acceptodds
Under review as a conference paper at ICLR 2027

GeneFlowV2: RNA-Controllable Histology Generation and Counterfactual Editing

12 views

Abstract

Spatial transcriptomics (ST) aligns gene expression with histopathological morphology, enabling the investigation of how molecular states translate into observable visual phenotypes. However, RNA-conditioned generative models may produce visually realistic tissue images while underutilizing transcriptomic inputs, limiting their conditional fidelity and controllability. We introduce GeneFlowV2, a rectified-flow framework for generating cellular-level histopathological images from single-cell gene expression profiles. GeneFlowV2 integrates gene-aware RNA encoding with global and spatial conditioning, and further introduces a matched-versus-shuffled RNA ranking objective to enhance the model's sensitivity to transcriptomic variation. The model generates high-quality H&E histopathology images and supports RNA-conditioned visual counterfactual generation by editing an observed single-cell RNA profile toward a target RNA profile. Under the same spatially guarded evaluation protocol, across three 100K-scale image–gene paired datasets, GeneFlowV2 consistently outperforms GeneFlow in all experiments, demonstrating improved image generation quality, and stronger responsiveness to RNA conditioning.

Primary area:
generative models
Keywords:
Spatial transcriptomics, RNA-conditioned generation, Histopathology image generation, Rectified flow, Counterfactual

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_or9NOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies histopathological image generation conditioned on single-cell RNA expression, focusing on the problem that a model may generate visually realistic images while underutilizing its RNA input. Building on GeneFlow's conditional rectified-flow framework, the authors introduce gene-level token encoding, global and spatial conditioning pathways, and a ranking objective that compares velocity-prediction errors under matched and shuffled RNA conditions. They also enable RNA-guided image editing through image inversion and condition replacement. Experiments use three Xenium skin-melanoma datasets with spatially guarded splits that prevent overlap between training and validation image crops. The authors evaluate image quality, RNA consistency, and controllability using shared-noise generation, within-cell-type RNA swaps, and magnitude-matched gene perturbations. The results support more effective use of RNA conditions by GeneFlowV2 than by GeneFlow.

Critical strengths

The paper identifies a clear problem and aligns its evaluation with its central claim. The authors directly test whether the generator uses sample-specific RNA information by varying RNA while holding the initial noise fixed. They also audit image-crop overlap under the original random-split protocol and retrain the compared models on the same spatially guarded partitions. These choices strengthen the credibility of the controllability comparisons.

Critical weaknesses

The main quantitative comparisons focus on GeneFlow, without evaluating a GeneFlow baseline augmented with ranking supervision under the same training protocol. Section 4.5 explicitly states that the "without RNA ranking" variant removes ranking from both training and checkpoint selection; its performance degradation therefore cannot be attributed solely to the training loss. This limits the evidence for the separate contributions of the ranking objective, model-selection strategy, and redesigned conditioning architecture. The gains over simpler conditional-modeling alternatives are consequently not yet fully established.

Questions for authors
  1. Were the samples used for the results in Tables 1-4 also used for checkpoint selection and early stopping under Eq. (20)? If a separate test set was used, please clarify its sample counts and spatial separation. If not, could the main results be evaluated on a spatial region that was not used for model selection?
  2. Could you isolate the effects of ranking during training and checkpoint selection by comparing configurations with and without each component, while holding the other settings fixed? This would materially affect my assessment of the ranking objective's independent contribution.
  3. Could you evaluate GeneFlow augmented with ranking supervision and report the training budgets and checkpoint-selection procedures for both models? This comparison would help determine whether gene-level encoding and dual-path conditioning provide additional benefits beyond ranking supervision.
3 comments
AuthorsOpenReview ↗

Authors' rebuttal, Part 1

We sincerely thank the reviewer for the thoughtful and constructive feedback. Changes in the revised PDF are marked in blue.

Weakness: We have addressed these concerns through a controlled 2×2 ablation separating ranking loss from checkpoint selection, together with a GeneFlow+ranking baseline under matched training and selection protocols. The results confirm the independent contribution of training-time ranking supervision and demonstrate that GeneFlowV2's advantages are not solely attributable to ranking supervision, but also benefit from gene-level encoding and dual-path conditioning (see Q2–Q3; newly added Table 6 and Appendix Table 25).

Q1. Independent test evaluation and checkpoint selection.

In the original submission, Eq. (20) was computed on the reported validation region, which was used for both checkpoint selection and evaluation. We have therefore conducted additional evaluations on strictly independent test set, confirming that the main improvements persist.

  • C1 (independent test set): Train 70,783 / val 7,905 / test 20,049 cells. The test cells form a spatially separated tissue region, with no overlapping 256×256 crops across partitions. All models are trained for 50 epochs without early stopping, with checkpoints selected exclusively on the validation set. The test set is never used for training or model selection.
  • HEST-1K → TENX125 (zero-shot): Models are trained and selected on 14 FFPE slides spanning 8 organs (train 828,287 / val 207,072 cells), then evaluated zero-shot on TENX125, an unseen tonsil slide from an independent patient (8,192 cells), without fine-tuning or checkpoint selection on the new patient.

Key results are summarized below. Results for the independent C1 test are summarized in the newly added Table 6, with complete results in Appendix Table 25, while TENX125 zero-shot results are reported in the updated Table 1.

Test setMethodUNI2-h FID ↓UNI2-h cosine ↑MRR@32 ↑Gene Pearson ↑
C1 independent testGeneFlow97.050.4400.125−0.001
C1 independent testGeneFlowV248.620.5540.3030.227
TENX125 zero-shotGeneFlow166.230.3510.126−0.001
TENX125 zero-shotGeneFlowV292.550.5160.1780.085

On the independent C1 test, GeneFlowV2 reduces UNI2-h FID from 97.05 to 48.62, improves UNI2-h cosine from 0.440 to 0.554, and increases MRR@32 from 0.125 to 0.303. These improvements also persist on the unseen patient, where UNI2-h FID decreases from 166.23 to 92.55, alongside improvements in RNA consistency.

These results confirm that GeneFlowV2's improvements in image quality and RNA consistency persist on an independent test set excluded from checkpoint selection and extend zero-shot to an unseen patient.

Q2. Isolating ranking loss from checkpoint selection.

We conducted a controlled 2×2 ablation to independently evaluate the contributions of the RNA-ranking training loss and ranking-based checkpoint selection. On C1, we crossed the ranking loss (λ_rank ∈ {0, 0.5}) with two checkpoint-selection criteria: the flow-matching loss alone (Flow) and the ranking-aware criterion (Eq. (20)), while keeping all other settings fixed. Training was conducted without early stopping, and checkpoints were selected exclusively on the validation set. Results on the independent test set (20,049 cells) are summarized below (newly added Table 6; complete results in Appendix Table 25).).

Ranking lossSelectionUNI2-h FID ↓UNI2-h cosine ↑MRR@32 ↑Gene Pearson ↑
×Flow65.550.4890.1460.036
×Eq. (20)67.790.4940.1440.028
✓Flow51.700.5520.3020.229
✓Eq. (20)48.620.5540.3030.227

Independent contribution of ranking supervision. Under identical Flow-only checkpoint selection, adding the ranking loss increases MRR@32 from 0.146 to 0.302 and Gene Pearson from 0.036 to 0.229, while reducing UNI2-h FID from 65.55 to 51.70. The improvements also persist under identical Eq. (20) selection (cFID: 67.79 → 48.62; Gene Pearson: 0.028 → 0.227). These results demonstrate the independent contribution of training-time ranking supervision.

Additional contribution of ranking-aware selection. With ranking supervision fixed, switching from Flow to Eq. (20) further reduces UNI2-h FID from 51.70 to 48.62 and improves UNI2-h cosine from 0.552 to 0.554, while maintaining comparable RNA consistency (MRR@32: 0.302 vs. 0.303; Gene Pearson: 0.229 vs. 0.227).

Together, these controlled experiments establish that RNA-ranking training loss independently drives substantial improvements in conditional fidelity, while ranking-aware checkpoint selection provides additional benefits in selecting a model with stronger image quality and comparable RNA consistency.

AuthorsOpenReview ↗

Authors' rebuttal, Part 2

Q3. GeneFlow with ranking supervision and training-budget comparison.

We augmented the GeneFlow with the same RNA-ranking loss used in GeneFlowV2 (λ_rank = 0.5), without introducing gene-level encoding or dual-path conditioning. Following the same C1 split and 50-epoch training protocol described above, checkpoints were selected exclusively on the validation set using either Flow or Eq. (20), with identical selection criteria applied to both architectures. Results are reported in the newly added Table 6, with complete results and training budgets in Appendix Table 25.

ModelRanking lossSelectionGPU-h/epochUNI2-h FID ↓UNI2-h cosine ↑MRR@32 ↑Gene Pearson ↑
GeneFlow×Flow1.097.050.4400.125−0.001
GeneFlow✓Flow1.851.770.5480.2660.182
GeneFlow✓Eq. (20)1.870.010.5160.2310.138
GeneFlowV2✓Flow3.251.700.5520.3020.229
GeneFlowV2✓Eq. (20)3.248.620.5540.3030.227

Ranking supervision improves the GeneFlow. Under identical Flow-only checkpoint selection, adding the ranking loss increases GeneFlow's MRR@32 from 0.125 to 0.266 and Gene Pearson from −0.001 to 0.182, while reducing UNI2-h FID from 97.05 to 51.77. This demonstrates that ranking supervision is effective even without our redesigned conditioning architecture.

GeneFlowV2 provides additional gains beyond ranking supervision. Under the same ranking loss and Flow-only checkpoint selection, GeneFlowV2 outperforms GeneFlow+ranking in MRR@32 (0.302 vs. 0.266) and Gene Pearson (0.229 vs. 0.182), with comparable image fidelity (UNI2-h FID: 51.70 vs. 51.77). Under the same ranking-aware selection criterion, GeneFlowV2 further achieves better UNI2-h FID (48.62 vs. 70.01), MRR@32 (0.303 vs. 0.231), and Gene Pearson (0.227 vs. 0.138).

Performance–compute trade-off. With ranking supervision, GeneFlowV2 requires approximately 3.2 H100 GPU-hours per epoch, compared with 1.8 for GeneFlow. Despite the additional cost, GeneFlowV2 achieves stronger RNA-conditional fidelity with improved image quality under matched checkpoint-selection criteria.

These comparisons demonstrate that ranking supervision substantially improves the original GeneFlow, while GeneFlowV2's gene-level encoding and dual-path conditioning provide additional benefits beyond the shared ranking objective.

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts to address my concerns, particularly through the independent test-set evaluation, controlled ablation studies, and the comparison with GeneFlow augmented by ranking supervision. These results help clarify the contributions of the proposed components and strengthen the experimental evidence. However, I believe some concerns regarding the extent of the improvements beyond the ranking-augmented baseline remain. Overall, my assessment of the paper remains unchanged, and I will maintain my original score.

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

GeneFlowV2 is a rectified-flow model for generating cellular-level histology from single-cell RNA profiles. It diagnoses a conditional-bypass failure in GeneFlow and addresses it using a gene-token Transformer, complementary global modulation and expression-routed spatial cross-attention, and a matched-versus-shuffled RNA ranking loss in the noise-dominated part of the trajectory. It also supports source-anchored editing through ODE inversion. Experiments on three Xenium melanoma datasets use spatially guarded splits and show stronger image-to-RNA consistency and fixed-noise controllability, with generally improved pathology-feature image metrics. Within-cell-type swaps and targeted gene-program analyses provide additional evidence of profile-level sensitivity beyond coarse cell labels.

Critical strengths
  1. The paper identifies and addresses a consequential failure of RNA-conditioned generation. Fixed-noise experiments show that GeneFlow can produce realistic histology while responding minimally to its RNA input, including with official checkpoints. GeneFlowV2 substantially strengthens conditional dependence: C1 Top-5@32 improves from 15.40% to 53.14%. This supports acceptance because the work addresses whether the model actually uses its biological condition, a prerequisite for meaningful controllable generation.
  2. The evaluation addresses important sources of misleading performance. The authors identify substantial train–validation crop overlap under random splitting and retrain both methods using spatially guarded partitions. Matched-noise comparisons, within-type controls, and an additional RNA evaluator further strengthen the evidence. These safeguards materially increase my confidence that the reported improvements reflect better conditional modeling rather than spatial leakage or a single favorable evaluation setting.
Critical weaknesses
  1. Biological fidelity remains primarily dependent on imperfect learned readouts. Most RNA-control endpoints rely on H&E-to-RNA predictors whose real-image mean gene correlations range from approximately 0.16 to 0.32. The H0-mini analysis is useful, but the exact within-type endpoint is not repeated with that evaluator. This limits confidence that stronger measured RNA responsiveness consistently corresponds to biologically faithful morphology, which is central to the intended application.
  2. External generalization is not established. All three datasets concern Xenium skin melanoma, and each model is evaluated using spatial holdouts within the corresponding specimen rather than independent-patient or cross-tissue transfer. Consequently, the experiments establish improvements in a restricted setting but do not show whether the approach remains effective under biological or technical distribution shifts. This limits the breadth of the contribution and argues against a stronger acceptance recommendation.
Questions for authors
  1. How well does GeneFlowV2 generalize to independent specimens or patients beyond the current within-specimen spatial holdouts? Please provide any available external validation or clarify the limits of the generalization claims.
  2. Do the within-cell-type swap and targeted gene-program results persist when evaluated with a second independently trained H&E-to-RNA readout? These endpoints would benefit from validation beyond the primary evaluator.
  3. Can the authors disentangle the contribution of the RNA-ranking training loss from ranking-based checkpoint selection, for example by comparing models with and without the ranking loss under the same checkpoint-selection criterion?
1 comment
AuthorsOpenReview ↗

Authors' rebuttal

We thank the reviewer for the positive assessment and the constructive suggestions. Changes in the revised PDF are marked in blue.

W1. Biological fidelity. The within-cell-type RNA swap and targeted gene-program results are independently replicated using H0-mini, with all four 95% confidence intervals excluding zero. This confirms that the observed RNA responsiveness is robust across independently trained H&E-to-RNA evaluators (see Q2; Appendix F.2, newly added Table 24).

W2. External generalization. GeneFlowV2 consistently outperforms GeneFlow on the 1M-scale, eight-organ HEST-1K dataset and transfers zero-shot to TENX125, an unseen slide from an independent patient. These results establish generalization beyond melanoma and within-specimen spatial holdouts (see Q1; updated Table 1).

Q1. External generalization.

Thank you for this suggestion. To evaluate beyond the three Xenium skin-melanoma datasets (C1, C2, P1), we extend our experiments to a ~1M-pair, multi-organ, non-melanoma subset of HEST-1K (14 FFPE slides, 8 organs, 377 genes; 207,072 validation cells). We further test zero-shot transfer to TENX125, an unseen tonsil slide from an independent patient (8,192 cells), without fine-tuning or checkpoint selection on the new patient.

DatasetMethodInception FID ↓UNI2-h FID ↓UNI2-h Cosine ↑RNA Top-5@32 (%) ↑Gene Pearson ↑
HEST-1K (multi-organ)GeneFlow46.47148.210.32015.610.000
HEST-1K (multi-organ)GeneFlowV222.8773.180.47949.190.224
TENX125 (zero-shot)GeneFlow92.49166.230.35116.14−0.001
TENX125 (zero-shot)GeneFlowV218.9492.550.51624.080.085

GeneFlowV2 consistently improves image quality and RNA consistency across the multi-organ HEST-1K subset. These improvements persist on TENX125 without adaptation, demonstrating cross-patient generalization to an unseen patient. We have updated Table 1 to include these new results.

Q2. Independent H&E-to-RNA readout validation.

We repeated both evaluations using an independently fine-tuned H0-mini H&E-to-RNA readout, keeping the generated images and evaluation metrics unchanged (newly added Table 24, Appendix F.2).

TaskCFGGeneFlowGeneFlowV2Difference [95% CI]
Targeted perturbation: dose specificity (Δρ) ↑10.0830.4060.322 [0.064, 0.581]
Targeted perturbation: dose specificity (Δρ) ↑20.0830.3940.311 [0.111, 0.508]
Within-type control: swap margin ↑10.0000.1210.121 [0.077, 0.152]
Within-type control: swap margin ↑20.0000.1770.177 [0.123, 0.215]

GeneFlowV2 consistently outperforms GeneFlow on both targeted perturbation and within-type control, with all four 95% confidence intervals excluding zero. Thus, both findings persist under a second independently trained H&E-to-RNA evaluator.

Q3. Ranking loss vs. ranking-based checkpoint selection.

We have added a new controlled 2×2 ablation to disentangle the contributions of RNA-ranking training loss and ranking-based checkpoint selection. We split C1 into spatially disjoint training (70,783 cells), validation (7,905), and independent test (20,049) sets. We then crossed the ranking loss (λ_rank ∈ {0, 0.5}) with two checkpoint-selection criteria: the flow-matching loss alone (Flow) and the ranking-aware criterion (Eq. (20)). Training was conducted without early stopping. The GeneFlowV2 test-set results are summarized below (newly added Table 6; complete results in Appendix Table 25).

Ranking lossSelectionUNI2-h FID ↓UNI2-h cosine ↑MRR@32 ↑Gene Pearson ↑
×Flow65.550.4890.1460.036
×Eq. (20)67.790.4940.1440.028
✓Flow51.700.5520.3020.229
✓Eq. (20)48.620.5540.3030.227

Independent contribution of ranking supervision. Under identical Flow-only selection, adding ranking improves UNI2-h FID (65.55 → 51.70) and Gene Pearson (0.036 → 0.229) . These gains persist under identical Eq. (20) selection (UNI2-h FID: 67.79 → 48.62; Gene Pearson: 0.028 → 0.227), confirming the independent contribution of training-time ranking supervision.

Additional benefit of ranking-aware selection. With ranking supervision fixed, switching from Flow to Eq. (20) further reduces UNI2-h FID from 51.70 to 48.62 and improves UNI2-h cosine from 0.552 to 0.554, while maintaining comparable RNA consistency (MRR@32: 0.302 vs. 0.303; Gene Pearson: 0.229 vs. 0.227).

Together, these controlled experiments establish that RNA-ranking training loss independently drives substantial improvements in conditional fidelity, while ranking-aware checkpoint selection provides additional benefits in selecting a model with stronger image quality and comparable RNA consistency.

Reviewer_2PozOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

GeneFlowV2 extends GeneFlow to generate 256 × 256 H&E patches (plus a DAPI channel) centered on a cell, conditioned on that one cell’s Xenium expression profile. It adds a gene-token RNA encoder, two conditioning paths, a CFG plus a ranking loss, and counterfactual editing capability, compared to GeneFlow.

Critical strengths

S1. Clear diagnosis of GeneFlow and careful evaluation design.

S2. Large gains with informative ablations. RNA consistency improves on all three datasets without hurting image quality.

Critical weaknesses

The task’s motivation, success criterion and single-cell framing are not established.

W1. No use case. The paper lists counterfactual editing as a main contribution but never explains what it is for, and does not demonstrate any downstream use. The only stated aim is to study how molecular states translate into visual phenotypes.

W2. No biological interpretation. The authors state that an edit is the model’s response to alternative RNA, “not an identified biological effect” (Sec. 3.5). This rules out the most natural use: seeing how a cell would look in another molecular state.

W3. No checkable definition of success.

  • In natural image editing, people can judge whether an edit followed the instruction and kept the rest. Here, the true counterfactual is never observed (Fig. 4 caption).
  • Success is therefore defined by a learned H&E-to-RNA predictor. It is weak even on real images (Gene Pearson 0.32, 0.30 and 0.16), and it is only used to check whether the edit lands in the target cell type. It does not demonstrate the utility and the reliability of counterfactual editing.

W4. The editing metrics ignore the source. RNA Top-1, UNI2-h Top-1 and cFID all reward target-category membership only. An edit that discarded the source and drew a typical target-type patch would score as well. No source-preservation metric is reported, and on P1 editing is barely above chance.

W5. “Single-cell” is overstated.

  • The condition is one cell’s RNA, but the output is a 256 × 256 patch that contains many cells (Figs. 3–4).
  • The paper still describes “cellular-level” generation and edits that predict the cell’s “morphology.”
  • In practice, the model maps one cell’s RNA to its whole neighborhood, most of which comes from other cells whose RNA is not given. Edits accordingly change the whole patch (Fig. 4b).
  • Since neighboring cells share expression and the predictor reads the whole patch, the results show a patch-level association that may mostly reflect tissue region, not single-cell control.
Questions for authors

Please see weaknesses.

4 comments
AuthorsOpenReview ↗

Authors' rebuttal, Part 1

We thank the reviewer for the thoughtful feedback and constructive comments. Changes in the revised PDF are marked in blue.

W2. Biological interpretation. The use case identified by the reviewer, visualizing how a cell might look under an alternative molecular state, is precisely what we intend counterfactual editing to support. The statement in Section 3.5 is meant to distinguish model predictions from experimentally established biological effects, not to rule out biological interpretation. Given an observed H&E patch and its RNA profile, GeneFlowV2 reconstructs the source image and then generates an edited image conditioned on an alternative observed RNA profile. The resulting image therefore represents the model's prediction under the alternative RNA condition, without claiming to establish the actual effects of a biological intervention.

W1. Intended use and demonstration. We agree with the reviewer that this is a meaningful use case for counterfactual editing. It allows researchers to examine transcriptomic-morphological relationships by comparing predicted morphological changes while retaining the original tissue context. Such comparisons may help generate hypotheses about how molecular differences relate to changes in nuclear morphology and tissue organization.

We demonstrate this capability in Figure 4b and further support it with the new Table 3. On C1, RNA-guided editing preserves source structure (SSIM = 0.906, compared with 0.294 for generation from noise) while reducing target RNA prediction error by approximately 16% relative to source reconstruction.

To clarify the intended use and interpretation, we have revised the sentence in Section 3.5 as follows:

“Here, counterfactual refers to the model-predicted morphology under alternative observed RNA, not an experimentally established causal effect.”

W3. Checkable definition of success.

  • Evaluation criteria without an observed counterfactual. We agree that the true counterfactual cannot be observed directly. However, two aspects of an edit can still be evaluated: whether the generated image reflects the target RNA profile and whether it retains the structure of the source image. The target RNA profile provides a reference for the first, while the original image provides a reference for the second. We also report supporting results in Tables 2–3.

  • Evaluation beyond the RNA predictor. We agree that relying solely on the H&E-to-RNA predictor would be insufficient. Importantly, the evaluation in Table 2 has only one column that uses this predictor. Both UNI2-h Top-1 and cFID are based on frozen, pretrained UNI2-h image features. We have also added SSIM to Table 2 to directly assess how much of the source image structure is retained after editing (see also our response to W4(a)).

W4(a). Source preservation in counterfactual editing.

We have added Source SSIM to Table 2 to measure source preservation across all datasets. GeneFlowV2 achieves SSIM of 0.906, 0.918, 0.947, and 0.936 on C1, C2, P1, and the newly added 1M-scale HEST-1K dataset, respectively.

We further evaluate GeneFlowV2 on C1 (new Table 3) by comparing the original image, its reconstruction, RNA-guided editing, and generation from random noise under the same target RNA. Target RNA error measures the smooth-L1 distance between the RNA predicted from each image and the target RNA profile (lower is better).

OutputSSIM to source ↑Target RNA error ↓
Original source image1.0000.454
Source-RNA reconstruction0.9670.452
Target-RNA editing0.9060.382
Target-RNA generation from noise0.2940.360

Generation from noise achieves a similar target RNA error to editing (0.360 vs. 0.382), but has much lower source SSIM (0.294 vs. 0.906). Compared with reconstruction, editing reduces target RNA error by 16% while preserving substantial source structure.

These results show that GeneFlowV2 changes the image toward the target RNA while retaining the source, rather than simply regenerating a target-type image.

AuthorsOpenReview ↗

Authors' rebuttal, Part 2

W4(b). P1 editing is near chance.

We traced the weak P1 results to a severe cell-type imbalance in the original train/validation split: 88% of epithelial cells were assigned to validation, leaving only 262 cells (12%) for training, which limited the model's ability to learn this category. We corrected the split ensuring 20–22% validation coverage per cell type. We then retrained both the generator and the H&E→RNA predictor and repeated all evaluations under the same protocol. On the new split, GeneFlowV2 outperforms GeneFlow: editing RNA Top-1 improves from 15.08% to 20.73%, and UNI2-h Top-1 from 20.73% to 25.40%, both above the chance level (14.29%). We have updated Tables 1–2 with the new P1 results and moved the original protocol and results to Appendix G. The other datasets remain unchanged.

W5. "Single-cell" is overstated.

We agree that the output is a multicellular tissue patch rather than an isolated cell. Our aim is to use the central cell's RNA profile to condition the generation of a histopathological patch, rather than to suggest that the entire patch represents that cell alone. This is also the setup used in previous methods, including GeneFlow and GE2Hist. Thus, "cellular-level" refers to the single-cell resolution of the conditioning profile, not to the spatial extent of the generated image.

To investigate how localized the resulting changes are, we measure pixel-level changes using Xenium nucleus and cell segmentation masks. The newly added Table 12 in Appendix D.4, shows that GeneFlowV2 produces 1.6× greater changes in the central nucleus than in the surrounding tissue, whereas GeneFlow produces very little change. Thus, although editing also affects the surrounding tissue, the morphological response is more pronounced around the central nucleus.

I thank the authors for the detailed response. The new analysis addresses part of my concern, but my main concerns still remain. “How a cell would look in another state” only works if the outputs can be trusted biologically, and nothing in the paper or rebuttal validates this against known biology. It remains unclear how well the evaluation metrics reflect true biology, and whether a model that scores well on those "good evaluation" metrics actually captures true biology. I therefore maintain my score.

We thank the reviewer for the follow-up and constructive feedback. Regarding the remaining concern about biological validation:

We conducted a new model-free evaluation against real tissue. Both GeneFlow and GeneFlowV2 were retrained using the same spatially guarded train/validation/test split reported in newly added Table 6, in which every training crop overlapping a held-out crop is excluded. All analyses use the 20,049 independent C1 test cells from this split, which were never used for training or checkpoint selection, and no evaluation parameters were fitted on these cells. Without any learned evaluator, we measured stain-based morphology by classical H&E stain densitometry: colour deconvolution into hematoxylin, eosin and a residual channel, summarised by five fixed-formula features (mean hematoxylin, mean eosin, mean residual-channel optical density, mean total optical density, and 90th-percentile hematoxylin). We applied this to real H&E images and to images generated from the same cells’ RNA profiles. All normalisation constants and reference directions were derived from real images only. Across 1,410 gene–morphology associations (282 genes × 5 features), GeneFlowV2 reproduces the relationships measured in real tissue (r = 0.91), with all 50 strongest real associations recovered in the correct direction (100% vs. 2% for GeneFlow). It also reproduces the known keratin–eosin association (ρ = 0.45 vs. 0.34 in real tissue and −0.01 for GeneFlow).

Table . Marker-gene associations with residual-channel optical density (Spearman ρ across 20,049 held-out cells; real images vs. images generated from the same cells’ RNA)

GeneLineageReal tissueGeneFlowGeneFlowV2
KRT17Epidermal0.52−0.010.58
DSPEpidermal0.57−0.020.62
TACSTD2Epidermal0.54−0.010.58
S100BMelanocytic−0.630.01−0.61
TYRMelanocytic−0.560.01−0.50
MLANAMelanocytic−0.530.00−0.49

These results show that GeneFlowV2 reproduces both positive and negative marker-gene associations observed in real tissue, whereas GeneFlow exhibits near-zero correlations.

Our evaluation metrics track this biology. The model ranked higher by our learned controllability metrics is the one that reproduces real-tissue biology under this model-free readout. Across the five model configurations we evaluated, the two rankings are identical (Spearman = 1.0).

The same biological agreement holds for counterfactual editing. RNA-guided edits move toward the measured stain-based morphology of the target cell type: the projection onto the real source-to-target direction is 0.35 real-cell SD vs. 0.00 for GeneFlow. The predicted “cell in another state” therefore shifts toward how real cells in that state look. Furthermore, graded in silico upregulation of a five-gene keratinocyte program (KRT5, KRT17, KRT15, DSP, TP63) in held-out myeloid RNA profiles shifts the generated morphology toward real epithelial morphology in a dose-dependent manner (ρ = 0.85 vs. −0.02 for GeneFlow), with a positive dose–response in all 12 held-out sources. At full dose, the generated morphology covers 30% of the real myeloid-to-epithelial morphological difference, compared with −7% for a neutral-gene control.

Scope. Since ground-truth counterfactuals are unobservable, our claims remain scoped to consistency with biology measured in real tissue rather than causal effects (Section 3.5), in line with prior counterfactual histopathology work (MoPaDi[1], His-MMDM[2]). We will include these analyses in the appendix of the final version.

[1] Žigutytė et al., Counterfactual Diffusion Models Provide Interpretable Explanations of Artificial Intelligence Models in Pathology. Cancer Research, 2026.

[2] Li et al., His-MMDM: Multi-Domain and Multi-Omics Translation of Histopathological Images with Diffusion Models. Advanced Science, 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.