acceptodds
Withdrawn from ICLR 2027

From Snapshots to Trajectories: Benchmarking Senescence-Conditioned Cell Morphology Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cellular senescence, a state in which cells stop dividing but remain in tissues, drives tissue ageing and is a central readout for the development of new treatments for age-associated diseases. Yet no existing cell morphology benchmark captures this phenotype. We introduce SenoMorph, a senescence-conditioned cell morphology benchmark that pairs each single-cell image with a continuous senescence score derived from its own transcriptome. SenoMorph comprises 81,255 real cross-sectional snapshot cells across three tissues and nine cell types. To our knowledge, SenoMorph is the first public benchmark dedicated to this problem. To understand how existing generators perform on SenoMorph, we assess 9 representative baselines spanning 7 Flow Matching (FM) variants and 2 non-FM generators. Our evaluation shows that these generators, despite attaining competitive distributional scores, reverse the expected morphology trend on individual cell types. We further present SenoFlow, an FM-based generator that enforces two geometric properties under continuous conditions, batch-invariant pairing of source-target supervision and score-Lipschitz, feature-aligned dynamics at integration. SenoFlow consistently outperforms all baselines, lowering FID from 14.2 to 11.6 and raising mean from 30.7% to 87.7%. It also follows real trajectories and preserves cell identity on tracked live-cell data. We additionally release a SenoFlow-generated synthetic companion set of 812,550 score-conditioned cells that lift snapshots into trajectories to support data-scarce downstream tasks like senescent-cell classification with +3.9 pp ACC and +1.9 pp AUC gains.

Primary area:
datasets and benchmarks
Keywords:
cellular senescence, cell morphology generation, flow matching, computational biology

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ed4jOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces SENOMORPH, a benchmark of 81,255 single-cell DAPI crops from three public Xenium sections, each paired with a transcriptome-derived cell-type label and a continuous DeepScence senescence score, together with a synthetic companion set. Given a source cell, the task is to generate its morphology at a target senescence score. The paper also proposes SENOFLOW, a conditional flow-matching model in a VAE latent space. During training, it pairs each source cell with a target sampled from the most similar k-means prototype within the target group (GPC). At inference, it interpolates velocities between two anchor scores and adds a projected gradient of nuclear area or DAPI intensity (CVC). The method reports the best FID/KID and mean ∣ρ∣|\rho| on all three tissues and the highest expert ratings. Additional experiments on tracked live-cell data and downstream classification are presented as evidence that the generator follows real trajectories, preserves cell identity, and provides synthetic cells that benefit data-scarce tasks.

Critical strengths
  1. Conditioning cell morphology on a continuous endogenous state is a meaningful and under-explored problem.
  2. Deriving per-cell senescence scores from aligned Xenium image–transcriptome data is a sensible way to construct such a benchmark.
Critical weaknesses
  1. The main fidelity metric cannot distinguish the expected direction of change from its reverse. The central failure mode highlighted in the paper is that baselines reverse the expected morphology trend, yet the main biological-fidelity metric, mean ∣ρ∣|\rho|, takes an absolute value. If nuclear area should decrease with the score and a model makes it strictly increase, the model still obtains ∣ρ∣=1|\rho| = 1. In addition, PbioP_{\mathrm{bio}} steers samples along the gradients of the same observables that ∣ρ∣|\rho| evaluates, so the method's prior and the evaluation criterion largely coincide. Signed correlations and metrics based on observables not used by PbioP_{\mathrm{bio}} are needed.

  2. The method does not establish the claimed ordering across target senescence scores. The paper aims to make nuclear area decrease and DAPI intensity increase as the target senescence score increases. However, the directional condition in Property (P2) concerns how these features change during the generation of a single image. This does not guarantee the correct ordering between images generated at different target scores. For example, both a low-score output and a high-score output could have smaller nuclei than the source cell, while the high-score output still has a larger nucleus than the low-score output. Both generations would shrink the nucleus, but their final outputs would follow the wrong trend across scores. Moreover, the bio-direction correction in Eq. (5) does not explicitly account for whether the requested score is above or below the source cell’s own score. The authors should explain how CVC ensures the expected ordering across target scores and allows changes toward less senescent morphology when a lower score is requested.

Questions for authors
  1. Can the authors report signed per-cell-type correlations diρd_i\rho and the fraction of trajectories with the expected sign for all methods?
  2. Since Eq. (5) does not depend on ss and gi⊤v≥0g_i^\top v \ge 0 only constrains change from x0x_0 along flow time tt, how does CVC ensure monotone change across target scores, especially when ss is below the source cell's own score?
Reviewer_QSkTOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents a new type of biological task based on cellular aging called SenoMorph, where they model senescence-conditioned morphological changes. The main challenge with this type of modeling is that no single cell can be observed across different senescence scores. We can only have snapshot pairs because computing senescence destroys the imaging sample. The authors propose SenoFlow, a flow-matching-based method, to generate what simulated real senescence images look like. The paper provides 81,255 image-senescence score pairs.

Critical strengths
  1. The paper presents a new biological task that the cellular phenotyping community can simulate using flow-matching methods.

  2. The image and transcriptome come from the same physical cell (Xenium), a real advantage over datasets that pair modalities across different cells.

  3. The baseline suite is broad (nine methods) and architecturally fair (same U-Net, same FiLM conditioning).

Critical weaknesses
  1. The paper claims that the SenoFlow method models trajectories successfully. That claim cannot be made unless the authors have a longitudinal component where a cell is measured across different senescence scores. In the current evaluation setup, there is no available data in the benchmark that a model could be tested on to validate simulated trajectories. The paper needs to restate its contribution as a claim that the flow-matching method can resemble cells at a given senescence score for a population.

    To make that claim, the paper needs more analysis to showcase how the cells resemble at the endpoints. The authors can use the transcriptome to train an image-to-marker predictor on real cells and apply it to generated cells. Classical feature extraction methods such as CellProfiler can also be used to showcase how biologically relevant features align between the real and generated images in terms of specific features.

  2. The method and its primary metric are underspecified, and the main result is not independent of the method. The fidelity metric |ρ| is defined only as an absolute Spearman correlation between the senescence score and a per-cell-type morphological observable; how it is computed, how the observable is chosen, and how cell types are weighted are not stated, no real-data reference value is given, and the absolute value discards the direction the paper’s claims depend on. More importantly, SenoFlow’s inference-time correction (P_bio) directly pushes the same observable that |ρ| scores, and the ablation shows the headline gain comes from this correction rather than from the learned model. The correction is training-free and model-agnostic but is never applied to the baselines, so the comparison conflates a better model with added post-processing. The other correction (I_smooth) bypasses the network’s score conditioning and reduces intermediate scores to a blend of two boundary generations, which is the endpoint interpolation where the paper in Section 2.1 says the task should go beyond. Finally, the equations, pseudocode, and appendix disagree on what was run, and GPC cannot be constructed as described for the rarer cell types. Detailed questions are listed in Questions for Authors.

  3. The paper states that the dataset contains 81,255 real cross-sectional snapshot cells (lines 18-19) but in figure 2 (dataset composition) they count the generated images (SenoFlow images) also as a part of the dataset. Synthetic data is cheap to obtain and it cannot be included in a benchmark dataset. It can serve as an external reference for what can be achieved with the real data but it cannot be part of the dataset contribution of the paper. In the spirit of this thought, the authors should remove lines 459-461 stating how generated score-conditioned cells are useful for a benchmark.

  4. The paper motivates the problem by stating that we can only get cross-sectional snapshots of senescence score and imaging. My question is why is it important to model senescence in a cell across time? The paper never demonstrates a use case that would require trajectories rather than just endpoints.

  5. The domain experts scoring the generations are not independent and have conflict of interests because they belong to the same lab. It would be helpful if the validation comes from external biologists.

Questions for authors
  1. Metrics

    • (a) Exactly how is ∣ρ∣|\rho| computed? The paper does not explain the metric clearly. How is the "most significant observable" chosen, on which data, and how are cell types weighted?
    • (b) What is ∣ρ∣|\rho| for real test cells? Why is the absolute value used when the paper claims that baselines reverse the expected trend (lines 22–24)?
  2. Section 3.2 (Global Prototype Coupling)

    • (a) Can the authors explain the quantile-based score bin bb? It is unclear how it is selected (number of bins, and whether quantiles are computed per tissue or per cell type).
    • (b) What happens to rare cell types (e.g. B cells, pericytes, smooth muscle cells) when forming the K=64K=64 clusters per (c,b)(c, b) group? Are there enough cells to support these partitions, and how are groups with fewer than KK cells handled?
  3. Section 3.3 (Continuous-score Velocity Correction)

    • (a) Does the Pbio\mathcal{P}_{\mathrm{bio}} term depend on ss or α\alpha in the implementation? If not, how does adding it alone raise ∣ρ∣|\rho| from 23.7% to 42.6% (Table 3a)?
    • (b) What is the differentiable proxy for nuclear area (defined in Section 2.4 as a mask pixel count), and is the gradient taken on xtx_t or on a predicted final image?
    • (c) Which version was run: Eq. 5 with the null-space projection and the AB3 solver (Appendix C), or Algorithm 1 without the projection and with Euler steps?
  4. Continuity interpolation. I\mathcal{I} (Eq. 4) blends the velocities obtained at two boundary scores, so the network's conditioning on ss is never used to generate the image at the requested score. How is this reconciled with Section 2.1, which states that the task requires smooth morphological change "across the full score range, not merely interpolate between two endpoints"? What are the boundary scores (the "boundary percentiles" in Algorithm 1), and do the test scores 0 and 1 fall inside them?

  5. Fair comparison. CVC is training-free and model-agnostic. Results for the baselines with CVC applied are needed to separate a better model from the added post-processing.

Reviewer_3fauOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper establishes the 1st public benchmark for senescence-conditioned cell morphology generation, addressing the gap that has so far prevented systematic study of an endogenous biological axis at the level of individual cells. The dataset pairs Xenium-derived single-cell DAPI crops with continuous, transcriptome-aligned senescence scores across multiple tissues and cell types, with a generated companion set that mitigates the scarcity of senescent cells in downstream tasks. Alongside the data, the authors provide a evaluation protocol, baselines spanning Flow Matching and non-FM models, and SENOFLOW, a Flow Matching framework that instantiates the coupling and velocity transformation of CFM with property-driven operators for continuous biological scores.

Critical strengths
  1. The proposed dataset is new and the paper has provided substantial supportive evidence to the benchmarking study.
  2. The proposed SENOFLOW has certain merits in terms of its soundness and contributions to the benchmarking framework.
  3. The paper has a good balance between the theoretical analysis, evaluation and benchmarking concepts.
Critical weaknesses
  1. This paper has pointed out the fact that, existing cell morphology datasets condition images on external perturbations, either chemical (BBBC021 [61], JUMP Cell Painting [13], the Cell Painting Gallery [79, 8, 12, 64, 39, 9, 10]) or genetic (RxRx3 [31], cpg0021 [13], OPS atlases [32, 71]), and observe each cell only once [72, 8, 12]. It seems that some of the papers contain much more cluttered images in their databases but the dataset proposed in this paper is fairly clean without much noise on the background. The database proposed in this paper may have less values in practical applications due to their cleanness that hold much higher expectation than the reality.
  2. This paper presents comparison results from FM and non-FM approaches. However, it is not clear whether or not the results from GAN and diffusion models demonstrate better performance than the proposed SENOFLOW in this paper. This lack of comparison reduces the value of the proposed framework.
  3. In the paper, section 3.3, the authors discuss the gradient drag cases, where the paper lacks detailed discussion of how the gradients help the system evolve in different directions, for example, in noisy or similar background images.
  4. In section 3.3, the strategy for obtaining and updating the weighted gradients is missing in this paper.
Questions for authors

Please address the major question raised in "critical weaknesses". In addition, if the images of cells in the scene become more complicated, e.g. noise and clutters, how confident the proposed SENOFLOW may achieve. This needs more justifications and discussions.

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

Related papers

Discussion (0)

Sign in to comment.