GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow

We thank the reviewer for their valuable feedback. The questions raised were extremely important for a more comprehensive evaluation and validation of our method. We have answered every question, performed more experiments, and explained the absence of spatial modelling which was the top concern by the reviewer. The complete set of results on all datasets could not fit the character limit. Those along with their corresponding images will be available in the final version or provided during the discussion period. Please let us know if further explanation or experiments needed to improve the quality of the final version of the paper.

Distinction between single-cell and multi-cell models is unclear, and generating full patches from single cells seems biologically questionable?

We acknowledge the potential confusion and will clarify this in the final version. The single-cell model generates 256×256 patches centered on a focal cell, using that cell’s gene expression as the conditioning input while preserving surrounding tissue context for realistic morphology, with preferential weighting over layout. In contrast, the multi-cell model aggregates gene expression from all cells in a patch using multi-head attention, capturing collective molecular states. While the core UNet architecture is shared, the RNA encoders differ: direct input for single-cell versus aggregated attention for multi-cell. Single-cell mode reflects how an individual cell’s transcriptome influences its immediate microenvironment, enabling study of cell-autonomous effects, whereas multi-cell mode models more complex intercellular interactions and heterogeneity. Training data preparation differs accordingly: single-cell patches are extracted centered on valid cells, yielding more homogeneous data enriched in melanocytes, whereas multi-cell patches use sliding windows over heterogeneous tissue. This distinction, alongside dataset biases, explains why single-cell models tend to perform better and produce purer cell-type reconstructions.

Evaluation relies too heavily on perceptual metrics like FID and SSIM; biological structural validation is missing?

Although FID uses Inception features trained on natural images, it provides useful relative benchmarks of image quality within our domain. We further explored emerging histopathology-specific evaluation metrics using UNI2-h foundational model:

UNI2-h FID: Domain-specific image quality assessment using a pathology foundation model.

UNI2-h embedding similarity: Comparing feature distributions of real and generated images.

Nuclear morphometric similarity: Quantification of circularity, eccentricity, and solidity from segmented nuclei, statistically compared between real/generated images.

Spatial energy similarity: Gray-level co-occurrence matrix energy.

Spatial complexity and feature magnitude: Derived from UNI2-h embeddings, capturing tissue structure.

Calculated these metrics across P1 C1 and C2 datasets but due to character constraints only presented C1 dataset comparison in the table below. Future work will integrate these metrics to better illustrate and diagnose specific multi-cell model limitations such as boundary definition and texture fidelity.

Table: Evaluation Metrics Comparison for C1 Dataset: Multi-cell and Single-cell Models

Benefits of rectified flow over simpler models are asserted but not empirically demonstrated; request ablation versus conditional UNet regression?

Rectified flow models inherently handle many-to-one gene-to-morphology mappings better than deterministic regressors, as supported by prior literature showing their simpler training and more efficient sampling compared to diffusion and autoencoders.

Per reviewer suggestion, we implemented a conditional UNet baseline with our gene encoder trained on direct regression losses (MSE, perceptual), results detailed in the included ablation table. The transformer-based RNA encoder outperforms simpler variants across all measured metrics (Section C2, Figure 7), and each encoder module is critical for interpretability (Section 4.3, Figure 4); Removing any module reduces gene importance detection and biological insight. These results justify the encoder’s complexity: slight performance trade-offs enable enhanced interpretability essential for understanding gene-morphology relationships.

We also extended experiments across the 59-sample HEST-1k dataset (12 organs, 1.6 million patches) to demonstrate scalability and cross-tissue generalization, with promising preliminary single-cell training underway. Cross-dataset tests highlight strong generalization despite limited gene overlap (~126/5000 shared genes), supporting that our model learns generalizable gene-morphology mappings rather than dataset-specific artifacts. Supplementary visuals corroborate these results.

Table: C1 Single-Cell Ablation Studies and Baseline | HEST-1k Dataset Multi-Cell

The model ignores spatial (x, y) coordinates of cells; this omission weakens biological plausibility and limits spatial modeling capacity?

We acknowledge this limitation and will clarify in the final version. Our approach prioritizes gene expression as the main determinant of morphology at the patch level, focusing on local tissue microenvironments where relative cellular positioning and molecular context suffice. Patch sizes (256×256 pixels) typically contain fewer than 50 cells with limited heterogeneity, reducing the necessity for explicit global coordinate modeling. Spatial coordinates are slide-specific, noisy, and vary across experiments, limiting their generalizability, while relative spatial relationships are inherently encoded in histology images and indirectly inferred by our model. Our preliminary attempts to incorporate explicit coordinate recovery degraded performance, highlighting the need for more sophisticated, invariant geometric representations. We are actively developing hierarchical frameworks integrating spatially aware tile generation and multi-scale reasoning to model both local and global tissue organization. Computational complexity demands careful architectural design; thus, we focused on this critical foundational step first.

Biologically meaningful generating a 256×256 histological image from single-cell gene expression? Downstream Application?

The single-cell generation serves as a proof-of-concept to isolate how individual transcriptomes influence cellular morphology and microenvironmental context, providing insights complementary to the multi-cell model’s tissue-level aggregation. This dual modality enables applications in drug mechanism studies, functional genomics, and biomarker discovery, where cross-modal prediction has demonstrated impact. Single-cell morphology prediction supports virtual cytomorphology for diagnosis and reveals how cell-intrinsic and extrinsic factors shape phenotypes, important in immunology and oncology for understanding paracrine signaling and tumor-immune interactions. Pathologist collaborators confirmed that even 256×256 images centered on a single cell contain morphological cues sufficient to broadly classify tumor categories, demonstrating biological relevance. This approach facilitates mechanistic insight and hypothesis testing challenging to obtain from bulk or patch-level analyses alone.

Thank you!

We are deeply grateful to Reviewer Uins for their thoughtful engagement and positive assessment of our work. We sincerely appreciate the time they invested in our manuscript and are especially thankful for intending to raise the score from 4 to 5.

Their excellent suggestions are invaluable, and we look forward to incorporating them to enhance our method's thoroughness in the final version. Should the reviewer have any additional questions, we would be delighted to address them promptly.

We thank the reviewer for their valuable feedback. The critiques mentioned were extremely fair and important for a more comprehensive evaluation and validation of our method. We have answered every question regarding novelty, performed more experiments with a new larger dataset, and explored biologically relevant metrics which was the top concern by the reviewer. The complete set of results on all datasets could not fit the character limit. Those along with new images will be available in the final version or provided during the discussion period. Please let us know if further explanation or experiments needed to improve the quality of the final version of the paper. With this, we hope the reviewer recognizes the quality, significance and originality of our research and consider to raise the score.

Paper assumes paired single-cell RNA images are always available?

Paired single-cell RNA and histology images are scarce due to technological constraints. We address this by using 10x Xenium datasets offering true subcellular transcriptomic profiles co-registered with images, enabling our framework to learn morphology from real paired data. Our method also supports near sc-level ST data with transcriptomics of mixed cells, leveraging spot cellularity modeling. We further trained and evaluated multi-cell model on 59 Xenium samples spanning 12 organs (HEST-1k dataset, 1.6 million patches post reprocessing), showing generalizability beyond melanoma and across tissues, with single-cell model results forthcoming due to longer training. The aim of our paper is a solution to this very problem.

Practical utility of solving the inverse problem?

Inverse model unlocks applications beyond morphology recovery for scRNA-seq data lacking spatial context. 1. Enables predictive visualization of morphological changes under gene perturbations (gene knockdown, drug effects), facilitating clinical diagnosis, in silico hypothesis generation and drug repurposing screens. 2. Gene importance analyses identify driver genes within clinically relevant pathways such as EMT and angiogenesis, supporting biomarker discovery. 3. Quality control tool in ST by flagging discrepancies between predicted and true histology that may signal artifacts or tissue issues. Ongoing collaborations with pathologists aim to validate these utilities. While existing work focuses on forward prediction from images to gene expression, our inverse framework complements and expands these research avenues with generative tissue modeling.

Novelty beyond the RNA encoder?

Novelty extends well beyond encoder design, four key contributions: 1. Pioneering ST inverse mapping with rectified flow dynamics, outperforming baselines; 2. Multi-head attention module for aggregating variable numbers of cells in spatial patches; 3. Gene relation network using low-rank factorization for interpretable, efficient gene-gene interaction modeling. This allows the model to learn both single and multi cell level information.; 4. Biological validation via pathway enrichment showing the model autonomously identifies clinically relevant signals.

Collectively, these innovations establish the first comprehensive framework tackling this inverse problem and set a new research direction linking generative modeling with spatial biology. The problem we solve is the most novel aspect of this research. Proven by the baseline comparisons which could not faithfully generate high quality cell/tissue level images.

Raw gene counts instead of pretrained models like scBERT?

Raw gene counts over pretrained embeddings like scBERT due to the specific requirements of ST data and multi-cell patch modeling. scBERT excels in single-cell embeddings for classification but lacks mechanisms for multi-cell spatial aggregation and gene-gene relational modeling critical to our image generation task. scBERT was trained on scRNA-seq with thousands of genes. Our ST datasets comprise smaller, variable targeted panels with limited gene overlap, reducing scBERT’s applicability. Our custom encoder preserves quantitative gene-level precision and interpretable gene relations tailored to morphology prediction, which pretrained embeddings designed for cell state representation do not capture.

Ablations on gene encoder?

We implemented and compared a conditional UNet baseline using our gene encoder trained with MSE and perceptual losses, confirming rectified flow’s advantages in capturing complex gene-to-image mappings. Our ablation study (Section C2, Figure 7) shows the transformer-based RNA encoder consistently outperforms simpler alternatives across all metrics, with each component critical for interpretability analyses (Section 4.3, Figure 4). Removing any module reduces gene importance detection and biological insight. These results justify the encoder’s complexity: slight performance trade-offs enable enhanced interpretability essential for understanding gene-morphology relationships.

Table: C1 Single-Cell Ablation Studies and Baseline | HEST-1k Dataset Multi-Cell

Improve evaluation using cross-modal metrics?

Evaluated cross-modal coherence by processing both ground-truth and generated images through the HE2RNA model (Schmauch et al., 2020) to predict RNA expression and computed Pearson correlations against actual RNA profiles. High correlations observed for validation sets: 0.991 ± 0.002 (C1) and 0.986 ± 0.001 (C2), indicating strong coherence between generated histology and transcriptomes. However, HE2RNA’s limited sensitivity to subtle histological variations suggests future improvements in forward models are vital for more discriminative evaluation. We will include these results in the final version.

FID on a natural image model, histopathology-specific evaluation metrics?

While FID leverages Inception features trained on natural images, it remains useful for relative comparisons within histology domains by capturing coarse texture and structural distributions. We further explored emerging histopathology-specific evaluation metrics using UNI2-h foundational model:

UNI2-h FID: Domain-specific image quality assessment using a pathology foundation model.

UNI2-h embedding similarity: Comparing feature distributions of real and generated images.

Nuclear morphometric similarity: Quantification of circularity, eccentricity, and solidity from segmented nuclei, statistically compared between real/generated images.

Spatial energy similarity: Gray-level co-occurrence matrix energy.

Spatial complexity and feature magnitude: Derived from UNI2-h embeddings, capturing tissue structure.

Calculated these metrics across P1 C1 and C2 datasets but due to character constraints only presented C1 dataset comparison in the table below. Future work will integrate these metrics to better illustrate and diagnose specific multi-cell model limitations such as boundary definition and texture fidelity.

Table: Evaluation Metrics Comparison for C1 Dataset: Multi-cell and Single-cell Models

Clearly highlight best results and metric directions?

We appreciate the suggestion to improve table clarity and will highlight all best-performing results in bold in the final version. Additionally, we will include clear metric direction indicators (↑ for higher is better, ↓ for lower is better) alongside each evaluation metric, facilitating intuitive interpretation across tables and speeding reader comprehension.

Dear Reviewer,

Thank you so much for your thoughtful follow-up and for taking the time to re-evaluate the paper. We're grateful for your reconsideration and for increasing the score! Your constructive suggestions have strengthened our paper.

The final version will incorporate the reviewers' comments and updated evaluations.

Best Regards,

Authors

We thank the reviewer for their valuable feedback and high score. We addressed all the raised questions, and performed all requested additional experiments. We have provided extended results for the C1 and new HEST-1k datasets below. Results for P1 and C2 will be available in the final version or provided during the discussion period. Please let us know if further explanation or experiments needed to improve the quality of the final version of the paper.

Multi-cell models worse than single-cell models? Model produces unrealistic outputs?

Better single-cell model performance (Tables 2–3) from both methodological and biological factors. Unlike multi-cell model using a sliding window, single-cell data is generated from patches centered on valid cells with transcript overlap above a threshold. We apply spatial weighting scheme which gradually decreases pixel weights from center toward margins, emphasizing target cell’s morphology while reducing peripheral influence. Patch sizes cover full cell areas accommodating size variability, yielding more homogeneous, pure-cell patches than multi-cell method. Melanoma-dominated samples compound this imbalance. Single-cell conditioning provides a focused gene expr context. Multi-cell patches include heterogeneous cell types, complicating gene-to-morphology mapping. In melanoma the diverse tumor immune and stromal microenvironment challenges the multi-cell model in resolving cell boundaries and texture consistency. We will include qualitative failure cases in the final version to illustrate these limitations.

We further explored emerging histopathology-specific evaluation metrics using UNI2-h foundational model:

UNI2-h FID: Domain-specific image quality assessment using a pathology foundation model.

UNI2-h embedding similarity: Comparing feature distributions of real and generated images.

Nuclear morphometric similarity: Quantification of circularity, eccentricity, and solidity from segmented nuclei, statistically compared between real/generated images.

Spatial energy similarity: Gray-level co-occurrence matrix energy.

Spatial complexity and feature magnitude: Derived from UNI2-h embeddings, capturing tissue structure.

Calculated these metrics across P1 C1 and C2 datasets but due to character constraints only presented C1 dataset comparison in the table below. Future work will integrate these metrics to better illustrate and diagnose specific multi-cell model limitations such as boundary definition and texture fidelity.

Table: Evaluation Metrics Comparison for C1 Dataset: Multi-cell and Single-cell Models

Comparisons to conditional UNet and ablation on encoder?

Diffusion models are appropriate generative baseline given the many-to-one gene-to-morphology relationship intrinsic to our task; deterministic regression models fail to capture this complexity. Rectified flow offers a simpler, faster alternative to traditional diffusion and regression methods. Following reviewer’s suggestion, we implemented a conditional UNet baseline using our gene encoder trained with MSE and perceptual losses; results in Table below.

Conducted an ablation study comparing transformer based RNA encoder with a simpler encoder in (Sec C2 Fig 7). Our ablation shows the transformer based RNA encoder consistently outperformed the simple encoder in all metrics. Regarding the individual components of our encoder, each of them is uniquely important for the interpretability analysis that we conduct in (Sec 4.3 Fig 4). Removing any component would limit the interpretable aspect of our encoder to identify the most important genes and their relationships. It is seen in the table below that the complexity of the RNA encoder modules is well justified as it takes a small hit in performance but allows for greater interpretability to analyse important gene relations.

Extended experiments to 59 human Xenium samples from 12 organs in the HEST-1k dataset, totaling 1.6M paired patches. Single-cell model is still training, multi-cell evaluations show robust generalization. Cross-dataset results (Tab 4) confirm that training on melanoma alone suffices for good performance despite limited gene overlap, indicating the model captures general gene-morphology mappings rather than dataset-specific artifacts. Figures 8–10 visually corroborate these findings.

Table: C1 Single-Cell Ablation Studies and Baseline | HEST-1k Dataset Multi-Cell

Downstream tasks cell types from morphology?

Collaborating with pathology partners to validate downstream applications, although these are beyond the current study’s scope centered on inverse mapping. Preliminary feedback confirms that pathologists recognize the high fidelity of generated images on melanoma datasets, successfully identifying tumor tissues, cancer-associated fibroblasts, vasculature, and immune infiltration comparable to ground truth. To probe downstream utility, apart from extra evaluations included in the above table, we performed unsupervised clustering using UNI2-h embeddings from generated images, yielding results highly consistent with clustering on real images. This supports that generated images preserve cell composition faithfully. Additionally, our gene importance analysis (Sec 4.3, Fig 4) highlights clinical pathways such as EMT, KRAS signaling, and angiogenesis (App D.1), suggesting the model can prioritize relevant biomarkers and therapeutic targets. The model can be used to predict expected tissue morphology from ST data, enabling consistency checks; deviations between predicted and true histology may flag technical artifacts, tissue quality issues, or biologically meaningful anomalies.

Diverse image outputs from the same gene input?

Evaluated diversity by generating multiple images conditioned on the same transcriptomic input (dataset C1). Mean LPIPS score 0.388, within the desirable range (0.2–0.5) balancing meaningful variation and biological plausibility. Mean SSIM variance 0.000263, mean PSNR variance 0.372, indicates maintained structural consistency, suggesting the model explores realistic morphological variability rather than producing random artifacts. Quantitative metrics confirm our model’s capability to generate biologically coherent diverse morphologies from identical molecular profiles. Representative diverse images will be included in the final version.

Changes in gene expr causally affect the generated image?

Our model can perform post-hoc generation under gene expr perturbations. Verification requires controlled perturbation ST datasets currently unavailable. Top 2–3 influential genes may not reflect a complete causal factor, however, causal disentanglement can be integrated in our framework. With the extensive HEST-1k dataset (~60 Xenium samples), future work could compare tumorous versus healthy tissues from the same organ to select candidate genes and verify causal effects. Image-level results of such controlled perturbations will be included in the final version.

Synthetic images be misinterpreted in pathology?

Xenium spatial transcriptomics data inherently exhibit sparsity and noise, measuring only a subset of genes defined by targeted panels—far fewer than scRNA-seq or bulk RNA-seq. Despite these challenges, our model robustly generates high-quality histological images, demonstrating resilience to gene expression bias, noise, dropout, and input sparsity. Notably, we trained on the C1/C2 datasets (~300 genes) and evaluated on the P1 dataset, which includes ~5,000 genes with only 126 overlapping, and observed strong generalization and performance (Table 4, Supplementary Figures 9 & 10). These results show the model’s practical applicability and robustness for real-world spatial transcriptomics workflows.

Thank you for your rebuttal and the additional experiments provided. The ablation study, the new deterministic baseline, the diversity quantification (LPIPS), and the deeper feature analysis have addressed my main technical concerns about the architectural justification and the rigour of the evaluation.

I understand the limitations of the gene perturbation study's data and the reasons behind the multi-cell model's performance. As noted in the rebuttal, these analyses have been deferred.

The authors have sufficiently addressed the critical points from my review. The paper is strengthened by these additions. I will maintain my original rating.