SuperSpot: coarse graining spatial transcriptomics data into metaspots

Abstract

Summary: Spatial Transcriptomics is revolutionizing our ability to phenotypically characterize complex biological tissues and decipher cellular niches. With current technologies such as VisiumHD, thousands of genes can be detected across millions of spots (also called cells or bins depending on the technologies). Building upon the metacell concept, we present a workflow, called SuperSpot, to combine adjacent and transcriptomically similar spots into "metaspots". The process involves representing spots as nodes in a graph with edges connecting spots in spatial proximity and edge weights representing transcriptomic similarity. Hierarchical clustering is used to aggregate spots into metaspots at a user-defined resolution. We demonstrate that metaspots reduce the size and sparsity of spatial transcriptomic data and facilitate the analysis of large datasets generated with the most recent technologies.

Availability and implementation: SuperSpot is an R package available at https://github.com/GfellerLab/SuperSpot and archived on Zenodo (https://doi.org/10.5281/zenodo.14222088). The code to reproduce the figures is available at https://github.com/GfellerLab/SuperSpot/tree/main/figures (https://doi.org/10.5281/zenodo.14222088).

Figure 1.

SuperSpot overview and applications. (A) Illustration of the SuperSpot pipeline to build metaspots in spatial transcriptomic data. (B) Splitting process of metaspots to guarantee both purity according to some annotation and spatial coherence. (C) 10x Visium mouse cortex dataset at the spot and metaspot (γ = 4) level. The spots and the metaspots are colored based on the layers of the cortex and white matter. The “Unknown” label corresponds to unannotated spots. (D) ARI scores of different spatial clustering methods with respect to the brain layer annotation at spot (red) and metaspot (blue) levels. The ARI score is computed as the mean over 10 runs with different seeds. (E) ARI scores between the clusters computed at the spot and metaspot levels by each clustering method (blue dots) and the clusters computed by different clustering methods at spot level (red dots). The ARI score is computed as the mean over 10 runs with different seeds. (F) Right slide of Nanos-tring CosMx human pancreas dataset at the metaspot (γ = 3.07) level. The spots and the metaspots are colored based on the cell types. (G) ARI scores of CellCharter with respect to the cell type annotation at spot (red) and metaspot (blue) levels on Nanostring CosMx human pancreas dataset. The ARI score is computed as the mean over 10 runs with different seeds. (H) VisiumHD Human Colorectal Cancer dataset. Colors show clusters obtained in transcriptomic analysis of metaspots and projected on the spatial coordinates of metaspots (γ = 64). Frames corresponds to the regions of interest for panel I. (I) Illustrations of clusters computed in 16 μm bins and in metaspots (γ = 64), compared to the H&E staining image. (J) Peak memory (GB) and elapsed time needed for computing normalization and spatially variable features in metaspots as a function of γ for the Visium Mouse Brain (3639 spots/33 538 genes, left), the Nanostring CosMx Human Pancreas (48 944 segmented cells/18 946 genes, center) and a 9·104x zoom of the VisiumHD Human Colorectal Cancer (H. CRC) (28 358 bins out of 8 731 400 2 μm bins/18 085 genes, right). Blue dots represent the final γ of 3.07 for the CosMx Human Pancreas after splitting impure metaspots. (K) Boxplot of the number of detected genes per spot (red) and metaspot (blue) for Visium Mouse Brain (γ = 4, left), Nanostring CosMx Human Pancreas (γ = 3.07, center), and VisiumHD Human Colorectal Cancer (γ = 64, right). Percentages above each boxplot correspond to the number of nonzero entries within the gene expressions matrix.