CoChat
Sign Up
Cited by 284Single-cell and spatial transcriptomicsCell Image Analysis TechniquesGene expression and cancer classification

The DOI matches its Crossref record; 75 of the 77 references it cites were found in a registry or an index.

BANKSY unifies cell typing and tissue domain segmentation for scalable spatial omics data analysis

Open access
Singhal, Vipul, Chou, Nigel, Lee, Joseph, Yue, Yifei, Liu, Jinyue, Chock, Wan Kee, Lin, Li, Chang, Yun-Ching, Teo, Erica Mei Ling, Aow, Jonathan, Lee, Hwee Kuan, Chen, Kok Hao, Prabhakar, Shyam·2024·Nature Genetics
DOI 10.1038/s41588-024-01664-3PMC PMC10937399

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Spatial omics data are clustered to define both cell types and tissue domains. We present Building Aggregates with a Neighborhood Kernel and Spatial Yardstick (BANKSY), an algorithm that unifies these two spatial clustering problems by embedding cells in a product space of their own and the local neighborhood transcriptome, representing cell state and microenvironment, respectively. BANKSY’s spatial feature augmentation strategy improved performance on both tasks when tested on diverse RNA (imaging, sequencing) and protein (imaging) datasets. BANKSY revealed unexpected niche-dependent cell states in the mouse brain and outperformed competing methods on domain segmentation and cell typing benchmarks. BANKSY can also be used for quality control of spatial transcriptomics data and for spatially aware batch effect correction. Importantly, it is substantially faster and more scalable than existing methods, enabling the processing of millions of cell datasets. In summary, BANKSY provides an accurate, biologically motivated, scalable and versatile framework for analyzing spatially resolved omics data.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41588-024-01664-3

Cites

75 of 75 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 BANKSY’s neighborhood-based feature augmentation strategy for clustering. a, The original gene–cell expression matrix (purple) was augmented with neighborhood-averaged expression matrices corresponding to the mean local expression (dark pink) and the AGF (light pink). Here, λ is a mixing parameter that controls the importance of cells’ own expression and neighborhood expression effects, G(r) is a radially symmetric Gaussian kernel that decays from magnitude 1 at r = 0, ‘expression’ refers to each gene’s expression level in each cell, the mean is taken over cells in the respective index cell’s neighborhood and the eiϕG(r) term confers gradient sensitivity to the AGF. b, Heatmap of the real and imaginary components of a gradient-sensitive AGF kernel. The plots show an unnormalized AGF kernel: the real part (cos(ϕ)) senses the gradient along the x axis and the imaginary part (sin(ϕ)) senses the gradient along the y axis. c, Simplified schematic of two distinct cell types in the neighbor-augmented space. The neighbor expression features, representing the local microenvironment, help to separate two clusters that would be difficult to separate based on the cells’ own expression alone. For simplicity, we show ‘pure’ microenvironments containing only a single cell type (cell type 1 in zone A and cell type 2 in zone B), although BANKSY is equally applicable to heterogeneous microenvironments containing mixtures of cell types (Extended Data Fig. 1).
Fig. 1 BANKSY’s neighborhood-based feature augmentation strategy for clustering. a, The original gene–cell expression matrix (purple) was augmented with neighborhood-averaged expression matrices corresponding to the mean local expression (dark pink) and the AGF (light pink). Here, λ is a mixing parameter that controls the importance of cells’ own expression and neighborhood expression effects, G(r) is a radially symmetric Gaussian kernel that decays from magnitude 1 at r = 0, ‘expression’ refers to each gene’s expression level in each cell, the mean is taken over cells in the respective index cell’s neighborhood and the eiϕG(r) term confers gradient sensitivity to the AGF. b, Heatmap of the real and imaginary components of a gradient-sensitive AGF kernel. The plots show an unnormalized AGF kernel: the real part (cos(ϕ)) senses the gradient along the x axis and the imaginary part (sin(ϕ)) senses the gradient along the y axis. c, Simplified schematic of two distinct cell types in the neighbor-augmented space. The neighbor expression features, representing the local microenvironment, help to separate two clusters that would be difficult to separate based on the cells’ own expression alone. For simplicity, we show ‘pure’ microenvironments containing only a single cell type (cell type 1 in zone A and cell type 2 in zone B), although BANKSY is equally applicable to heterogeneous microenvironments containing mixtures of cell types (Extended Data Fig. 1).CC BY · PubMed Central
Fig. 2 BANKSY enables unsupervised cell-type clustering of mouse cerebellum Slide-seq data. a, Spatial map of one of the four clusters in the Slide-seq data and three of the eight clusters in the Slide-seq v.2 data (for additional clusters, see Supplementary Figs. 2 and 3). Top row, BANKSY clustering. Second row, nonspatial clustering (BANKSY with λ = 0). Third row, RCTD4 weights. The r values indicate the point-biserial correlation of cluster assignments with RCTD weights. Bottom, Averaged expression of top marker genes from the scRNA-seq reference data. b, Spatial map of four cell type clusters in the Slide-seq dataset (granular layer, Purkinje neurons, MLIs and the oligodendrocytes and polydendrocytes cluster). The MLI cluster is absent in nonspatial clustering (λ = 0). c, Top, Computation of normalized spatial cross-connectivity score between two clusters (A and B). Gray lines, spatial proximity relationships connecting cells from the two clusters; red lines, edges between cells from the same cluster. Bottom, NCC score matrix between clusters localized to distinct layers in the Slide-seq dataset. Higher scores indicate greater intermingling between layers.
Fig. 2 BANKSY enables unsupervised cell-type clustering of mouse cerebellum Slide-seq data. a, Spatial map of one of the four clusters in the Slide-seq data and three of the eight clusters in the Slide-seq v.2 data (for additional clusters, see Supplementary Figs. 2 and 3). Top row, BANKSY clustering. Second row, nonspatial clustering (BANKSY with λ = 0). Third row, RCTD4 weights. The r values indicate the point-biserial correlation of cluster assignments with RCTD weights. Bottom, Averaged expression of top marker genes from the scRNA-seq reference data. b, Spatial map of four cell type clusters in the Slide-seq dataset (granular layer, Purkinje neurons, MLIs and the oligodendrocytes and polydendrocytes cluster). The MLI cluster is absent in nonspatial clustering (λ = 0). c, Top, Computation of normalized spatial cross-connectivity score between two clusters (A and B). Gray lines, spatial proximity relationships connecting cells from the two clusters; red lines, edges between cells from the same cluster. Bottom, NCC score matrix between clusters localized to distinct layers in the Slide-seq dataset. Higher scores indicate greater intermingling between layers.CC BY · PubMed Central
Fig. 3 BANKSY enables unsupervised cell-type clustering of mouse hypothalamus MERFISH and human colorectal cancer MERSCOPE data. a, Overview of the mouse hypothalamus dataset2, colored using BANKSY cluster labels (Bregma 0.16 mm and 0.26 mm, as defined in the original study). b, Corresponding UMAP representation of the BANKSY embedding, colored using BANKSY cluster labels. The orange and red clusters (boxed) correspond to the spatially separated mature oligodendrocyte subclusters. c, Spatial maps of all MODs (nonspatial clustering, λ = 0) and two MOD subtypes (BANKSY): two z-slices (Bregma 0.26 mm and 0.16 mm). d, Heatmap of genes differentially expressed between the two MOD subtypes, showing MERFISH expression values z-scaled across the mature oligodendrocytes. e, Average expression of MOD DEGs from d, plotted in the first two components of BANKSY’s neighbor-augmented product space (x axis, own expression; y axis, mean neighborhood expression). f, Clustering of oligodendrocytes in scRNA-seq data from Moffitt et al.2, using the DEGs in d and their top 25 scRNA-seq guilt-by-association (highly correlated) genes. g, UMAP visualization of cell clusters identified in MERSCOPE human colorectal cancer data. Cells are colored using BANKSY and nonspatial clustering (λ = 0) labels. Cluster colors were chosen to indicate correspondence between BANKSY and nonspatial clusters. h, Spatial maps for nonspatial (bottom) versus BANKSY (top) clustering, colored as in g. Red arrows, cycling epithelial cell cluster detected only by BANKSY. i, BANKSY UMAP plot colored according to the expression levels of the cell cycle metagene. j, Violin plot showing expression of the cell cycle metagene in BANKSY (top) and nonspatial (bottom) clusters. Red box, cycling cluster indicated in g–i. CAF, cancer-associated fibroblast.
Fig. 3 BANKSY enables unsupervised cell-type clustering of mouse hypothalamus MERFISH and human colorectal cancer MERSCOPE data. a, Overview of the mouse hypothalamus dataset2, colored using BANKSY cluster labels (Bregma 0.16 mm and 0.26 mm, as defined in the original study). b, Corresponding UMAP representation of the BANKSY embedding, colored using BANKSY cluster labels. The orange and red clusters (boxed) correspond to the spatially separated mature oligodendrocyte subclusters. c, Spatial maps of all MODs (nonspatial clustering, λ = 0) and two MOD subtypes (BANKSY): two z-slices (Bregma 0.26 mm and 0.16 mm). d, Heatmap of genes differentially expressed between the two MOD subtypes, showing MERFISH expression values z-scaled across the mature oligodendrocytes. e, Average expression of MOD DEGs from d, plotted in the first two components of BANKSY’s neighbor-augmented product space (x axis, own expression; y axis, mean neighborhood expression). f, Clustering of oligodendrocytes in scRNA-seq data from Moffitt et al.2, using the DEGs in d and their top 25 scRNA-seq guilt-by-association (highly correlated) genes. g, UMAP visualization of cell clusters identified in MERSCOPE human colorectal cancer data. Cells are colored using BANKSY and nonspatial clustering (λ = 0) labels. Cluster colors were chosen to indicate correspondence between BANKSY and nonspatial clusters. h, Spatial maps for nonspatial (bottom) versus BANKSY (top) clustering, colored as in g. Red arrows, cycling epithelial cell cluster detected only by BANKSY. i, BANKSY UMAP plot colored according to the expression levels of the cell cycle metagene. j, Violin plot showing expression of the cell cycle metagene in BANKSY (top) and nonspatial (bottom) clusters. Red box, cycling cluster indicated in g–i. CAF, cancer-associated fibroblast.CC BY · PubMed Central
Fig. 4 BANKSY accurately identifies anatomically distinct cell subtypes around the mouse hippocampus. a, BANKSY separates CA3 neurons from SSC neurons. b–h, Comparisons with nonspatial clustering (λ = 0) and several spatial clustering methods (b, nonspatial; c, MERINGUE; d, Giotto; e, BayesSpace; f, SpiceMix; g, SpaGCN; h, FICT). When an algorithm merged the CA3 or SSC neurons with those in the CA1 or dentate gyrus, the merged clusters are shown in blue (BayesSpace and FICT), light green and brown (SpiceMix) or light blue (SpaGCN). i, Example DEGs that distinguish CA3 neurons from SSC neurons. Red, high expression; white, medium expression; blue, low expression. j, BANKSY separates the fornix (green) and thalamic (purple) oligodendrocytes. k–q, Comparisons to other methods (as in b–h). r, Example DEGs distinguishing fornix and thalamic oligodendrocytes. s, Schematic showing a coronal section of the hippocampal region shown in a–r. t–w, Corroboration using scRNA-seq data. t, UMAP showing CA3 and cortical neurons from the scRNA-seq data. u, Blue box, expression levels of DEGs determined from comparing the CA3 and SSC BANSKY clusters. Brown box, expression levels of the top two corresponding scRNA-seq guilt-by-association genes. v, UMAP showing fornix and thalamic oligodendrocytes in the scRNA-seq data. w, Blue box, expression levels of DEGs determined from comparing the fornix and thalamic oligodendrocyte BANSKY clusters. Brown box, expression levels of the top two corresponding scRNA-seq guilt-by-association genes.
Fig. 4 BANKSY accurately identifies anatomically distinct cell subtypes around the mouse hippocampus. a, BANKSY separates CA3 neurons from SSC neurons. b–h, Comparisons with nonspatial clustering (λ = 0) and several spatial clustering methods (b, nonspatial; c, MERINGUE; d, Giotto; e, BayesSpace; f, SpiceMix; g, SpaGCN; h, FICT). When an algorithm merged the CA3 or SSC neurons with those in the CA1 or dentate gyrus, the merged clusters are shown in blue (BayesSpace and FICT), light green and brown (SpiceMix) or light blue (SpaGCN). i, Example DEGs that distinguish CA3 neurons from SSC neurons. Red, high expression; white, medium expression; blue, low expression. j, BANKSY separates the fornix (green) and thalamic (purple) oligodendrocytes. k–q, Comparisons to other methods (as in b–h). r, Example DEGs distinguishing fornix and thalamic oligodendrocytes. s, Schematic showing a coronal section of the hippocampal region shown in a–r. t–w, Corroboration using scRNA-seq data. t, UMAP showing CA3 and cortical neurons from the scRNA-seq data. u, Blue box, expression levels of DEGs determined from comparing the CA3 and SSC BANSKY clusters. Brown box, expression levels of the top two corresponding scRNA-seq guilt-by-association genes. v, UMAP showing fornix and thalamic oligodendrocytes in the scRNA-seq data. w, Blue box, expression levels of DEGs determined from comparing the fornix and thalamic oligodendrocyte BANSKY clusters. Brown box, expression levels of the top two corresponding scRNA-seq guilt-by-association genes.CC BY · PubMed Central
Fig. 5 BANKSY accurately segments tissue domains. a, Left, Reference annotations for sample no. 151673 of the Visium human DLPFC dataset. Right, Spatial maps of DLPFC clusters from nonspatial clustering, Giotto’s HMRF, BayesSpace, SpiceMix, SpaGCN, STAGATE, GraphST and BANKSY. Clusters are colored according to their closest matching reference annotation; the numbers in parentheses represent the ARI for each sample computed against the reference annotations. b, Box plot of domain segmentation accuracy on the 12 DLPFC datasets, quantified using the ARI metric. For the deep-learning methods (SpaGCN, STAGATE, GraphST) and BANKSY, each of the 12 data points represents the median across 11 random seeds. Center line, median of the 12 data points; red dotted line, mean of the 12 data points; height of the box, interquartile range (IQR); whiskers, 1.5 × IQR. *P < 0.05, ***P < 0.001, NS P > 0.05 (paired one-sided Wilcoxon signed-rank test). Exact P values from left to right: 0.00024, 0.00024, 0.010, 0.00049, 0.00024, 0.72, 0.48. c, Domain segmentation comparison on the STARmap 1,020-gene mouse visual cortex dataset (ARI in parentheses). L1–6, six neocortical layers. d, BANKSY domain segmentation identified known brain regions in MERFISH mouse hypothalamus data matching annotations from Moffitt et al.2 (Allen Reference Atlas-Mouse Brain52 annotations shown in parentheses when they differ). 3V (V3), third ventricle; BNST (BST), bed nucleus of the stria terminalis; MPA (MPO), medial preoptic area; MPN, medial preoptic nucleus; PaAP (PVH), paraventricular hypothalamic nucleus; PVA (PVT), paraventricular thalamic nucleus.
Fig. 5 BANKSY accurately segments tissue domains. a, Left, Reference annotations for sample no. 151673 of the Visium human DLPFC dataset. Right, Spatial maps of DLPFC clusters from nonspatial clustering, Giotto’s HMRF, BayesSpace, SpiceMix, SpaGCN, STAGATE, GraphST and BANKSY. Clusters are colored according to their closest matching reference annotation; the numbers in parentheses represent the ARI for each sample computed against the reference annotations. b, Box plot of domain segmentation accuracy on the 12 DLPFC datasets, quantified using the ARI metric. For the deep-learning methods (SpaGCN, STAGATE, GraphST) and BANKSY, each of the 12 data points represents the median across 11 random seeds. Center line, median of the 12 data points; red dotted line, mean of the 12 data points; height of the box, interquartile range (IQR); whiskers, 1.5 × IQR. *P < 0.05, ***P < 0.001, NS P > 0.05 (paired one-sided Wilcoxon signed-rank test). Exact P values from left to right: 0.00024, 0.00024, 0.010, 0.00049, 0.00024, 0.72, 0.48. c, Domain segmentation comparison on the STARmap 1,020-gene mouse visual cortex dataset (ARI in parentheses). L1–6, six neocortical layers. d, BANKSY domain segmentation identified known brain regions in MERFISH mouse hypothalamus data matching annotations from Moffitt et al.2 (Allen Reference Atlas-Mouse Brain52 annotations shown in parentheses when they differ). 3V (V3), third ventricle; BNST (BST), bed nucleus of the stria terminalis; MPA (MPO), medial preoptic area; MPN, medial preoptic nucleus; PaAP (PVH), paraventricular hypothalamic nucleus; PVA (PVT), paraventricular thalamic nucleus.CC BY · PubMed Central
Fig. 6 BANKSY is scalable to large datasets and faster than existing spatial methods. Runtimes of BayesSpace, FICT, Giotto’s HMRF module, GraphST, MERINGUE’s spatial clustering module, SpaGCN, SpiceMix, STAGATE, nonspatial clustering (Seurat) and BANKSY for increasing cell numbers, up to 2 million cells. All methods were benchmarked on a 16-CPU 128-GB machine. Runtimes are shown up to the maximum cell number accommodated by each method, with a cutoff of 16 h.
Fig. 6 BANKSY is scalable to large datasets and faster than existing spatial methods. Runtimes of BayesSpace, FICT, Giotto’s HMRF module, GraphST, MERINGUE’s spatial clustering module, SpaGCN, SpiceMix, STAGATE, nonspatial clustering (Seurat) and BANKSY for increasing cell numbers, up to 2 million cells. All methods were benchmarked on a 16-CPU 128-GB machine. Runtimes are shown up to the maximum cell number accommodated by each method, with a cutoff of 16 h.CC BY · PubMed Central
Extended Data Fig. 1 Detailed schematic to aid conceptual understanding of BANKSY. How BANKSY may lead to more accurate cluster assignments relative to non-spatial clustering (λ=0) by accounting for local microenvironmental information. (a) Cells near the decision boundary in own expression space alone get better separated in the neighbor-augmented space. We have used only the mean neighborhood expression for illustration purposes, but the same arguments apply to the AGF feature space. (b) Top – Schematic of tissue slice shows physical locations of two cell types (1 and 2) in two zones (A and B). Zone A contains mainly cell type 1 but some of cell type 2. Zone B contains mainly cell type 2 but some of cell type A. Bottom – Effect of increasing λ: at λ =0 (bottom left), only the cell’s transcriptome is used, making it identical to conventional unsupervised clustering analysis. As λ increases (bottom middle), subsets of cells in different environments with subtly different transcriptomes become easier to tease out. For instance, the blue circles – filled and empty – correspond to the same cell type in two different neighborhoods, possessing subtle transcriptomic differences. A similar effect is also shown for the green cell type. At higher lambdas (bottom right), a zone or spatial-domain segmentation effect occurs where zones representing different microenvironments are clustered separately and can comprise multiple cell types per zone.
Extended Data Fig. 1 Detailed schematic to aid conceptual understanding of BANKSY. How BANKSY may lead to more accurate cluster assignments relative to non-spatial clustering (λ=0) by accounting for local microenvironmental information. (a) Cells near the decision boundary in own expression space alone get better separated in the neighbor-augmented space. We have used only the mean neighborhood expression for illustration purposes, but the same arguments apply to the AGF feature space. (b) Top – Schematic of tissue slice shows physical locations of two cell types (1 and 2) in two zones (A and B). Zone A contains mainly cell type 1 but some of cell type 2. Zone B contains mainly cell type 2 but some of cell type A. Bottom – Effect of increasing λ: at λ =0 (bottom left), only the cell’s transcriptome is used, making it identical to conventional unsupervised clustering analysis. As λ increases (bottom middle), subsets of cells in different environments with subtly different transcriptomes become easier to tease out. For instance, the blue circles – filled and empty – correspond to the same cell type in two different neighborhoods, possessing subtle transcriptomic differences. A similar effect is also shown for the green cell type. At higher lambdas (bottom right), a zone or spatial-domain segmentation effect occurs where zones representing different microenvironments are clustered separately and can comprise multiple cell types per zone.CC BY · PubMed Central
Extended Data Fig. 2 BANKSY cell typing on a CosMX colon tissue dataset. a) Non-spatial and BANKSY UMAPs colored by their respective clusters, with cell type labels for BANKSY clusters. b) Spatial plots for the full dataset and a single ROI colored by BANKSY clusters. c) Spatial distribution of macrophages, with the Mac.2 cluster enriched in the submucosa. Violin plots show DE genes between the two macrophage subpopulations. Mac.1 n = 1177, Mac.2 n = 214.
Extended Data Fig. 2 BANKSY cell typing on a CosMX colon tissue dataset. a) Non-spatial and BANKSY UMAPs colored by their respective clusters, with cell type labels for BANKSY clusters. b) Spatial plots for the full dataset and a single ROI colored by BANKSY clusters. c) Spatial distribution of macrophages, with the Mac.2 cluster enriched in the submucosa. Violin plots show DE genes between the two macrophage subpopulations. Mac.1 n = 1177, Mac.2 n = 214.CC BY · PubMed Central
Extended Data Fig. 3 Benchmarking of domain segmentation (a-c) and cell type clustering (d-f) on data simulated from the STARMAP 1020 gene mouse visual cortex dataset. Data were simulated using scDesign2, with gene expression characteristics modeled after the 16 cell types in the STARmap data. Cell type compositions across the 7 domains (vertical strips) were modeled after cell type compositions in each domain in the STARmap data. (a) ARIs (median of 3 pseudorandom number generator seeds) of each method for domain segmentation as a function of gene number. The methods were used to generate domain labels by setting the number of requested clusters to 7. (b) Ground truth annotations for domain segmentation. (c) Qualitative comparisons across methods for a representative 1020 gene dataset. (d) ARIs (median of 3 seeds) of each method for cell type clustering as a function of gene number. The methods were used to generate cell type labels by setting the number of requested clusters to 16. (e) Ground truth annotations for cell type clustering. (f) Qualitative comparisons across methods for a representative 1020 gene dataset.
Extended Data Fig. 3 Benchmarking of domain segmentation (a-c) and cell type clustering (d-f) on data simulated from the STARMAP 1020 gene mouse visual cortex dataset. Data were simulated using scDesign2, with gene expression characteristics modeled after the 16 cell types in the STARmap data. Cell type compositions across the 7 domains (vertical strips) were modeled after cell type compositions in each domain in the STARmap data. (a) ARIs (median of 3 pseudorandom number generator seeds) of each method for domain segmentation as a function of gene number. The methods were used to generate domain labels by setting the number of requested clusters to 7. (b) Ground truth annotations for domain segmentation. (c) Qualitative comparisons across methods for a representative 1020 gene dataset. (d) ARIs (median of 3 seeds) of each method for cell type clustering as a function of gene number. The methods were used to generate cell type labels by setting the number of requested clusters to 16. (e) Ground truth annotations for cell type clustering. (f) Qualitative comparisons across methods for a representative 1020 gene dataset.CC BY · PubMed Central
Extended Data Fig. 4 Comparing BANKSY and non-spatial clustering results on the VeraFISH data to MERINGUE, Giotto (HMRF) and BayesSpace (HMRF). The number of clusters was varied by adjusting parameters related to clustering resolution (number of nearest neighbors, k, in the graph for MERINGUE, and the number of clusters, N, for SpaGCN and FICT). Each column represents one clustering run for the indicated method at a given effective resolution. Blue arrows indicate the runs compared in Fig. 4. Green and red boxes indicate success and failure to identify the known neuronal or oligodendrocyte subclusters (Fig. 4; Supplementary Fig. 18). Conventional non-spatial clustering and MERINGUE (at any number of clusters) failed to separate both the oligodendrocyte subclusters and the neuronal subclusters. While Giotto separated the CA3 neurons (orange) from the somatosensory cortex neurons (brown), and at some resolutions was able to identify the fornix oligodendrocytes (dark green), it failed to identify the thalamic oligodendrocytes (purple) at all resolutions. Importantly, Giotto tended to cause clusters to form spatially disconnected groups of cells (‘patches’) that do not correspond to any known anatomical structures in the thalamus or hippocampal formation (dashed red boxes). BayesSpace merged several neuronal clusters: at N = 16 - 18, it merged the somatosensory cortex neurons with the CA3, CA1 and dentate gyrus neurons (blue clusters; all anatomical regions labeled in Fig. 4 in the main text), and at N = 19 - 20, it separated out the CA3 neurons, from the somatosensory cortex neurons, but merged the CA1 and dentate gyrus neurons. It also failed to separate out the thalamic oligodendrocytes from the fornix oligodendrocytes (red boxes around the dark green cluster).
Extended Data Fig. 4 Comparing BANKSY and non-spatial clustering results on the VeraFISH data to MERINGUE, Giotto (HMRF) and BayesSpace (HMRF). The number of clusters was varied by adjusting parameters related to clustering resolution (number of nearest neighbors, k, in the graph for MERINGUE, and the number of clusters, N, for SpaGCN and FICT). Each column represents one clustering run for the indicated method at a given effective resolution. Blue arrows indicate the runs compared in Fig. 4. Green and red boxes indicate success and failure to identify the known neuronal or oligodendrocyte subclusters (Fig. 4; Supplementary Fig. 18). Conventional non-spatial clustering and MERINGUE (at any number of clusters) failed to separate both the oligodendrocyte subclusters and the neuronal subclusters. While Giotto separated the CA3 neurons (orange) from the somatosensory cortex neurons (brown), and at some resolutions was able to identify the fornix oligodendrocytes (dark green), it failed to identify the thalamic oligodendrocytes (purple) at all resolutions. Importantly, Giotto tended to cause clusters to form spatially disconnected groups of cells (‘patches’) that do not correspond to any known anatomical structures in the thalamus or hippocampal formation (dashed red boxes). BayesSpace merged several neuronal clusters: at N = 16 - 18, it merged the somatosensory cortex neurons with the CA3, CA1 and dentate gyrus neurons (blue clusters; all anatomical regions labeled in Fig. 4 in the main text), and at N = 19 - 20, it separated out the CA3 neurons, from the somatosensory cortex neurons, but merged the CA1 and dentate gyrus neurons. It also failed to separate out the thalamic oligodendrocytes from the fornix oligodendrocytes (red boxes around the dark green cluster).CC BY · PubMed Central
Extended Data Fig. 5 Comparing BANKSY and non-spatial clustering results on the VeraFISH data to SpiceMix (HMRF), SpaGCN and FICT (HMRF). Similar to Extended Data Fig. 4. As before, the number of clusters was varied by adjusting parameters related to clustering resolution. Each column represents one clustering run for the indicated method at a given resolution (controlled by the number of clusters input parameter, N). Blue arrows indicate the runs compared in Fig. 4. Green and red boxes indicate success and failure (respectively) to identify the known neuronal or oligodendrocyte subclusters (Fig. 4; Supplementary Fig. 18). SpiceMix correctly separated the thalamic and fornix oligodendrocytes, merged the somatosensory cortex neurons with either the CA3 neuron (orange clusters) or the dentate gyrus neurons (brown and light green clusters). SpaGCN was able to identify the CA3 neurons, but merged the somatosensory cortex neurons with cells in the dentate gyrus (brown cluster), and was unable to separate the thalamic and fornix oligodendrocytes (red boxed green clusters). FICT, like Giotto in Extended Data Fig. 4, tended to group cells into disconnected spatial patches that do not conform to known anatomical structures (dashed red boxes), and tended to merge the dentate gyrus, CA1, CA3 and somatosensory cortex neurons into one or two clusters (light and dark blue clusters).
Extended Data Fig. 5 Comparing BANKSY and non-spatial clustering results on the VeraFISH data to SpiceMix (HMRF), SpaGCN and FICT (HMRF). Similar to Extended Data Fig. 4. As before, the number of clusters was varied by adjusting parameters related to clustering resolution. Each column represents one clustering run for the indicated method at a given resolution (controlled by the number of clusters input parameter, N). Blue arrows indicate the runs compared in Fig. 4. Green and red boxes indicate success and failure (respectively) to identify the known neuronal or oligodendrocyte subclusters (Fig. 4; Supplementary Fig. 18). SpiceMix correctly separated the thalamic and fornix oligodendrocytes, merged the somatosensory cortex neurons with either the CA3 neuron (orange clusters) or the dentate gyrus neurons (brown and light green clusters). SpaGCN was able to identify the CA3 neurons, but merged the somatosensory cortex neurons with cells in the dentate gyrus (brown cluster), and was unable to separate the thalamic and fornix oligodendrocytes (red boxed green clusters). FICT, like Giotto in Extended Data Fig. 4, tended to group cells into disconnected spatial patches that do not conform to known anatomical structures (dashed red boxes), and tended to merge the dentate gyrus, CA1, CA3 and somatosensory cortex neurons into one or two clusters (light and dark blue clusters).CC BY · PubMed Central
Extended Data Fig. 6 Clustering performance of each method on all 12 DLPFC datasets, quantified using 3 metrics. (a) Adjusted Rand index (ARI); (b) normalized mutual information (NMI); and (c) Matthews correlation coefficient (MCC). Significance is computed using paired one-sided Wilcoxon signed rank test (*: p≤0.05, **: p≤0.01, ***: p≤0.001). Exact p-values from left to right for (a) 0.00024, 0.00024, 0.010, 0.00049, 0.00024, 0.72, 0.48, (b) 0.00024, 0.00024, 0.0061, 0.00073, 0.00024, 0.81, 0.52, (c) 0.00024, 0.00024, 0.0034, 0.0061, 0.00049, 0.79, 0.57. Center line: median, dotted line: mean, height of box: interquartile range (IQR), whiskers: 1.5 x IQR. In all boxes, n = 12 samples.
Extended Data Fig. 6 Clustering performance of each method on all 12 DLPFC datasets, quantified using 3 metrics. (a) Adjusted Rand index (ARI); (b) normalized mutual information (NMI); and (c) Matthews correlation coefficient (MCC). Significance is computed using paired one-sided Wilcoxon signed rank test (*: p≤0.05, **: p≤0.01, ***: p≤0.001). Exact p-values from left to right for (a) 0.00024, 0.00024, 0.010, 0.00049, 0.00024, 0.72, 0.48, (b) 0.00024, 0.00024, 0.0061, 0.00073, 0.00024, 0.81, 0.52, (c) 0.00024, 0.00024, 0.0034, 0.0061, 0.00049, 0.79, 0.57. Center line: median, dotted line: mean, height of box: interquartile range (IQR), whiskers: 1.5 x IQR. In all boxes, n = 12 samples.CC BY · PubMed Central
Extended Data Fig. 7 STARmap mouse visual cortex 1020 gene dataset. (a) Top: cell-type annotations; Bottom: implied domain annotations. (b) Domain segmentation comparison (boxplot) on the STARmap 1,020-gene mouse visual cortex dataset, showing performance variability across random seeds. GraphST, STAGATE, SpaGCN and BANKSY were run with n = 11 different random seeds and ARI relative to the reference domain annotation was calculated in each case. (c) STARmap benchmark: alluvial plot showing correspondence between cluster labels assigned by spatial algorithms (left) and manual domain annotations (right). Correspondence is quantified using ARI, MCC and NMI.
Extended Data Fig. 7 STARmap mouse visual cortex 1020 gene dataset. (a) Top: cell-type annotations; Bottom: implied domain annotations. (b) Domain segmentation comparison (boxplot) on the STARmap 1,020-gene mouse visual cortex dataset, showing performance variability across random seeds. GraphST, STAGATE, SpaGCN and BANKSY were run with n = 11 different random seeds and ARI relative to the reference domain annotation was calculated in each case. (c) STARmap benchmark: alluvial plot showing correspondence between cluster labels assigned by spatial algorithms (left) and manual domain annotations (right). Correspondence is quantified using ARI, MCC and NMI.CC BY · PubMed Central
Extended Data Fig. 8 BANKSY tissue segmentation and community identification on CODEX multiplexed imaging data of the intestine. a). BANKSY UMAP and spatial plots colored by tissue segment annotations. b). BANKSY UMAP and spatial plots colored by BANKSY clusters. c). Non-spatial UMAP and spatial plots colored by non-spatial clusters. d). Comparison of BANKSY and non-spatial clusters with community domain annotations. e). Marker genes for smooth muscle and mature epithelial communities identified by BANKSY. Boxplots are colored as in (d). Center line: median, dotted line: mean, height of box: interquartile range (IQR), whiskers: 1.5 x IQR. CD66 Annotation: n = 3179 (mean = 1.47, median = 0.636) vs 6189 (mean = 0.107, median = -0.167) cells, BANKSY: n = 3435 (mean = 1.70, median = 0.935) vs 5232 (mean = -0.184, median = -0.249), aSMA BANKSY: n = 3222 (mean = 1.26, median = 1.04) vs 4826 (mean = 0.245, median = 0.138). All p-values < 2.2e-16.
Extended Data Fig. 8 BANKSY tissue segmentation and community identification on CODEX multiplexed imaging data of the intestine. a). BANKSY UMAP and spatial plots colored by tissue segment annotations. b). BANKSY UMAP and spatial plots colored by BANKSY clusters. c). Non-spatial UMAP and spatial plots colored by non-spatial clusters. d). Comparison of BANKSY and non-spatial clusters with community domain annotations. e). Marker genes for smooth muscle and mature epithelial communities identified by BANKSY. Boxplots are colored as in (d). Center line: median, dotted line: mean, height of box: interquartile range (IQR), whiskers: 1.5 x IQR. CD66 Annotation: n = 3179 (mean = 1.47, median = 0.636) vs 6189 (mean = 0.107, median = -0.167) cells, BANKSY: n = 3435 (mean = 1.70, median = 0.935) vs 5232 (mean = -0.184, median = -0.249), aSMA BANKSY: n = 3222 (mean = 1.26, median = 1.04) vs 4826 (mean = 0.245, median = 0.138). All p-values < 2.2e-16.CC BY · PubMed Central
Extended Data Fig. 9 Benchmarking different integration methodologies. Performance of non-spatial clustering, BayesSpace and BANKSY with and without batch correction on joint clustering of spots from n = 12 DLPFC Visium datasets. ARI: adjusted Rand index. NMI: normalized mutual information. MCC: Matthew’s correlation coefficient. Center line: median, red dotted line: mean, height of box: interquartile range (IQR), whiskers: 1.5 x IQR.
Extended Data Fig. 9 Benchmarking different integration methodologies. Performance of non-spatial clustering, BayesSpace and BANKSY with and without batch correction on joint clustering of spots from n = 12 DLPFC Visium datasets. ARI: adjusted Rand index. NMI: normalized mutual information. MCC: Matthew’s correlation coefficient. Center line: median, red dotted line: mean, height of box: interquartile range (IQR), whiskers: 1.5 x IQR.CC BY · PubMed Central
Extended Data Fig. 10 GPU runtimes of GraphST, SpaGCN, SpiceMix and STAGATE. All methods were benchmarked using a NVIDIA T4 GPU with 16GB of memory. Runtimes are shown up to the maximum cell number accommodated by each method.
Extended Data Fig. 10 GPU runtimes of GraphST, SpaGCN, SpiceMix and STAGATE. All methods were benchmarked using a NVIDIA T4 GPU with 16GB of memory. Runtimes are shown up to the maximum cell number accommodated by each method.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 77 references

Research works printed but not linked: 3

  1. Huuki-Myers, L. et al. Integrated single cell and unsupervised spatial transcriptomic analysis defines molecular anatomy of the human dorsolateral prefrontal cortex. Preprint at bioRxiv https://doi.org/10.1101/2023.02.15.528722 (2023). · DOI
  2. Gabor, D. Theory of communication. J. Inst. Electrical Engineering 93, 429–441 (1946).
  3. Csardi, G. & Nepusz, T. The igraph software package for complex network research. Int. J. Complex Syst. 1695, 1–9 (2006).

Cited by

Sorted by citations

See all 284

Related

Metadata

Authors
Singhal, V., Chou, N., Lee, J., Yue, Y., Liu, J., Chock, W. K., Lin, L., Chang, Y.-C., Teo, E. M. L., Aow, J., Lee, H. K., Chen, K. H., & Prabhakar, S.
Affiliations
  • Agency for Science, Technology and Research (SG)
  • National University of Singapore (SG)
Corresponding author
  • Chen, Kok Hao
  • Prabhakar, Shyam
ORCID
Venue
Nature Genetics
Publisher
Springer Science and Business Media LLC
Published
2024-02-27
Volume
56
Issue
3
Pages
431-441
ISSN
  • 1061-4036 (ISSN-L)
  • 1546-1718
DOI
10.1038/s41588-024-01664-3
PMC
PMC10937399
OpenAlex
W4392198237
Cited by
282
References
77
Citations by year
2026: 124 · 2025: 103 · 2024: 47
FWCI
37.44
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Genetics (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DR-NTU (Nanyang Technological University) (submittedVersion, cc-by)
Funders
  • MOH | National Medical Research Council (OFIRG21jun-0090, OF-YIRG18nov-0014, OFIRG-000618-00)
  • Agency for Science, Technology and Research (#H18/01/a0/020, 202D800010, H18/01/a0/020, I1801E0029)
  • National Research Foundation Singapore (NRF-CRP25-2020-0001)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Biophysics › Cell Image Analysis Techniques
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gene expression and cancer classification
Keywords
spatial omics, cell typing, spatial clustering, mouse brain, batch effect correction, scalable data analysis
MeSH terms
Data Analysis, Algorithms, Animals, RNA, Benchmarking, Gene Expression Profiling, Mice, Transcriptome

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review