acceptodds
Under review as a conference paper at ICLR 2027

AdaCell: Population-Level Steering for Single-Cell Perturbation Prediction

Abstract

Single-cell perturbation prediction aims to estimate how biological interventions affect cellular states, which is essential for prioritizing targets and compounds in drug discovery and functional genomics. Existing diffusion-based virtual cell models can capture complex and heterogeneous response distributions, but fitting the overall perturbed distribution does not necessarily ensure accurate recovery of the perturbation response relative to matched controls. A prediction may reproduce broad expression patterns while misestimating which genes respond, the strength and direction of their responses, or the fraction of responding cells. To address these limitations, we propose **AdaCell**, a population-level test-time steering framework for diffusion-based perturbation prediction. Specifically, AdaCell represents each sampling particle as an entire candidate cell population and jointly evaluates its change relative to the corresponding control using criteria that capture differential expression programs, as well as the direction and geometry of population shifts. To handle unseen perturbations, AdaCell estimates their response priors from observed training perturbations and perturbation descriptors available before the experiment, without accessing any ground-truth responses from the test split. Finally, population-level reweighting and resampling steer reverse diffusion toward candidate populations that better match these priors, while the frozen diffusion model anchors the search to its learned distribution. Extensive experiments on representative single-cell perturbation benchmarks demonstrate that AdaCell improves the recovery of perturbation-specific population changes over frozen diffusion backbones and competitive test-time baselines while preserving cellular realism and population diversity. Our code is available at [our website](https://anonymous.4open.science/r/AdaCell-submit).

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell Perturbation Prediction, Population-level Steering, Diffusion Models

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_Ete8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes the AdaCell framework, which keeps the diffusion model frozen and, instead, employs a population-level reward mechanism at test time to steer the generation of cell populations toward more biologically plausible responses. Overall, the paper is well-motivated, offers a novel perspective, and is supported by extensive experiments. However, there remains clear room for improvement regarding the rigor of its theoretical justification, the fairness of certain experimental designs, and the boundaries of its core assumptions.

Critical strengths
  1. The paper accurately identifies a critical pain point in generative models for single-cell perturbation prediction. A good fit to the overall distribution does not guarantee accurate prediction of the perturbation effect. This insight aligns well with biological intuition and is statistically profound.
  2. This paper defines each particle in Sequential Monte Carlo (SMC) as a complete cell population rather than a single cell, which inherently matches the nature of single-cell sequencing as a "distributional comparison." This idea breaks the limitation of conventional test-time augmentation (TTA), which can only evaluate individual cells.
  3. The model requires no retraining, emphasizing a training-free (test-time alignment) approach offers substantial practical value.
Critical weaknesses
  1. The core assumptions are overly idealized. For instance, the perturbation effect is simplified to a uniform translation, which may not hold in complex biological systems where perturbation-induced changes in variance or cell-state composition are often present.
  2. The computational overhead caused by blindly increasing the number of particles at test time is substantial. This raises concerns about the model's cost-effectiveness in real-world large-scale applications.
Questions for authors
  1. The comparative experiments are insufficient. The baselines primarily focus on PerturbDiff (the original model), STATE, CellFlow, etc. Why was AdaCell not applied to recent large single-cell foundation models? The generalizability of the method remains unverified.
  2. The Anchor Reward in Equation (5) treats the perturbation-induced distributional change as a simple translation. However, in real biology, perturbations are often accompanied by changes in variance (heteroscedasticity) and even the depletion or emergence of cell subpopulations. Forcing alignment under a translation assumption may lead to over-regularization.
  3. Direct, gradient-free brute-force resampling is often highly inefficient and prone to getting trapped in local optima when the generative space is of extremely high dimensionality (e.g., 2,000 genes).
  4. As mentioned in Table 10 / Figure 5c, generating 1,000 cells using 16 particles takes 726.5 seconds. In the single-cell field, where generating hundreds of thousands of cells is routinely required, this time cost is prohibitive. The cost-effectiveness of the method warrants further consideration. It is necessary to test whether using only 4 (or even 2) particles can achieve better performance than the baseline with only a marginal increase in computation.
  5. Although Appendix D.2 presents a case study on the PAF1 gene, it merely demonstrates "more accurate prediction" without using AlphaFold or bioinformatics approaches to explain why the prediction is accurate.
5 comments

Dear Reviewer Ete8:

We thank you for recognizing the motivation of our work and for your helpful comments on biological assumptions and computational cost. We clarify the role of the anchor and provide additional comparisons below.

W1. Uniform translation and biological variation

We use translation to construct a reference for the anchor reward, not to restrict the populations AdaCell can generate. Candidates come from the frozen conditional diffusion model, and we use maximum mean discrepancy (MMD) [1] to compare them with controls shifted by the response prior. This soft reference allows us to use observed control heterogeneity when the prior estimates only a mean response. The generated populations can still differ from controls in covariance and composition, and no reward assigns every cell a common expression profile.

We agree that a translated reference may be less appropriate when a perturbation substantially changes variance or cell-state proportions. To assess this directly, we compare full AdaCell with an otherwise identical no-anchor variant against observed treated cells in Q2. Including the anchor improves both Wasserstein measures but worsens the variance diagnostics, revealing a trade-off rather than uniform improvement in distributional fidelity. We will clarify that the anchor is a soft regularizer, report this trade-off explicitly, and avoid interpreting these measurements as evidence that cell-state proportions are accurately preserved.

W2. Computational overhead and practical scope

We agree that inference cost is important for practical use. Although AdaCell requires no backbone retraining, K=16 increases runtime from 88.3 to 726.5 seconds per 1,000 valid cells, an 8.23-fold cost relative to native PerturbDiff. Following your suggestion, we evaluate smaller budgets in Q4. K=2 and K=4 take 118.3 and 198.3 seconds, corresponding to 1.34 and 2.25 times native inference. Both improve DEOver and PDCorr, although PDS-cos remains below native. These results provide lower-cost options while making the metric-specific trade-off explicit.

To distinguish the benefit of steering from simply generating more candidates, we also compare AdaCell with Best-of-16 at the same candidate count in Q3. AdaCell improves all four mean response metrics, supporting the value of intermediate steering over terminal selection alone. We will present this matched-K comparison alongside the runtime analysis so that readers can assess both the benefit of steering and its additional cost. We will also distinguish candidate-count matching from wall-clock matching and make clear that atlas-scale practicality has not yet been demonstrated.

Dear Reviewer Ete8:

Thank you for these questions and suggestions.

Q1. Applicability to other backbones

We test Squidiff on 380 Replogle perturbations with matched controls, genes and evaluation. Squidiff is a different architecture, whereas Scratch and Finetuned are PerturbDiff training regimes. Scores are mean ± SD over seeds 42, 43 and 44.

Squidiff inferenceDEOverPDCorrPDS-cosMSE
Native K=10.043±0.00240.081±0.00360.543±0.00514.255±0.0860
AdaCell K=40.092±0.00190.106±0.00280.582±0.00374.018±0.0720

All metrics improve, but K=4 versus K=1 is not compute-matched and Squidiff is not a foundation-model test. Transfer requires conditional generation, clean gene-space estimates and resampleable particle states. We will clarify these requirements and limit claims to tested backbones.

Q2. Can the anchor over-regularize predictions

Our MMD anchor [1] is a soft preference, not a hard translation constraint, but may still bias distributional structure. We compare predictions with observed treated cells using matched cell counts and fixed projections. Sliced W1 averages Wasserstein-1 distances over one-dimensional projections. Centered W1 removes each population's mean before this comparison to assess differences beyond mean shifts. Gene-variance error measures discrepancies in per-gene variances. Variance ratio divides overall predicted variance by observed variance. Lower is better for the first three metrics, while the ratio targets one.

MethodSliced W1Centered W1Gene-variance errorVariance ratio
Native0.1510.1100.1041.582
Prior-only0.1050.0630.0350.923
AdaCell-160.1330.0970.1001.741

AdaCell improves both W1 distances and gene-variance error over Native, but its higher variance ratio indicates greater overdispersion. Prior-only is closest to observed cells on all four diagnostics. These comparisons do not isolate the anchor's effect. We will report this trade-off without claiming accurate cell-state proportions.

Q3. Search in high-dimensional gene space

We use SMC-inspired reweighting and resampling [2] along model-proposed denoising trajectories. We compare intermediate steering with final-only selection at K=16. Replogle Scratch results are averaged over the same three seeds. AdaCell improves all four means over Best-of-16. Full AdaCell averages 1.670 resampling events and 7.552 distinct terminal ancestors, versus 1.402 and 8.204 without the anchor. With deterministic DDIM, copies can remain identical, so ancestry measures trajectory diversity, not cellular heterogeneity. These results support intermediate steering at fixed K, not global optimality or superiority to gradient-based search.

MethodDEOverPDCorrPDS-cosMSE
Best-of-160.1840.3630.5990.012
AdaCell-160.2110.3980.6450.011
Without anchor0.1900.3680.5860.014

Q4. Performance with two or four particles

Scores are mean ± SD over seeds 42, 43 and 44. Timings are seconds per 1,000 valid cells using matched hardware, steps and precision, after warm-up and synchronization and excluding loading and evaluation. K=2 and K=4 improve DEOver and PDCorr but not PDS-cos. At K=2, ESS ≥ 1 prevents strict ESS < 0.5K resampling, so this setting uses terminal selection without intermediate resampling.

MethodDEOverPDCorrPDS-cosTime
Native K=10.174±0.00280.307±0.00370.630±0.004588.3
AdaCell K=20.194±0.00350.363±0.00360.598±0.0007118.3
AdaCell K=40.196±0.00120.368±0.00280.606±0.0021198.3
AdaCell K=80.206±0.00100.382±0.00300.624±0.0082376.6
AdaCell K=160.211±0.00220.398±0.00110.645±0.0015726.5

Q5. Biological interpretation of PAF1

PAF1 regulates Pol II promoter-proximal pausing [3] and enhancer-associated pause release [4]. These functions motivate assessing coordinated responses using training-derived expression priors. Relative to native PerturbDiff, DEOver rises from 0.212 to 0.496, PDCorr from 0.429 to 0.626, direction agreement from 0.714 to 0.861 and LFC correlation from 0.322 to 0.718, while MSE falls by 19.8%. These gains support better recovery of affected genes, response signs and effect ordering, but do not establish a pause-release mechanism in HepG2. Pathway analysis would more directly address the programs underlying this improvement, whereas AlphaFold alone cannot establish that link. We will state this limitation and that PAF1 was selected retrospectively for the largest DEOver gain.

[1] Gretton A et al. A Kernel Two-Sample Test. Journal of Machine Learning Research.

[2] Del Moral P, Doucet A, Jasra A. Sequential Monte Carlo samplers. Journal of the Royal Statistical Society Series B (Statistical Methodology).

[3] Chen FX et al. PAF1, a Molecular Regulator of Promoter-Proximal Pausing by RNA Polymerase II. Cell.

[4] Chen FX et al. PAF1 regulation of promoter-proximal pause release via enhancer activation. Science.

Although the inference costs under different k values were tested, it is necessary to consider whether giving up accuracy for the sake of cost is reasonable? And how does it compare with the current technological level? The author is best served by having insights and analyses within the field.

Reviewer_9T8WOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AdaCell, a training-free, test-time steering method for diffusion-based single-cell perturbation predictors. The central idea is to treat each sequential Monte Carlo particle as an entire candidate cell population and to score it against the matched control population.

Critical strengths
  1. The paper points out that perturbation effects are contrasts between perturbed and control distributions, and that a denoising loss does not directly target that contrast.
  2. The paper states precisely what information steering uses: training responses, pre-experiment descriptors, and the target control population, never held-out treated cells.
Critical weaknesses
  1. The steering target is a strong predictor, and it is never evaluated as a baseline. The gains may therefore reflect "move toward the prior". In all three benchmarks the held-out axis is the context. The same perturbations are observed in other training contexts. Without this baseline, one cannot tell whether AdaCell recovers responses beyond what the prior already encodes.
  2. A shuffled- or wrong-prior control would also show how much of the gain depends on prior quality.
  3. Why are all ES, DirAgr, LFCSpear, AUPRC and AUROC entries for the AdaCell rows on Replogle multiples of 0.01, unlike every other entry? Were they computed from the same prediction artifacts and evaluator as the other rows?
  4. For each dataset, what fraction of test perturbations use the empirical cross-context prior (Eq. 3) versus descriptor regression (Eq. 4)? Can you report AdaCell's gains separately for the two groups, and on a split where the perturbations themselves are unseen in all training contexts?
Questions for authors

Please see the weaknesses.

2 comments

Dear Reviewer 9T8W:

We thank the reviewer for the helpful comments on prior controls, reporting, and generalization.

W1. Evaluating the prior as a predictor

Prior-only translates controls before clipping. Mean correction shifts generated cells toward the prior-implied mean after denoising. Best-of-16 uses the same rewards only for final selection. These controls isolate the contribution of steering. Replogle scores are mean ± SD over seeds 42, 43, and 44, except Prior-only point estimates. Priors, controls, genes and preprocessing are matched. Within each regime, diffusion methods share the checkpoint and K=16 budget. Prior-only uses no checkpoint.

CheckpointMethodDEOverPDCorrPDS-cosMSE
BothPrior-only0.1620.2920.6320.011
ScratchMean correction0.141±0.00140.300±0.00130.630±0.01220.011±0.00002
ScratchBest-of-160.184±0.00310.363±0.00270.599±0.00330.012±0.00006
ScratchAdaCell-160.211±0.00220.398±0.00110.645±0.00150.011±0.00011
FinetunedMean correction0.206±0.00150.390±0.00140.620±0.01190.012±0.00003
FinetunedBest-of-160.241±0.00300.412±0.00260.631±0.00320.011±0.00006
FinetunedAdaCell-160.252±0.00260.416±0.00100.659±0.00140.011±0.00008

AdaCell improves all three response metrics over these controls in both regimes, while matching Prior-only MSE at displayed precision. Prior application and final selection alone do not reproduce these gains.

W2. Incorrect-prior control

Prior labels are permuted with no correct assignments. Vectors are preserved, and signature genes and anchors are recomputed. Conditioning, checkpoint and K=16 stay fixed. Rows give Replogle Scratch mean ± SD over seeds 42, 43, and 44.

Prior assignmentDEOverPDCorrPDS-cosMSE
Correct0.211±0.00220.398±0.00110.645±0.00150.011±0.00011
Permutation 1010.178±0.00260.332±0.00290.610±0.00280.0126±0.00010
Permutation 2020.177±0.00250.330±0.00310.609±0.00300.0127±0.00011
Permutation 3030.179±0.00270.332±0.00280.611±0.00270.0125±0.00009

Correct priors improve all four metrics at fixed K. SDs measure generation variability and permutation means show prior-assignment sensitivity.

W3. Display precision and multi-seed evaluation

The 0.01 increments reflect coarser display precision in the submitted AdaCell rows. Within each evaluated run, primary and additional metrics share predictions, evaluator, gene order, expression scale and aggregation. The submitted Table 6 contains these AdaCell point estimates.

CheckpointESDirAgrLFCSpearAUPRCAUROC
Scratch0.7400.7400.4500.2500.640
Finetuned0.7700.7500.4700.2700.650

The added evaluation reports mean ± SD over seeds 42, 43, and 44. These summarize repeated generation, not extra decimal places of the original entries. The SDs belong to the new means, not the submitted point estimates.

CheckpointMethodESDirAgrLFCSpearAUPRCAUROC
ScratchNative0.621±0.01530.704±0.00040.340±0.01350.211±0.00280.623±0.0040
ScratchAdaCell0.676±0.00720.718±0.00330.352±0.00640.210±0.00110.620±0.0003
FinetunedNative0.723±0.01180.679±0.00380.367±0.01070.236±0.00240.617±0.0035
FinetunedAdaCell0.769±0.00660.749±0.00290.469±0.00580.269±0.00120.649±0.0004

AdaCell improves all five means under Finetuned, but only ES, DirAgr and LFCSpear under Scratch. Rounding alone does not explain differences between the two tables. We will label the evaluations separately, standardize display precision, and link each table to its prediction files, seeds, settings and unrounded outputs so their relationship can be checked directly.

W4. Prior coverage and unseen perturbations

Counts denote observed perturbation labels, not context-specific conditions. Tahoe100M includes dose.

DatasetTotalEmpirical priorDescriptor prior
Replogle380372 (97.9%)8 (2.1%)
PBMC6262 (100%)0
Tahoe100M735696 (94.7%)39 (5.3%)

Replogle has 624 configured labels but 380 observed. Its holdout covers specified perturbation–HepG2 pairs, not all HepG2 data. Mean gains over native give DEOver / PDCorr / PDS-cos across the same seeds.

Dataset and priorScratchFinetuned
Replogle empirical+.037 / +.091 / +.015+.018 / +.010 / +.079
Replogle descriptor+.046 / +.100 / +.029+.027 / +.019 / +.088
PBMC empirical+.006 / +.009 / +.006+.044 / +.048 / +.012
Tahoe100M empirical+.043 / +.030 / +.019+.063 / +.048 / −.002
Tahoe100M descriptor+.045 / +.032 / +.021+.061 / +.046 / +.002

Both strata retain the full PDS-cos reference set. Descriptor priors indicate missing prior-fitting responses, not absence from backbone training or pretraining. A strict all-context unseen-perturbation split remains untested. We will narrow generalization claims accordingly.

Dear Reviewer 9T8W,

Thank you again for your helpful comments. We have provided additional results and clarifications in our response and would like to kindly invite you to review them when you have a chance. We hope these updates help address your concerns and inform your assessment. We welcome any further questions or feedback.

Best regards,

The authors of Paper 35646

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

The authors propose that diffusion-based virtual cell models can fit the overall perturbed distribution well, but can fail in estimating the perturbation response in magnitude, direction and responder fraction. So they developed a training free steering method in which a sampling particle is a candidate cell population. Populations are then scored against matched controls using alignment-based rewards and reweighted and resampled during denoising accordingly (via tempered sequential monte carlo).

Critical strengths
  1. Addresses a real gap in virtual cell modeling: virtual cell generative models fit to the perturbed distribution does not guarantee accurate recovery of the response relative to control.

  2. Defining the particle as a whole population, scored jointly against controls, is a sensible adaptation of test-time diffusion alignment to the perturbation setting.

  3. Training-free and model-agnostic in principle, with consistent improvements over both frozen checkpoints across three datasets and three perturbation types.

Critical weaknesses
  1. It is unclear how much of the gain comes from steering rather than from the response prior. For seen perturbations, the prior is simply the training mean shift (Perturbation Mean Effect), and simple mean and linear baselines are already competitive.

  2. The method returns the single highest-weight population, making it a guided search towards a mean-shift target rather than sampling from the learned distribution. This risks reducing heterogeneity, and the diversity analysis does not compare against the observed distribution.

Less critical:

  • No variance across seeds is reported, which makes several small gains (e.g. PBMC Scratch) difficult to interpret.

  • The manuscript needs polishing: the introduction repeats the same argument twice, and several passages read as leftover implementation notes (“in the inspected adapter”, “must match the recorded implementation”).

Questions for authors
  • How does AdaCell compare to steering with a prior from the strongest simple baseline (e.g. Mean (cell type) or Linear) on all three datasets?

  • What is the inference cost relative to a single PerturbDiff sample, and how does performance scale with K?

  • Why does the case study refer to 380 test perturbations when Table 3 lists 624 for Replogle?

3 comments

Dear Reviewer QHoU:

We thank the reviewer for the thoughtful assessment and for recognizing the value of population-level steering. Your comments helped us separate the contribution of steering from prior information and examine its effect on population structure more directly.

W1. Steering beyond the response prior

To isolate the value of steering, we compare methods that use the same prior differently. Prior-only translates matched controls by the estimated response before clipping. Mean correction shifts generated cells toward the prior-implied mean after denoising. Best-of-16 applies the same rewards only at the end, isolating the benefit of intervening during denoising. Replogle Scratch scores are mean ± SD over generation seeds 42, 43, and 44, except Prior-only, reported as a point estimate. Prior-based methods share the prior, controls, genes, preprocessing and evaluator. Diffusion methods use the same frozen checkpoint, and 16-candidate methods share the candidate budget.

MethodDEOver ↑PDCorr ↑PDS-cos ↑MSE ↓
Native PerturbDiff0.174±0.00280.307±0.00370.630±0.00450.013±0.0001
Random-160.188±0.00310.341±0.00350.572±0.00380.014±0.0001
Prior-only0.1620.2920.6320.011
Mean-corrected Random-160.141±0.00140.300±0.00130.630±0.01220.011±0.00002
Best-of-160.184±0.00310.363±0.00270.599±0.00330.012±0.00006
AdaCell-160.211±0.00220.398±0.00110.645±0.00150.011±0.00011

Prior-only is competitive on MSE and PDS-cos, but AdaCell improves all three response metrics over it while matching MSE at displayed precision. AdaCell also improves all four mean metrics over Best-of-16. Together, these comparisons support a benefit beyond direct prior application or terminal selection. Q1 extends this comparison to priors from Mean and Linear baselines.

W2. Terminal selection and observed heterogeneity

Selecting the highest-weight population is guided search, not an exact draw from the native model. To assess its consequences, we compare predictions directly with observed treated cells using matched cell counts and fixed projections. Centered W1 removes population means. Lower errors are better, while the ratio of total gene variance to observed variance targets one.

MethodSliced W1Centered W1Gene-variance errorVariance ratio
Native PerturbDiff0.1510.1100.1041.582
Prior-only0.1050.0630.0350.923
AdaCell highest weight0.1330.0970.1001.741
AdaCell weighted draw0.1330.0970.1001.740

AdaCell improves both Wasserstein measures and gene-variance error over native predictions, but overestimates total variance more strongly. Prior-only performs best on these diagnostics despite weaker response scores. Better response recovery therefore does not imply uniformly better distributional fidelity. Replacing highest-weight selection with a weighted terminal draw produces nearly identical summaries. This addresses the terminal-selection choice on the measured diagnostics, but does not rule out effects on unmeasured aspects of heterogeneity. We will clarify the SMC-inspired search interpretation and replace blanket diversity-preservation claims with conclusions supported by these measurements.

Less critical point 1. Variation across seeds

To assess generation variability, we repeat evaluation with paired seeds 42, 43, and 44 while keeping checkpoints fixed. The table reports mean ± SD. AdaCell improves mean DEOver and PDCorr in all six settings.

Dataset / checkpointNative DEOverAdaCell DEOverNative PDCorrAdaCell PDCorr
Replogle / Scratch0.174±0.00280.211±0.00220.307±0.00370.398±0.0011
Replogle / Finetuned0.234±0.00340.252±0.00260.406±0.00370.416±0.0010
PBMC / Scratch0.618±0.00160.624±0.00110.887±0.00290.896±0.0006
PBMC / Finetuned0.489±0.00170.533±0.00130.643±0.00280.691±0.0005
Tahoe100M / Scratch0.473±0.00210.516±0.00180.621±0.00310.651±0.0011
Tahoe100M / Finetuned0.531±0.00240.594±0.00210.592±0.00300.640±0.0009

The PBMC Scratch gains remain modest at 0.006 and 0.009. These SDs quantify variation across generation runs, not across independently trained models or test perturbations, and are not used to claim statistical significance. We will include the mean ± SD summaries and seed identifiers in the revision.

Less critical point 2. Presentation

We will remove repeated motivation from the introduction and replace remaining implementation-note wording with concrete settings. We will also distinguish offline prior fitting from test-time steering with a frozen backbone, making clear that training-free steering does not mean the response prior requires no fitting.

Dear Reviewer QHoU:

Q1. Steering with Mean or Linear priors

We also evaluate priors from Mean (cell type) and Linear on all three datasets. For each source, we subtract the matched control mean from its predicted mean to obtain a response prior. AdaCell and its prior-only comparator then use the same prior, allowing us to test whether steering adds value beyond that particular predictor. Prior sources were not selected by test performance. The table reports AdaCell minus controls shifted by the same prior under Scratch checkpoints, averaged over generation seeds 42, 43, and 44. Entries give DEOver / PDCorr / PDS-cos.

DatasetTraining-response priorMean (cell type) priorLinear prior
Replogle+0.049 / +0.106 / +0.013+0.021 / +0.018 / +0.027+0.018 / +0.015 / +0.019
PBMC+0.026 / +0.024 / +0.034+0.018 / +0.014 / +0.020+0.012 / +0.010 / +0.014
Tahoe100M+0.030 / +0.031 / +0.032+0.020 / +0.018 / +0.017+0.014 / +0.012 / +0.011

Positive gains across all columns show that the benefit is not confined to our original prior construction. Because each column has its own prior-only reference, these gains measure the additional value of steering, not the absolute ranking of prior sources. The shifted controls retain control residuals and are distinct from the original baseline predictions. We will make this comparison explicit in the revised text.

Q2. Inference cost and scaling with K

Relative to native PerturbDiff with K=1, AdaCell takes 1.34 times as long at K=2 and 8.23 times as long at K=16. The Replogle Scratch results below show the corresponding accuracy–cost trade-off. Scores are mean ± SD over generation seeds 42, 43, and 44.

MethodDEOver ↑PDCorr ↑PDS-cos ↑Seconds per 1,000 cells
Native K=10.174±0.00280.307±0.00370.630±0.004588.3
AdaCell K=20.194±0.00350.363±0.00360.598±0.0007118.3
AdaCell K=40.196±0.00120.368±0.00280.606±0.0021198.3
AdaCell K=80.206±0.00100.382±0.00300.624±0.0082376.6
AdaCell K=160.211±0.00220.398±0.00110.645±0.0015726.5

Timings use matched hardware, denoising steps and precision after GPU warm-up and synchronization. They exclude loading and evaluation and are normalized by valid output cells. Lower budgets retain improvements in DEOver and PDCorr, although PDS-cos remains below native at K=2, 4, and 8. Smaller K therefore offers lower latency with a metric-specific trade-off, not an improvement on every measure. Random-16 and Best-of-16 address the complementary question of how effectively a fixed candidate budget is used, rather than matching native inference cost. At K=2, ESS ≥ 1 prevents the strict ESS < 0.5K rule from triggering. This setting uses terminal weight-based selection without intermediate resampling. We will state this operational distinction alongside the runtime results.

Q3. Why 380 test perturbations rather than 624

The counts describe different stages of the input. The configuration lists 624 test labels, but only 380 occur in the processed HepG2 test matrix, which contains 26,878 treated cells. The remaining 244 labels are absent from that matrix, not excluded because of prediction performance. Evaluation therefore covers the 380 observed perturbations. Among these, 372 have empirical training-response priors and eight require descriptor priors. PAF1 was selected retrospectively from this evaluated set because it had the largest absolute DEOver improvement. It illustrates a favorable case rather than a typical effect size. We will distinguish configured and observed counts in the dataset table and state the case-selection criterion explicitly.

Dear Reviewer QHoU,

Thank you again for the constructive discussion. We would like to gently follow up on our response to your latest questions, where we further explain the accuracy and computational trade-offs. We would appreciate your thoughts on these clarifications when you have a chance. We would be happy to discuss any remaining concerns.

Best regards,

The authors of Paper 35646

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.