Spatial transcriptomics at subspot resolution with BayesSpace

Abstract

Recent spatial gene expression technologies enable comprehensive measurement of transcriptomic profiles while retaining spatial context. However, existing analysis methods do not address the limited resolution of the technology or use the spatial information efficiently. Here, we introduce BayesSpace, a fully Bayesian statistical method that uses the information from spatial neighborhoods for resolution enhancement of spatial transcriptomic data and for clustering analysis. We benchmark BayesSpace against current methods for spatial and non-spatial clustering and show that it improves identification of distinct intra-tissue transcriptional profiles from samples of the brain, melanoma, invasive ductal carcinoma and ovarian adenocarcinoma. Using immunohistochemistry and an in silico dataset constructed from scRNA-seq data, we show that BayesSpace resolves tissue structure that is not detectable at the original resolution and identifies transcriptional heterogeneity inaccessible to histological analysis. Our results illustrate BayesSpace's utility in facilitating the discovery of biological insights from spatial transcriptomic datasets.

Fig. 1. The BayesSpace workflow.

a, The BayesSpace workflow begins with preprocessed ST or Visium data. Data are spatially clustered to infer regions with similar expression profiles. These clusters can be refined via enhanced clustering to provide a higher-resolution spatial map. Enhanced clustering also provides the basis for predicting gene expression at the higher resolution, which can be used in further differential expression analyses. b, From geometric representations of spatial distribution of spots in the ST and Visium technologies, neighbors can be identified for each spot based on shared edges (top). Each spot can be subdivided into subspots, which again have natural edge-based neighbors (bottom).

Fig. 2. BayesSpace improves computational resolution of layers in the DLPFC.

a, Ground truth. We highlight the manually annotated six DLPFC layers and white matter (WM) in sample 151673 from the spatialLIBD dataset. Annotated layers for the remaining samples can be found in the original publication. b, Summary of clustering accuracy in all twelve samples. The ARI is used to compare similarity between cluster labels from each method against the manually annotated layers for all twelve samples. In the boxplot, the center line, box limits and whiskers denote the median, upper and lower quartiles and 1.5× interquartile range, respectively. c, Cluster assignments generated by non-spatial (top) and spatial (bottom) methods for sample 151673.

Fig. 3. Enhanced-resolution clustering identifies tumor-proximal lymphoid tissue in a melanoma sample.

a, The original histopathological annotations of H&E-stained tissue (N = 1 tissue section, n = 293 spots) revealed a section of melanoma (black) adjacent to tumor-proximal lymphoid tissue (yellow) and a region of stroma (red), separating these from a larger section of tumor-distal lymphoid tissue (yellow). Adapted from ref. with permission from the American Association for Cancer Research. Spatial clustering (b) and enhancement (c) generate biologically meaningful spatial domains corresponding to the original annotations. Enhanced-resolution clustering identified tumor-proximal lymphoid tissue (cluster 4, yellow), which was not resolved at spot-level clustering. d, Differential expression analysis between the four clusters highlighted spatial differences in the expression of immune genes, cancer markers and genes encoding extracellular matrix proteins. e, For each of the five major cell types, the observed total spot-level expression (as measured by the summed log-normalized counts) of the defined marker genes (left) is shown alongside the corresponding enhanced-resolution expression (right). We show spatial expression plots for tumor cells (PMEL), fibroblasts (COL1A1), macrophages (CD14, FCGR1A, FCGR1B), B cells (CD19, MS4A1) and T cells (CD2, CD3D, CD3E, CD3G, CD7).

Fig. 4. Immunohistochemistry validates BayesSpace enhancement in an IDC sample and an OC sample.

a, Average intensity of the anti-CD3 immunofluorescent stain in the IDC. Intensity was scaled to the range (0, 1) for visualization. b, Log-normalized gene expression of CD3E measured on the Visium platform (left, ‘spot’) and enhanced with BayesSpace (right, ‘subspot’). c, Dichotomized clustering of Visium gene expression values. After clustering the tissue section into ten clusters, the clusters were binned by their median anti-CD3 stain intensity into CD3 ‘high’ and CD3 ‘low’ clusters, shown here. White squares outline three ROI where the enhanced clustering revealed areas of increased heterogeneity. d, Zoomed-in views of the n = 3 ROI. Each panel shows a 1-mm² area of the immunofluorescence image. DAPI intensity is shown in blue, and anti-CD3 intensity is shown in green. Overlaid on each panel in the top row is the spot-level clustering. Each circle corresponds to the position and size (55-μm diameter) of a spot on the Visium array and is colored based on whether it belongs to a CD3 ‘high’ (yellow) or CD3 ‘low’ (blue) cluster. The bottom row contains a similar overlay of the enhanced-resolution subspot clustering, where the circles are now subdivided into six wedges corresponding to the positions of subspots in the BayesSpace model. As in the spot overlay, the subspots are colored based on their cluster membership. e, Summary of subspot reassignment after enhancement. On the left, we show a contingency table describing the number of subspots (n = 17,574) that belong to a CD3 ‘high’ or ‘low’ cluster at the spot level and at the subspot level. Using two-sided Wilcoxon rank-sum tests, we also show that anti-CD3 intensity in subspots that are reassigned to a ‘high’ cluster is significantly higher (P < 2.22 × 10⁻¹⁶) thanthat inthose that remain in a ‘low’ cluster (center) and that subspots that are reassigned to a ‘low’ cluster have a significantly lower (P < 2.22 × 10⁻¹⁶) anti-CD3 intensity than that in those that remain in a ‘high’ cluster (right). f–j, Panels for the OC mirror those for the IDC, with anti-CD45 intensity replacing anti-CD3 intensity and PTPRC (CD45) gene expression replacing that of CD3E. In e, we show n = 12,246 subspots. In i, we show n = 3 ROI. In j, using two-sided Wilcoxon rank-sum tests, we show that anti-CD45 intensity in subspots that are reassigned to a ‘high’ cluster is significantly higher (P < 2.22 × 10⁻¹⁶) than that in those that remain in a ‘low’ cluster (center) and that subspots that are reassigned to a ‘low’ cluster’ have a significantly lower (P = 2.9 × 10⁻¹¹) anti-CD45 intensity than that in those that remain in a ‘high’ cluster (right). All reported P values are unadjusted values.

Fig. 5. BayesSpace identifies transcriptional heterogeneity within an IDC.

a, Immunofluorescent imaging of the tissue section (N = 1 tissue section, n = 4,727 spots) and histopathological annotations. DAPI intensity is shown in blue, anti-CD3 intensity is shown in green, and the Visium fiducial frame is shown in red. Annotated regions of IC are outlined in red, those of carcinoma in situ are outlined in yellow, those of benign hyperplasia are outlined in green, and those of unclassified tumor are outlined in gray. b, Enhanced BayesSpace clustering. c, Spatial expression of genes coding for HER2 (ERBB2) and ER (ESR1) and PR (PGR). d, Spatial expression of immune genes PTPRC (CD45), CD4, CD8A, CD14, CD68 and IGHG3. e, Spatial expression of proliferation marker MKI67 (Ki-67), markers of tumor progression MUC1 and COL1A2, the oncogene ZNF703, GRB2 (coding for the growth factor receptor protein) and BAMBI (coding for transforming growth factor (TGF)-β pseudoreceptor).

Fig. 6. BayesSpace outperforms spatial and non-spatial clustering methods with simulated data.

a, In N = 8 replicates simulated from the melanoma sample and N = 8 replicates simulated from the OC sample, BayesSpace spot-level clustering outperforms other clustering methods. b, In N = 20 replicates for the simulation performed at the subspot level, BayesSpace enhanced clustering outperforms the optimal spot-level clustering (red dotted line). c, In the third simulation using single-cell data, the ground truth is derived from expert annotation of an immunofluorescence staining image corresponding to the OC sample (top left). Examples of clustering partitions generated by BayesSpace at the spot and subspot levels as well as by the next best method (Giotto) are also shown. d, BayesSpace clustering at the spot level slightly outperforms competing methods, while BayesSpace enhancement to the subspot level generally provides substantially higher performance than that of other methods in recapturing ground-truth clusters among the N = 8 simulation replicates. In all boxplots, the center line, box limits and whiskers denote the median, upper and lower quartiles and 1.5× interquartile range, respectively.