a, Abundance of different gates as fraction of total BM.b, Quantification of cell type composition for each FACS gate shown in main Fig.1a.

Source data

For full lists of marker genes, see Supplementary Table1. n=7497 cells are shown.

Source data

For full lists of marker genes, see Supplementary Table1. n=7497 cells are shown.

Source data

a, Enrichment of gene expression signatures of haematopoietic populations in immune cell transcriptomes published by the ImmGen consortium (data source: GEO GSE109125)⁵⁸. Data points in the same row correspond to experimental replicates.b, Enrichment of gene expression signatures of non-haematopoietic populations in published transcriptomes of populations defined by genetic markers^5,7,8,9,59; seemethods for specification of data sources, and see the Supplementary Note1 for a detailed evaluation of the algorithm used.c, The optimal number of mesenchymal cell clusters was determined using the SOUP method¹⁵, a semi-soft clustering algorithm designed to distinguish between distinct cell types and transition states between cell types. Error bars indicate 90% confidence intervals across 10 independent SOUP runs, central values correspond to medians.d, Main cluster identity from SOUP highlighted on the t-SNE from Fig.1b (mesenchymal cell types only, n=2223 cells).e, Comparison of clusters identified by Seurat (Fig.1b) to clusters identified by SOUP (panel d) demonstrates strong overlap between both methods.

Source data

a,b, Additional scRNAseq data was generated as described, except that BM was derived by flushing bones and subjected or not subjected to enzymatic digestion. Data was projected to two dimensions using t-SNE and cell type labels were assigned using the anchoring approach implemented in seurat3⁵³. N=2411 cells are shown.c,d, Single-cell RNA-seq data from a recent study of different genetically labelled populations from flushed BM²⁶ was projected to two dimensions using t-SNE and cell type labels were assigned using the anchoring approach implemented in seurat3⁵³. N=6617 cells are shown.e, Comparison of cell type frequencies between various published datasets. For Tikhonova et al.²⁶, n=6617 cells, Baryawno et al.²⁵, n=25933 cells, Baccin et al., n=7497 cells.

Source data

a, Boxplots comparing the number of genes observed per sample in different protocols. All samples were down-sampled to 1 million reads for comparison. For the dataset presented in main Fig.3, the protocol relying on random priming was used. N represent numbers of samples or cells as indicated on the graph. Boxplot elements are defined in the Methods section (section on data visualization). Smart-Seq2 data is from ref. 21, bulk transcriptome data is from ref. 7.b, Representative images of samples collected for LCM-seq; scale bar: 100µm.c, Immunofluorescence staining of a BM arteriole stained for Col1a1, Pdpn and CD31. Scale bar: 20 µm. The experiment was repeated independently 4 times with similar results.d, Schwann cell markers were lowly expressed across all nichese, haematopoietic markers were highly expressed across all niches. In panel d and e, sample sizes are as follows: Arteriolar niches, n=28, endosteal niches, n=12, sinusoidal niches, n=14, non-vascular niches, n=11, sub-endosteal niche, n=11. Boxplot elements are defined in the Methods section (section on data visualization).

Source data

a, Heatmap of population-specific marker genes used for the algorithm.b,c, Simulations to assess the ability of CIBERSORT to decompose individual samples; see Supplementary Note1 for detail. n=15 sampling runs were performed. Boxplot elements are defined in the Methods section (section on data visualization).d, CIBERSORT estimates of cell type composition of total BM, compared to the cell type composition estimate from 10x genomics (see Extended Data Fig.1b). Central values indicate means, error bars correspond to standard error of the mean of n=3 samples.e,f, Simulations to assess the ability of CIBERSORT to identify changes in population frequencies across multiple samples; see Supplementary Note1 for detail. n=15 samples runs were performed.g, FACS was used to assemble N=8 different pools of B220⁺ B-cells, CD3⁺ T/NK-T cells and Gr1⁺SSC^high neutrophils. Each pool contained a total of 100 cells at predefined ratios of B cells, T cells and neutrophils. Pools were then fixed and processed using the LCM-seq protocol, and CIBERSORT was used to decompose their composition. Estimates for T and NK cells, as well as different B-cell subpopulations, were summed for the display. Pearson correlation coefficients are shown.h, Simulations to assess the ability of CIBERSORT to discriminate between similar cell types; see Supplementary Note1 for detail. Red squares highlight pairs of similar cell types. Pearson correlation coefficients across N=100 simulated samples are shown.i, Stability of the CIBERSORT estimates from main Fig.3e with regard to re-sampling of the marker gene lists used; see Supplementary Note1 for detail.

Source data

a, Whole-mount imaging data of a Cxcl12–GFP bone section stained for Alpl and Emcn was segmented in 3D using the imaris software. Large Alpl⁺ surfaces (red, corresponding to endosteum and arteries) were identified and any GFP+ spots with <20µm proximity to these structures were excluded from further analysis (yellow spots). The strong Alpl signal in these regions prevents a quantitative assessment. Remaining GFP⁺ spots were classified as within 15µm of sinusoidal vessels (purple dots), of away from sinusoidal vessels (cyan dots). GFP⁺ spots were further classified as Alpl⁺ (right panels, red spots) or Alpl^- (right panels, green spots). Scale bar left panel: 200µm; right panels: 50µm.b, Additional whole-mount imaging data of a femur from a Cxcl12–GFP mouse, stained for Alpl and the arteriolar marker Sca1. Arrowheads point to Alpl⁺Cxcl12⁺ cells near, but not overlapping with, Sca1⁺ arteriolar endothelial cells. See also Fig.4f. In ROI 3, asterisk correspond to GFP⁺Alpl⁺ protrusions on, but clearly distinct from, Sca1⁺ arteriolar endothelial cells. In ROI 4, various z-sections of a highly reticulate Cxcl12–GFP⁺Alpl⁺ cell are shown. Scale bars in ROIs: 20µm; scale bar in main panel: 300µm.

a, Overview of the receptor–ligand list used. Seemethods for data sources, and Supplementary Table3 for the complete list.b, Stability of the RNA-Magnet location estimate for different choices of the fuzzification parameters k and x₀. For each parameter set, RNA-Magnet location estimates were summarized per cell type, and compared to the summarized location estimate displayed in Fig.6c. The asterisk indicates the parameter set used in Fig.6c.c, Choice of local neighbourhoods. As detailed in the methods section, RNA-Magnet works by identifying interactions specific to a single cell compared to similar cells. The figure displays the size of local neighbourhoods for four representative cells demarked by a black triangle.d, Detailed comparison of location estimates obtained from LCM-seq and RNA-Magnet. See also main Fig.6c. Sample sizes are as follows: Arteriolar niches, n=28, endosteal niches, n=12, sinusoidal niches. Mean values and standard error of the mean are shown.e, Fully labelled display of the network from main Fig.8a.f, Expression of selected cytokines and growth factors involved in bone remodelling.

a, Sorting scheme used.b, Expression of key marker genes confirm the cell type assignment obtained by scmap, cf. main Fig.7c.

Source data

Supplementary Notes 1 and 2. Supplementary notes on the computational methods used. Supplementary Note 1 describes computational and experimental validations of the CIBERSORT algorithm. Supplementary Note 2 describes the RNA-Magnet algorithm.

Supplemetary Table 1. Marker genes and gene ontology analysis. Supplementary Table 2. Differential expression between niches. Supplementary Table 3. List of receptor–ligand and receptor–receptor pairs used by RNA-Magnet. Supplementary Table 4. Antibodies used in this study. Supplementary Table 5. Antibody combinations used in whole-mount imaging.

Processed genomics data as R and Seurat objects. Single-cell RNA-seq data is stored as a Seurat v3 object and additionally contains all metadata, as well as clustering and dimensionality reduction results. LCM-seq data is stored as a data matrix and metadata is made available as a data frame. See Readme file for detailed description.

Statistical Source Data

Statistical Source Data

Statistical Source Data

Statistical Source Data

Statistical Source Data

Statistical Source Data

Statistical Source Data

Statistical Source Data

Statistical Source Data

Statistical Source Data

Statistical Source Data

Statistical Source Data

Statistical Source Data

Statistical Source Data