Systematic benchmarking of high-throughput subcellular spatial transcriptomics platforms across human tumors

Abstract

Recent advancements in spatial transcriptomics technologies have significantly enhanced resolution and throughput, underscoring an urgent need for systematic benchmarking. Here, we generate serial tissue sections from colon adenocarcinoma, hepatocellular carcinoma, and ovarian cancer samples for systematic evaluation. Using these uniformly processed samples, we generate spatial transcriptomics data across four high-throughput platforms with subcellular resolution: Stereo-seq v1.3, Visium HD FFPE, CosMx 6K, and Xenium 5K. To establish ground truth datasets, we profile proteins on tissue sections adjacent to all platforms using CODEX and perform single-cell RNA sequencing on the same samples. Leveraging manual nuclear segmentation and detailed annotations, we systematically assess each platform's performance across capture sensitivity, specificity, diffusion control, cell segmentation, cell annotation, spatial clustering, and concordance with adjacent CODEX. The uniformly generated and processed multi-omics dataset could advance computational method development and biological discoveries. The dataset is accessible via SPATCH, a user-friendly web server for visualization and download.

Fig. 1. Evaluation of gene detection sensitivity across ST platforms.

a Experimental workflow. For each tumor type (COAD, HCC, OV), samples were divided into three parts: (1) FFPE blocks were used for Visium HD FFPE, CosMx 6K, and Xenium 5K; (2) fresh-frozen OCT-embedded tissue was used for Stereo-seq v1.3; (3) dissociated tissue was subjected to scRNA-seq. Sections adjacent to each ST slide were profiled by 16-plex CODEX for spatial proteomics. b H&E staining and EPCAM expression from ST data, along with PanCK staining from adjacent CODEX sections of COAD samples across the four ST platforms. Color intensity reflects the transcript count per 8 μm bin. Scale bars, 1 mm. c Mean transcript count per 8 μm bin for selected marker genes, computed across all bins with non-zero expression values over the entire tissue sections. d Pearson correlation of gene expression levels between ST data and scRNA-seq data. For each gene, the total transcript counts across three cancer types were averaged and log₁₀ transformed. Each data point represents one gene. The diagonal red line indicates a slope of 1, and color intensity corresponds to relative gene counts. R denotes the correlation coefficient, and n indicates the number of genes included in the analysis. e Log₂-transformed total transcript count per gene across the ten selected regions (400 × 400 μm each) in HCC and OV. Each data point represents one gene (n = 17,134 for Stereo-seq v1.3 and Visium HD FFPE, n = 6175 for CosMx 6K, n = 5001 for Xenium 5K). Center lines indicate the median value, and lower and upper hinges represent the 25th and 75th percentiles, respectively. The whiskers denote 1.5× the interquartile range. f Log₂-transformed gene and transcript counts per 8 μm bin within the ten selected regions in HCC and OV. Each data point represents one bin. Center lines indicate the median value, and lower and upper hinges represent the 25th and 75th percentiles, respectively. The whiskers denote 1.5× the interquartile range. g Mean sequencing saturation across the 10 selected regions for human transcripts detected by Stereo-seq v1.3 and Visium HD FFPE, calculated at stepwise increasing sequencing depths. Panel a created with BioRender.com. Source data are provided as a Source Data file.

Fig. 2. Evaluation of false positives.

a Spatial distribution of gene calls (left) and negative control calls (right) in COAD samples profiled by CosMx 6K and Xenium 5K. Color intensity indicates mean call count per probe in each 8 × 8 μm bin. b Total call counts and Moran’s I for common genes (2552), platform-specific genes (3623 for CosMx 6K and 2449 for Xenium 5K), negative probes (NegProbe, 20 for CosMx 6K and 40 for Xenium 5K), and negative codes (NegCode, 324 for CosMx 6K and 609 for Xenium 5K) detected by CosMx 6K and Xenium 5K across the shared regions shown in (a). Each data point represents one target. Center lines indicate the median value, and lower and upper hinges represent the 25th and 75th percentiles, respectively. The whiskers denote 1.5× the interquartile range. c Histogram showing the log₁₀-transformed signal proportions of individual negative probes and codes, pooled across COAD, HCC, and OV tissues. d H&E staining and transcript distribution inside and outside COAD tissue regions. Red dashed line outlines the Stereo-seq v1.3 region used for diffusion analysis; solid lines mark extra-tissue regions with high transcript levels. Color intensity indicates mean-normalized transcript count of each 8 × 8 μm bin. Scale bars, 1 mm. e Evaluation of transcript diffusion in COAD. The x-axis and y-axis represent the mean-normalized transcript counts and the distance to tissue edge for bins outside the tissue, respectively. Color intensity indicates the number of bins. f Ratio of the mean transcript count in extra-tissue bins to that in intra-tissue bins. Hollow circles indicate the ratio calculated for each of the three cancer types (n = 3). Data are presented as mean values +/− SEM. Source data are provided as a Source Data file.

Fig. 3. Evaluation of transcript-protein correlation.

a Representative immune-enriched regions (each 500 × 500 μm) from COAD samples profiled using all four ST platforms. Left to right: H&E-stained histology, spatial distribution of transcriptomic signature scores for B cells, CD4⁺ T cell, and CD8⁺ T cells derived from ST data, and corresponding CODEX staining images (DAPI, CD20, CD8, CD4). For ST data, color intensity represents the corresponding signature score of each 8 × 8 μm bin. Scale bars, 100 μm. b, c Spatial correlation between CODEX-inferred cell counts and ST-derived signature scores for different cell types over the spatial grids. Panel b shows the correlations for immune and stromal signature scores, while panel c shows the correlations for epithelial signature scores. Pearson correlation coefficients are reported. Hollow circles indicate individual correlation values obtained under different grid sizes (n = 5). Data are presented as mean values +/− SEM. Source data are provided as a Source Data file.

Fig. 4. Comparison of cell segmentation.

a Comparison of platform-derived automatic cell segmentation and manual nuclear segmentation across Stereo-seq v1.3, CosMx 6K, and Xenium 5K HCC sections. For each platform, a 250 × 250 μm region is shown. H&E staining for Stereo-seq v1.3 and multi-channel immunofluorescent staining for CosMx 6K and Xenium 5K are shown. Left column, automatically segmented cell boundaries. Middle column, manually segmented nuclear boundaries. Right column, overlay of automatic and manual segmentations, where white polygons denote automatic segmentations, and blue-filled masks in Stereo-seq v1.3 and yellow polygons in CosMx 6K and Xenium 5K indicate manual segmentations. Scale bars, 50 μm. b Number of automatically segmented cells and manually segmented nuclei per 100 × 100 μm bin across platforms (n = 125 bins per platform per cancer type). Each data point represents one bin. Center lines indicate the median value, and lower and upper hinges represent the 25th and 75th percentiles, respectively. The whiskers denote 1.5× the interquartile range. c Log₂-transformed transcript and gene counts per cell across platforms. For ST platforms, the platform-derived automatic segmentations were used. Left: all detected genes included. Right: only retained genes shared across scRNA-seq, CosMx 6K, and Xenium 5K. Each data point represents one cell. Center lines indicate the median value, and lower and upper hinges represent the 25th and 75th percentiles, respectively. The whiskers denote 1.5× the interquartile range. d Joint density plots showing the expression of exclusive marker pairs within cells in COAD. Only cells with ≥1 transcript of either marker gene were included. Color intensity indicates the density of cells. e Expression correlation of 36 gene pairs expected to be exclusively expressed in distinct major lineages. Pearson correlation was computed across either cells or 8 × 8 μm bins. Each data point represents one marker pair (n = 36). Lower values indicate better separation of marker pairs. Center lines indicate the median value, and lower and upper hinges represent the 25th and 75th percentiles, respectively. The whiskers denote 1.5× the interquartile range. Source data are provided as a Source Data file.

Fig. 5. Comparative analysis of cell clustering, cell type annotation, and spatial alignment with adjacent CODEX.

a Uniform Manifold Approximation and Projection (UMAP) of scRNA-seq and ST data for COAD samples. Each point represents a single cell (for scRNA-seq, Stereo-seq v1.3, CosMx 6K, and Xenium 5K) or an 8 × 8 μm bin (for Visium HD FFPE). Colors denote clusters identified by unsupervised clustering applied independently to each dataset based solely on transcriptomic profiles. b Average silhouette width (ASW) of unsupervised clustering results across platforms, with higher scores indicating better separation between distinct cell states. c Consistency of automated cell type annotations across five reference-based annotation tools. Bars represent the proportion of cells annotated as the same cell type by one to five tools. d Spatial distribution of annotated cell types in ST and CODEX data. Colors denote major cell types. Each ST platform is compared to its adjacent CODEX section. e, f Spatial correlation between CODEX-inferred and ST-inferred cell counts for different cell types over the spatial grids. Panel e shows the correlations for immune and stromal cells, while panel f shows the correlations for epithelial cells. Pearson correlation coefficients are reported. Hollow circles indicate individual correlation values obtained under different grid sizes (n = 5). Data are presented as mean values +/− SEM. g Representative immune-enriched regions (500 × 500 μm) from COAD sections. H&E staining, ST-derived annotations, CODEX-derived annotations, and multiplexed CODEX staining for CD20, CD8, and CD4 are shown. For ST data, each point represents a single cell (for Stereo-seq v1.3, CosMx 6K, and Xenium 5K) or an 8 × 8 μm bin (for Visium HD FFPE), colored by annotated cell type as shown in the legend. Scale bars, 100 μm. Source data are provided as a Source Data file.

Fig. 6. Comparison of spatial clustering and malignant cell distributions.

a Spatial clustering of COAD tissue sections profiled using four ST platforms and adjacent CODEX. Colors represent spatial clusters identified within each dataset. b Pearson correlation between cluster proportions in ST and CODEX data across spatial grids. Hollow circles indicate individual correlation values obtained under different grid sizes (n = 5). Data are presented as mean values +/− SEM. c. H&E staining and spatial localization of malignant cells at tumor boundary and core regions, as defined by unsupervised spatial clustering within each ST platform. Source data are provided as a Source Data file.