Central-to-Local Adaptive Generative Diffusion Framework for Improving Gene Expression Prediction in Data-Limited Spatial Transcriptomics
Authors: Yaoyu Fang, Jiahe Qian, Xinkun Wang, Lee A. Cooper, Bo Zhou
Organizations: Department of Radiology, Northwestern University. · Department of Cell and Developmental Biology, Northwestern University. · Department of Pathology, Northwestern University.
Spatial Transcriptomics (ST) provides spatially resolved gene expression profiles within intact tissue architecture, enabling molecular analysis in histological context. However, the high cost, limited throughput, and restricted data sharing of ST experiments result in severe data scarcity, constraining the development of robust computational models. To address this limitation, we present a Central-to-Local adaptive generative diffusion framework for ST (C2L-ST) that integrates large-scale morphological priors with limited molecular guidance. A global central model is first pretrained on extensive histopathology datasets to learn transferable morphological representations, and institution-specific local models are then adapted through lightweight gene-conditioned modulation using a small number of paired image-gene spots. This strategy enables the synthesis of realistic and molecularly consistent histology patches under data-limited conditions. The generated images exhibit high visual and structural fidelity, reproduce cellular composition, and show strong embedding overlap with real data across multiple organs, reflecting both realism and diversity. When incorporated into downstream training, synthetic image-gene pairs improve gene expression prediction accuracy and spatial coherence, achieving performance comparable to real data while requiring only a fraction of sampled spots. C2L-ST provides a scalable and data-efficient framework for molecular-level data augmentation, offering a domain-adaptive and generalizable approach for integrating histology and transcriptomics in spatial biology and related fields.
Figures & tables
Figure 1: Overview of the C2L-ST framework for central-to-local adaptive generative modeling in spatial transcriptomics. The framework integrates large-scale public histopathology data with limited local ST data to enable molecular-level generation across centers worldwide. The central generative diffusion model is first pretrained on large-scale histopathology image patches to learn diverse tissue morphological priors. Institution-specific local models from different centers are then adapted using a small number of paired image–gene samples (red), aligning morphological features with molecular information to produce synthetic histology patches conditioned on gene expression (green). The resulting synthetic data expand the effective training distribution without requiring the exchange of real patient data across institutions. The adapted local models are further applied to unsampled tissue regions (blue) to predict gene expression directly from histopathology images. This central-to-local strategy alleviates data scarcity by generating synthetic samples, reduces domain discrepancies through shared morphological priors, and respects institutional data privacy by allowing each center to adapt models locally without data exchange.
Figure 2: Detailed workflow of the C2L-ST framework integrating central pretraining and local adaptation. The central generative diffusion model is first trained on large-scale public histopathology datasets to learn transferable tissue morphological priors through iterative noising and denoising processes. Each institution/center then performs local adaptation using its private ST dataset, where histology patches and corresponding gene expression values (red) guide lightweight adaptation of the pretrained model to align morphology with molecular information. After adaptation, the local model generates synthetic histology patches conditioned on gene expression (green), expanding the effective dataset. These synthetic samples are combined with limited real data to train gene prediction networks that learn the mapping from histopathology images to gene expression. The trained predictors are subsequently applied to unsampled tissue regions (blue) to infer gene expression profiles directly from histology images. This unified pipeline enables molecular-level data augmentation, improves prediction performance under data-limited conditions, and maintains data confidentiality across centers.
Figure 3: Molecular and morphological correspondence in generated histology patches for the bowel (ZEN48) dataset. The heatmap shows hierarchical clustering of representative genes, revealing four distinct molecular clusters (A–D). Each cluster is associated with characteristic histology patches from synthetic images, shown below with color-coded borders corresponding to the cluster identity. Dashed borders indicate synthetic images, and solid borders indicate real images. The PCA projection of gene expression highlights the separation among clusters in molecular space. Representative synthetic patches within each cluster display distinct structures consistent with their molecular signatures, indicating that the model captures the relationship between transcriptional programs and local morphology.
Figure 4: Feature embedding similarity between real and synthetic histology patches across organs. t-SNE visualization of feature embeddings extracted from the pretrained Conch pathology encoder for real (brown) and synthetic (blue) histology patches in four spatial transcriptomics (ST) centers: bowel (ZEN48), breast (TENX13), skin (MEND40), and lung (MEND90). In each organ, the synthetic patches exhibit distributions that closely overlap with those of the real samples while also extending the overall embedding space, indicating both high fidelity and enhanced diversity of the generated data. This embedding-level consistency demonstrates that the C2L-ST model effectively captures local tissue heterogeneity and morphological diversity across distinct anatomical domains.
Figure 5: Comparison of cellular composition and morphological fidelity between real and synthetic histology patches. Violin plots on the left show the distribution of six major cell types, including background, neoplastic, inflammatory, connective, dead, and epithelial, quantified using HoverNet segmentation for real (orange) and synthetic (blue) images from bowel (ZEN48) and skin (MEND40) ST datasets. The segmentation visualizations on the right display corresponding real and synthetic tissue patches with color-coded cell-type annotations, where blue indicates non-neoplastic epithelial cells, red indicates neoplastic epithelial cells, yellow indicates inflammatory cells, and green indicates connective cells. The synthetic images reproduce cell-type distributions and spatial organization that closely resemble those of real samples, while also expanding the diversity of local cellular configurations. These results demonstrate that the C2L-ST framework preserves realistic cellular composition and structural integrity across distinct tissue domains.
Dataset (Organ)
Type
Background
Neoplastic
Inflammatory
Connective
Dead
Epithelial
ZEN48 (Bowel)
Real
0.060 ± 0.066
0.411 ± 0.176
0.010 ± 0.049
0.085 ± 0.118
0.013 ± 0.042
0.421 ± 0.209
Synthetic
0.086 ± 0.079
0.449 ± 0.247
0.015 ± 0.053
0.133 ± 0.154
0.017 ± 0.060
0.299 ± 0.240
TENX13 (Breast)
Real
0.007 ± 0.019
0.972 ± 0.071
0.003 ± 0.018
0.004 ± 0.025
0.001 ± 0.010
0.014 ± 0.052
Synthetic
0.009 ± 0.029
0.944 ± 0.134
0.007 ± 0.036
0.011 ± 0.057
0.004 ± 0.033
0.025 ± 0.089
MEND40 (Skin)
Real
0.0002 ± 0.002
0.0026 ± 0.015
0.0000 ± 0.000
0.0000 ± 0.000
0.0000 ± 0.000
0.997 ± 0.015
Synthetic
0.0002 ± 0.003
0.0031 ± 0.016
0.00002 ± 0.001
0.00007 ± 0.002
0.0002 ± 0.003
0.996 ± 0.017
Table 1: Comparison of cell-type composition between real and synthetic histology patches. Proportions (mean ± SD) of six major cell types identified by HoverNet across four representative ST datasets.
Dataset
Training setup
Synthetic samples per real sample
Real only
1×
2×
3×
4×
5×
6×
7×
8×
9×
10×
ZEN48 (Bowel)
w/ Synthetic
–
0.4078
0.4011
0.3972
0.3863
0.3819
0.3762
0.3692
0.3641
0.3627
0.3604
w/ Real + Synthetic
0.4115
0.3993
0.3888
0.3789
0.3770
0.3699
0.3657
0.3607
0.3569
0.3539
0.3546
TENX13 (Breast)
w/ Synthetic
–
0.3933
0.3804
0.3727
0.3674
0.3585
0.3526
0.3523
0.3458
0.3447
0.3419
w/ Real + Synthetic
0.3903
0.3755
0.3636
0.3628
0.3559
0.3511
0.3463
0.3462
0.3405
0.3386
0.3391
MEND40 (Skin)
w/ Synthetic
–
0.5492
0.5333
0.5241
0.5148
0.5084
0.5057
0.5193
0.5040
0.5033
0.5058
Table 2: Mean absolute error (MAE) of gene expression prediction with varying amounts of synthetic data. Each model was trained using either real data only, synthetic data only, or a combination of both. Synthetic-to-real ratios range from 0× (real only baseline) to 10× per real sample.
Figure 6: Performance improvement in gene expression prediction enabled by the C2L-ST framework. (A) Gene prediction performance evaluated across four local ST centers, including bowel (ZEN48), breast (TENX13), skin (MEND40), and lung (MEND90). Each plot shows the mean absolute error (MAE) as a function of the number of real samples per gene when training with real data only (orange) or with both real and synthetic data (blue). In all organs, incorporating synthetic samples leads to consistently lower prediction errors, particularly under data-limited conditions. All subsequent analyses, including panels (B) and (C), are based on models trained with real data combined with synthetic data at a synthetic-to-real ratio of 10 × . (B) Gene-wise comparison of prediction errors demonstrates that adding synthetic data reduces MAE for most genes, as indicated by the majority of points below the diagonal line. (C) Spatial expression maps of representative genes are shown for four organs: B2M in bowel, COX6C in breast, IGKC in skin, and SFTPC in lung. Each panel compares the ground truth, predictions with real and synthetic data, and predictions with real data only. The inclusion of synthetic data improves spatial coherence and preserves fine-grained molecular heterogeneity across distinct tissue domains.
Figure 7: Scaling analysis of gene expression prediction performance with varying synthetic-to-real data ratios. (A–B) MAE of gene expression prediction for bowel (ZEN48) and breast (TENX13) datasets under different training configurations. Each group of bars represents models trained with synthetic data only (blue), real data only (red dashed line, baseline), and a combination of real and synthetic data (orange). Results are shown for four levels of real data availability, including 5%, 10%, 25%, and 50% of the full dataset, across synthetic-to-real ratios ranging from 1 × to 10 × . In both datasets, adding synthetic data consistently reduces prediction error compared to models trained on real data alone, and performance continues to improve as the synthetic-to-real ratio increases.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Molecular and morphological correspondence in generated histology patches for the breast (TENX13) dataset. The heatmap shows hierarchical clustering of representative genes, revealing four distinct molecular clusters (A–D). Each cluster is associated with characteristic histology patches from synthetic images, shown below with color-coded borders corresponding to the cluster identity. Dashed borders indicate synthetic images, and solid borders indicate real images. The PCA projection of gene expression highlights the separation among clusters in molecular space. Representative synthetic patches within each cluster display distinct structures consistent with their molecular signatures, indicating that the model captures the relationship between transcriptional patterns and local morphology.
Figure 9: Molecular and morphological correspondence in generated histology patches for the skin (MEND40) dataset. The heatmap shows hierarchical clustering of representative genes, revealing four distinct molecular clusters (A–D). Each cluster is associated with characteristic histology patches from synthetic images, shown below with color-coded borders corresponding to the cluster identity. Dashed borders indicate synthetic images, and solid borders indicate real images. The PCA projection of gene expression highlights the separation among clusters in molecular space. Representative synthetic patches within each cluster display distinct morphologies consistent with their molecular signatures, indicating that the model captures the relationship between transcriptional patterns and local morphology.
Figure 10: Molecular and morphological correspondence in generated histology patches for the lung (MEND90) dataset. The heatmap shows hierarchical clustering of representative genes, revealing four distinct molecular clusters (A–D). Each cluster is associated with characteristic histology patches from synthetic images, shown below with color-coded borders corresponding to the cluster identity. Dashed borders indicate synthetic images, and solid borders indicate real images. The PCA projection of gene expression highlights the separation among clusters in molecular space. Representative synthetic patches within each cluster display distinct structures consistent with their molecular signatures, indicating that the model captures the relationship between transcriptional patterns and local morphology.
Figure 11: Comparison of cellular composition and morphological fidelity between real and synthetic histology patches in additional organs. Violin plots on the left show the distribution of six major cell types, including background, neoplastic, inflammatory, connective, dead, and epithelial, quantified using HoverNet segmentation for real (orange) and synthetic (blue) images from breast (TENX13) and lung (MEND90) ST datasets. The segmentation visualizations on the right display corresponding real and synthetic tissue patches with color-coded cell-type annotations, where blue indicates non-neoplastic epithelial cells, red indicates neoplastic epithelial cells, yellow indicates inflammatory cells, and green indicates connective cells. The synthetic images reproduce cell-type distributions and spatial organization that closely resemble those of real samples, while also expanding the diversity of local cellular configurations. These results demonstrate that the C2L-ST framework preserves realistic cellular composition and structural integrity across distinct tissue domains.
Figure 12: Robustness of C2L-ST to gene dropout in conditioning inputs on the MEND40 (skin) dataset. Representative real histology patches and synthetic images generated under two conditions are shown. The first synthetic set is conditioned on the complete set of 250 genes, while the second is generated after randomly dropping five genes from the conditioning input to mimic the common dropout phenomenon observed during ST experiments. The generated images remain visually consistent with real tissues, preserving cellular organization and structural integrity, which demonstrates the robustness of the C2L-ST framework to incomplete molecular supervision.
Figure 13: Evaluation of synthetic data for improving class-level prediction performance on the ZEN48 (bowel) dataset. Classification models were trained with synthetic data only (blue), real data only (red dashed line, baseline), and a combination of real and synthetic data (orange). Performance was assessed using accuracy, F1 macro, recall macro, precision macro, and area under the curve (AUC) for both one-vs-one and one-vs-rest configurations. Each x-axis represents the number of synthetic samples per real sample, corresponding to synthetic-to-real ratios ranging from 1 × to 10 × . Across all evaluated metrics, incorporating synthetic data improved classification performance compared with models trained on real data alone, demonstrating that molecularly guided synthetic histology generated by the C2L-ST framework can enhance not only molecular prediction but also class-level discriminative tasks.
Spatial transcriptomics (ST) enables genome-wide gene expression profiling while preserving tissue architecture, but its cost and limited scalability remain major bottlenecks. This has motivated models that predict spatial expression directly from routine histology. Despite promising results, most existing approaches operate at the gene level without leveraging established transcriptomic modeling practices and rely on heterogeneous gene selection strategies, which complicates fair comparison across methods. We propose to reformulate morphology-to-transcriptomics prediction as conditional generation in transcriptional program space, thereby exploiting coordinated transcriptional variation instead of predicting genes independently. Using consensus non-negative matrix factorization (cNMF), we extract a low-dimensional set of transcriptional programs capturing coordinated expression variation in the training data, and train a conditional diffusion model to generate program activations from histology. This formulation exploits coordinated transcriptional variation and substantially lowers the dimensionality of the conditional generative task.
Ruyter Swann, Dorent Reuben, Racoceanu Daniel
Sorbonne Université, CNRS, Inserm, AP-HP, Inria, Paris Brain Institute – ICM, Paris, France
Spatial Transcriptomics (ST) measures gene expression while preserving spatial context, but its high cost and low throughput leave public datasets small. Inferring expression directly from widely available Hematoxylin and Eosin (H&E) stained histology offers a cost-effective alternative. However, existing approaches face several limitations: regression methods over-smooth toward the conditional mean, while generative methods are faithful but require slow multi-step inference; most methods treat genes as independent and equally important, ignoring inter-gene dependencies and heterogeneous gene informativeness; and most are tailored to a single resolution, either spot-level or cell-level. To address these issues, we propose DriftST, a unified framework for inferring spatially resolved gene expression from H&E images. DriftST builds on a Cellular Drifting generative model that learns a direct drift from a histology-conditioned source to the expression distribution, retaining generative expressiveness while enabling efficient one-step generation. To capture gene structure, we introduce the STransformer, which combines a co-expression attention module for inter-gene dependencies with a gene residual gate for differential gene importance. Operating on a generic gene-panel representation, DriftST applies directly to both spot-level and cell-level data in one framework, and extensive experiments across diverse tissues and platforms show that it achieves state-of-the-art performance at both resolutions.
Yuhang Yang, Yonggan Bu, Shengyuan Zhou +2
University of Science and Technology of China, Hefei, China · Peking University Cancer Hospital, Beijing, China
Spatial Transcriptomics (ST) has transformed biomedical research by enabling the spatial mapping of gene expression across tissue sections. However, high operational costs, specialized equipment requirements, and sensitivity to experimental noise limit the accessibility and scalability of ST. Recent computer vision approaches aim to overcome these limitations by predicting spatial gene expression directly from histopathology images. While effective, current approaches often suffer from gene expression over-smoothing and overly uniform predictions across tissue regions, suggesting that further progress depends on learning representations that reflect the hierarchical and asymmetric structure of gene regulation and tissue morphology. To address these issues, we propose Hyperbolic Contrastive Learning with Entailment for Spatial Transcriptomics (HyCLoST), a hyperbolic contrastive learning model that captures the intrinsic hierarchical relationships within ST data. By leveraging hyperbolic geometry and a gene-to-image entailment loss, HyCLoST learns structured, biologically grounded representations that improve gene expression prediction accuracy, achieving a 6% reduction in MSE and an 8% increase in PCC across 26 ST datasets, over previous methods. Our source code is publicly available at https://github.com/BCV-Uniandes/HyCLoST
Daniela Vega, Paula Cárdenas, Hannah Ceballos +2
Center for Research and Formation in Artificial Intelligence Universidad de los Andes, Colombia