STP-BENCH: A Unified Systematic Benchmark for Virtual Spatial Transcriptomics from Histopathology Images
Authors: Youngmin Chung, Ji Hun Ha, Andrew H. Song, Cristina Almagro-Pérez, Chaeyoung Seo, Won Jun Suh, Jeong Won Beom, Kyoung Bin Oh, Eytan Ruppin, Faisal Mahmood, Joo Sang Lee
Organizations: Department of Artificial Intelligence, Sungkyunkwan University, Suwon, Republic of Korea · Department of Pathology, Mass General Brigham, Harvard Medical School, Boston, MA, USA · Department of Precision Medicine, School of Medicine, Sungkyunkwan University, Suwon, Republic of Korea · Department of Translational Molecular Pathology and Division of Pathology and Laboratory Medicine, The University of Texas MD Anderson Cancer Center, Houston, TX, USA · Data Science Program, Dana-Farber Cancer Institute, Boston, MA, USA · Cancer Program, Broad Institute of Harvard and MIT, Cambridge, MA, USA · Harvard-MIT Division of Health Sciences and Technology, Massachusetts Institute of Technology, Cambridge, MA, USA · Translational Research Unit, Cedars-Sinai Medical Center, Los Angeles, CA, USA · Harvard Data Science Initiative, Harvard University, Cambridge, MA, USA · Department of Digital Health, Samsung Advanced Institute of Health Science and Technology, Sungkyunkwan University, Seoul, Republic of Korea
Spatial transcriptomics (ST) provides unprecedented insights into tumor heterogeneity by capturing spatially resolved gene expression, yet its high experimental cost hinders large-scale adoption. Consequently, computational approaches that predict spatial gene expression directly from hematoxylin and eosin slides, termed virtual ST, have rapidly emerged. Despite this progress, assessing advances in the field remains difficult due to insufficient benchmarking: prior studies rely on small, heterogeneous datasets, inconsistent training and inference pipelines, and limited evaluation of biological interpretability and model robustness. To address these gaps, we present STP-BENCH, a standardized benchmark for virtual ST models. STP-BENCH comprises six cancer types spanning two ST platforms (Visium and Xenium), with each training dataset containing more than 30,000 spots and at least 15 slides to ensure statistical reliability. We evaluate 21 predictive approaches, re-implemented with a unified pathology foundation model as the morphological encoder when architecturally applicable. Beyond conventional benchmarks that report average predictive accuracy on highly variable genes, we systematically examine which genes and gene sets are recoverable from histomorphology. We further evaluate the downstream biological utility of predicted profiles through cell-type deconvolution and spatial domain identification, and assess model reliability under domain shifts and data scaling. Notably, unified morphological encoding substantially re-orders model rankings established in prior studies, indicating that architectural innovations and image encoding have been conflated in previous evaluations. We publicly release STP-BENCH to support reproducibility and serve as a community benchmark at https://github.com/NEXGEM/STP-Bench.
Figures & tables
Figure 1: Overview of STP-Bench and benchmarked ST prediction models . (a) Virtual ST is generated by predicting spatial gene expression from H&E-stained histology image patches. (b) Distribution of the large-scale dataset that combines Visium and Xenium platforms, categorized into STP-Bench-Internal (609,024 spots across 6 cancer types) and STP-Bench-External (185,831 spots across 6 cancer types) for training and testing, respectively. (c) Workflows of the three methodological categories (a total of 21 methods) evaluated on STP-Bench : Regression models (14 models), Bi-modal learning & Retrieval models (4 models), and Generative models (3 models). Diagrams depict the training and inference procedures for each category. (d) Extensive evaluations within STP-Bench . This includes overall predictive performance comparison and impact of different patch image encoders. (e) In-depth analysis of gene and gene set level prediction quality, including a gene-wise performance comparison across all benchmarked models, gene set analysis linking representative marker genes to functional gene sets, and per gene set Pearson Correlation Coefficient (PCC) comparisons. (f) Evaluation at different biological granularities, demonstrating the utility of virtual ST profiles for downstream tasks including spatial domain identification and cell type deconvolution. (g) Robustness and scalability analyses, comprising a cross-institution correlation between internal validation and external test performance, and data scaling experiments under intra-cohort and inter-cohort settings.
Figure 2: Performance benchmark of virtual ST baselines on STP-Bench-Internal . (a) The average PCC of virtual ST models evaluated on high-mean highly-variable genes (HMHVG) (top) and 16 marker genes (bottom), across 7 evaluation datasets. For each model, the patch features were extracted with UNIv2 (w/ PFM) unless a model required a specific patch feature extractor (w/o PFM, diagonal lines). The baseline (dashed line) represents the linear probing performance with UNIv2 image features. Error bars indicate the standard deviation across cross-validation folds. (b) The comparison between virtual ST models when using UNIv2 feature extractor and the native feature extractors from respective original studies. (c-d) Evaluation on representative NCCHE-LUAD-Xenium (c) and NCCHE-LUAD-Visium (d) datasets. Averaged gene counts are shown for each gene across datasets. The bar charts (left) rank the models based on their PCC performance on each specific dataset, with error bars indicating the standard deviation across cross-validation folds. The box plots (right) illustrate the distribution of prediction accuracy (PCC) across five gene expression levels (Very Low to Very High), with Kruskal-Wallis test results indicating statistical significance. (e) PCC for ten selected models that utilize different patch encoders on the union of HMHVGs and marker genes: H-Optimus1, Virchow2, UNIv2, CTransPath, and ResNet50.
Figure 3: Individual gene and gene set analysis . (a) Gene-wise performance heatmap across 20 benchmarked methods. Each row and column represents individual genes (union of HMHVGs and marker genes) and prediction models, respectively, with the color bar below indicating the methodological category of each model. The color scale indicates the gene-wise PCC. Whenever applicable, UNIv2 was used as the patch encoder. (b) Performance comparison for genes exhibiting high absolute performance (PCC ≥0.4 ) and the highest predictive margin over the linear probing baseline. Bars represent the PCC for the linear baseline (LinearProb) and the selected models. Red asterisks indicate genes for which the selected model exceeded the baseline by more than 0.1. (c-e) Gene set scoring prediction performance evaluated via singscore [ 22 ] on NCCHE-LUAD-Xenium. Predicted gene expression profiles were used to compute per spot gene set scores using singscore, which were then compared against ground-truth scores derived from observed expression. (c) Bar chart showing averaged PCC across all evaluated gene sets. (d) Heatmap showing per gene set PCC across all benchmarked methods. (e) Bar charts showing gene set level results for representative pathways.
Figure 4: Evaluation at multiple biological granularities across three cohorts (NCCHE-LUAD-Xenium, NCCHE-LUAD-Visium, and WUSTL-BRCA-Visium) . In all panels, reference annotations are derived by applying the respective downstream analysis directly to ground-truth gene expression profiles, serving as the upper bound for performance evaluation. (a) Performance for cell type abundance estimated by Cell2location [ 37 ] . Heatmaps display per-cell-type PCC between predicted and ground-truth cell type abundance estimates for six representative models, with row-wise averages summarizing overall deconvolution accuracy. (b-c) Performance for spatial domain identification, where the domains are identified with SpaGCN [ 29 ] . Adjusted Rand Index (ARI) scores were computed between the spatial domains identified from the ground truth (Reference) and virtual ST from each method. (b) Visual comparison of representative tissue sections with reference and predicted spatial domain maps, where each color in the legend denotes a distinct spatial domain. (c) Bar charts summarizing mean ARI scores across all test samples, with ARI annotated within each bar and sample-wise standard deviations indicated as error bars.
Figure 5: Robustness and Scalability Analysis . (a) Scatter plots illustrating the correlation between performance on STP-Bench-Internal (x-axis) and STP-Bench-External (y-axis) under three generalization scenarios: Cross-Institution (left), Cross-Tissue (center), and Cross-Platform (right). The displayed external performances are averaged across all dataset pairs (i.e., specific train-test data combinations) within each scenario. To assess the consistency of the model rankings, the Spearman correlation coefficient was used. Models are color-coded by category (Regression, Bi-Retrieval, Generative). The black dotted line indicates equal internal and external performance. The red dotted line indicates the fitted linear regression model to quantify the relationship between internal and external performance. Whenever applicable, UNIv2 was used as the patch encoder. (b) Bar charts quantifying the generalization gap between the STP-Bench-External and STP-Bench-Internal performance for each model across the same scenarios as (a): Cross-Institution (left), Cross-Tissue (center), and Cross-Platform (right). The “No Gap” line (red) indicates no discrepancy between internal and external test performance. The models are ordered by the increasing PCC on STP-Bench-Internal , also represented by the color code. Error bars represent standard deviation of the performance differences across all dataset pairs. (c) Bar chart showing the PCC performance of the selected models (LinearProb, STFlow, BLEEP, EGN, TRIPLEX, DeepSpot) when trained on subsets of increasing size (Small: 33%, Medium: 66%, Full) within the same dataset (WUSTL-BRCA-Visium). The error bars represent the standard deviation across all the samples in each cohort. (d) Bar chart showing the impact of expanding training data across multiple cancer types (BRCA only vs. BRCA+LUAD vs. BRCA+LUAD+PRAD) on model performance.
Extended Data Fig. 1: Gene-wise spatial expression prediction performance stratified by predictability for HMHVG . Gene-wise Pearson correlation coefficient (PCC) distributions of five ST prediction models (STFlow, BLEEP, EGN, TRIPLEX, and DeepSpot) across gene predictability bins. 200 HMHVG genes per dataset are ranked by their mean PCC across all models and partitioned into five equal-frequency bins (Very Low to Very High). Each box shows the interquartile range of per-gene PCC for a given model within each bin; outliers are omitted for clarity.
Extended Data Fig. 2: Dataset-wise performance of virtual ST models on STP-Bench-Internal . Evaluation results corresponding to Fig. 2c–d, extended to all remaining datasets in STP-BENCH-INTERNAL. For each dataset, averaged gene counts per gene are shown. The bar charts (left) rank the models based on their Pearson correlation coefficient (PCC) performance for each dataset, with error bars indicating the standard deviation across cross-validation folds. The box plots (right) illustrate the distribution of prediction accuracy (PCC) across five gene expression levels (Very Low to Very High), with statistical significance assessed using the Kruskal–Wallis test.
Extended Data Fig. 3: Performance benchmark of virtual ST baselines on STP-Bench-Internal using SSIM . (a) The average Structural Similarity Index Measure (SSIM) of virtual ST models evaluated on high-mean highly-variable genes (HMHVG) (top) and 16 marker genes (bottom), across 7 evaluation datasets. For each model, the patch features were extracted with UNIv2 (w/ PFM) unless a model required a specific patch feature extractor (w/o PFM, diagonal lines). The baseline (dashed line) represents the linear probing performance with UNIv2 image features. Error bars indicate the standard deviation across cross-validation folds. (b) The comparison between virtual ST models when using UNIv2 feature extractor and the native feature extractors from respective original studies, evaluated by SSIM. (c-d) Evaluation on representative NCCHE-LUAD-Xenium (c) and NCCHE-LUAD-Visium (d) datasets using SSIM. Averaged gene counts are shown for each gene across datasets. The bar charts (left) rank the models based on their SSIM performance on each specific dataset, with error bars indicating the standard deviation across cross-validation folds. The box plots (right) illustrate the distribution of prediction accuracy (SSIM) across five gene expression levels (Very Low to Very High), with Kruskal-Wallis test results indicating statistical significance. (e) SSIM for ten selected models that utilize different patch encoders on the union set of HMHVG and marker genes: H-Optimus1, Virchow2, UNIv2, CTransPath, and ResNet50.
Extended Data Fig. 4: Gene- and gene-set-level prediction performance across additional datasets . (a) Gene-wise performance heatmaps across benchmarked methods for additional datasets (HEST-PRAD, HEST-CCRCC, SNU-GBM, and WUSTL-PDAC). Each row and column represent individual genes (union of HMHVGs and marker genes) and prediction models, respectively, with the color bar indicating the methodological category of each model. The color scale denotes the gene-wise Pearson correlation coefficient (PCC). (b) Gene-set Pearson evaluation in NCCHE-LUAD-Visium and WUSTL-BRCA-Visium. For each cohort, bar plots show model-wise average gene-set Pearson correlations, and heatmaps show per-gene-set PCCs across all benchmarked methods.
Extended Data Fig. 5: Impact of pathology foundation model choice on internal-to-external performance correlation of ST prediction models . (a-e) Scatter plots illustrating the correlation between performance on STP-Bench-Internal (x-axis) and STP-Bench-External (y-axis) under three generalization scenarios (Cross-Institution, Cross-Tissue, and Cross-Platform), evaluated separately for five patch encoders: H-Optimus1, Virchow2, UNIv2, CTransPath, and ResNet50. Only models compatible with interchangeable patch encoders are included. The Spearman correlation coefficient was used to assess consistency of model rankings. Models are color-coded by category (Regression, Bi-Retrieval, Generative). The black dotted line indicates equal internal and external performance; the red line indicates the fitted linear regression.
Extended Data Fig. 6: Impact of pathology foundation model choice on the generalization gap of ST prediction models . (a-e) Bar charts quantifying the generalization gap under three generalization scenarios (Cross-Institution, Cross-Tissue, and Cross-Platform), evaluated separately for five patch encoders: H-Optimus1, Virchow2, UNIv2, CTransPath, and ResNet50. Only models compatible with interchangeable patch encoders are included. Models are ordered by increasing Pearson correlation coefficient on STP-BENCH-INTERNAL, also represented by the color code. The red line indicates no generalization gap. Error bars represent standard deviation across all dataset pairs within each scenario.
Extended Data Fig. 7: Effect of Pathology Foundation Model Choice on Data Scaling . Scaling performance of selected spatial transcriptomics prediction models evaluated across different pathology foundation models (PFMs) used as patch encoders. The figure shows intra-cohort scaling results on WUSTL-BRCA-Visium using progressively larger training subsets and inter-cohort scaling results with progressively expanded training data composition. Model performance is reported as Pearson correlation, with error bars indicating the standard deviation across samples. These analyses assess whether the observed scaling trends are consistent across different PFM choices.
Dataset
Source
Cancer type
Platform
Slides
Patients
Total spots
STP-Bench-Internal
HEST-PRAD-Visium
HEST-1K
PRAD
Visium
23
2
62,710
WUSTL-BRCA-Visium
WUSTL
IDC
Visium
47
21
157,138
NCCHE-LUAD-Xenium
NCCHE
LUAD
Xenium
16
16
34,519
NCCHE-LUAD-Visium
NCCHE
LUAD
Visium
43
30
96,954
HEST-CCRCC-Visium
HEST-1K
RCC
Visium
24
24
74,220
Extended Data Table 1: STP-Bench benchmark dataset composition. Summary of training ( STP-Bench-Internal ) and independent test ( STP-Bench-External ) cohorts, detailing data source, cancer type, ST platform, number of tissue slides and patients, and total capture spots per dataset. The internal cohort spans seven datasets across six cancer types and two ST platforms (Visium and Xenium). The external cohort comprises datasets and samples that are non-overlapping with the training data to enable unbiased generalizability assessment.
Extended Data Table 2: Comparison of ST prediction models. FM indicates whether a foundation model is used. HEG: highly expressed genes; HVG: highly variable genes; HMHVG: high-mean highly-variable genes; ACG: all considered genes.
Model
Pretrain size
Pretrain dataset detail
Magnification
SSL
Model architecture
Model size
Year
H-Optimus1
1M+
50 organs, 800,000 patients
20x
DINOv2
ViT-g/14
1.1B
2025
Virchow2
3.1M
150+ tissues, 225,401 patients, H&E + IHC
5x, 10x, 20x, 40x
DINOv2
ViT-H/14
632M
2024
UNIv2
350k+
20+ organs, H&E + IHC
20x
DINOv2
Modified ViT-H/14
681M
2024
CTransPath
30k
32 cancer subtypes, 10,953 patients
10x
SRCL
CNN+Swin-Transformer
28M
2021
Extended Data Table 3: Pathology foundation model (PFM) description. Information about each PFM (parameters, pretraining data size, release year, etc.).
Spatial transcriptomics (ST) enables gene expression mapping within anatomical context but remains costly and low-throughput. Hematoxylin and eosin (H&E) staining offers rich morphology yet lacks molecular resolution. We present \ours (Spatial Transcriptomics and histOlogy Representation Model), a foundation model trained on 1.2 million spatially resolved transcriptomic profiles with matched histology across 18 organs. Using a hierarchical architecture integrating morphological features, gene expression, and spatial context, STORM bridges imaging and omics through robust molecular--morphological representations. STORM enhances spatial domain discovery, producing biologically coherent tissue maps, and outperforms existing methods in predicting spatial gene expression from H&E images across 11 tumor types. The model is platform-agnostic, performing consistently across Visium, Xenium, Visium HD, and CosMx. Applied to 23 independent cohorts comprising 7,245 patients, STORM significantly improves immunotherapy response prediction and prognostication over established biomarkers, providing a scalable framework for spatially informed discovery and clinical precision medicine.
Jinxi Xiang, Siyu Hou, Yuchen Li +17
Department of Radiation Oncology, Stanford University School of Medicine, Stanford, CA, USA · Department of Statistics and Data Science, Yale University, New Haven, CT, USA · Department of Medicine (Oncology), Stanford University School of Medicine, Stanford, CA, USA +4
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
Spatial transcriptomics (ST) links gene expression with tissue morphology but remains expensive and low-throughput, motivating surrogates that infer expression from routine histology. Whole-slide H&E-to-ST inference pairs a gigapixel image with gene measurements at a sparse, irregular set of locations, making multiscale modeling challenging without incurring dense-grid overhead or quadratic token mixing. We propose HiST, a hierarchical sparse transformer that treats measured locations as a lattice-indexed sparse field and builds a dyadic encoder--decoder directly on the active tissue footprint. HiST combines sparse window attention for local geometric correspondence with resolution-changing operators for rapid multiscale context integration. For a fixed window size, the dominant runtime and memory scale with the number of observed locations rather than the dense slide area. To mitigate slide-specific acquisition variation, HiST adds a bottlenecked global conditioning pathway via a slide calibration token that summarizes slide-level context and conditions local representations. On a multi-organ benchmark spanning diverse tissues and acquisition sources, HiST improves predictive performance over recent baselines while reducing runtime and peak memory.
Weiyi Wu, Xinwen Xu, Xingjian Diao +4
Dartmouth College · Mass General Hospital · New Jersey Institute of Technology +1