acceptodds
Under review as a conference paper at ICLR 2027

Responder-Weighted Flow Matching for Cell Morphology

Abstract

Generative models of cell morphology, such as CellFlux, learn to turn images of control cells into images of treated cells. They are trained on every imaged cell of a perturbation, although often only some of these cells respond. We show that a perfectly trained flow then reproduces this mixture, so its predicted change of any feature equals the responder effect times the fraction of responding cells. No single classifier-free guidance strength undoes this, because the required strength grows as one over that fraction. We therefore estimate for every treated cell the probability that it responded. A classifier compares the cell with control cells of the same plate and never sees the cell's own well during training. We then weight each cell's flow-matching loss by this probability. Estimated responder fractions in public datasets are often significantly smaller than one, especially CRISPR knockouts. In a benchmark where only one in ten training cells is a real treated cell, a CellFlux-style model recovers none of the perturbation effect on morphology. Responder weighting recovers 83% of it, close to the 89% of an oracle given the true labels. On real data, the proposed weighting mechanism lowers the energy distance to real treated cells by 29% on BBBC021 and 49% on RxRx1. On JUMP, where by our estimate the median perturbation changes fewer than one in five cells, it raises the distance, and only milder weighting improves the fit.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
virtual cell, perturbation prediction, high-content screening, flow matching

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_tKyuOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes responder-weighted flow matching for generating cell morphology images under heterogeneous perturbation responses. It estimates per-cell response scores using a classifier cross-fitted over wells, then weights the flow-matching loss to emphasize likely responders. A controlled dilution benchmark shows recovery of effects lost by standard training, while tests on BBBC021, RxRx1, and JUMP show gains on some datasets but mixed results on JUMP.

Critical strengths

1 The paper gives a clear derivation of mixture-induced effect shrinkage and shows how calibrated response weights change the flow-matching target.

2 The dilution benchmark has known responder labels, and the real-data evaluation includes well-level cross-fitting, multiple metrics, ablations, and the negative JUMP result.

Critical weaknesses

1 The responder model assumes non-responders follow the same distribution as same-plate controls. The paper’s own diagnostics (Table 4) show that low-scoring treated cells remain distinguishable from controls on BBBC021 and RxRx1. The authors consequently interpret the weights as emphasizing more affected cells rather than identifying a clean responder subset. Thus, real-data results support a reweighting heuristic, but do not validate the estimated responder fractions or posterior probabilities as such.

2 The effect of the weighting exponent is dataset-dependent. On JUMP, (\gamma=1) raises the energy distance to the full treated population to 5.88 times the unweighted value, while (\gamma=0.25) improves it (Table 17). The paper does not establish how to choose (\gamma), or whether the intended target for a new dataset is the responder population or the full treated population.

Questions for authors

Given the Table 4 diagnostics, should (\hat{\rho}) and (s(x)) on BBBC021 and RxRx1 be interpreted as responder probabilities, or only as scores that rank cells by perturbation effect? Please clarify which real-data claims rely on each interpretation.

How should a practitioner select (\gamma) for a new dataset without using test-set outcomes, and how should they decide whether to model responders alone or the full treated-cell mixture?

3 comments
AuthorsOpenReview ↗

Weakness 1 and Question 1.

Thank you for your careful review and for recognizing the "clear derivation of mixture-induced effect shrinkage". Following your comments, the revised paper states more clearly how to interpret ρ^\hat\rho and ss on BBBC021 and RxRx1, and it now explains how to choose the target and γ\gamma.

We agree with your point. Our responder model assumes that each treated cell either responds or looks exactly like a control. On BBBC021 and RxRx1, even the treated cells with the lowest scores still differ slightly from controls (Table 4), so there the response is graded rather than all-or-none. In that case ss is not the probability that a cell responded but a score that ranks cells by how strongly they differ from controls, and ρ^\hat\rho counts the cells that differ clearly from controls. On JUMP the same check stays near 0.5, as the assumption predicts. We now state this in Section 3.2 and in the Limitations paragraph of Section 6.

Three statements need calibrated probabilities: that γ=1\gamma=1 learns exactly the responders (Section 3.4), that ρ^\hat\rho is the share of responders (Section 5.1), and that mixture mode rebuilds the treated population (Section 5.5). The gains of weighting on real data need two properties of ss, and neither requires it to be an exact probability: ss must rank the cells well, and it must give control-like cells only the floor weight smin⁡s_{\min}. The table below removes one property at a time. Shuffled weights put the weights of ss on random cells of the same perturbation, which removes the ranking. The raw classifier output hh ranks cells like ss, but gives a control-like cell about half the weight of a clear responder when few cells respond. Values are mean ± standard error over three seeds, with one seed for shuffled weights on BBBC021.

MethodBBBC021 IID, MoA acc. ↑RxRx1, top-5 acc. ↑JUMP, mAP ↑
γ=0\gamma=00.618 ± 0.0010.154 ± 0.0020.148 ± 0.002
Shuffled weights0.6280.150 ± 0.0020.190 ± 0.003
Raw score hh0.636 ± 0.0010.165 ± 0.0030.178 ± 0.002
γ=1\gamma=1, score ss (ours)0.633 ± 0.0070.171 ± 0.0020.252 ± 0.002

Without the ranking, weighting keeps only part of the gain, and none of it on RxRx1. Without the floor, it keeps the gain on BBBC021, where almost all cells respond, but loses most of it on JUMP, where few do. So only the three statements above need calibrated probabilities. The gains on real data need just the ranking and the floor.

AuthorsOpenReview ↗

Weakness 2 and Question 2.

The paper did not say how to choose the target or γ\gamma, and we have added a two-step procedure to Section 5.5. First, the target follows from the question. It is the whole treated population when the generated cells should stand in for a well, for example to predict well-level profiles, and the responders when they should show what a perturbation does to the cells it affects, for example to identify its mechanism. Second, γ\gamma is chosen on validation wells with a metric that matches the target: the energy distance to all treated cells for the population, and MoA accuracy on BBBC021, top-5 accuracy on RxRx1 or mAP on JUMP for the responders. So no test outcomes are needed.

To test the second step, we split the evaluation wells 20 times at random into two halves, chose γ\gamma on one half, and checked whether it was also the best γ\gamma on the other half, so the choice never sees the wells it is checked on. The candidates are γ\gamma of 0, 0.25, 0.5, 1 and 2 on BBBC021 and RxRx1, and 0, 0.25 and 1 on JUMP, where two seeds give the same counts. The table below counts the splits, and we have added it to Appendix E.

DatasetValidation metricγ\gamma chosenAlso best on the other half
BBBC021Energy distance to all treated2 (19), 0.5 (1)18
MoA accuracy2 (20)20
RxRx1Energy distance to all treated1 (20)20
Top-5 accuracy1 (14), 0.5 (3), 0.25 (3)7
JUMPEnergy distance to all treated0.25 (20)20
mAP1 (20)20

Validation wells help to pick γ\gamma reliably. For the population, they pick mild weighting on JUMP, where few cells respond, and strong weighting on BBBC021 and RxRx1, where most do. For the responders, they pick γ=1\gamma=1 on JUMP (mAP) and γ=2\gamma=2 on BBBC021 (MoA accuracy). With these metrics, the choice misses the best γ\gamma in only 2 of 100 splits, losing about 1% of the gain over γ=0\gamma=0. The exception is top-5 accuracy on RxRx1, which is too noisy on half of the wells to rank γ\gamma from 0.25 to 2. There the choice matches the best γ\gamma in only 7 of 20 splits, but it never picks γ=0\gamma=0 and keeps about half of the gain.

The response addressed my comments. I will keep my score at 3.

Reviewer_23VmOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper addresses partial response in image-based virtual cell models: within a treated well, often only a fraction of cells respond to the perturbation. The authors propose estimating, for each treated cell, a probability of response using a well-cross-fitted classifier against same-plate controls, and weighting each cell’s flow-matching loss by this score.

Critical strengths
  1. Identifies an important and under-addressed problem in virtual cell modelling: not all cells respond to a perturbation, whether for technical reasons or biological resistance.

  2. Proposes a simple, model-agnostic approach (responder-weighted training) with clear theoretical motivation and a well-designed dilution benchmark in which the true effect is known.

Critical weaknesses
  1. Whilst addressing the fact that not all cells respond to perturbation (whether that’s due to technical transfection effect or actual biological resistance); fundamentally, the training objective seems a bit reductionist in response to the complexity of perturbation response simulation. Response given an untreated cell motivates a necessary position in evaluation (correctly predicting the proportion of cells that should respond) as opposed to regularising the training of a model/pre-selecting a response-only population. In a model that simulates phenotypic response to perturbation, one would expect the simulated phenotype to encompass response/no response as an emergent property of a well-calibrated generative model. Predicting whether a cell should even respond at all is a very important component of these virtual cell models, as it allows us to simulate effects at inference on cell lines (with differing genetic backgrounds), in which simulating no response for a population of cells is a biologically correct outcome. By removing cases in which there is no response, we also remove the ability of a generative model to correctly predict that resistant cells shouldn’t respond in the first place.

  2. This statement, “Our weighting changes only the training target, so it applies to any of them”, regarding other perturbation prediction models, was not proven, and no evaluation was extended to all of these models addressed in the related work section.

Questions for authors
  1. Would these principles hold in other generative frameworks, such as diffusion models or GANs?

  2. How would the responder fraction be obtained at inference for unseen perturbations or cell lines?

3 comments
AuthorsOpenReview ↗

Weakness 1.

We appreciate your feedback and respond to your comments point by point below.

We agree that whether a cell responds is part of what a virtual cell model must predict. Our approach does not remove it. It splits the treated population into two parts: the share of responders ρ\rho, and how the responders look, pRp_R (Eq. 1). Mixture mode puts the two parts back together: it generates a responder with probability ρ^\hat\rho and keeps a control cell otherwise. A resistant population is then simply ρ\rho near zero, and ρ^\hat\rho is reported for every perturbation (it rises with dose for 13 of 15 compounds, Figure 3b).

Non-response does not emerge reliably from the unweighted model. The table below gives the share of responders among real and generated cells, rated by the same scorer, on the dilution benchmark, where the true share is qq, and on BBBC021. The unweighted model (γ=0\gamma=0) generates too few responders at every level, and the rarer the responders, the more of them it loses. The weighted model (γ=1\gamma=1) generates almost only responders, as intended, and mixture mode adds the non-responders back at rate ρ^\hat\rho. On BBBC021, where most cells respond, γ=1\gamma=1 is also closer to the real share than γ=0\gamma=0. We have added this table to Appendix D.

Dilution, q=0.1q=0.1Dilution, q=0.25q=0.25Dilution, q=0.5q=0.5Dilution, q=0.75q=0.75Dilution, q=0.9q=0.9BBBC021
Real treated cells0.100.250.500.750.900.978
γ=0\gamma=00.0170.0380.2560.6570.8680.911
γ=1\gamma=1 (ours)0.9550.9740.9790.9951.0000.942

On all three datasets, non-response remains part of the prediction. Our method estimates the share of responders explicitly for every perturbation, whereas the unweighted model, from which this share should emerge, generates too few responders in every column of the table above.

AuthorsOpenReview ↗

Weakness 2 and Question 1.

Thank you for pointing this out. We agree that the sentence stated more than we tested, and we have softened it in Related work (Lines 100-101): weighting applies to models trained with a loss averaged over treated cells. Proposition 3 needs only such a loss and a weight that depends on each cell alone (proof in Appendix B), so it holds for diffusion models trained by denoising score matching. The same argument applies to GANs, whose real cells enter only through the discriminator loss: weighting them there makes the discriminator compare the generated cells with the weighted population, which then becomes the generator's optimum. We have tested diffusion, but not GANs.

On the dilution benchmark with 10% responders, we trained a diffusion model with and without weighting. It shares the latents, scores, DiT-S/2 backbone, conditioning, optimizer and training steps of our flow, but uses the standard noise-prediction loss (1,000 steps, linear schedule), starts from pure noise instead of a control cell, and is sampled with 100 DDIM steps at our guidance strength of 1.2 (Appendix H). The table below gives the recovered effect rr (1 is the full effect) and FID.

ModelRecovered effect rr, γ=0\gamma=0Recovered effect rr, γ=1\gamma=1FID ↓, γ=0\gamma=0FID ↓, γ=1\gamma=1
Flow matching (ours)−0.0010.85864.435.8
Diffusion0.1451.186106.379.3

Weighting moves the diffusion model in the same direction as our flow: without weighting it misses the responders, and with weighting it recovers them. It overshoots the effect, and both diffusion models give poor images, since we did not tune them, so we read this result as the direction of the effect only.

AuthorsOpenReview ↗

Question 2.

The trained model does not need ρ\rho at inference: it turns control cells into responders. ρ^\hat\rho is used in two places only: during training, to turn the classifier output into weights, and in mixture mode, to rebuild the whole treated population. For example, on held-out BBBC021 compounds, weighting raises MoA accuracy from 0.472 to 0.493 without any ρ\rho at inference (Appendix E).

During training, ρ^\hat\rho needs only the treated and control wells of a perturbation. A logistic regression on frozen MorphEm embeddings, cross-fitted over wells, separates the treated cells from controls of the same plate, and its outputs give ρ^\hat\rho (Eq. 2). For an unseen perturbation, only mixture mode needs ρ\rho, and there it would have to be predicted, for example from the same features the model is conditioned on.

For a new cell line, both ρ\rho and the responders can change. Predicting either needs control and perturbed cells from several cell lines, for any model. This is open for virtual cell models in general: in transcriptomics, deep models do not yet beat simple linear baselines on unseen perturbations and unseen context [1]. Virtual Cell Challenge 2025 found models "not yet consistently outperforming naive baselines across all metrics" [2].


[1] Ahlmann-Eltze, Huber and Anders. Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Nature Methods, 2025.

[2] Arc Institute. Virtual Cell Challenge 2025 Wrap-Up: Winners and Reflections, 2025, and The 2026 Virtual Cell Challenge: predicting perturbation responses in cell contexts a model has never seen, 2026

Reviewer_QyM6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose adapting flow-matching training to cell morphology data, where only a fraction of treated cells typically respond to a perturbation. They argue that a model trained on all treated cells captures the full distribution, from unaffected to strongly affected cells, rather than the perturbation effect of interest. They estimate for each treated cell how likely it is to have responded, using a classifier against same-plate controls on frozen image embeddings, and weight each cell's flow-matching loss by a scaled version of this score. The method is evaluated on a dilution benchmark and on BBBC021, RxRx1 and JUMP, where it improves recovery of the perturbed-cell response on two of the three datasets compared to an unweighted flow-matching baseline.

Critical strengths
  • The observation and how it is modeled. That only a fraction of treated cells respond to a perturbation is reasonable, and the way it is modeled is meaningful. This is consequential because it points to a limitation that affects any generative morphology model trained on all treated cells.
  • Well-designed experiments. The dilution benchmark, which injects controls as artificial non-responders so the true responder fraction is known, directly tests the central claim and is reusable by others. The scoring and method design is appropriate without being overly complex, well effects are carefully separated from the response assessment, experiments span several datasets, splits and runs, and the code is released. Together, these make the central claim credible in the controlled setting and the work reproducible.
Critical weaknesses
  • Motivation and evaluation. The motivation for targeting responders rather than the whole treated population could be stated more clearly; it is currently limited to one sentence about virtual cells. Relatedly, the main real-data metric compares to all treated cells, which favors the population target. This may explain the weaker JUMP results, but it makes it hard to judge whether the method helps on real data for its intended purpose.
  • Comparisons to simpler alternatives. The scoring approach is close to HiDDEN, so the contribution would be stronger with comparisons to simpler alternatives, such as weighting by the raw classifier score, a properly trained filtering baseline on all datasets, or per-perturbation guidance.
Questions for authors

Clarifications and requested analyses

  • I understand why one would want the full possible response, but why is the population effect not of interest as well?
  • Can you report the recovered-effect metric on the real datasets, or another evaluation that targets the responder distribution?
  • Please improve the captions of Figures 3 to 5, as it is not immediately clear what each figure shows.

Suggestions for future work

  • Consider using the response score as an input to the flow conditioning rather than as a loss weight.
3 comments
AuthorsOpenReview ↗

Motivation and evaluation.

We are grateful to the reviewer for their insightful comments and suggestions, which have contributed to strengthening our work.

Motivation. We agree that one sentence is not enough, and we have added a motivation paragraph to the Introduction. In short: A flow trained on all treated cells reproduces their mixture, so every predicted feature change equals ρ\rho times the change of the responders (Proposition 1). For the treated population this is correct, but an average over generated cells then cannot tell a strong effect on few cells from a weak effect on many. In practice, the unweighted model does not even reproduce the population. The table below shows this on the dilution benchmark, where a share qq of the training cells are real treated cells and the rest are injected controls. The upper part gives the share of responders among the cells, rated by the same scorer, and the lower part the recovered effect rr. The unweighted model (γ=0\gamma=0) generates fewer responders than its training data contain, and the rarer the responders, the more of them and of their effect it loses, so it does not reproduce even the mixture it was trained on. The weighted model (γ=1\gamma=1) generates almost only responders and recovers close to the full effect at every qq, and mixture mode adds the non-responders back at rate ρ^\hat\rho.

q=0.1q=0.1q=0.25q=0.25q=0.5q=0.5q=0.75q=0.75q=0.9q=0.9
Share of respondersTraining data0.100.250.500.750.90
γ=0\gamma=00.0170.0380.2560.6570.868
γ=1\gamma=1 (ours)0.9550.9740.9790.9951.000
Recovered effect rr ↑Training data0.1040.2470.4830.7300.879
γ=0\gamma=0−0.0010.0850.3200.6710.847
γ=1\gamma=1 (ours)0.8580.9340.9751.0241.011

Evaluation. We agree that the latent energy distance, like FID and KID, compares to all treated cells and therefore favors the population target. To judge the method for its intended purpose, we measure specificity, namely mechanism-of-action (MoA) accuracy on BBBC021, top-5 accuracy on RxRx1 and mean average precision (mAP) on JUMP, the distance to real references that target the responders, and the recovered effect of CellProfiler features. We have added these evaluations to Appendix E. The specificity table below gives mean ± standard error over three seeds. In addition, these results can be found in the paper (Figure 5: a, d, e).

BBBC021, MoA acc. ↑RxRx1, top-5 acc. ↑JUMP, mAP ↑
γ=0\gamma=00.618 ± 0.0010.154 ± 0.0020.148 ± 0.002
γ=1\gamma=1 (ours)0.633 ± 0.0070.171 ± 0.0020.252 ± 0.002

The reference table below gives the energy distance to two real references, divided by that of γ=0\gamma=0, so values below 1 are closer to the reference than γ=0\gamma=0. The first reference is all treated cells. The second is the top half by score, that is, in each perturbation the half of the treated cells with the highest response score ss. As a control, shuffled weights train γ=1\gamma=1 with the same weights permuted at random among the cells of each perturbation, which keeps the amount of reweighting but not the choice of the cells that receive weight.

Energy distance relative to γ=0\gamma=0 ↓

DatasetModelAll treatedTop half by score
BBBC021γ=1\gamma=1 (ours)0.750.74
Shuffled weights0.860.88
RxRx1γ=1\gamma=1 (ours)0.470.55
Shuffled weights0.650.69
JUMPγ=0.25\gamma=0.25 (ours)0.480.37
γ=1\gamma=1 (ours)5.904.58
Shuffled weights1.721.50

Weighting raises specificity on all three datasets, and the best strength follows the responder fraction. On BBBC021 and RxRx1, where most cells respond, γ=1\gamma=1 is closer than γ=0\gamma=0 and than shuffled weights to both references, so the gain comes from placing the weight on the responders, not from reweighting as such. On JUMP, where the median perturbation changes fewer cells, γ=1\gamma=1 moves far from both references, while mild weighting (γ=0.25\gamma=0.25) is closer than γ=0\gamma=0 to both and is reliably selected on held-out wells.

AuthorsOpenReview ↗

Comparisons to simpler alternatives.

We totally agree that these comparisons are valuable, and we trained and evaluated all three alternatives. We have added them to Appendix F. For the raw classifier score, we trained γ=1\gamma=1 with the classifier output hh in place of our score ss. For filtering, we trained γ=0\gamma=0 only on the cells with s>0.5s>0.5, for as many steps as our model.

MethodBBBC021 IID, energy ↓BBBC021 IID, MoA acc. ↑BBBC021 OOD, energy ↓BBBC021 OOD, MoA acc. ↑
γ=0\gamma=00.086 ± 0.0020.618 ± 0.0010.332 ± 0.0110.472 ± 0.003
Raw score hh0.069 ± 0.0020.636 ± 0.0010.300 ± 0.0270.478 ± 0.005
Hard filter, s>0.5s>0.50.066 ± 0.0030.638 ± 0.0030.286 ± 0.0220.497 ± 0.010
γ=1\gamma=1, score ss (ours)0.061 ± 0.0040.633 ± 0.0070.304 ± 0.0130.493 ± 0.010
MethodRxRx1, energy ↓RxRx1, top-5 acc. ↑JUMP, energy ↓JUMP, mAP ↑
γ=0\gamma=00.260 ± 0.0050.154 ± 0.0020.0061 ± 0.00020.148 ± 0.002
Raw score hh0.137 ± 0.0040.165 ± 0.0030.0049 ± 0.00110.178 ± 0.002
Hard filter, s>0.5s>0.50.745 ± 0.0090.169 ± 0.0020.0804 ± 0.00040.257 ± 0.009
γ=1\gamma=1, score ss (ours)0.133 ± 0.0040.171 ± 0.0020.0356 ± 0.00280.252 ± 0.002

Both methods hh and ss are close on BBBC021, where most cells respond, while on JUMP, where few do, hh acts as a mild tilt and keeps less than a third of the gain in mAP. The hard filter matches weighting on BBBC021, where it keeps 85% of the training cells, and matches or exceeds it in specificity on RxRx1 and JUMP, but there its energy distance is an order of magnitude above that of γ=0\gamma=0 on JUMP and almost three times on RxRx1. A hard threshold removes every uncertain cell from training, while weighting keeps them with a smaller weight.

Per-perturbation guidance. By Proposition 2, guidance recovers the responders only with a strength proportional to 1/ρ1/\rho. We therefore sampled γ=0\gamma=0 with αu=α⋆/ρ^u\alpha_u=\alpha^\star/\hat\rho_u per perturbation (capped at 8), for α⋆=1\alpha^\star=1 and α⋆=2.55\alpha^\star=2.55 (measured on the dilution benchmark), and compared it with one strength for all perturbations, with and without weighting. For the weighted model on JUMP and RxRx1, we chose this strength on held-out wells: α=4\alpha=4 on RxRx1, and α=2\alpha=2 on JUMP, the strongest α\alpha at which γ=0.25\gamma=0.25 is still as close to all treated cells as γ=0\gamma=0 at the default strength. The tables below are in Appendix F.

ModelGuidanceBBBC021 IID, MoA acc. ↑BBBC021 IID, FID ↓BBBC021 OOD, MoA acc. ↑BBBC021 OOD, FID ↓
γ=0\gamma=0one strength, α=1.2\alpha=1.20.6189.730.46618.70
γ=0\gamma=0per perturbation, α⋆=1\alpha^\star=10.60911.030.46118.76
γ=0\gamma=0per perturbation, α⋆=2.55\alpha^\star=2.550.72116.390.51922.34
γ=0\gamma=0one strength, α=2\alpha=20.71810.190.50820.12
γ=2\gamma=2 (ours)one strength, α=1.2\alpha=1.20.6428.710.49117.93
γ=2\gamma=2 (ours)one strength, α=2\alpha=20.7438.760.52218.45
ModelGuidanceRxRx1, top-5 acc. ↑RxRx1, FID ↓JUMP, mAP ↑JUMP, FID ↓
γ=0\gamma=0one strength, α=1.2\alpha=1.20.15422.380.1513.36
γ=0\gamma=0per perturbation, α⋆=1\alpha^\star=10.19521.20.2813.5
γ=0\gamma=0per perturbation, α⋆=2.55\alpha^\star=2.550.28318.10.4416.7
γ=0\gamma=0one strength, α=3\alpha=30.25417.280.2612.98
γ=0.25\gamma=0.25 (ours)one strength, α=2\alpha=2––0.2512.82
γ=1\gamma=1 (ours)one strength, α=1.2\alpha=1.20.17019.090.2514.87
γ=1\gamma=1 (ours)one strength, α=3\alpha=30.27520.060.4815.00
γ=1\gamma=1 (ours)one strength, α=4\alpha=40.30020.910.5315.21

One strength combined with weighting is the better use of guidance. Per-perturbation guidance raises specificity only at α⋆=2.55\alpha^\star=2.55, and there it costs image quality. With one strength and weighting, our model has the highest specificity on all three datasets, at a lower FID than per-perturbation guidance on BBBC021 and JUMP and with an almost three times lower energy distance on RxRx1, where γ=0\gamma=0 at α=3\alpha=3 keeps the lowest FID. At every strength, weighting raises specificity, so weighting and guidance add up rather than replace each other.

AuthorsOpenReview ↗

Questions and suggestions.

Question 1.

It is of interest, and our approach captures it. Under the responder model the treated population is ρpR+(1−ρ)p0\rho p_R + (1-\rho) p_0, so a model of the responders together with ρ^\hat\rho gives the population back. Mixture mode does exactly this: it replaces each evaluation cell by a generated responder with probability ρ^\hat\rho and by a real same-plate control otherwise. The reverse does not work, because the population effect is ρ\rho times the responder effect and an average cannot separate the two factors (Proposition 1). The table below gives the energy distance of mixture mode to all treated cells, divided by that of γ=0\gamma=0, together with a mixture built from the model with shuffled weights. We have added it to Appendix E.

Energy distance to all treated cells, relative to γ=0\gamma=0 ↓

ModelBBBC021RxRx1JUMP
Mixture mode, γ=1\gamma=1 (ours)0.75 ± 0.051.12 ± 0.022.06 ± 0.13
Mixture mode, shuffled weights0.911.29 ± 0.022.55 ± 0.07

Mixture mode is closer to the treated population than γ=0\gamma=0 on BBBC021, but not yet on RxRx1 and JUMP. It is closer than the mixture with shuffled weights on every dataset, so the ranking of cells by ss also helps to rebuild the population. When the population itself is the target on JUMP, mild weighting is the better choice: γ=0.25\gamma=0.25 is closer to all treated cells than γ=0\gamma=0.

Question 2.

Yes. The table below reports both forms of the recovered effect in Eq. 7 on all three datasets, and the reference table in our answer to W1 compares the models with real references that target the responders. We have added both to Appendix E. The recovered effect uses CellProfiler features, which neither the scorer nor the generator sees, and we report it for the perturbations with a detectable effect (50 on BBBC021, 16 on JUMP and 95 on RxRx1). Here rprojr_{\mathrm{proj}} projects the generated effect of each perturbation onto the real one and takes the median over perturbations, so each perturbation counts once and errors on its strongest features cost the most. The more lenient rr, the measure of our dilution benchmark, takes the median ratio of generated to real change over all pairs of perturbation and feature. Reconstructions give the most that a model in this latent space can reach.

ModelBBBC021, rprojr_{\mathrm{proj}} ↑BBBC021, rr ↑RxRx1, rprojr_{\mathrm{proj}} ↑RxRx1, rr ↑JUMP, rprojr_{\mathrm{proj}} ↑JUMP, rr ↑
γ=0\gamma=00.6470.8180.0620.5580.7830.947
γ=0.25\gamma=0.25 (ours)––––0.7810.997
γ=1\gamma=1 (ours)0.7150.8560.2810.7380.7700.973
γ=2\gamma=2 (ours)0.7190.8570.2130.7850.7370.940
Shuffled weights0.6730.8410.2330.6600.5250.883
Reconstructions0.9021.0020.5940.9710.9381.019

On BBBC021, the recovered effect grows with the strength of weighting in both forms. Shuffled weights reach part of this gain, so it comes partly from reweighting as such and partly from placing the weight on the responders. On RxRx1, weighting raises the recovered effect well above γ=0\gamma=0 in both forms, and above shuffled weights except for γ=2\gamma=2 in rprojr_{\mathrm{proj}}. On JUMP, weighting changes the recovered effect little in either form.

Question 3.

Thank you for the comment. We have rewritten the captions of Figures 3 to 5.

Suggestion for future work.

Thank you for the promising idea. A flow conditioned on ss could generate both the responders and the whole population with one model. For a new perturbation it would still need the distribution of ss, which is the same unknown as ρ\rho. We have added it as future work to Discussion and conclusion section.

Comments on the paper

1 comment

Dear reviewers, Thank you for your feedback and comments. We hope that the new experiments and analyses addressed your concerns. We also edited the text to make it easier to read, redrew the figures and uploaded the revised manuscript. Thank you again for your time and effort!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.