Main

Embryonic development necessitates the precise timing and location of numerous molecular, cellular and tissue-level processes1,2,3,4,5. These events are directed via spatiotemporal control of gene expression that orchestrates cell type specification, migration and localization6,7,8,9. Any disruption of this regulation often results in embryonic lethality or developmental defects5,10,11. At the end of gastrulation and the onset of organogenesis (embryonic days (E) 8.5–9.5), tissues experience major morphological changes, such as heart looping, brain compartmentalization and neural tube folding, to guarantee proper structure and function12,13,14. Through neurulation, epithelial cells in the neural plate fold to form a morphologically defined tube, which exhibits a stratified gene expression signature along its dorsoventral (DV) axis, which is necessary for subsequent neuronal subtype diversification15,16,17,18,19,20,21. Many genes involved in this process have been identified, but the precise gene regulatory networks governing these patterns remain under investigation. Recent single-cell studies have begun to provide a deeper understanding of the topography of fate specification and highlighted some molecular mechanisms underlying these cell state transitions22,23,24,25,26,27,28. One limitation of dissociation-based approaches is their inability to preserve tissue structure, which precludes expression analysis within the native context. Recent advances in spatial transcriptomic technologies have begun to fill this gap, aiming to explore the organization of cell types within adult tissues and developing embryos29,30,31,32,33,34,35,36,37,38.

In this study, we used Slide-seq, a technology that generates transcriptome-wide gene expression data at 10-µm spatial resolution33,39, to build maps of whole embryos during early mouse organogenesis. Our data enabled the exploration of spatial gene expression patterns, cell state distributions, the reconstruction of three-dimensional (3D) transcriptomic maps for ‘virtual’ gene expression analysis and mutant phenotype dissection. We specifically leveraged the data to identify regionalized gene expression and differentiation trajectories in space, focusing on neural tube formation and patterning. Overall, we provide a comprehensive, high-resolution spatial atlas together with an accessible and ready-to-use visualizer to explore gene expression patterns in the developing mouse embryo.

Results

Spatial transcriptomic maps to construct 3D virtual embryos

To spatially map cell identities during early organogenesis, we used Slide-seq on two representative E8.5, one E9.0 and three partial E9.5 embryos (Fig. 1a and Extended Data Fig. 1a). For the two E8.5 embryos, we obtained 15 and 17 sagittal sections (10-µm thickness), respectively, with approximately 30-µm intervals between them. For the E9.0 embryo, 26 sagittal sections with 20-µm intervals, and for the three E9.5 embryos, 13 slices from the midline were obtained (Fig. 1a and Supplementary Table 1). In total, we recovered 533,116 high-quality beads with a median value of 1,798 transcripts and 1,224 genes per bead (Extended Data Fig. 1b–d). To ascertain the cell states assigned to each bead, we computationally mapped beads to a previously generated single-cell reference (Extended Data Fig. 1e,f)26. With this information, we extracted from each sequenced bead (1) spatial coordinates, (2) associated gene expression profile and (3) cell state assignment. Overall, we observed good alignment of cell states and spatial restriction of marker genes, such as Ttn (heart), T (tail bud), Meox1 (somites) and Sox2 (neural tube, brain), among others (Extended Data Fig. 2a–e). Additionally, we observed high reproducibility in recovering a comparable embryo composition and gene expression patterns among replicates (Extended Data Fig. 2c–f).

Fig. 1: Embryo-wide profiling of gene expression with spatial coordinates using Slide-seq.
figure 1

a, Schematic of the experimental workflow and data analysis. Sagittal sections of mouse embryos at E8.5, E9.0 and E9.5 were obtained for Slide-seq. The dotted lines indicate the approximate position of the embryonic sections. b,c, 3D-reconstructed E8.5 (b) and E9.0 (c) stage embryos with six cell states highlighted (brain, heart, neural tube, somites, NMPs, PSM); the caudal marker gene Cdx2 and heart tube marker gene Ttn are shown (normalized gene expression) in a vISH of the reconstructed E8.5 embryo1. Each dot corresponds to a single bead. Point of view is denoted by the eye symbol. Scale bars, 100 µm. d, 3D view of an E8.5 embryo showing the indicated cell states and anatomical features in brightfield (left) or mapped onto the E8.5 embryo1 (right). Different orientations are displayed. Scale bars, 500 µm. e, Schematic of the strategy used to identify localized gene expression in tissues. f, Heatmap of localization scores for the top 20 spatially differentially expressed genes in 3D for each analyzed tissue in the E8.5 embryo1 (see Supplementary Table 4 for the list). Top 20 genes (row z-score-normalized) in the brain are highlighted. A, anterior; LPM, lateral plate mesoderm; P, posterior; D, dorsal; V, ventral; LPM, lateral plate mesoderm; NMPs, neuromesodermal progenitors; PSM, pre-somitic mesoderm.

Full size image

To translate our two-dimensional data into a 3D embryo, we developed sc3D, a computational method that enables the alignment of individual spatial transcriptomic arrays for 3D reconstruction. Specifically, we used sc3D to align serial Slide-seq ‘z’ samples from E8.5 and E9.0 embryos, which allowed us to capture the spatial distribution and morphologies of the emerging tissues at the onset of organogenesis (Extended Data Fig. 3a,b and Supplementary Figs. 1 and 2). This in turn enabled quantitative measurements of their volumes (250–39,264 × 103 µm3), which were reproducible across individual replicate embryos (Extended Data Fig. 3c–e, Supplementary Fig. 1, Supplementary Movie 1 and Supplementary Table 2). We further showed that the reconstruction remained consistent as the interval between slices increased, with very minimal distortion in the rotation axes (Extended Data Fig. 3f). Importantly, sc3D also allows 3D reconstruction of other spatial transcriptomic datasets with high precision, speed and robustness to reduced spatial resolution40 (Extended Data Fig. 4a–c and Supplementary Table 3).

We next used the reconstructed embryos to perform ‘virtual’ in situ hybridization (vISH) of over 27,000 genes on a quantitative scale, with the opportunity to query gradient gene expression along any given body axis, inclination plane and rotation angle (Fig. 1b–d, Supplementary Figs. 1 and 2, and Supplementary Movies 1–4). To further increase sc3D accessibility, we developed sc3D-viewer, a user-friendly and interactive environment for exploring the Uniform Manifold Approximation and Projection (UMAP) representation, spatial cell state maps, and vISH for single and dual gene combinations (methods in Supplementary Information). To investigate spatial gene expression within each tissue type, we computed the genome-wide correlations between tissue volumes and densities of expressing cells to generate a localization score that allowed us to rank genes within each tissue based on their spatial restriction in expression. This analysis identified a set of highly informative, tissue-specific, regionalized genes within the embryo and across replicates as well as developmental stages (Figs. 1e,f and 2a, Extended Data Fig. 4d–g and Supplementary Table 4). For instance, we found high localization scores for genes such as Nppa, Tdgf1, Cck and Sfrp5 in the developing heart tube (Extended Data Fig. 5a–d and Supplementary Movie 5). Mapping these genes onto our digital embryo revealed that their expression domains mark specific developmental axes (anterior–posterior (AP), DV and right-left) and delineate presumptive anatomical structures, such as the primitive ventricles and atria, the outflow tract, the cardiomyocyte jelly and the venous pole, respectively41,42,43, as opposed to spatially ubiquitous gene expression seen in the blood (Extended Data Fig. 5c–e). Furthermore, Cck and Sfrp5 expression spatially distinguish differentiated from undifferentiated cardiomyocytes domains (Extended Data Fig. 5d)41,42,43. Taken together, this demonstrates our ability to identify regionalized markers and study the distinct domain organization within complex developing tissues along developmental axes.

Fig. 2: Localized gene expression patterns in the developing brain.
figure 2

a, Schematic (left) and normalized gene expression spatial plot (right) of selected genes (highlighted in Fig. 1f) in the 3D brain of E8.5 embryo1, E9.0 and E9.5 (array E9.5_3). Each dot represents a single bead. b, Spatial plot of RNA velocity in the E9.5 (array E9.5_3) brain region. Vector direction indicates the trajectory and length denotes the magnitude. Low-velocity regions are indicated as R1, R2, R3 and R4. R1, neural ridge prosencephalon; R2, telencephalon–diencephalon boundary; R3, diencephalon–mesencephalon boundary; R4: mesencephalon–hindbrain boundary; and R5, hindbrain–spinal cord boundary. c, Spatial plot showing the brain boundaries (top) and top enriched spatially differentially expressed genes along the R2, R3 and R4 brain boundaries (represented as row z-score-normalized expression) at E9.5 (array E9.5_3). The full list of genes can be found in Supplementary Table 5. Scale bar for all plots, 50 µm. C, caudal; D, dorsal; R, rostral; V, ventral.

Full size image

Delineating molecular boundaries in the developing brain

Between E8.5 and E9.5, the most anterior portion of the neural tube develops into three distinct vesicles (prosencephalon–forebrain, mesencephalon–midbrain and rhombencephalon–hindbrain), which together form the primordial brain20,21,44,45,46,47. Our high-resolution spatial transcriptomic map of the forebrain–prosencephalon in E9.5 embryos allowed us to identify presumptive telencephalon and diencephalon regions and to delineate DV patterning of the diencephalon–midbrain (Extended Data Fig. 6a–e)48. To study the emergence of such patterns, we analyzed the transcriptome of E8.5, E9.0 and E9.5 brains and found that genes such as Foxg1, Barhl2, Otx2, En1 and Egr2 show regionalized expression patterns already at E8.5 (Figs. 1f and 2a). While Foxg1 was confined to the rostral prosencephalon, defining the presumptive telencephalon, Barhl2 was expressed caudally, already marking the presumptive diencephalon49,50,51. Rax, a marker of the developing eye, exhibited spatial gene expression confinement between E8.5 and E9.5, defining the future optic cup (Fig. 2a). It appears therefore that spatial restriction of gene expression precedes anatomical segregation. To further explore the relationship between cell fate commitment and spatial restriction of emerging structures, we used unsupervised spatial RNA velocity without prior knowledge of cell states52. We recovered distinct ranges of velocity dynamics with either converging or diverging trajectories, potentially corresponding to stepwise transitions or cellular steady states (Fig. 2b and Extended Data Fig. 7a,b). A closer inspection into low-velocity regions (defined by the low-velocity length and confidence of vector directionality), combined with the expression of known marker genes, revealed the presence of progenitor field domains (R1 anterior neural ridge) as well as differentiated neuronal territories (R5 hindbrain–spinal cord boundary) (Extended Data Figs. 6b,c and 7c). Additionally, areas with diverging trajectories highlighted boundary regions, such as the R4 mesencephalon–rhombencephalon boundary, which was marked by the restrictive and exclusive expression patterns of Otx2 in the mesencephalon and Gbx2 in the rhombencephalon, as well as Fgf8 at the boundary (Extended Data Fig. 7c)53.

Although vector ends do not necessarily represent a terminally differentiated state, such a relationship might be inferred when spatial trajectories are known to match with developmental patterning processes. For instance, the trajectories observed at R4 resemble the lineage specification of the mid–hindbrain progenitors that will generate neuronal cells that subsequently populate the whole midbrain and anterior hindbrain areas21. In addition, our high-resolution map enabled a more granular view of the molecular determinants of developing boundaries before the formation of anatomical constriction (Fig. 2c and Supplementary Table 5). We analyzed the three boundaries demarcating the main brain regions to identify features that might shed light on the regulatory mechanisms involved in brain regionalization. In R2 and R4, we identified several signaling molecules (WNT, FGF) and downstream effectors ratifying the role of these boundaries as signaling centers instructing the patterning of the adjacent structures (the zona limitans intrathalamica (ZLI) and the isthmic organizer). While the interplay between WNT and FGF signaling in the mid–hindbrain boundary (R4) is well known, their role at the telencephalon–diencephalon boundary (R2), where SHH signaling has a major role54, has been less studied. In this study, we show the spatial constriction of Wnt7b expression to the boundary together with Wnt8b (Extended Data Fig. 7d). Although Wnt7b is expressed in the rostral and dorsal parts of the diencephalon55, its co-expression and colocalization with Wnt8b, known to be expressed within the ZLI56, has not been described yet. Compared to the other analyzed regions, the diencephalon–midbrain boundary (R3) is characterized by low signaling molecule activity, high neuronal marker expression (neurofilament proteins and neuronal genes) and convergent RNA velocity signature, suggesting the presence of a more mature neuronal rather than a progenitor domain (Fig. 2c and Supplementary Table 5). Therefore, our analysis can help identify relevant molecules such as CDH8 (Fig. 2c and Supplementary Table 5), which is known to compartmentalize the diencephalon–midbrain boundary together with other cadherins57. Moreover, within the forebrain region, we mapped known as well as uncharacterized gene expression distributions, including those involved in eye development (Extended Data Fig. 7e,f). By examining such patterns during early brain development, we stratified the spatial emergence of anatomical structures.

Cell identity and spatial distance in the trunk

While the anterior neural tube develops into the brain, the posterior portion generates the future spinal cord during trunk elongation. The embryonic trunk consists of morphologically diverse structures with distinct developmental origins that ensure correct axial elongation and body plan segmentation (Fig. 3a)58,59,60. As the embryo develops, axial progenitor cells acquire a bipotent differentiation potential that allows the generation of neuroectodermal and mesodermal derivatives58,59,60. Specifically, these cells, known as neuromesodermal progenitors (NMPs), generate the posterior portion of the neural tube and the paraxial mesoderm via the determination front (pre-somitic (PSM) and somitic mesoderm) (Fig. 3a)58,59,60. To profile gene expression regionalization in the developing trunk in 3D, we first mapped the cell states corresponding to the NMPs, PSM and somites (somitogenesis trajectory) onto our E8.5 virtual embryo (Fig. 3b). The vISH of Tbx6, Ripply2 and Meox1 further confirmed the organized gene expression patterns along the AP and right-left symmetry involved in somitogenesis (Fig. 3b). Next, to understand how a regionalized gene expression signature impacts developmental dynamics, we profiled trunk developmental trajectories by combining transcriptional pseudotime measurements with spatial information. UMAP and spatial visualization revealed a continuum of transcriptomic states along the somitic and neural trajectories (Fig. 3c,d and Extended Data Fig. 8a–c). We next investigated this relationship in more detail and calculated the transcriptional and spatial distances for every bead to every other bead within each trunk tissue in an E9.5 embryo (‘pseudotime’ and ‘spatial’ distances, respectively) (Fig. 3e)61,62,63. Interestingly, we discovered that changes in transcriptomes are not necessarily proportional to cell–cell distances and that their relationship is tissue-specific. In particular, progenitor cells (NMPs and PSM) occupy a small region of the embryo, showing low spatial distance distributions concurrent with low transcriptional variability (Fig. 3e). On the other hand, more differentiated cells (neural tube and somites) are widely distributed along the trunk region (high distance distribution), even in cases of low transcriptional differences (Fig. 3e). We also observed a group of proximal cells in the neural tube characterized by notable transcriptional variability that, when mapped to spatial arrays, represented local DV patterning (Fig. 3e dotted line and Fig. 3f). Together, our findings show that during trunk development, progenitors differentiate into subsequent cell states in a restricted spatial domain before dispersal.

Fig. 3: Spatial organization of the embryonic trunk.
figure 3

a, Schematic of the spatial organization of cell states in the embryonic trunk region at E8.5, and close-up view on the somitogenesis process. b, Schematic and 3D spatial plot of E8.5 (embryo2), showing selected cell states and vISH of Tbx6, Ripply2 and Meox1 in the trunk region. Each dot denotes a bead and the color corresponds to the indicated state. Scale bar, 200 µm. c, UMAP showing beads from E8.5 and E9.5 embryos corresponding to the NMPs, PSM, somites and neural tube states (left) and the corresponding spatial distribution (right) in an E9.5 embryo (array E9.5_2). Each dot denotes a bead and the color corresponds to the indicated state. Scale bar, 100 µm. d, UMAP showing pseudotime analysis along the somitic and neural differentiation trajectories (left) and the corresponding spatial distribution (right) in an E9.5 embryo (array E9.5_2). Each dot denotes a bead and the color corresponds to the assigned pseudotime value. Scale bar, 100 µm. e, Density plot displaying the computed pseudotime difference (y axis) versus the measured spatial distance (x axis) between all beads of the same cell state. Each dot is a pairwise comparison. The dotted line delineates cells with low spatial distance and large transcriptional divergence in the neural tube. f, Spatial plot showing the beads (top) within the dotted line (Fig. 2e) and the distribution of pseudotime values in an E9.5 embryonic trunk (array E9.5_3) that reflects neural tube patterning. Each dot represents a bead. Scale bars, 100 µm.

Full size image

Emerging patterns along the neural tube axes

As the trunk develops and tissues extend in space, the transcriptional differences along their length determine the positional identity of various cellular states. For example, between E8.5 and E10.5, the neural tube folds from the neural plate and undergoes patterning to establish the cellular stratification for the future spinal cord16,17,19,20,21,62,63,64. Regionalized gene expression programs guarantee further diversification of neuronal types along the AP and DV axes16,17,19,20,21,62,63,64. To understand the genetic programs involved in establishing the discrete progenitor domains along the neural tube, we isolated the corresponding beads and searched for spatial co-expression patterns along the AP (approximately 4,600 µm) and DV (approximately 320 µm) axes (‘axis profiling’) (Fig. 4a and Extended Data Fig. 8d). We identified distinct expression patterns along the neural tube AP axis for genes involved in several cellular and molecular functions (Extended Data Fig. 8e). As expected, Hox genes were among the most highly localized genes within this axis (Fig. 4b)65,66. HOX factors interact with each other to regulate transcriptional programs. To identify putative functionally collinear groups, we performed Hox gene module analysis combined with spatial resolution and found six distinct modules of Hox gene expression, from the most anterior module (Hox module I, comprising Hoxb2 and Hoxa3) to the most posterior one (Hox module VI, containing Hoxd8 and Hoxa9 (Fig. 4c and Extended Data Fig. 8f,g). Next, we examined the DV axis of the neural tube and were able to resolve and annotate well-studied structures like the notochord, the floor plate, the actual neural tube and the roof plate based on their spatially restricted transcriptional signature (Fig. 4a,d and Extended Data Fig. 9a,b)29. Among the most spatially constrained genes, we found well-known lineage-defining markers like Zic1, Pax3, Olig2, Nkx6-1 and Nkx2-9, which we further confirmed using RNA–fluorescence in situ hybridization (FISH) (Fig. 4d–f and Extended Data Fig. 9c–e). We also identified 43 additional genes in the early mouse neural tube at the E9.5 stage that appear to exhibit a patterned expression along the DV axis in the ventricular zone containing progenitors (Fig. 4d–f, Extended Data Fig. 9f and Supplementary Table 7). Our results show the utility of the axis profiling tool in detecting well-studied gene expression patterns along the neural tube axis and its application in the de novo discovery of genes with locally restricted expression.

Fig. 4: Neural tube profiling along the AP and DV axes.
figure 4

a, Schematic of an E9.5 stage embryo, with beads corresponding to the neural tube (green) and the denoted axes along which profiling was performed (array E9.5_10). b, Heatmap showing the top 160 genes with expression patterns (row z-score-normalized, filtered for a log fold change greater than 0.05 and a false discovery rate (FDR) < 0.01; see Supplementary Table 7 for the list of genes) in the 25 generated spatial bins along the AP axis. Expression of Hox genes is highlighted (right). Scale bar, 184 µm. c, Spatial plots showing the normalized expression (bin z-score) of Hox modules (I, II, III, IV, V, VI) in the E9.5 (array E9.5_10) neural tube. Two representative genes for each module are indicated. Scale bar, 100 µm. d, Heatmap showing the top 160 genes with expression patterns (row z-score-normalized, filtered for a log fold change greater than 0.05 and FDR < 0.01; see Supplementary Table 7 for the list of genes) in the eight generated spatial bins along the DV axis. Expression of known patterned and our identified genes are highlighted. Scale bar, 40 µm. e, RNA–FISH showing validation for known (top) and our identified (bottom) patterns along the DV axis of the neural tube. A single tissue slice obtained from the posterior portion of the neural tube is shown in the images. n = 3 embryos with reproducible staining pattern for each profiled target from WT embryos. Scale bars, 50 µm. f, Schematic showing the spatial patterns of known (within the tube) and our identified (outside the tube) genes along the DV axis of the E9.5 stage neural tube.

Full size image

Conflicting identity of ectopic neural tubes in Tbx6 mutants

After cataloging the spatial transcriptome underlying the developing neural tube axes, we investigated a classic embryonic mutant phenotype where ectopic neural tubes arise. The T-box transcription factor TBX6 is expressed in the PSM and is required for somite segmentation and specification10. In embryos lacking Tbx6, ectopic neural tubes arise at the expense of the somitic compartments (Fig. 5a)11,67. To assess the precise molecular identity of the ectopic neural tubes, we used CAS9-based disruption of Tbx6 in zygotes and performed Slide-seq on E9.5 wild-type (WT) and Tbx6 mutant (Tbx6 knockout (KO)) transversal embryo sections, focusing on the posterior segment of the trunk region where multiple tubes arise in the absence of Tbx6 (Fig. 5a and Extended Data Fig. 10a–c)26,68,69,70. As expected, we observed an overrepresentation of beads assigned to the neural tube cluster in Tbx6 KO embryos compared to WT controls, along with a commensurate lack of somitic cells (Fig. 5a).

Fig. 5: Slide-seq profiling of Tbx6 KO embryos.
figure 5

a, Transverse section of DAPI stained embryos (top) and spatial plot (bottom) showing the annotated tissue morphologies (somites, neural tube) and corresponding cell states in E9.5 WT and KO embryo. The KO experiment was independently performed five times in total with n > 5 embryos per experiment, and consistently yielded the same phenotype. The Slide-seq experiment was performed on one representative transversal section for WT and two for KO embryos. b, Spatial grid map showing the organization of neural tube 1, 2 and somitic cells in WT and Tbx6 KO embryos. c, UMAP showing the projection of cells assigned to the indicated clusters on the trunk trajectory (Fig. 3c), with the size of the dots representing the degree of uncertainty of mapping to the respective position. d, Dot plot showing the expression of the indicated genes in the three clusters (dot size is percentage of cells per cluster; color is cluster average normalized expression). e, RNA–FISH of the indicated genes in a transversal section of an E9.5 WT and KO embryos. The dotted lines in the schematic denote the AP position within the trunk from which sections were obtained. A representative section from WT and KO embryos at E9.5 is shown. The expression pattern was verified in two of the KO experiments with n > 3 embryos per experiment. Scale bars, 50 µm. f, Schematic showing the transcriptional identity and patterning characteristic of the ectopic neural tubes that arise in the absence of Tbx6 expression, highlighting their conflicting transcriptional identity.

Full size image

Next, we reclustered the beads having a neural and somitic identity to more closely inspect differences between the neural tubes in the Tbx6 KO embryos. We found that neural tube cells are subsequently resolved into four transcriptional subclusters, which we labeled as neural crest, neural plate and two main neural tube clusters (neural tubes 1 and 2, Extended Data Fig. 10d)71,72,73,74,75. Interestingly, when we spatially assigned the two neural tube cluster cells on WT and Tbx6 KO arrays, we discovered that neural tube 1 cells mapped to the central tube in both genotypes. In contrast, neural tube 2 cells mapped to both sides of the central tube exclusively in the Tbx6 KO, suggesting that ectopic tubes are characterized by a distinct transcriptional state (Fig. 5b). Specifically, cells of ectopic tubes display a transcriptional identity that is in between the somitic and neural cells, despite their predicted neural identity (Fig. 5c and Extended Data Fig. 10e,f). We observed high levels of mesodermal-specific genes including Aldh1a2 and Mest (Fig. 5d,e). Concomitant with the expression of mesodermal markers, ectopic tube-assigned cells had reduced or absent expression of classic neural tube patterning marker genes such as Olig2, Gm38103, Sox3, Nkx6-1 and Shh (Fig. 5d,e). Nonetheless, the mesodermal signature was overlayed with the conflicting expression of other characteristic neural tube patterning genes, such as Foxa2 and Pax6 (Extended Data Fig. 10g)11.

In summary, our spatial transcriptomic analysis of the ectopic tubes in Tbx6 mutant embryos identified cells that acquire a mixed transcriptomic identity, characterized by the expression of mesodermal genes and partial DV gene expression patterning (Fig. 5f).

Discussion

We performed embryo-wide spatial transcriptomic profiling using Slide-seq to decipher the tightly regulated gene expression patterns of approximately 27,000 genes in developing embryos at the onset of organogenesis. The reconstruction of digital 3D embryos using sc3D enabled the quantitative exploration of gene expression patterns and gradients on a virtual in situ basis. Combined with the development of sc3D-viewer (a napari plugin)76, an accessible, interactive and user-friendly visualization platform to register and explore 3D spatial genomic data, we facilitate the rapid and seamless exploration of cell type distribution and gene expression patterns along any given developmental axis, including the possibility to reconstruct tissues from other spatial transcriptomic datasets.

Spatially resolved single-cell sequencing methods continue to evolve rapidly37,38,40,77,78,79. The increase of available datasets will require faster and more precise computational approaches to take full advantage of the added spatial information. sc3D contributes toward the in-depth analysis of the topology and geometry of gene expression and co-expression patterns, providing the infrastructure to start modeling gene expression diffusion, and their interaction in embryonic tissues. The sc3D data structure has been purposely designed to be close to the one of imaging-based cell tracking algorithm outputs. This similarity will ease the porting of cell tracking-based inter-sample alignment algorithms, such as Tardis, to the spatial transcriptomic field31.

The dynamics at which transcriptomes evolve as progenitors differentiate in their spatial distribution has been challenging to explore. The analysis of differentiation trajectories in the embryonic brain and trunk regions revealed several discrete domains in which transcriptional changes converge or diverge spatially, indicating a non-linear and tissue-specific relationship. We observed that spatially dependent relationships were frequently associated with regions characterized by high intercellular signaling gradients. These changes may reflect the specification of sublineages, progenitor pool migration, differentiated subtypes maturation or regionalization of specific fates. Additional studies combining lineage recording, spatial information and single-cell transcriptomic profiling might help resolve the causal relationship between gene expression programs, spatial allocation and cell fate specification.

The transcriptional rewiring along the AP and DV axes of the developing neural tube defines discrete gene expression domains instrumental in controlling future cellular diversification16,17,19,20,21,62,63,64. We discovered several interesting genes that display regionalized gene expression patterns along these axes, including epigenetic and metabolic regulators that require further investigation to determine their molecular role and functional implications.

As an example for leveraging spatial information in a perturbation experiment, we provided a detailed transcriptomic characterization of the molecular identity of the ectopic neural tubes that arise in Tbx6 mutant embryos. Unexpectedly, what has been historically assigned as additional neural tubes are morphologically tubular structures with incomplete patterning and continued expression of mesodermal genes that are usually associated with non-epithelial and mesenchymal cell identities. This suggests a decoupling of transcriptional programs and morphogenetic outcomes during embryonic development, with signaling gradients, extracellular matrix and mechanical clues probably playing crucial roles80. While Tbx6 mutant embryos exhibit a distinct and well-characterized phenotype, many other genes may cause less obvious morphological changes and hence will benefit from a similar spatiotemporal characterization to define their developmental roles.

Our accessible resource of spatial transcriptomic maps at the onset of organogenesis and supporting computational tools will help the continued exploration of mammalian development. Furthermore, the framework presented in this study could be implemented to conduct molecular spatial phenotyping on many additional perturbations. Lastly, by combining lineage mapping and multi-omic analysis, a comprehensive map of the gene regulatory network during embryogenesis could be developed based on this work.

Methods

All experiments described in this article comply with the relevant ethical regulations at the respective institutions. All experiments were approved by the Landesamt für Gesundheit und Soziales. All animal procedures were performed according to animal welfare guidelines and regulations approved by the Max Planck Institute for Molecular Genetics (G0243/18-SGr1_G and ZH120).

Animal work and embryo preparation (WT and KO embryos)

WT E8.5, E9.0 and E9.5 embryos were dissected from the uteri of naturally mated CD-1 mice in 1× HBSS (catalog no. 14175053, Gibco) on ice. Embryos were staged based on morphology, size and somite number (3–5 somite pair stage for E8.5, 10–12 somite pair stage for E9.0 and 15–18 somite pair stage for E9.5). Extra-embryonic tissues were removed from E9.5 embryos before processing. Embryos were washed in cold 1× HBSS with 2 U ml−1 RNase inhibitor (catalog no. N8080119, Thermo Fisher Scientific) and embedded in O.C.T. solution (catalog nos. 23-730-571 and 23-730-572, Thermo Fisher Scientific). Embryos in O.C.T. were oriented under a stereoscope, immediately placed on dry ice for flash-freezing and ultimately stored at −80 °C. Tbx6 mutant embryos were generated by a previously established protocol26,68,69,70. Briefly, in vitro fertilized (IVF) zygotes were electroporated with Alt-R CRISPR–Cas9 RNP complex with guides targeting three different exons of Tbx6 (Supplementary Table 10). Embryos that developed to blastocyst stage were retransferred to CD-1 pseudo-pregnant surrogate animals as described. All Tbx6 mutant embryos isolated at E9.5 showed the mutant phenotype (enlarged tail bud, ectopic neural tubes). The trunk region was dissected from WT and mutant embryos (removed parts above the limb bud and heart), embedded in O.C.T. solution and frozen at −80 °C.

Animals were kept under specific pathogen-free conditions in individually ventilated cages at 22 ± 2 °C, 55 ± 10% humidity with a 12-h light–dark cycle (6:00–18:00). IVF was performed with B6D2F1 oocyte donors (aged 7–9 weeks; Envigo) and sperm was isolated from B6.CAST F1 males (aged 2 months, generated in-house by breeding C57BL/6J females and CAST/EiJ males). For the embryo transfer experiments, pseudopregnant CD-1 female mice (Hsd:ICR; 9–12 weeks old; 21–25 g; Envigo) were mated with vasectomized males (Swiss Webster; older than 13 weeks; Envigo).

Cryosectioning for Slide-seq V2

Fresh-frozen O.C.T. blocks with mouse embryos were equilibrated to −20 °C in a cryostat (CM1950, Leica Biosystems), mounted onto a cutting block with O.C.T., sliced at a 10-μm thickness and then overlaid and melted onto sequenced spatial arrays33,39. Sagittal sections for whole embryos were collected at the following intervals: E8.5 embryos, 30-μm distance; E9.0 embryos, 20-μm distance; and for E9.5, sections from mid-volume were collected from three independent embryos. One for the head region and another for the thoracic and trunk region from each embryo (Fig. 1a and Supplementary Table 1). An E9.5 transversal section of the trunk region (posterior trunk corresponding to the somite-neural tube region) was collected for WT and Tbx6 mutant embryo with a 10-µm section thickness.

Whole-embryo RNA–FISH

Whole-embryo RNA–FISH was performed according to the protocol from Molecular Instruments with some modifications. Briefly, embryos fixed overnight with 4% paraformaldehyde (PFA) at 4 °C were washed three times for 10 min each with 1× PBS with 0.1% Tween 20 (PBST) at 4 °C. Embryos were dehydrated in an increasing concentration series of methanol + PBST washes, for 10 min each wash at 4 °C (25% methanol; 50% methanol; 75% methanol; 100% methanol). Embryos were stored at −20 °C overnight or longer. Next, embryos were rehydrated in a decreasing concentration series of methanol + PBST washes, for 10 min each wash at 4 °C (100% methanol; 75% methanol; 50% methanol; 25% methanol; 100% PBST). After two washes for 10 min at 4 °C in PBST, embryos were bleached with 6% hydrogen peroxide (for endogenous peroxidase activity in blood cells) at 4 °C for 20 min. After two washes in PBST for 10 min each at 4 °C, embryos were treated with 10 µg ml−1 proteinase K (catalog no. EO0491, Thermo Fisher Scientific) for the indicated time at room temperature (E9.5: 10 min). After two washes with PBST for 15 min each, embryos were postfixed in 4% PFA for 15 min at room temperature and washed three times in PBST for 15 min each step. Embryos were then prepared for hybridization by incubating in hybridization buffer at 37 °C for 1 h. Probes were resuspended in hybridization buffer at a concentration of 1 pM and incubated with embryos overnight at 37 °C. Embryos were washed four times with probe wash buffer for 15 min each wash at 37 °C, followed by three washes in 5× SSCT hybridization buffer + 0.1% Tween 20. Fluorescent hairpins were prepared as described by the manufacturer at a concentration of 0.06 µM each hairpin in amplification buffer. Embryos were then incubated in amplification buffer before incubation with hairpin probes overnight at room temperature in the dark. Excess probes were removed by five washes of 15 min each step in 5× SSCT at room temperature in the dark. Nuclei were counterstained by incubation with 2 µg ml−1 4,6-diamidino-2-phenylindole. The buffers and probe sequences used in this study are available at Molecular Instruments and their unique ID can be found in Supplementary Table 10. After hybridization, embryos were embedded in O.C.T. and frozen at −80 °C. O.C.T. blocks were sectioned to obtain transversal sections of the trunk region and neural tube at 30-µm thickness. Embryos and sections were imaged with a ZEISS LSM-880 confocal microscope at 10×, 20× magnification, averaging four times per frame and 10-µm z-stacks. Images were processed with the ImageJ software. The Plot Profile function was used to perform the signal intensity along a user-defined axis for each fluorescent channel.

Slide-seq V2

The Slide-seq V2 protocol was used to generate all the sequencing libraries. Bead synthesis, array sequencing, image processing and base calling were performed33,39 as described below. Briefly, the 10-μm barcoded beads were synthesized in-house by ChemGenes with a 14-bp spatial barcode separated by a 14-bp linker sequence, an 8-bp unique molecular identifier (UMI) sequence and a 20-bp poly(T) tail. The bead arrays were prepared by resuspending the synthesized beads in 10% dimethylsulfoxide at a concentration of 20,000–50,000 beads per microliter. Then, 10 μl of the bead solution was pipetted into each position on the gasket. The coverslip-gasket filled with bead solution was centrifuged at 750g for at least 30 min at 40 °C until the surface was dry. To extract the spatial barcodes, arrays were sequenced using Bioptechs FCS2 flow cells with an RP-1 peristaltic pump (Rainin) and a modular valve positioner (Hamilton MVP). During sequencing, flow rates between 1 and 3 ml min−1 were used. Imaging was obtained with Nikon Plan Apo 10×/0.45 objective. Sequencing was performed using a sequencing-by-ligation approach. Base calling from the images was performed using the custom MATLAB package PuckCaller (https://github.com/MacoskoLab/PuckCaller).

Slide-seq V2 library generation

The complete protocol for the Slide-seq V2 library preparation can be found at https://www.protocols.io/view/library-generation-using-slide-seqv2-81wgb7631vpk/v1?version_warning=no. Briefly, arrays covered with freshly cut tissue sections were transferred to tubes containing 6× SSC supplemented with RNAase inhibitor (1:20 concentration, catalog no. 30281-2, NxGen, Lucigen) and incubated for 15 min at room temperature. Arrays were then dipped in 1× reverse transcriptase buffer and then transferred to tubes containing the reverse transcription mix (Maxima 1× reverse transcriptase buffer, 1 mM deoxynucleoside triphosphates (dNTPs), 2 U ml−1 RNase inhibitor, 2.5 mM template switch oligonucleotides (catalog no. 339414YCO0076714, QIAGEN) and 10 U ml−1 Maxima H minus reverse transcriptase) for 30 min at room temperature followed by a 90-min incubation at 52 °C. Proteinase K (1:50 concentration) and tissue clearing solution were added to the same tube and incubated at 37 °C for 30 min. Beads were then removed from the glass slide by pipetting up and down a few times and resuspended in TE-TW solution (10 mM Tris, pH 8.0, 1 mM EDTA, 0.01% Tween 20) subjected to two TE-TW washes followed by 2-min centrifugation at 3000g. After removing the supernatant, beads were resuspended in 200 μl of exonuclease I mix (20 μl of 10× ExoI buffer, 10 μl of ExoI, New England Biolabs) and incubated at 37 °C for 50 min. Beads were then washed twice in TE-TW, followed by a 5 min incubation in 0.1 N NaOH at room temperature and another TE-TW wash. Second-strand synthesis was performed on beads by adding 200 μl second-strand mix (Maxima 1× reverse transcription buffer, 1 mM dNTPs, 10 mm dN-SMRT oligonucleotides, 0.125 U ml−1 Klenow fragment) and incubating at 37 °C for 60 min. Next, beads were washed three times in TE-TW before amplification with whole transcriptome amplification PCR (1× Terra Direct PCR mix buffer, 0.25 U ml−1 Terra polymerase, 2 mM TruSeq PCR handle primer and 2 mM SMART PCR primer) with the following conditions: 95 °C 3 min; 4 cycles of 98 °C for 20 s, 65 °C for 45 s, 72 °C for 3 min and 9 cycles of 98 °C for 20 s, 67 °C for 20 s, 72 °C for 3 min and 72 °C for 5 min. The PCR product was cleaned up by 0.6× solid-phase reversible immobilization twice and resuspended to a final volume of 10 μl. Then, 1 μl of the library was quantified on either an Agilent Bioanalyzer High Sensitivity DNA chip or Agilent TapeStation High Sensitivity D500 DNA screenTape. Then, 600 pg of the PCR product was used as input to generate Illumina sequencing libraries by tagmentation with an Illumina Nextera XT kit (catalog no. FC-131-1096). The library was amplified with TruSeq 5 and N700 series barcoded index with the following conditions: 72 °C for 3 min; 95 °C for 30 s; 12 cycles of 95 °C for 10 s, 55 °C for 30 s, 72 °C for 30 s and 72 °C for 5 min. After cleaning up, final libraries were sequenced on a NovaSeq S2 or S4 flowcells with approximately 300 million reads per array for E8.5_Embryo_1 and E9.5 embryos, approximately 200 million reads per array for E8.5_Embryo_2 and E9.0 and approximately 50 million reads per array for WT and Tbx6 KO transversal sections.

Slide-seq data processing and cell state annotation

The sequenced reads were processed using the Slide-seq tools pipeline (https://github.com/MacoskoLab/slideseq-tools) to generate the gene count matrix and match the bead barcode between array and sequenced reads. Most of the downstream analysis was performed in R (v.4.1.0), except the RNA velocity analysis performed in Python (v.3.8.3, v.3.9.0, v.3.10). The gene count matrix and bead spatial coordinates were processed using Seurat (v.3.0.0, v.4.0.2)81. Beads with more than 200 counts and less than 20% mitochondrial gene counts were retained for further analysis. The data from each stage were merged due to minimal batch effect and analyzed together. For the E8.5 replicate and E9.5 data, the top 3,000 highly variable genes were used in FindVariableFeatures and the top 40 principal components from the principal component analysis (PCA) (RunPCA from Seurat). Seurat label transfer was used to obtain cell state annotation, with the functions FindTransferAnchors and TransferData. For the second E8.5 replicate and E9.0 data, we used robust cell type decomposition (RCTD)82 to annotate the cell state because it is more robust to lower UMI counts. Data from the E8.5 stage in our previous study were used as reference data for the Seurat label transfer function and the RCTD function26. Plots were generated with ggplot2 (v.3.1.0).

Reclustering of cell types in the brain

To obtain a better understanding of the different cell types in the brain of E9.5 embryos, we performed de novo clustering from the beads corresponding to the E9.5 stage arrays, further annotating and refining the identities using the Seurat pipeline (resolution = 1), and manually annotated each of the 30 clusters using known marker genes and label transfer results.

Differential expression analysis on brain boundaries

First, we combined marker gene spatial velocities to identify genes expressed within the brain boundaries, Fgf8 for the mid–hindbrain boundary, Foxg1, Barhl2 and Wnt8b for the telencephalon–diencephalon boundary, and Barhl2 and Pax6 for the diencephalon–midbrain boundary. Next, we calculated the distance between each bead to the boundaries and selected beads within a 300-µm distance. Genes whose expression correlated with the distance to the brain boundaries were identified using an edgeR (v.3.34.1) quasi-likelihood model (glmQLFit function)83. The top 40 genes ranked according to FDR were selected for heatmap visualization (Complex heatmap, v.1.99.5) in Fig. 2c.

Identification of spatially variable genes in the developing eye

The beads associated with developing eye were identified in a semi-supervised way. First, the genes correlated with the known marker genes Rax, Vax1 and Six6 were extracted using gene co-expression analysis. These genes were used as input for the PCA and clustering pipeline in Seurat; the top 5 principal components were used. Next, we used dbscan to spatially refine the clustering results and remove a few outliers in the eye cluster. Then, we compared the eye cluster to the forebrain cluster using the FindMarkers function and searched for new marker genes specifically expressed in the developing eye regions.

3D reconstruction and identification of spatially variable genes in 3D

The sc3D reconstruction and associated analysis is described in detail in the Supplementary Information and can also be found at https://github.com/GuignardLab/sc3D.

RNA velocity analysis on Slide-seq data

We adapted the scVelo (v.0.2.4) package84 to analyze the RNA velocity at the spatial axis. Using the tutorial at https://scvelo.readthedocs.io/, exonic and intronic counts from each bead were extracted from the data and used as input. The stochastic model with default parameters was used to compute the velocity of each bead. Next, the velocity vectors were projected to the physical space. For visualization, the velocity vectors were computed according to a 50 × 50 μm grid; the 50 nearest neighbors were selected in each grid to calculate the average velocity. The length of velocity with regard to the speed of the transcriptomic changes was calculated by taking the average length of the velocity vector from the neighbors (n = 50) of a bead. The velocity confidence represents the coherence of the velocity direction. It was calculated by taking the sum of cosine angle between the velocity vector and its neighbors (n = 50). The function rank_velocity_genes was used to identify and rank genes that contribute to the vector field, which means that genes are actively transcribed and have more nascent mRNA as cells differentiate.

Trajectory analysis in the trunk region

Beads annotated as NMPs, somites and neural tube from E8.5 and E9.5 were selected. Next, beads with a prediction score lower than 0.6 were discarded to remove cell mixtures. The remaining beads from E8.5 and E9.5 were integrated using Harmony (v.0.1.0)85 with default parameters; then, UMAP dimensionality reduction (runUMAP) was performed based on an integrated matrix. Next, we used Monocle3 (v.1.0.0) to calculate the pseudotime from the UMAP output, according to the tutorial and using default parameters (https://cole-trapnell-lab.github.io/monocle3/docs/trajectories/). A generalized linear model with quasi-likelihood dispersion estimators from edgeR (v.3.34.1)83 was used to find the genes that correlated with the pseudotime trajectory. Briefly, as with the tutorial86 instructions, we used estimateDisp to estimate the gene-wise negative binomial dispersions, followed by glmQLFit or glmQLFTest to test genes that were significantly correlated with the pseudotime value, which was used as a covariate in the design matrix. Genes with an FDR < 0.01 and a log fold change greater than 0.05 were selected and plotted into a heatmap.

Analysis of the relationship between transcriptional dynamics and spatial distance

After trajectory analysis in the trunk region, the pairwise distance between beads within each cell state were calculated based on their spatial distance and pseudotime difference. This analysis results in the generation of a spatial and pseudotime distance matrices. Next, the spatial distance and pseudotime distance of each pair were compared and plotted.

Identification of spatial differentially expressed genes in the neural tube

The beads assigned to have a neural tube identity in array E9.5_10 were used for this analysis. Using Slingshot::slingshot (v.2.0.0)87, we calculated the principal curves from the beads’ spatial location and ordered the cells along the principal curves as their anterior to posterior distance. The distance of each bead to the convex hull of the neural tube was computed as the dorsal to ventral distance and split into eight equally spaced bins. We used SPARK (v.1.1.1)88 with default parameters to find spatially variable genes. We took the intersection of spatially variable genes from SPARK and the variable genes from Seurat as the input for the spatial module analysis. Dynamic time warping was applied to find genes with coherent spatial patterns that we defined as modules (in the Hox genes analysis). The Dtwclust (v.5.5.6) (https://CRAN.R-project.org/package=dtwclust) package was used for the analysis with the tsclust function (k = 8). The centroid of each module was calculated as their spatial pattern. Next, we used the same edgeR pipeline as the trajectory analysis to find more genes correlated with each spatial pattern. We selected genes with the same threshold and visualized them into a heatmap. The Hox genes were grouped as described in the previous section and the average expression of each module was smoothed using the gam::gam (v.1.2.0) function for visualization. To determine the protein classes that are enriched in the differentially and spatially variable genes, we used PANTHER (v.17.0)89.

Analysis of Tbx6 mutant Slide-seq data

The data from the transversal sections of WT and Tbx6 mutant embryos were merged and analyzed due to the low batch effect. The neural tube and somite regions were extracted based on marker gene expression, cell type labels and tissue morphology. Because of the small sample size, the extracted beads were processed using the Seurat pipeline with different parameters. The parameters used were the following: select 1,000 genes using the FindVariableFeatures function; use the first 20 principal components in the FindNeighbors function; and consider 15 neighbors in the k-nearest neighbor calculation, using of a resolution of 0.8 in the FindClusters function and setting n.neighbors = 20, min.dist = 0.3 in the RunUMAP function. After de novo clustering, we annotated each cluster by its marker genes and excluded cluster 0 because it mainly contained low-quality beads. Clusters 4 and 5 corresponded to the neural crest and neural plate, respectively, ruling out further analysis. We computed the raster spatial density of each cluster using MASS::kde2d (v.7.3–54) and combined them by taking the highest density for each position, which is plotted in Fig. 5b. We mapped the WT and Tbx6 mutant data to the E9.5 trunk dataset as a reference by calculating their mutual nearest neighbor using the fast mutual nearest neighbors correction90 method. For each bead in the Tbx6 mutant data, we selected its ten nearest neighbors in the integrated PCA space as anchors. We computed the variance of the dimensionality reduced matrix from neighbors as a metric of projection uncertainty. The averaged UMAP coordinates of ten nearest neighbors for each bead were projected on the reference UMAP, with size representing the projection uncertainty. The FindAllMarkers function was used to find the marker genes for each cluster. We then applied the FindMarkers function to compare two clusters (somite versus central neural tube and central neural tube versus ectopic neural tube), with the fold change of each gene.

Statistics and reproducibility

All attempts at replicating the observations were successful (indicated below). Preselection of samples was performed, if indicated (below). No samples or data were excluded from the analysis, unless otherwise stated in the Methods. All comparisons (Tbx6 KO) were performed with control samples from the same experiment. Sequencing and downstream processing and analysis were independent of human intervention. Blinding was not relevant because this was not an intervention study; pipelines were executed uniformly across all samples, allowing unbiased analysis. No statistical methods were used to predetermine sample sizes, but our samples are similar to those reported in previous publications29,30,31,32,33,34,35,36,37,38,39.

A two-sided Wilcoxon rank-sum test was used to identify marker genes and Bonferroni correction was used for the multiple comparisons in Fig. 2c, Extended Data Figs. 6d and 10f, Supplementary Tables 5 and 6, and Supplementary Tables 8 and 9. All other tests are described in the figure legends. The FDR and P values used (FDR < 0.01 and log fold change greater than 0.05) are indicated in the legends.

In Extended Data Fig. 4c, elements in the box plots are as follows: middle line, median; box plot limits, upper and lower quartiles; whiskers, s.d.

The Slide-seq experiments involving WT embryos were performed on two whole E8.5, one E9.0 and 13 partial sections from three embryos at the E9.5 stage. Embryos were obtained from at least three independent isolation experiments and staged for somite count (3–5 somite pair stage for E8.5; 10–12 somite pair stage for E9.0; 15–18 somite pair stage for E9.5), which are the representative embryos shown in Extended Data Fig. 1a. The Slide-seq experiment involving WT and Tbx6 KO embryos was performed on one (WT) and two (Tbx6 KO) transversal sections. The Tbx6 KO experiment was performed independently five times (in total) to verify the phenotype, which was reproduced in every embryo across all experiments. Sections were obtained from the posterior part of the trunk (the representative image of the section collected before Slide-seq is shown in Fig. 5a). Whole-mount in situ hybridization was performed once in WT E9.5 stage embryos for the indicated number (n = 3) and showed reproducible results (Extended Data Fig. 2b). The RNA–FISH experiments were performed from two of the independent experiments with the indicated number (n = 3 embryos) and showed reproducible results, with one representative image shown in Extended Data Figs. 5c, 7f, 9f and 10g, and Figs. 4e and 5e.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.