Embryo-scale, single cell spatial transcriptomics

Abstract

Spatial patterns of gene expression manifest at scales ranging from local (e.g. cell-cell interactions) to global (e.g. body axis patterning). However, current spatial transcriptomics methods either average local contexts or are restricted to limited fields of view. We introduce sci-Space, which retains single cell resolution while resolving spatial heterogeneity at larger scales. Applying sci-Space to developing mouse embryos, we capture approximate spatial coordinates and whole transcriptomes of ~120,000 nuclei. We identify thousands of genes exhibiting anatomically patterned expression, leverage spatial information to annotate cellular subtypes, show that cell types vary substantially in their extent of spatial patterning, and reveal correlations between pseudotime and the migratory patterns of differentiating neurons. Looking forward, we anticipate sci-Space will facilitate the construction of spatially-resolved single cell atlases of mammalian development.

Single cell methods have the potential to transform our understanding of organismal development. In diverse models of embryogenesis, we and others have performed “whole organism” profiling of gene expression or chromatin accessibility at single cell resolution (1–7), yielding a richer view of the emergence and trajectories of cell types than was previously available or possible (8, 9).

Cells’ spatial organization plays a central role in normal development and homeostasis, as well as in pathophysiology. However, a key limitation of most single cell molecular profiling methods is that they operate on disaggregated cells or nuclei. Although in situ methods can measure the expression of many or all genes while retaining spatial information (10), these methods also have limitations (fig. S1). Some, including the original “spatial transcriptomics” (11) and Slide-seq (12) methods, barcode and then count mRNAs derived from positions across patterned arrays. Although these methods can be implemented at a range of spatial scales, the boundaries of spots have no natural correspondence to the boundaries of cells. As such, they yield aggregate profiles of small regions encompassing multiple cells and/or portions of cells, rather than truly resolving individual cells. Other methods, including MERFISH (13), seqFISH (14), and FISSEQ (15), rely on in situ hybridization or sequencing to measure the expression of many genes while retaining single cell (or even subcellular) resolution within each field of view. However, such methods are typically limited by long image acquisition times and complex instrumentation requirements. In sum, existing techniques necessitate tradeoffs such that assaying the whole transcriptomes of individual cells over large territories remains impractical.

Spatial labeling of nuclei with hashing oligonucleotide

Previously, we developed sci-Plex, a method for labeling or “hashing” nuclei using unmodified DNA oligos prior to single cell RNA-seq with combinatorial indexing (sci-RNA-seq) (16). To leverage this workflow to capture spatial information, we spatially arrayed unique combinations of hashing oligos, and then transferred these oligos to nuclei within a tissue slice by diffusion (17). These hashing oligos, recovered in association with sci-RNA-seq profiles, capture each cell’s approximate tissue coordinates upon sequencing.

As a proof-of-concept of this “sci-Space” approach, hashing oligos were spotted onto glass slides coated with dried agarose. These grids contained 7,056 uniquely barcoded spots spanning a 18mm-by-18mm area (mean radius of 73.2 ± 14.1μm; mean spot-to-spot center distance of 222 ± 7.5μm; fig. S2). About 5% of spots, constituting an identifiable pattern, were also loaded with SYBR green fluorescent dye. After transferring the oligos to the tissue, the grid could be registered with an image of the tissue using these concurrently imaged fluorescent fiduciaries (figs. S2, S3). After optimization of hashing oligo concentrations and dissociation protocols (figs. S4, S5), our protocol comprised four steps: 1) fresh-frozen tissue is sectioned; 2) sectioned tissue is permeabilized with a solution containing a slide-specific oligo and physically juxtaposed to a glass slide bearing the spatially gridded hashing oligos; 3) during oligo transfer, the assembly is imaged; and 4) nuclei from the tissue on the slide are extracted, fixed and subjected to sci-RNA-seq (Figs. 1A, S6, S7).

Figure 1. sci-Space recovers single cell transcriptomes while recording spatial coordinates.

(A) Arrayed single-stranded oligos are transferred onto permeabilized nuclei in fresh-frozen tissue sections and imaged. Cells from each slide are also labeled with a section-identifying barcode so that multiple sections can be pooled prior to sci-RNA-seq. (B) Joint embedding of E14.0 single cell transcriptomes from this study and published data spanning E9.5 to E13.5 (1). Major trajectories are labeled. (C) UMAP embedding of 121,909 cells from sectioned E14.0 mouse embryos. Cell types are denoted by color and number in legend.

Spatially-resolved single cell sequencing of the mouse embryo

We applied sci-Space to profile fourteen sagittal sections derived from two E14.0 mouse embryos (C57BL/6N). After sequencing, quality filtering (18), and assignment of each cell to a slide (on the basis of its slide-specific hashing oligo), our dataset comprised 121,909 spatially resolved single cell transcriptomes (mean 2,514 unique molecular identifiers (UMIs) and mean 1,231 genes detected per cell), without apparent batch effects between slides or embryos (figs. S8, S9). This corresponds to capture of 164 nuclei/mm² of tissue on average, or sampling of 2.2% of the estimated nuclei present (fig. S10). Rather than annotating cell types ab initio, we co-embedded (19, 20) these data with published, non-spatial sci-RNA-seq “mouse organogenesis cell atlas” (MOCA) dataset spanning E9.5 to E13.5 (1) and cells from a developing mouse brain atlas (DMBA) spanning E13.5 to E14.5 (21). Reassuringly, these E14.0 data integrated well with both datasets (Figs. 1B, S11). Draft cell type annotations, inferred by nearest neighbor label transfer, were highly concordant with those recovered by Garnett (22), a semi-supervised annotation algorithm (fig. S11). We refined the annotations by manual inspection of differentially expressed genes (Fig. 1C; File S1).

Images of sectioned embryos and sequencing data were co-registered using SYBR waypoints (22, 24) (Figs. 2A, S12, S13). Each nucleus was mapped to the position matching its highest combination of spot and sector oligos within the imaged section. For ~9% of nuclei, the top assignment was not located near any other nuclei of the same cell type; for such “outliers”, alternative mappings were considered. Altogether, nuclei were well-localized (fig. S8C–F) to one of 15,102 spatial positions across 14 sections; on average, each spatial position was assigned 8.1 nuclei (10.5 s.d.) (fig. S14). To quantify the anatomical distribution of cell type annotations, each section was segmented by organ (Figs. 2B, S15), aided by immunostaining of adjacent sections (25) (fig. S16). Neurons mapped largely within the spinal cord and brain outlined by cells of the developing meninges, cardiomyocytes within the heart, and white blood cells throughout the organism (Figs. 2C, 2D, S18, S17). Analysis with Giotto (26), an unsupervised tool for segmenting spatial transcriptomic images into tissue “domains” of similar cell type composition, revealed 22 domains shared across sci-Space slides. In addition to detecting boundaries between major organs, Giotto was able to automatically recognize more complex domains with distributions extending throughout the embryo (e.g. mesenchymal tissue and cartilage) (fig. S19).

Figure 2. sci-Space captures spatially and cell type resolved gene expression across the embryo.

(A) Co-registered DAPI stained section image and oligo array, superimposed. SYBR waypoints are highlighted in green. (B) Anatomical regions of Slide 1 (left) and an adjacent immunostained serial section aligned to Slide 1 (right). (C) Highlighted cell types mapping to a single slide. (D) UMAP embedding colored by anatomical regions. (E) Gene expression of dopamine transporter Slc6a3 from sci-Space data (left) and published (27) section/stage matched in situ (right). (F) Smoothed percentage of gene expression for Fgf1, Fgfr1, Fgfr2 and Fgfr3 in cardiomyocytes (top) or endothelial cells (bottom). Color scaled to maximum percentage within each gene. Scale bars in panels (A-C) = 0.5 mm.

Finally, wherever cells are captured, sci-Space data enables the visualization of any gene in the transcriptome akin to an in situ hybridization, albeit at lower spatial resolution. For example, a sci-Space “digital in situ” of the dopamine transporter Slc6a3 highlights a cluster of dopaminergic neurons at the midbrain-hindbrain boundary, consistent with stage- and section-matched whole-mount in situs (27) (Figs. 2E, S20). Unlike conventional in situ data, sci-Space data also resolves gene expression by cell type. For example, in the heart, both cardiomyocytes and endothelial cells express the growth factor Fgf1, while only cardiomyocytes express the growth factor receptors Fgfr1, Fgfr2, and Fgfr3 (Fig. 2F).

Contrasting sci-Space and spatial transcriptome capture (STC) methods.

A key distinction between sci-Space and STC methods (fig. S1, left) is that because STC methods capture transcripts from lysed tissue sections (11, 12), each spot can include RNA from multiple cells and/or portions of cells. As such, STC methods are limited in their ability to resolve gene expression variation within individual cells or cell types. To quantify the consequences of this limitation, we spatially aggregated transcriptomes from sci-Space nuclei, effectively creating a “mock-STC” dataset. Cell types and clusters readily discernible in the single cell data were challenging to resolve in the mock-STC data (fig. S21A–C). One approach to ameliorate this limitation involves mapping cells profiled by non-spatial single-cell RNA-seq onto a STC “scaffold”, thereby imputing cellular locations within a tissue (20, 28). To assess the viability of this approach, we applied such imputation to mock-STC data, and compared it to the single nucleus data used to derive it (fig. S21D,E). The mean distance between a nucleus’ imputed position and its measured position was 11.6 spots (~2.5mm), a measure which varied by cell type (fig. S21F). Thus, despite sampling fewer transcripts than STC methods (fig. S21G–I), sci-Space’s single cell resolution--that is, its ability to unambiguously ascertain sets of transcripts expressed in the same cell--represents a key advantage over STC methods.

Analogously, we posited that sci-Space data could serve as a scaffold for the imputation of locations of cells profiled by non-spatial single-cell RNA-seq, but with the advantage of matching single cell transcriptomes. To test this, we aligned neurons from sci-Space and the developing mouse brain atlas (DMBA). Reassuringly, transfer of dissection-based DMBA anatomic annotations were consistent with sci-Space coordinates (fig. S23). Furthermore, this alignment mapped hundreds of DMBA transcriptional clusters to specific positions, many spatially restricted (fig. S23). For example, of 193 transcriptomic clusters from La Manno and colleagues (21) that mapped to slides 13 and 14, 94 and 115 clusters displayed statistically significant focal enrichment, respectively (FDR < 0.01; Getis-Ord Local G).

Spatial patterns of gene expression across cell types

To systematically examine these data for spatially patterned, cell type-specific gene expression across the E14.0 embryo, we quantified spatial autocorrelation--the degree to which the cells expressing a given gene are spatially proximate. Testing each annotated cell type separately, we identified hundreds to thousands of genes exhibiting positive spatial autocorrelation per cell type (File S2; Moran’s test, FDR < 0.001). Amongst the cell types analyzed, connective tissue progenitors and neurons had the most spatially autocorrelated genes detected (Fig. 3A; mean 12,150 ± 2,270 and 8,623 ± 3,846 genes per slide, respectively).

Figure 3. Spatially restricted gene expression in developing neurons.

(A) Number of spatially significant (Moran’s test, FDR < 0.001) autocorrelated genes within each slide (color) and cell type. Only cell types with more than 50 cells per slide were included. (B) Log-log (log10) plot of autocorrelation in UMAP embedding (x-axis) versus autocorrelation in spatial coordinates (y-axis) for each gene. Computed on excitatory neurons from Slide 1. Moran’s I values close to 1 indicate perfect spatial correlation, while a value of 0 indicates a random spatial distribution. Hox genes are highlighted. (C) log10-scale boxplot of Moran’s I statistic for Hox genes displayed in (B) versus all other expression level-matched genes (p-value < 0.001, two sided t-test). (D) Gene expression of HoxC cluster in Slide 1. (E) Similar to (B), log-log (log10) plot of autocorrelation in UMAP embedding (x-axis) versus autocorrelation in spatial coordinates (y-axis) for each gene with genes in different regimes highlighted for Slide 1. (F) Expression patterns across Slide 1 for other spatially restricted genes that are not restricted to a single neuron subcluster. (G) Comparison of sci-Space (Slide 14) and RNA FISH (RNAscope) detected Cyp26b1 patterns of expression (gray) and coexpression with markers (colors as indicated in key) for neuronal and supportive cell types.

One explanation for such spatial autocorrelation of genes within a cell type would be the presence of spatially-restricted, unannotated cell subtypes. For example, upon sub-clustering of connective tissue progenitors, we can indeed find spatially restricted cell subtypes (fig. S24A,B), such that the genes defining these subtypes are expected to be spatially autocorrelated. Alternatively, a gene’s spatial pattern of expression could arise from spatially restricted gene expression contributed by multiple cell subtypes. To distinguish between these scenarios, we calculated each gene’s spatial autocorrelation (Spatial Moran’s I) and compared this value to its autocorrelation in UMAP space (UMAP Moran’s I), a proxy for gene expression driven by a single cell subtype. Across all genes, the two measures were reasonably well-correlated (Pearson’s rho 0.49, p-value < 2e-16). However, a subset of genes, particularly in neurons, displayed higher spatial autocorrelation in the tissue context than in UMAP space (Figs. 3B, S25A,B). For example, Hox genes, a class of homeotic transcription factors that specify the body plan, featured prominently in this subset, consistent with spatial patterning that could not be explained solely by spatial restriction of a single cell subtype (Figs. 3B,C, S25A–D). Expression of HoxA cluster genes paralleled the establishment of the spinal cord’s anterior-posterior axis (29) (Figs. 3D, S25) with no cell subtype restriction observed across datasets (30) or modalities (fig. S26A–C).

Additional non-Hox genes also displayed excess spatial autocorrelation (Figs. 3E,F, S26D). One such gene, Cyp26b1, encodes an enzyme that metabolizes the developmental morphogen retinoic acid (RA). The spatial distribution of morphogens like RA can play a critical role in tissue patterning (29). Sci-Space data localized the focal expression of Cyp26b1 to the brainstem with expression observed in multiple neuronal subclusters (figs. S25A,B, S26A,B). This result was validated by RNA FISH for Cyp26b1 and neuronal-subtype specific genes (Figs. 3G, S26C). Together, these data identify Cyp26b1 expression both in progenitors (radial glia) lining the hindbrain and in their progeny, spatially-adjacent post-mitotic neurons, suggesting that the expression of Cyp26b1 is retained as these cells differentiate. This example illustrates how sci-Space can distinguish between spatial patterns of gene expression driven by a single cell type from those present across multiple cell types.

Quantifying the explanatory power of spatial position

To ask how each cell type’s transcriptome varied globally across the embryo, we calculated the angular distance between the transcriptomes of pairs of cells of the same type separated by varying distances. For many cell types, as the physical distance between cells increased, so did the angular distance between their transcriptomes. However, this trend varied considerably, e.g. particularly pronounced in radial glia, neurons, and endothelial cells (Figs. 4A, S27A,B).

Figure 4. Quantifying the variance in gene expression attributable to spatial position.

(A) Pairwise angular distance (radians) between global transcriptomes of cells of the indicated cell types. Cell pairs are grouped by distance on the physical array (mm) (** p-value < 0.001, *** p-value < 0.0001, Wald linear regression test). (B) Proportion of non-technical variance, explained within cells of each type by spatial position. Point size indicates number of cells and point color indicates the slide of origin. (C) Recovered positions of chondrocytes from Slides 6, 11 and 14 colored by subcluster. Arrows indicate focal concentrations of craniofacial mesenchyme (green) and digit condensate subclusters (red). Insets to the right of each plot 9 show parts of each image with similarly positioned arrows. White text of each inset labels the anatomic structure displayed. Scale bars in panel (C) = 0.25 mm.

To quantify the contribution of spatial context to variation in gene expression across individual cells, we developed a statistical approach. Briefly, we first partitioned cells into groups on the basis of cell type and spatial location. Then, we computed the angular distance between each cell and the average expression profile for cells of that same type in the same spatial bin. After estimating and subtracting the technical variance attributable to data sparsity, we quantified how much of the remaining biological variance was due to each cell’s type and/or spatial position. To validate this approach, we decomposed the variance attributed to lineage in the developing C. elegans embryo, where there is a relationship between two cells’ lineage relationship and their observed gene expression variance (2). Reassuringly, we found that our variance decomposition approach followed the Law of Total Variance (fig. S28C) and in agreement with Packer et al. (2), showed that variance attributable to a cell’s lineage relationship peaked around generation 7 (fig. S28D).

Applying this variance decomposition approach to the sci-Space data, we estimated that sparse UMI sampling accounted for 95.1% of the observed variance in global gene expression across cells and in subsequent analyses, we focused on the remaining 4.9% “non-sampling” variance. Of this non-sampling variance in global gene expression, cell type alone accounted for 19%. However, a joint model that included both cell type and spatial position accounted for 50% (fig. S29). The implication--that spatial information explains as much, if not more, of non-sampling gene expression variance as major cell type--was supported by the recovery of cell type and spatial gene modules of similar size and composition (fig. S30). However, some cell types’ transcriptomes appeared more sensitive to spatial position than others (Fig. 4B). For example, chondrocytes were influenced by position within the embryo, reflecting the ongoing development of various connective tissue lineages at E14.0, and explained at least in part by subclusters that appear to correspond to digit condensates and craniofacial mesenchyme (Figs. 4C, S31). Other such cell types included neurons and their precursors, the radial glia.

Pseudotemporal sci-Space trajectories reflect neuronal migration dynamics

To explore how spatial context might relate to gene expression heterogeneity in a developing cell lineage, we focused on radial glia and neurons. In particular, we hypothesized that we might be able to detect and localize the coordinated processes of neuronal differentiation and migration (31). UMAP dimensionality reduction of these cell types revealed the presence of three distinct trajectories originating in radial glia and leading to neurons (Fig. 5A). Gene expression dynamics along these three branches were consistent with neuronal differentiation -- the upregulation of cell cycle genes followed by expression of genes involved in migration (fig. S32). Each branch was strongly enriched for subtype specific marker gene expression -- Pou4f1+/Pax3+ tectal neurons, Isl1+/Lhx6+ cortical interneurons, and Emx1+/Neurod6+ cortical pyramidal neurons -- indicating that the embedding captures their specification from radial glia (Fig. 5B). We examined how these trajectories were anatomically distributed, by segmenting the brain from each section using the Allen Institute’s Anatomical Reference Brain Atlas (www.atlas.brain-map.org) as a guide. Cells from each trajectory overwhelmingly occupied a distinct brain region (Fig. 5C). To quantify progression through differentiation, we calculated pseudotime for each branch using radial glia as the root (Fig. 5D, S32). Intersecting pseudotime and spatial information, we observed that cells early in differentiation clustered around the ventricles in the forebrain and developing midbrain, while those farther away exhibited a more differentiated transcriptome (Figs. 5E, S33).

Figure 5. Pseudotemporal spatial trajectories capture migratory patterns in the developing brain.

(A-D) UMAP embedding displaying the neural trajectories colored by (A) cell type, (B) specific gene expression, (C) cortical region, or (D) pseudotime. (E) Neurons and radial glia in the cortex colored by pseudotime or otherwise colored grey. Insets of caudal (above) and rostral (below) brain outlines are shown to the right of each slide (M-Midbrain, P-Pallium, SP-Sub Pallium, V-Ventricle). (F) Scaled and centered gene expression for genes (rows) significantly varying over pseudotime in all three trajectories. Enriched Gene Ontology Biological Processes terms (GO BP) are displayed next to clustered genes.

The spatial gradients of cellular maturity estimated with sci-Space data are consistent with well-documented coordination of cellular differentiation and neuronal migration. In the pallium, immature neurons migrate and differentiate radially outwards leading to the inside-out development of the cortical layers (31). In the sub pallium, cortical interneurons born in the ganglionic eminences migrate tangentially to populate the developing cortex and olfactory bulb (32). Our data also identify a third major pattern of migration in which precursors emanate from the dorsal aspect of the ventricular zone in the developing midbrain. These midbrain neurons seem to migrate both radially, towards the pial surface, and tangentially, parallel to the pial surface, to populate this region (Fig. 5E - slides 8, 13, and 14; fig. S33 - slides 4 and 7). Although radial and tangential migration are generally discussed as mutually exclusive phenomena, our data -- in line with some prior work (33, 34) -- suggests otherwise in the developing midbrain. Furthermore, these cells share a common transcriptional program with differentiating and migrating neurons in the pallium and subpallium (Fig. 5F; File S3).

Discussion

In summary, sci-Space is a method for spatial transcriptomics that retains single cell resolution while capturing spatial information at a scale specified by a patterned array of cell hashing oligos. As a proof-of-concept, we applied sci-Space to retrieve the approximate spatial coordinates of transcriptionally profiled cells across whole mount sections from E14.0 mouse embryos. The sci-Space data are readily integrated with non-spatial single-cell RNA-seq data collected from mouse embryos at adjacent timepoints (1, 21), enabling rapid annotation of diverse cell types and visualization of cell type-specific, spatially patterned gene expression, i.e. digital in situs. We identify examples, some expected and others novel, of genes expressed in an anatomically patterned manner within cells of a given type.

The spatial resolution of sci-Space is presently limited by the patterned array of hashing oligos, here to approximately 200 microns. Although increasing spot density and decreasing spot size are a straightforward path to increasing resolution, sci-Space is unlikely to detect effects arising from interactions between adjacent cells. This is limited by recovery of only a fraction of cells from each serial section, such that we obtain a “survey”, rather than achieving dense measurements. Nonetheless, sci-Space fills a need not addressed by other technologies. Like other spatial transcriptome capture methods (e.g. Slide-seq), sci-Space can be applied, routinely and efficiently to large regions, e.g. whole embryo serial sections. However, unlike these methods, sci-Space recovers single cell transcriptomes. It can therefore capture patterns of spatial gene regulation private to specific cell types and estimate the contribution of each cell type to the expression of morphogens and other signalling molecules, both within and across anatomical regions. Moreover, sci-Space data can serve as a spatial “scaffold” for conventional, non-spatial single-cell RNA-seq atlases, which may be more challenging to map onto tissues profiled by spatial profiling methods that lack single-cell resolution.

Finally, using this spatially-resolved single cell data, we developed a statistical approach to identify cell types in the developing embryo that exhibit spatially regulated gene expression. Closer analysis of radial glia and neurons revealed gradients of developmental maturity in different regions of the brain indicative of known and novel patterns of neuronal migration. Together with data from complementary technologies, we anticipate that the further application of sci-Space to serial sections spanning entire embryos from many timepoints will facilitate the construction of a set of highly time- and space-resolved 4-dimensional atlases of gene expression across mammalian development.

Supplementary Material

File S1

File S1: Curated gene list for cluster annotation

File S3

File S3: Pseudotime-dependent genes for neurons of the brain

File S2

File S2: Spatially autocorrelated genes

Data and materials availability:

Processed and raw data can be downloaded from NCBI at GSE166692. Additional data, and code is hosted at https://github.com/cole-trapnell-lab/sci-Space (35). Laboratory protocols can be found on protocols.io (36–38).