReview ↗

Comment (1/4): Pair Composition and the Training Signal of the Bracket Scale

Dear Reviewer tD53,

Thank you for the detailed and constructive review. We address each point below, beginning with how OpPert composes a pair and how training learns the bracket scale ss. Line references [L] point to the submitted paper and [R] to the revised paper.

Pair composition during training and evaluation

…Neither expression contains s… The authors need to show which actual forward computation produced the reported fitted value.

…Please write the implemented pair forward pass and combination loss, including the exact occurrence of s, and show its response-loss gradient or equivalent implementation.

…Are pairs evaluated as exact products or truncated BCH predictions?…

  • We would like to first clarify how OpPert predicts a pair and the role of the bracket scale ss. OpPert represents each single perturbation as a rotation of the cell's latent state, and it predicts a pair by combining the two single rotations through the truncated Baker-Campbell-Hausdorff formula of Eq. 3, in which ss scales the bracket term [L207-212]. During training, the model reconstructs every measured combination cell from the pair rotation, and since the reconstruction depends on ss, the reconstruction loss updates ss jointly with the rotation lookup, as Appendix B.6 describes and the gradient check below confirms [L999-1001].

  • As you point out, the submitted Eq. 24 wrote the training term with the exact product RpRqR_pR_q of the two single rotations, which Section 3.3 uses to derive Eq. 3 [L197-212, L1215-1222].

  • Concretely, for a measured combination cell xix_i with perturbations pp and qq, the forward pass forms the training pair generator θ~pq\tilde{\theta}_{pq} from the two dose-scaled single-perturbation generators θp\theta_p and θq\theta_q and their bracket, and then decodes the encoded cell after rotating it by the exponential of that generator, as Eq. 23, Eq. 24 and Appendix N.6 state [R1260-1276, R2353-2363]

    \widehat{x}_i(s)=g_\omega\big(R(\tilde{\theta}_{pq})\\,f_\phi(x_i)\big),$$ where $d_p$ and $d_q$ are the learned dose scales of the two perturbations, $f_\phi$ is the encoder, $g_\omega$ the decoder and $R(\cdot)$ the blockwise exponential that maps a generator to its rotation. We compare $\widehat{x}_i(s)$ with $x_i$ using the gene reconstruction loss of Eq. 22 and its term weighted towards differentially expressed genes. Moreover, the pair generator is the only part of the objective that contains $s$ [L1191-1203, R1270-1276]. At evaluation, we compose the raw generators of each held-out pair through the truncated Baker-Campbell-Hausdorff generator of Eq. 33, $\theta_{pq}(s)=\theta_p+\theta_q+\frac{s}{2}[\theta_p,\theta_q]$, with the stored $s$ [L1994-1997], then rotate the mean encoded control state before decoding.

Prompted by this comment, in the revised manuscript we have now written the pair generator with ss in Section 3.3 and in the forward pass of Eq. 23 and Eq. 24, listed in Eq. 6 every term that the implementation optimises, and stated the dose scaling of the training pair generator in Appendix N.6, to show explicitly how the training objective depends on ss [R199-216, R254-261, R1260-1276, R2353-2363].

Response-loss gradient

…Adding s to AdamW does not supply a response-dependent learning signal.

  • Since the reconstruction loss depends on ss through the pair generator, as described above, we verified this dependence directly, as you suggest, by varying only ss on the same training combination cells with every optimiser step disabled. The table reports the change in reconstruction loss relative to the fitted value and the gradient of the full Stage 1 objective with respect to ss.

    Gradient check on 512 Norman training combination cells in units of 10−410^{-4}

    Bracket scale ss122.415 fitted3
    Change in gene reconstruction loss6.60.501.7
    Gradient of the Stage 1 objective−9.3−2.70.55.4
  • As the table above shows, the reconstruction loss changes with ss while the optimiser takes no step, and the gradient of the full objective changes sign near the fitted value. Since no other parameter changes during this check, the change in loss comes from ss alone, which establishes a learning signal for ss from the measured responses of the combination cells.

Prompted by this comment, in the revised manuscript we have now added this gradient check to Appendix N.1 and Table 26 [R2268-2282].

AuthorsOpenReview ↗

Comment (2/4): Pair Order and Norm Preservation

Continuing from our first comment, we address the order of the two genes and the bracket scale, followed by norm preservation and Lemma 7.

Pair order and the paired test of the bracket scale

…the intervention {p,q} has no intrinsic order, but reversing the operands changes the sign of the bracket. A canonical ordering, an averaging rule, or a specified sequential-intervention task is needed to define the prediction…

…What happens if p and q are exchanged, and what is the paired DA difference between learned s and s=0 with the rest of the model fixed?

…The 0.5-point difference requires a paired analysis…

  • This is an interesting point for discussion, since simultaneous perturbations have no intrinsic order and exchanging the two genes reverses the sign of their Lie bracket [θp,θq][\theta_p,\theta_q], as [θq,θp]=−[θp,θq][\theta_q,\theta_p]=-[\theta_p,\theta_q]. However, this cannot change the interaction score in our case, as Eq. 4 uses only the norm of the bracket, which is the same in either order [L213-220]. For the prediction itself, we use alphabetical gene order during training and evaluation, which defines one prediction per pair.

  • Following your suggestion, we exchanged the genes in each held-out Norman pair with the checkpoint and flow seeds unchanged, and to isolate the bracket contribution, we restored alphabetical order and set s=0s=0 at inference [R1766-1770, R2283-2289]. The table reports directional accuracy, DA, on the top differentially expressed genes, together with the pp-value of a two-sided Wilcoxon signed-rank test over the 79 pairs, each averaged over the three flow seeds and excluding zero differences.

    Norman pair DA in percent from one trained model

    DE genesOpPertExchanged orders=0s=0OpPert minus exchangedpOpPert minus s=0s=0p
    Top-2092.6292.4592.81+0.170.56−0.190.48
    Top-5092.4892.2292.42+0.260.50+0.060.92
  • As the table above shows, neither change makes a significant difference to pair DA at either cutoff within this trained model, which suggests that the ordering convention and the bracket scale have no significant effect on the accuracy that OpPert reports, consistent with Section 3.3 [R213-214]. We expect this, since both changes modify only the second-order term of the generator, which is small for most learned blocks [L205-212, L319-321]. Moreover, setting s=0s=0 leaves the interaction score of Eq. 4 unchanged, because the score contains no ss [L1722-1723].

Prompted by this comment, in the revised manuscript we have now i) stated the alphabetical order in Section 3.3, ii) reported the paired test of s=0s=0 in the caption of Table 17, and iii) added both paired tests to Appendix N.2 and Table 27 [R214-216, R1766-1770, R2283-2298].

Norm preservation

…Figure 2 says norm preservation bounds decoded predictions to the training distribution. A rotation can move a point from a thin training ellipse into a direction never observed while keeping its norm unchanged.

  • We also considered this case when we chose the operator, and in our setting the learned rotations are small, with nearly all eigenangles below π/4\pi/4, which limits how far a rotation moves a latent state from its starting point [L319-321]. In addition, a rotation keeps the norm of every latent state zz it acts on, ∣Rpz∣=∣z∣\\|R_pz\\|=\\|z\\| as Lemma 3 shows [L966-968]. However, in general, norm preservation does not constrain the direction of a latent state, as your example of a thin ellipse shows.

Prompted by this comment, in the revised manuscript we have now stated the guarantee in terms of the norm of the decoder input, in Section 3.2, the contribution statement and the Figure 2 caption, to be more precise and avoid confusion [R079, R148-149, R178-179].

Lemma 7

…The lemma could be replaced by a suitably local statement. Orthogonality remains valid…

  • Thank you for providing this interesting example with Ω=2πJ12\Omega=2\pi J_{12}, which assumes a generator that completes a full rotation, and since its exponential is the identity, this rotation commutes with every other rotation, whereas Ω\Omega itself does not commute with every generator. However, such a generator does not arise in our case, as every block of the generators we fitted on the Norman dataset has eigenangles inside (−π,π)(-\pi,\pi), with nearly all below π/4\pi/4 [L319-321, R1075-1076].
  • Under this condition, the principal logarithm [1, Theorem 1.31] writes each generator as a polynomial in its rotation, and commuting rotations therefore have commuting generators.

In the revised manuscript we have now stated this eigenangle condition in Lemma 7 and based its proof on the principal logarithm [R1068-1076].

[1] Higham. SIAM (2008).

AuthorsOpenReview ↗

Comment (3/4): Norman Counts and Gene Cutoffs

Continuing from our second comment, we reconcile the pair results and address the evaluation protocol.

Figure 4b and the full set of pairs

…Section 4.2 and Figure 4 report about 68% DA when neither constituent gene was observed… Table 20 reports 93.4 +/- 1.7 for seventeen pairs in that category…

  • Thank you for spotting this, and we would like to clarify where the 68% in Figure 4b and the 93.4% in Table 20 come from. As the revised Figure 4b and Section 4.2 show, the 68% bar belongs to the pairs with both genes observed, whereas the pairs with neither gene observed score higher on the same genes [R306-311, R363-365]. In addition, Figure 4b scores only the genes that OpPert and PerturBench share, whereas Table 20 scores the full Norman gene panel, and the table below lists the pairs, genes and cutoff behind each value.

    Evaluation behind each reported value

    SourceCategorynGenesCutoffDA
    Figure 4bUnseen 052,382 sharedTop-2068
    Figure 4b pairsUnseen 05Full panelTop-2098.0
    Figure 4bUnseen 272,382 sharedTop-2092
    Submitted Table 20Unseen 217Full panelTop-20093.4 ± 1.7
    Revised Table 20Unseen 217Full panelTop-20091.6 ± 0.2
    Revised Table 20Unseen 217Full panelTop-2089.9 ± 0.1
  • As the table above shows, the 68% and the 93.4% describe different categories, and the pairs with neither gene observed score similarly under every evaluation, including the top-200 cutoff that Appendix G.8 states for the submitted Table 20 [L1897, R1841-1847]. In addition, the five pairs behind the 68% reach 98.0% on the full gene panel, which shows that this value reflects the shared gene set rather than the pairs.

In the revised manuscript we have now updated the group labels of Figure 4b and Section 4.2, and Table 20 reports each category over all 79 scored pairs at the top-20, top-50 and top-200 cutoffs, with its model and selection in the caption [R306-311, R363-365, R1836-1846].

…A single per-target score file should identify the DEG cutoff, target eligibility, seeds and model setting for every result…

  • To make each score traceable, the supplementary file norman_per_perturbation_scores.csv lists these fields for every target and reproduces Table 20 and the tests of our second comment.

In the revised manuscript we have now added this file to the supplementary and referred to it in Section 4.2 and in the caption of Table 20 [R309-311, R1836-1839].

Norman counts

…The main figures label the Norman test set as 105 perturbations, while Table 6 lists 110, comprising 26 singles and 84 pairs.

  • We score every held-out perturbation whose genes all have a gene descriptor, which excludes 1+IGDCC3+TGFBR2, 1+ZBTB10 and 2+ZBTB10, since each contains a numeric token rather than a gene, and the table below gives the counts [R1442].

    Norman setSinglesPairsTotal
    Submitted Table 62684110
    Held out in the GEARS split2682108
    Scored2679105

Prompted by this comment, in the revised manuscript we have now updated Table 6 to the 108 held-out perturbations of the GEARS split and stated the 105 scored perturbations in the dataset description [R275, R1477-1480].

Gene cutoffs and baseline protocol

…Section 4.1 specifies directional accuracy on the top-20 DEGs, whereas Appendix D.5 and the later ablation/analysis sections use top-200.

…The baseline scores need to use the same declared protocol.

  • We agree that the baselines need a shared protocol, and we therefore already train and evaluate every learned method in Table 1 on the GEARS-standard splits under one pipeline, with the same held-out targets and the metrics of Section 4.1 [L280-282, R288-289]. Similarly, in Figure 4b we restrict every method, including OpPert, to the shared pairs and genes described above, since comparing methods on different pairs or genes would confound the method with the difficulty of the targets [R306-311, R363-364].
  • We use the top-20 DE genes for the main comparison, since not all genes show regulation in response to a perturbation [2], and a metric on the most strongly changed genes therefore reflects the response to the perturbation rather than the unchanged background [L280-282]. For Figure 5, we use the top-50 DEGs [R397-400, R414-415].

Prompted by this comment, in the revised manuscript we have now stated the DE-gene cutoff of each result, in the Eq. 7 text, the Figure 5 caption, the dataset appendix and Appendix F, whose tables use the top-50 DE genes [R397-400, R414-415, R1465-1466, R1642-1644].

[2] Nadig et al. Nat. Genet. (2025).

AuthorsOpenReview ↗

Comment (4/4): Confidence without Measured Held-Out Expression

Continuing from our third comment, we address confidence without measured held-out expression.

Confidence without measured held-out expression

…Which uncertainty statistic generated Figure 5… Can it be computed without measured held-out expression? Please provide the confidence-error association and risk-coverage curves using an outcome-independent gene set…

  • In OpPert, we obtain the confidence signal from the flow itself without a dedicated uncertainty module, as for each unseen perturbation we integrate independent noise draws conditioned on its descriptor and average the resulting generators to form the prediction [L239-242, L376-378]. Since the flow maps each draw deterministically, the spread of the decoded expression shifts across draws reflects how uncertain the flow is about that perturbation, and Eq. 7 measures this spread as their coefficient of variation [L402-405]. A perturbation with a larger coefficient of variation therefore receives a lower confidence, and we expect its prediction to have a larger error.

  • In Figure 5, we evaluate this score on the top-50 DEGs of each perturbation, which come from the measured response and match the genes on which we compute the directional accuracy shown on its y-axis [R397-400, R414-415]. However, we do not use that response to predict any unseen perturbation, and only the choice of genes for scoring depends on it. In Appendix I.3, we describe the variance of the sampled generators, which is a different statistic from the score of Figure 5 [L1992].

  • To compute the score without any measured held-out expression, as suggested, we recomputed Eq. 7 on all genes for K562 and Norman, and on the 50 genes with the largest predicted change for Norman, since neither set requires measured held-out expression. To isolate gene selection, we kept the predictions, flow samples and targets unchanged [R2058-2062]. The table reports the Spearman correlation of each score with prediction error, SPCC, the DA of the most confident 10% and 50% of targets, and the rank agreement of each score with the Figure 5 score.

    Confidence score in percent for 270 K562 and 105 Norman targets

    DatasetGene setSPCCDA at 10% coverageDA at 50% coverageRank agreement
    K562Figure 5 top-50 DEGs47.585.383.2100
    K562All genes39.385.081.998.1
    NormanFigure 5 top-50 DEGs30.698.494.5100
    NormanAll genes25.998.293.797.1
    NormanTop-50 predicted25.298.093.597.5
  • As the table above shows, the score retains a positive association with prediction error on every gene set, and the most confident predictions reach the highest accuracy on both datasets, also when we compute the score on gene sets that use no measured held-out expression. In addition, these scores rank the targets almost exactly as the Figure 5 score does. Moreover, the risk-coverage curves of Figure 9 remain close to those of Figure 5 at every coverage [R420-422, R2058-2062]. We can therefore use this score to prioritise predictions before measuring the held-out response, since neither gene set requires it.

Prompted by this comment, in the revised manuscript we have now i) stated the gene set of the score in the Eq. 7 text and the Figure 5 caption, ii) named Eq. 7 as the score of Figure 5 in Appendix I.3, iii) added Figure 9 with the score on gene sets that need no measured held-out expression, and iv) added the rank agreement in Appendix N.5 and Table 30 [R397-400, R414-415, R2009-2010, R2058-2062, R2334-2350], with the per-target scores in the supplementary files uncertainty_per_pert.json and norman_uncertainty_top50_predicted_degs.csv.

In summary, your detailed comments led to additional analyses, which show that i) the reconstruction loss trains ss, ii) exchanging the two genes leaves the interaction score unchanged and does not change pair accuracy significantly, iii) the pairs with neither gene observed score similarly under every evaluation, and iv) the confidence score ranks predictions without measured held-out expression. We are grateful for these suggestions, which have led us to i) revise Sections 3.2 and 3.3, Eq. 6, Eq. 24, Lemma 7, Figure 4b, Table 6 and Table 20, ii) add Appendix N with the gradient check, the paired tests and the confidence rank agreement, and iii) add per-target score files to the supplementary.

We hope these clarifications and revisions address your concerns, and we would be glad to answer any further questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Additional Analyses and Clarifications

Dear Reviewers and Area Chair,

We thank you for your reviews and for the analyses you suggested, which we have added to the paper and the supplementary and collected in Appendix N. Several of the points raised are also answered by details that the submitted paper already contained, and we now state these details more clearly in the sections that we summarise below.

  • The rotation outperforms a general linear map. To provide further evidence for skew-symmetric rotations, we trained a general linear control under the same training procedure, which also lets the effect depend on each cell and has more than twice the parameters. It scores below the rotation in every training run, which shows that the accuracy of OpPert comes from the rotation itself rather than from its dependence on the cell or its parameter count, and the supplementary includes its code (Appendix N.3). (Reviewers ASVu\color{blue}{\text{ASVu}}, DM9H\color{orange}{\text{DM9H}})
  • OpPert predicts pairs without training pairs. To test whether OpPert needs measured pairs, we trained it on single perturbations alone, and it predicts the held-out pairs with nearly the same accuracy, which shows that it composes pairs from the singles (Appendix N.4). (Reviewers ASVu\color{blue}{\text{ASVu}}, DM9H\color{orange}{\text{DM9H}})
  • The confidence score needs no held-out data. To test the score without measured held-out expression, we recomputed it on gene sets chosen without that data, where it still separates reliable from unreliable predictions and ranks the targets almost exactly as before (Section 4.4, Appendices I.8 and N.5). (Reviewers tD53\color{red}{\text{tD53}}, ASVu\color{blue}{\text{ASVu}}, DM9H\color{orange}{\text{DM9H}})
  • Training learns the bracket scale. To make the training computation explicit, we wrote the training pair generator with its learned dose scales and added a gradient check, which confirms that the measured combination cells train ss (Section 3.3, Appendices N.1 and N.6). (Reviewer tD53\color{red}{\text{tD53}})
  • The gene order does not change the predictions. To define one prediction per pair, we stated the alphabetical gene order, and paired tests show that the pair predictions stay stable when the two genes are exchanged. The interaction score is the same in either order, since it uses only the norm of the bracket (Section 3.3, Appendix N.2). (Reviewers tD53\color{red}{\text{tD53}}, ASVu\color{blue}{\text{ASVu}}, DM9H\color{orange}{\text{DM9H}})
  • The in-vitro validated drug was never seen in training. Belinostat is one of the held-out drugs, which OpPert predicts from its chemical structure alone without any drug-class label (Figure 6). (Reviewers ASVu\color{blue}{\text{ASVu}}, DM9H\color{orange}{\text{DM9H}})
  • Every pair result under one protocol. To make the pair results directly comparable, Table 20 now reports every category at three cutoffs, with a per-target score file in the supplementary that reproduces it, and we clarified that the 68% bar of Figure 4b belongs to the pairs with both genes observed (Section 4.2, Table 20). (Reviewers tD53\color{red}{\text{tD53}}, ASVu\color{blue}{\text{ASVu}})
  • Counts, cutoffs and replicates for every result. To make every number traceable, we stated the Norman counts of the GEARS split in Table 6, the DE-gene cutoff of each result and the RT-qPCR replicate count (Appendices D and F, Methods). (Reviewers tD53\color{red}{\text{tD53}}, ASVu\color{blue}{\text{ASVu}})
  • Explicit conditions for the theory. To make the theory explicit, we state the guarantee for the norm of the decoder input and the eigenangle condition of Lemma 7, which every learned block satisfies (Section 3.2, Appendix B.7). (Reviewers tD53\color{red}{\text{tD53}}, DM9H\color{orange}{\text{DM9H}})

We hope that these additional analyses and clarifications address your concerns, and we would be glad to answer any further questions.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.