Pith. sign in

REVIEW 5 major objections 5 minor 37 references

CellFlux: Simulating Cellular Morphology Changes via Flow Matching

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read CellFlux claims cellular morphology prediction is a distribution-to-distribution flow-matching problem, and reports large FID and mode-of-action gains over prior control-conditioned baselines.

desk verdict Solid, clearly written application of flow matching to perturbation response prediction; main results are plausible, but the evaluation would be stronger with an identity/no-perturbation baseline and error bars. read the letter →

arxiv 2502.09775 v3 pith:THLBREA2 submitted 2025-02-13 q-bio.QM cs.CVcs.LGq-bio.BMq-bio.CB

Yuhui Zhang , Yuchang Su , Chenyu Wang , Tianhong Li , Zoe Wefers , Jeffrey Nirschl , James Burgess , Daisy Ding
show 3 more authors
Alejandro Lozano Emma Lundberg Serena Yeung-Levy
This is my paper · ORCID
classification q-bio.QMcs.CVcs.LGq-bio.BMq-bio.CB
keywords flowmatchingcellularmorphologyperturbationpredictionvirtualcellbatcheffectsimagegenerationmodeofactiondistribution-to-distribution
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

CellFlux argues that predicting how cells change shape under a chemical or genetic perturbation should be treated as a distribution-to-distribution problem: learn a continuous transformation from the distribution of unperturbed control cells in a batch to the distribution of treated cells in the same batch, rather than as a per-image translation. Because fluorescence imaging is destructive, no paired before-and-after images of the same cell exist, so the paper's key move is to use same-batch control images as a proxy for the unobservable basal state and to learn the perturbation as a flow of one distribution into the other. The payoff is that true perturbation effects can in principle be separated from experimental batch artifacts, and the learned flow gives continuous and reversible trajectories between cell states. If the approach holds up, it offers an in silico way to screen drugs and genetic perturbations and to study how cell morphology changes over time.

What carries the argument

The core mechanism is conditional flow matching between control and perturbed image distributions: straight-line interpolation $x_t = (1-t)x_0 + t x_1$ between randomly paired control and treated images, with the U-Net trained to predict the constant velocity $x_1 - x_0$ under an $\ell^2$ loss and used to solve the ODE $dx_t = v_\theta(x_t,t,c)\,dt$ at inference. Classifier-free guidance and Gaussian noise augmentation on the source samples keep the velocity field smooth and the generated images close to the target distribution, while a learnable embedding of the perturbation condition steers the flow. Proposition 1 supplies the theoretical justification: the same-batch control image carries information about the batch and the basal state, so conditioning on it reduces the expected KL divergence to the true per-cell response compared with conditioning on the perturbation alone.

What would settle it

On a dataset with true paired live-cell images of the same cells before and after perturbation, or using BBBC021's dose-response structure, check whether the flow-matching trajectory from control to the final treated state passes through states matching real intermediate timepoints or doses; if the intermediate generated states do not match real intermediates, the claim that the learned velocity field captures the biological transformation is falsified.

Watch

Extended reading notes

Core claim

CellFlux's central claim is that flow matching solves cellular morphology prediction: a neural velocity field $v_\theta(x_t,t,c)$ transports same-batch control images $x_0 \sim p_0$ to perturbed images $x_1 \sim p_1(\cdot|c)$ along rectified-flow paths $x_t = (1-t)x_0 + t x_1$ whose target velocity is $x_1 - x_0$, with the ODE integrated at inference. The paper shows (Proposition 1) that conditioning on a same-batch control image is never worse, in expected KL divergence, than conditioning on the perturbation alone when $X_1$ is conditionally independent of the control variable given the true basal state and the perturbation. Empirically, CellFlux reports overall FID scores of 18.7 on BBBC021, 33.0 on RxRx1, and 9.0 on JUMP, improvements of 21-45% over the GAN and diffusion baselines, and a MoA classification accuracy of 71.1% on BBBC021 against 63.7% for the best baseline, close to the 72.4% achieved on real perturbed images. The paper also claims generalization to held-out perturbations, batch-effect calibration through same-batch control initialization, and bidirectional interpolation between cellular states.

Load-bearing premise

The load-bearing premise is that the unobservable basal state of treated cells in a batch follows the same distribution as the observed same-batch control cells, so randomly pairing a control image with a treated image and learning the flow between them teaches the true perturbation effect rather than a batch artifact or an arbitrary coupling.

Editorial extensions

If this is right

  • Generated images are more realistic than those from control-conditioned baselines on all three benchmarks: overall FID drops by 21-45% and KID drops correspondingly across BBBC021, RxRx1, and JUMP.
  • Biological fidelity is preserved: a mode-of-action classifier trained on real perturbed images scores 71.1% accuracy on CellFlux-generated images, nearly matching the 72.4% on real images and beating the 63.7% best baseline.
  • The model generalizes to unseen perturbations, and conditioning on same-batch rather than other-batch controls improves conditional FID by 14% and MoA accuracy by 48%, evidence that control conditioning is what separates perturbation effects from batch artifacts.
  • Because the velocity field is continuous and reversible, CellFlux can interpolate between control and perturbed states and run the trajectory backward, enabling visualization of morphological transitions.
  • Each component matters: ablating the perturbation condition, classifier-free guidance, or noise augmentation raises FID substantially, indicating the design is jointly necessary for the reported quality.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The biological validity of the interpolated trajectories is not tested in this paper; if later experiments with dose or timepoint data show that intermediate flow states match real intermediate morphologies, the same velocity field becomes a dynamical model of perturbation response rather than just a generator.
  • The same distribution-to-distribution framing should transfer to any destructive assay where controls and treated samples share a batch but paired observations are impossible, such as differentiation, senescence, or patient-derived cell responses, provided a control population can be defined.
  • Because training pairs are random, the method guarantees per-distribution fidelity, not per-cell correspondence; single-cell applications would need paired tracking data or a latent alignment step to assign generated morphologies to specific input cells.
  • A direct testable extension suggested by the paper itself is to use multi-dose datasets to check whether interpolation from control to the highest dose recreates the real middle-dose images, which would validate the trajectory interpretation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. CellFlux proposes a flow-matching model for cellular morphology prediction, framed as a distribution-to-distribution transformation from same-batch control images to perturbed images under a given chemical or genetic perturbation. The method is evaluated on BBBC021, RxRx1, and JUMP, reporting FID/KID improvements over PhenDiff and IMPA, a MoA classification accuracy of 71.1% on BBBC021, ablations, an out-of-distribution study, a batch-effect analysis, and bidirectional interpolation trajectories. The paper argues that conditioning on control images from the same batch helps separate perturbation effects from batch artifacts.

Significance. If the empirical claims hold, CellFlux is a useful step toward in-silico perturbation-response modeling. The paper is clearly written, uses public benchmarks, and includes a broader evaluation than many prior works: ablations, OOD generalization, batch-effect conditioning, and CellProfiler-based morphological checks. The formal motivation for conditioning on same-batch controls is a genuine contribution, and the flow-matching formulation is simple and well matched to the task. However, the central empirical claim that CellFlux captures perturbation-specific morphology rather than batch appearance is not fully supported without a no-perturbation baseline and error bars; additionally, the paper overstates the conditional per-cell distribution that the deterministic flow-matching procedure actually estimates. These issues are fixable but require additional experiments and claim revision.

major comments (5)
  1. [Section 2.1 and Section 3.1] The objective in Section 2.1 is to learn the conditional distribution p(x1|x0,c), but the flow-matching objective in Section 3.1 is trained with independently sampled pairs (x0,x1) and, at inference, solves a deterministic ODE starting from a fixed x0. For a fixed input image, the model returns a single output, not a sampled conditional distribution; the learned object is a transport map from p0 to p1(·|c), with the specific coupling determined by the random pairing used in training. The evaluation metrics (FID/KID and a classifier applied to pooled images) are distribution-level and do not validate the per-cell conditional claim made in Proposition 1. The authors should either revise the claim to distribution-to-distribution transport or describe an explicit stochastic decoding procedure that produces a distribution over x1 for each x0.
  2. [Section 4.3, Table 1a, Appendix D] The headline FID gains are not separated from batch matching. Appendix D states that same-batch control and treated images cluster tightly together and that batch effects dominate perturbation effects; in that setting, a model that returns the same-batch control image (or a near-identity transformation) can achieve deceptively low FID without modeling the perturbation. The paper never reports this no-perturbation baseline. I ask for two concrete additions: (i) FID/KID/MoA of same-batch control images used directly as predictions, and (ii) evaluation on a no-effect condition such as DMSO, which appears in Table 3, to show the model does not invent morphology when the perturbation is biologically inert. Without these, the central claim that CellFlux learns perturbation-specific morphology rather than batch appearance is not established.
  3. [Section 4.2, Tables 1 and 2] All headline numbers are point estimates from a single run. FID on 5K generated images and MoA accuracy have substantial sampling variability, and the improvements cited in the abstract are averages over heterogeneous per-dataset gains (e.g., RxRx1 conditional FID improves by only 0.8% over IMPA). Without error bars or significance tests, the reported margins—especially the MoA gap of 71.1% versus 63.7%—could be within noise. The authors should report means and confidence intervals over at least three seeds, or an equivalent significance analysis, and use those in the abstract and main claims.
  4. [Algorithm 1 and Section 2.3] The theoretical setup in Section 2.3 frames the transformation as p(x̃0|b) → p(x1|c,b), i.e., same-batch conditioning, and the batch-effect claims depend on this. However, Algorithm 1 samples x0 ∼ p0 and x1 ∼ p1 without any explicit batch variable or same-batch constraint, and Sections 3.1–3.4 likewise omit b. If training pairs are not restricted to the same batch, the model is not learning the batch-calibrated transition described in Section 2.3; if they are, the algorithm and preprocessing need to state this. The authors should clarify the batch sampling procedure, including how the 'other batch' initialization study in Table 2c was constructed.
  5. [Section 2.3] The assertion that the unobservable basal state X0 and the control state X̃0 follow the same distribution p(·|b) is presented as a fact, but it is an untestable assumption. Control and treated wells may differ in cell density, seeding, or early responses to treatment, and the PCA visualization in Appendix D does not establish equality of distributions. This assumption should be stated explicitly as a limitation, and, where possible, checked with a no-effect perturbation (e.g., DMSO) or with negative controls. The current wording obscures that the batch-effect correction capability depends on this equality.
minor comments (5)
  1. [Abstract and Section 4.3] The abstract's '35% improvement' is an average over datasets and baselines; the text should state per-dataset improvements relative to the best baseline and clarify that the improvement is not uniform (e.g., RxRx1 conditional FID improves by only 0.8%).
  2. [Algorithm 1 and Section 3.4] Notation is inconsistent: the algorithm uses noise injection probability p_n, while Section 3.4 uses p_e; the noise is N(0,I) in Algorithm 1 but N(0,σ²I) in Section 3.4; the condition-drop null token appears as ϕ in pseudocode and ∅ elsewhere.
  3. [Table 5] Table 5 mixes results for BBBC021 and JUMP without clear row or column labels, and the sample-size columns are duplicated; please label the dataset and sample size for each block.
  4. [Figure 11] The cross-dataset transfer claim is supported only by selected qualitative examples; a quantitative metric, even a limited one, would strengthen the claim.
  5. [Section J] The appendix appropriately acknowledges that interpolation trajectories are not biologically validated; the main text should reflect this caveat more prominently in Section 4.4, where the interpolation capability is introduced.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the empirical FID/MoA claims rest on external benchmarks against real images, and Proposition 1 is a valid information-theoretic motivation, not a derivation of the reported results.

full rationale

CellFlux's derivation chain splits into a theoretical part and an empirical part, and neither reduces to its own inputs. The theoretical part (Section 2.3, Appendix A) justifies conditioning on same-batch controls: Proposition 1's proof rewrites the claimed inequality as I(X1;X0|C) = I(X1;X0|X~0,C) + I(X1;X~0|C) under the graphically assumed conditional independence X1 ⊥ X~0 | X0, C, so the conclusion is a theorem from stated assumptions, with strictness only when I(X1;X~0|C) > 0, which the paper says 'generally holds true when batch effect exists.' This proposition does not produce any reported number. The empirical part is externally benchmarked: FID/KID compare generated images to real held-out treated images for a model trained on velocity MSE, not on FID; the MoA classifier is trained on real perturbed images and tested on generations (Table 2a); CellProfiler nuclear-size features (Table 7) are independent measurements; and Table 5 shows the FID advantage persists across sample sizes. No fitted parameter is renamed as a prediction, although Appendix C's model selection on validation FID modestly weakens the headline FID as an unbiased estimate, which is a correctness, not circularity, issue. The only references overlapping the author list (Bunne et al. 2024 and Johnson et al. 2023, both co-authored by E. Lundberg) support the 'virtual cell' motivation and are not load-bearing for the method or results. Several validity risks are real but fall outside the circularity patterns: Appendix D concedes that control and treated images from the same batch 'cluster together, rather than forming separate control and target clusters,' so a missing identity/no-perturbation baseline leaves open that part of the FID gain reflects batch-appearance fidelity; Section 2.3's assumption that unobservable basal states follow p(·|b) and the random unpaired coupling in Algorithm 1 mean the learned transport need not equal the biological conditional p(x1|x0,c); and Appendix J plus the Impact Statement explicitly concede that interpolation trajectories remain biologically unverified, which is flagged and weighed here as honest hedging that supports, rather than undermines, the non-circularity finding. Overall, the central claims are self-contained against external benchmarks; score 1 reflects only the marginal, non-load-bearing self-citations.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

CellFlux introduces no new physical or biological entities. The latent basal state X0 is a statistical variable in the graphical model, not a postulated entity with independent falsifiable handles. The free parameters are standard generative model hyperparameters tuned on validation FID. The axioms consist of standard flow matching theory plus domain assumptions about batch-conditional distributions and unpaired coupling that are structurally necessary for the method to identify perturbation effects.

free parameters (5)
  • Classifier-free guidance strength alpha = 1.2
    Chosen by validation; interpolates between conditional and unconditional velocity to improve fidelity (Section 3.3, Appendix C).
  • Condition drop probability pc = 0.2
    Hyperparameter for classifier-free guidance, set in Appendix C.
  • Noise injection probability pn = 0.5
    Probability of adding Gaussian noise to source samples for smoother velocity field (Section 3.4, Appendix C).
  • Noise scale sigma = 1 (implied by epsilon ~ N(0, I))
    Used in noise augmentation; the scale is not explicitly tuned in the paper.
  • Model architecture and training hyperparameters = hidden size 128, 4 stages, batch size 128, lr 1e-4, 100 epochs
    Manual choices selected by lowest FID on the validation set (Appendix C).
assumptions (4)
  • standard math The conditional flow matching / rectified flow objective trains the marginal velocity field that transports p0 to p1 (Lipman et al. 2023, Liu et al. 2023).
    Used as the training objective in Section 3.1 without re-derivation.
  • domain assumption The graphical model satisfies conditional independence X1 _|_ tilde-X0 | X0, C, and the batch-conditional distributions of X0 and tilde-X0 are identical.
    Invoked in Section 2.3 and Appendix A to justify conditioning on control images; if controls do not share the basal distribution of treated cells, the proposition does not apply.
  • domain assumption Random pairing of independently sampled x0 and x1 during flow matching yields a valid transport between distributions.
    Training samples pairs independently (Algorithm 1); the ODE maps the control marginal to the treated marginal, but individual trajectories are not grounded in paired observations.
  • domain assumption The U-Net has sufficient capacity to approximate the true velocity field.
    Standard neural network approximation assumption; not verified beyond empirical performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CellFlux: Simulating Cellular Morphology Changes via Flow Matching." pith.science (2026). https://pith.science/paper/THLBREA2

@misc{pith2026250209775,
  author       = {Pith},
  title        = {Pith review of: CellFlux: Simulating Cellular Morphology Changes via Flow Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/THLBREA2}},
  note         = {Machine review of arXiv:2502.09775}
}
read the original abstract

Building a virtual cell capable of accurately simulating cellular behaviors in silico has long been a dream in computational biology. We introduce CellFlux, an image-generative model that simulates cellular morphology changes induced by chemical and genetic perturbations using flow matching. Unlike prior methods, CellFlux models distribution-wise transformations from unperturbed to perturbed cell states, effectively distinguishing actual perturbation effects from experimental artifacts such as batch effects -- a major challenge in biological data. Evaluated on chemical (BBBC021), genetic (RxRx1), and combined perturbation (JUMP) datasets, CellFlux generates biologically meaningful cell images that faithfully capture perturbation-specific morphological changes, achieving a 35% improvement in FID scores and a 12% increase in mode-of-action prediction accuracy over existing methods. Additionally, CellFlux enables continuous interpolation between cellular states, providing a potential tool for studying perturbation dynamics. These capabilities mark a significant step toward realizing virtual cell modeling for biomedical research. Project page: https://yuhui-zh15.github.io/CellFlux/.

Figures

Figures reproduced from arXiv: 2502.09775 by the authors.

Figure 1
Figure 1. Overview of CellFlux. (a) Objective. CellFlux aims to predict changes in cell morphology induced by chemical or gene perturbations in silico. In this example, the perturbation effect reduces the nuclear size. (b) Data. The dataset includes images from high-content screening experiments, where chemical or genetic perturbations are applied to target wells, alongside control wells without perturbations. Control wells p… view at source ↗
Figure 2
Figure 2. CellFlux algorithm. (a) Training. The neural network vθ learns a velocity field by fitting trajectories between control cell images (x0 ∼ p0) and perturbed cell images (x1 ∼ p1). At each training step, intermediate states xt are sampled along the linear interpolation between x0 and x1, with t ∼ U[0, 1]. The network minimizes the loss L, which measures the difference between the predicted velocity vθ(xt, t, c) and th… view at source ↗
Figure 3
Figure 3. Qualitative comparisons. CellFlux generates significantly more accurate images that reflect the actual biological effects of perturbations compared to baselines. For example, Floxuridine inhibits DNA replication, leading to reduced cell density; AZ138 is an Eg5 inhibitor, causing cell death and shrinkage; Demecolcine destabilizes microtubules, resulting in smaller, fragmented nuclei. Columns 1–5, 6–7, and 8–9 corres… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: CellFlux enables new capabilities. (a.1) Batch effect calibration. CellFlux initializes with control images, enabling batch￾specific predictions. Comparing predictions from different batches highlights actual perturbation effects (smaller cell size) while filtering out…
Figure 5
Figure 5. Figure 5: visualizes three batches of BBBC021 images using PCA, showing that each batch forms a distinct cluster. Notably, control (ctrl) and perturbed (trt) images from the same batch cluster together, rather than forming separate control and target clusters. This illustrates t…
Figure 6
Figure 6. Figure 6: More qualitative comparisons of generated samples on BBBC021. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: More qualitative comparisons of generated samples on RxRx1. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: More qualitative comparisons of generated samples on JUMP. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: (1/2) Bidirectional interpolation trajectory in BBBC021. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_9.png]
Figure 10
Figure 10. Figure 10: (2/2) Bidirectional interpolation trajectory in BBBC021. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_10.png]
Figure 11
Figure 11. Figure 11: Cross-dataset transfer of CellFlux. Although CellFlux is trained solely on BBBC021, it demonstrates zero-shot generalization to two unseen datasets—RxRx1 and JUMP. Notably, it can predict morphological changes induced by perturbations (AZ138 and Demecolcine) that are …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 26 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Phendiff: Revealing subtle phenotypes with diffusion models in real images

    Bourou, A., Boyer, T., Gheisari, M., Daupin, K., Dubreuil, V., De Thonel, A., Mezger, V., and Genovesio, A. Phendiff: Revealing subtle phenotypes with diffusion models in real images. In MICCAI, 2024

    work page 2024

  3. [3]

    B., et al

    Bunne, C., Roohani, Y., Rosen, Y., Gupta, A., Zhang, X., Roed, M., Alexandrov, T., AlQuraishi, M., Brennan, P., Burkhardt, D. B., et al. How to build the virtual cell with artificial intelligence: Priorities and opportunities. Cell, 2024

    work page 2024

  4. [4]

    D., Walls, R

    Caie, P. D., Walls, R. E., Ingleston-Orme, A., Daya, S., Houslay, T., Eagle, R., Roberts, M. E., and Carragher, N. O. High-content phenotypic profiling of drug response signatures across distinct cancer cells. Molecular Cancer Therapeutics, 2010

    work page 2010

  5. [5]

    Carpenter, A. E. Image-based chemical screening. Nature Chemical Biology, 2007

    work page 2007

  6. [6]

    N., Ackerman, J., Alix, E., Ando, D

    Chandrasekaran, S. N., Ackerman, J., Alix, E., Ando, D. M., Arevalo, J., Bennion, M., Boisseau, N., Borowa, A., Boyd, J. D., Brino, L., et al. Jump cell painting dataset: morphological impact of 136,000 chemical and genetic perturbations. BioRxiv, pp.\ 2023--03, 2023

    work page 2023

  7. [7]

    J., Skewes-Cox, P., and Godinez, W

    Cook, S., Chyba, J., Gresoro, L., Quackenbush, D., Qiu, M., Kutchukian, P., Martin, E. J., Skewes-Cox, P., and Godinez, W. J. A diffusion model conditioned on compound bioactivity profiles for predicting high-content images. bioRxiv, pp.\ 2024--10, 2024

    work page 2024

  8. [8]

    Scaling rectified flow transformers for high-resolution image synthesis

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., M \"u ller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024

    2024

Show all 37 references
  1. [9]

    M., Kraus, O., Victors, M., Arumugam, L., Vuggumudi, K., Urbanik, J., Hansen, K., Celik, S., Cernek, N., Jagannathan, G., et al

    Fay, M. M., Kraus, O., Victors, M., Arumugam, L., Vuggumudi, K., Urbanik, J., Hansen, K., Celik, S., Cernek, N., Jagannathan, G., et al. Rxrx3: Phenomics map of biology. Biorxiv, pp.\ 2023--02, 2023

    2023

  2. [10]

    Generative adversarial nets

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. Generative adversarial nets. In NIPS, 2014

    2014

  3. [11]

    J., Oktay, D., Lin, Z., Verkuil, R., Tran, V

    Hayes, T., Rao, R., Akin, H., Sofroniew, N. J., Oktay, D., Lin, Z., Verkuil, R., Tran, V. Q., Deaton, J., Wiggert, M., et al. Simulating 500 million years of evolution with a language model. Science, 2025

    2025

  4. [12]

    and Salimans, T

    Ho, J. and Salimans, T. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  5. [13]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. In NeurIPS, 2020

    2020

  6. [14]

    Z., Zhang, C

    Hung, A. Z., Zhang, C. J., Sexton, J. Z., O'Meara, M. J., and Welch, J. D. Lumic: Latent diffusion for multiplexed images of cells. bioRxiv, pp.\ 2024--11, 2024

    2024

  7. [15]

    T., Agmon, E., Akamatsu, M., Lundberg, E., Lyons, B., Ouyang, W., Quintero-Carmona, O

    Johnson, G. T., Agmon, E., Akamatsu, M., Lundberg, E., Lyons, B., Ouyang, W., Quintero-Carmona, O. A., Riel-Mehan, M., Rafelski, S., and Horwitz, R. Building the next generation of virtual cells to understand cellular biology. Biophysical Journal, 2023

    2023

  8. [16]

    Digital twins for health: a scoping review

    Katsoulakis, E., Wang, Q., Wu, H., Shahriyari, L., Fletcher, R., Liu, J., Achenie, L., Liu, H., Jackson, P., Xiao, Y., et al. Digital twins for health: a scoping review. npj Digital Medicine, 2024

    2024

  9. [17]

    V., Gupta, A., Hahn, M., Hauth, A., Hendon, D., Martinez, A., Minnen, D., Sirotenko, M., Sohn, K., Yang, X., Adam, H., Yang, M.-H., Essa, I., Wang, H., Ross, D

    Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Schindler, G., Hornung, R., Birodkar, V., Yan, J., Chiu, M.-C., Somandepalli, K., Akbari, H., Alon, Y., Cheng, Y., Dillon, J. V., Gupta, A., Hahn, M., Hauth, A., Hendon, D., Martinez, A., Minnen, D., Sirotenko, M., Sohn, K...

    2024

  10. [18]

    Revealing invisible cell phenotypes with conditional generative modeling

    Lamiable, A., Champetier, T., Leonardi, F., Cohen, E., Sommer, P., Hardy, D., Argy, N., Massougbodji, A., Del Nery, E., Cottrell, G., et al. Revealing invisible cell phenotypes with conditional generative modeling. Nature Communications, 2023

    2023

  11. [19]

    T., Ben-Hamu, H., Nickel, M., and Le, M

    Lipman, Y., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. In ICLR, 2023

    2023

  12. [20]

    T., Lopez-Paz, D., Ben-Hamu, H., and Gat, I

    Lipman, Y., Havasi, M., Holderrieth, P., Shaul, N., Le, M., Karrer, B., Chen, R. T., Lopez-Paz, D., Ben-Hamu, H., and Gat, I. Flow matching guide and code. arXiv preprint arXiv:2412.06264, 2024

    2024 arXiv

  13. [21]

    Flowing from words to pixels: A framework for cross-modality evolution

    Liu, Q., Yin, X., Yuille, A., Brown, A., and Singh, M. Flowing from words to pixels: A framework for cross-modality evolution. arXiv preprint arXiv:2412.15213, 2024

    2024 arXiv

  14. [22]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Liu, X., Gong, C., and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023

    2023

  15. [23]

    L., and Carpenter, A

    Ljosa, V., Sokolnicki, K. L., and Carpenter, A. E. Annotated high-throughput microscopy image sets for validation. Nature Methods, 2012

    2012

  16. [24]

    F., and Altschuler, S

    Loo, L.-H., Wu, L. F., and Altschuler, S. J. Image-based multivariate profiling of drug responses from single cells. Nature Methods, 2007

    2007

  17. [25]

    A., Liu, L., Carpenter, A

    Navidi, Z., Ma, J., Miglietta, E. A., Liu, L., Carpenter, A. E., Cimini, B. A., Haibe-Kains, B., and Wang, B. Morphodiff: Cellular morphology painting with diffusion models. In ICLR, 2025

    2025

  18. [26]

    Gpt-4 technical report, 2024

    OpenAI. Gpt-4 technical report, 2024. URL https://arxiv.org/abs/2303.08774

    2024 arXiv

  19. [27]

    J., and Lotfollahi, M

    Palma, A., Theis, F. J., and Lotfollahi, M. Predicting cell morphological responses to perturbations using generative modeling. Nature Communications, 2025

    2025

  20. [28]

    J., Mohamed, S., and Lakshminarayanan, B

    Papamakarios, G., Nalisnick, E., Rezende, D. J., Mohamed, S., and Lakshminarayanan, B. Normalizing flows for probabilistic modeling and inference. JMLR, 2021

    2021

  21. [29]

    E., Slack, M

    Perlman, Z. E., Slack, M. D., Feng, Y., Mitchison, T. J., Wu, L. F., and Altschuler, S. J. Multidimensional drug profiling by automated microscopy. Science, 2004

    2004

  22. [30]

    U-net: Convolutional networks for biomedical image segmentation

    Ronneberger, O., Fischer, P., and Brox, T. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015

    2015

  23. [31]

    M., Schaff, J

    Slepchenko, B. M., Schaff, J. C., Macara, I., and Loew, L. M. Quantitative cell biology with the virtual cell. Trends in cell biology, 2003

    2003

  24. [32]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015

    2015

  25. [33]

    and Ermon, S

    Song, Y. and Ermon, S. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019

    2019

  26. [34]

    R., et al

    Sypetkowski, M., Rezanejad, M., Saberian, S., Kraus, O., Urbanik, J., Taylor, J., Mabey, B., Victors, M., Yosinski, J., Sereshkeh, A. R., et al. Rxrx1: A dataset for evaluating experimental batch correction methods. In CVPR, 2023

    2023

  27. [35]

    Pixel recurrent neural networks

    Van Den Oord, A., Kalchbrenner, N., and Kavukcuoglu, K. Pixel recurrent neural networks. In ICML, 2016

    2016

  28. [36]

    X., Barzilay, R., Jaakkola, T., and Uhler, C

    Yang, K., Goldman, S., Jin, W., Lu, A. X., Barzilay, R., Jaakkola, T., and Uhler, C. Mol2image: improved conditional flow models for molecule to image synthesis. In CVPR, 2021

    2021

  29. [37]

    Zheng, Q., Le, M., Shaul, N., Lipman, Y., Grover, A., and Chen, R. T. Guided flows for generative modeling and decision making. arXiv preprint arXiv:2311.13443, 2023

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.