A spatiotemporal atlas of mouse gastrulation and early organogenesis to explore axial patterning and project in vitro models onto in vivo space

Summary

During gastrulation, mouse epiblast cells form the three germ layers that establish the body plan and initiate organogenesis. While single-cell atlases have advanced our understanding of lineage diversification, spatial aspects of differentiation remain poorly defined. Here, we applied spatial transcriptomics to mouse embryos at embryonic (E) E7.25 and E7.5 days and integrated these data with existing E8.5 spatial and E6.5–E9.5 single-cell RNA-seq atlases. This resulted in a spatiotemporal atlas of over 150,000 cells with 82 refined cell-type annotations. The resource enables exploration of gene expression dynamics across anterior-posterior and dorsal-ventral axes, uncovering spatial logic guiding mesodermal fate decisions in the primitive streak. We also developed a computational pipeline to project additional single-cell datasets into this framework for comparative analysis. Freely accessible through an interactive web portal, this atlas offers a valuable tool for the developmental and stem cell biology communities to investigate mouse embryogenesis in a spatial and temporal context.

Graphical abstract.

Introduction

Gastrulation is a critical phase in early embryonic development during which a relatively simple, bi-layered arrangement of cells undergoes rapid proliferation, migration, and differentiation into various mesoderm, endoderm, and ectodermal precursors that establish the body plan and give rise to all major organs.^1,2 Although significant progress has been made in profiling these early stages of murine development at both single-cell and molecular levels,^3–9 the intricate spatial interactions between cells and tissues as specified and patterned during gastrulation remain underexplored. Spatial omics technologies, including those that allow profiling of tissues at the resolution of hundreds of genes, have yet to be applied to the study of these early stages of mouse gastrulation, representing an important gap in our understanding.

We recently generated a comprehensive, time-resolved single-cell RNA sequencing (scRNA-seq) atlas, spanning embryonic (E) 6.5 to E9.5 days of mouse development, encompassing gastrulation and early organogenesis.⁸ This atlas, comprising over 400,000 single-cell transcriptomes, defines 88 transcriptionally distinct cell types across embryonic and extraembryonic tissues. Using a previously published atlas³ spanning E6.5–E8.5, we identified 351 marker genes (Table S1) that distinguish cell types in mouse embryos across this time span, selected for their suitability in spatial transcriptomics via sequential fluorescent in situ hybridization (seqFISH).¹⁰ In our previous work, we examined the expression patterns of these 351 markers in sagittal sections of three E8.5 mouse embryos.¹¹

In this study, we transformed the previous static snapshot spatial data into a temporally resolved spatial atlas of mouse gastrulation and early organogenesis by applying the same seqFISH approach to sagittal sections (five optical sections/embryo) of four additional mouse embryos captured during gastrulation at E7.25 and E7.5.¹² This advanced atlas provides insights into the heterogeneous transcriptional profiles that align with spatial locations throughout mouse development. Additionally, it enables exploration of gene expression dynamics along anterior-posterior (AP) and dorsal-ventral (DV) axes, shedding light on rapid, region-specific mesodermal fate decisions within the primitive streak at E7.25. Finally, we developed a computational pipeline to project additional scRNA-seq datasets into this spatiotemporal atlas. As a proof of concept, we mapped the AP distribution of cell types in gastruloid mouse models,¹³ showcasing the atlas as a powerful tool for benchmarking developmental processes in vitro.

Results

Exploring gene expression patterns during mouse gastrulation using seqFISH

To explore gene expression patterns during murine gastrulation, we performed seqFISH¹⁰ on a total of 20 optical sagittal sections from four embryos (five optical sections/embryo) that were collected at E6.5 and E7.5 (Figures 1A–1E and S1A; Videos S1 and S2). Embryos collected at E6.5 (1 and 2) contained primitive streaks that extended to the distal tip of the egg cylinder, consistent with late-streak stage embryos, which are typically classified as E7.25.¹² After sample preparation, imaging, cell segmentation, and mRNA dot calling, we computed normalized gene expression levels for 351 genes across 14,794 cells (Figure 1E). Before clustering, we performed Seurat rPCA integration,¹⁴ first between embryos at the same developmental stage (e.g., embryos 1 and 2) and then across stages (E7.25 and E7.5). As a quality control step after the first round of integration, we excluded clusters with abnormal RNA/feature counts as well as a cluster from the most distal region of the ectoplacental cone in E7.25 embryos (Figure S1B, red clusters). In total, 17 transcriptionally distinct clusters were identified, which we visualized in a low-dimensional UMAP (uniform manifold approximation and projection) and their spatial context (Figures 2A–2C and S1C–S1G). The cell types were annotated based on marker gene expression patterns and spatial localization (Figures 2B–2F and S1C–S1G; Table S1).

Figure 1. Spatial transcriptomic mapping of early mouse development using seqFISH.

(A and C) 3D illustrations of Theiler stage (TS)10 (A) and TS11 (C) mouse embryos, adapted from eMouse Atlas. Dotted red lines mark the estimated positions of sagittal optical tissue sections shown in (B) and (D). Orientation abbreviations: D, distal; V, ventral; R, right; L, left; A, anterior; P, posterior; PR, proximal.

(B and D) Tile scans of 4-μm sagittal sections from two independently sampled E7.25 embryos (B) and one E7.5 embryo (D), imaged using seqFISH with DAPI nuclear staining (white).

(E) Schematic outline of the seqFISH pipeline adapted from Lohoff et al. 2022.¹¹ See also Table S1.

Figure 2. High-resolution spatial profiling of gene expression during mouse gastrulation.

(A and B) UMAP projections generated from integrated seqFISH expression data. In (A), cells are colored by their embryonic origin, and in (B), by cell type.

(C) Spatial maps of E7.25 and E7.5 embryos, with cells colored according to their cell types. The black dotted line indicates the ExE-Em boundary. BI, blood islands; EPC, ectoplacental cone; HEP, hematoendothelial progenitors.

(D) Representative visualization of normalized log expression counts of selected genes, measured by seqFISH to validate the performance in E7.25 (top) and E7.5 (bottom) embryos.

(E) Stacked bar chart showing the proportion of cell types per seqFISH embryo.

(F) Dot plot displaying the average gene expression for marker genes across different cell types identified in the E7.25 and E7.5 seqFISH embryos. See also Table S1.

Murine gastrulation begins at approximately E6.25 when the primitive streak initiates formation at the prospective posterior of the embryo near the extraembryonic-embryonic (ExE-Em) junction. By E7.25, the primitive streak extends to the distal tip of the egg cylinder, at which point the distal half of the embryo consists of three developing germ layers: an outer visceral endoderm, an inner epiblast, and an emerging middle layer of nascent mesoderm and definitive endoderm. The mesodermal and definitive endoderm layers originate from epiblast cells migrating through the primitive streak. In embryos 1 and 2, we identified an inner epiblast cell layer expressing Dnmt3a, Dnmt3b, Pou5f1, and Slc7a3 (Figures 1H–1K and S1H), as well as a primitive streak population marked by Sp5, T (Brachyury), and Lhx1 (Figure S1H). Cells expressing Lhx1, Sox17, Sfrp1, Foxa2, and Gpc4 visible along the axial midline of embryo 2 were labeled axial mesendoderm, and cells with a similar molecular signature, located in the distal tip of the egg cylinders in embryos 1 and 2, that intermingled with the overlying visceral endoderm, were annotated definitive endoderm (DE) precursors (Figures 2C–2F, S1D, S1F, and S1H; Table S1).

At E7.25, several mesodermal populations were observed, including extraembryonic (ExE) mesoderm cells expressing Dlk1 and mesodermal wings at the anterior ExE-Em boundary in embryo 2, which expressed Foxf1, Hand1, Tbx3, and Msx1 (Figures 2C–2F, S1D, S1F, and S1H). Hematoendothelial progenitors, marked by Runx1, Gata1, and Tal1, were detected in developing yolk sac (YS) blood islands (Figures 1H–1K, S1D, S1F, and S1H). Various Col4a1+ visceral endoderm populations surrounding the egg cylinder were identified, including an ExE endoderm population expressing Clic6 and Soat2 (Figures 2C–2F, S1D, S1F, and S1H). Finally, distinct ExE ectodermal populations displayed unique gene expression patterns, in line with their proximal-distal locations relative to the ExE-Em junction (Figures 2C–2F, S1D, S1F, and S1H).

By E7.5, the primitive streak has elongated to the distal tip of the egg cylinder, and the allantoic bud begins arising from the posterior-most region of the streak and extending toward the chorionic layer. In seqFISH embryo 3, a prominent allantoic bud composed of ExE mesoderm expressing Tbx4 was observed (Figures 2C–2F). Additional extraembryonic tissues in embryos 3 and 4 contained ExE mesoderm, including the visceral YS, amnion, and chorion (Figures 1H, S1C, S1E, and S1G). E7.5 embryos also contained hematoendothelial cell populations (#1 and #2) expressing Runx1, Gata1, Lmo2, and Fgf3, adjacent to ExE endoderm in YS blood islands (Figures 2C–2F, S1C, S1E, S1G, and S1I). In contrast to E7.25 embryos, distal halves of E7.5 embryos consisted of caudal epiblast cells expressing Cdx1, Hoxb1, and Nkx1-2 and reduced expression of E7.25 epiblast markers (Figures 2C–2F, S1C, S1E, S1G, and S1I) as well as rostral ectoderm cells marked by expression of Cxcl12 (Figures 2C–2F and S1C). Altogether, these analyses highlight the utility of seqFISH and our 351-gene probe set to resolve transcriptional diversity associated with spatially arranged cell types at the onset of murine gastrulation.

Generating an integrated spatiotemporal transcriptional atlas covering murine gastrulation and early organogenesis

To fully leverage these new spatial transcriptomic datasets, we integrated them with three additional E8.5 seqFISH embryo datasets¹¹ (Figures S2A–S2D) and applied StabMap¹⁵ and reducedMNN¹⁶ to align all quality-controlled seqFISH cells (Figures S1B and S2A) with a time-resolved scRNA-seq (scRNA) atlas⁸ spanning E6.5 to E9.5 of murine embryogenesis (Figures 3A–3C). After integration, we performed clustering, UMAP visualization, cell-type label transfer, and imputed missing gene expression patterns in the seqFISH embryos. Clusters with poor representation across both datasets, referred to as “poor joint clusters,” were identified and filtered prior to further analyses (Figure S3A). For instance, YS endothelium and YS mesothelium clusters were composed almost entirely (>98%) of scRNA cells, in line with their absence from seqFISH samples. Additionally, we identified seqFISH cells that mis-aligned with expected embryonic stages (~12%) (Figure S3A), which may represent interesting cell populations that are either further or less differentiated (see metadata “poor stage mapping”).

Figure 3. Integration of spatial and single-cell transcriptomics to construct a spatiotemporal atlas of mouse gastrulation and early organogenesis.

(A) Overview of the single-cell transcriptomic datasets (seqFISH in blue and scRNA-seq in red) integrated to create a spatiotemporal transcriptional atlas of mouse gastrulation and early organogenesis (E6.5–E9.5). Gray bars represent the total number of cells per dataset.

(B and C) UMAP projection of cells from the extended gastrulation atlas (scRNA) and seqFISH into a unified reduced-dimensional space. Cells are colored by embryonic stage (A) and refined cell type (B). See also Figures S3–S6 and Table S1.

Cells were initially annotated by assigning joint clusters with the highly voted scRNA labels. These annotations were assessed and, in some cases, further refined by analyzing marker gene expression patterns and visually inspecting the spatial distribution of cell types by seqFISH embryos (see STAR Methods for details). Altogether, our approach generated a comprehensive spatiotemporal transcriptional atlas of mouse gastrulation and early organogenesis, encompassing 82 refined cell-type annotations across more than 150,000 cells (with good alignment across both seqFISH and scRNA datasets) spanning developmental stages E6.5 to E9.5 (Figures 3B and 3C). This new resource enables exploration of cellular diversity in both spatial and temporal contexts during mouse embryogenesis (Figures 4A and S3–S5), providing new single-cell resolved spatial maps during gastrulation at E7.25–E7.5 and an increase in cell-type resolution at early stages of organogenesis from ~30 to 55 cell types at E8.5.

Figure 4. Exploring cell type distributions in space across time points during mouse embryogenesis.

(A) Spatial maps of E7.25, E7.5, and E8.5 embryos, organized by germ layer, with cells colored by refined cell type labels.

(B) Heatmap showing the proportion of cell types across embryonic stages in the extended gastrulation atlas (scRNA) and seqFISH samples. Proportions are row-scaled, and the columns are hierarchically clustered. ExE, extraembryonic; PGC, primordial germ cell; LPM, lateral plate mesoderm; FHF, first heart field; SHF, second heart field; EMP, erythron-myeloid progenitors; MEP, myeloid-erythroid progenitor; NMP, neuro-mesodermal progenitor; Chorioall.-der., chorioallantoic-derived. See also Figures S3–S6 and Table S1.

A comparison of the cellular composition of seqFISH embryos with scRNA embryo pools from different stages revealed that E7.25 seqFISH embryos closely align with E6.5–E7.25 scRNA samples, while E7.5 and E8.5 seqFISH embryos correspond to E7.5 and E8.5–E9.5, respectively (Figure 4B). It is worth noting that our seqFISH datasets do not comprehensively capture the full medial-lateral axis of the embryos. As a result, the observed cell type proportions reflect only a subset of the total embryonic composition. The cell counts per cell type across all seqFISH embryos (embryos 1–7) are provided in Figures S3C–S3F. Additionally, the spatial distribution of the refined cell types with key marker gene expression patterns from both seqFISH and scRNA cells is presented by embryonic stage and germ layer in Figures S4–S6, Video S3; Tables S1. The integrated dataset is accessible through a user-friendly web portal (see data and code availability section), allowing researchers to perform virtual dissections, investigate both raw and imputed seqFISH gene expression patterns, and identify differentially expressed genes in this context.

Accounting for spatial location enables refined cell-type annotation

Consistent with our analyses in Figure 2, E7.25 and E7.5 seqFISH embryos contained various epiblast-derived cell populations (E6.5–E7.25 scRNA samples), including cells labeled primitive streak, nascent mesoderm (Mesp1, Lefty2, Mixl1, and Eomes),^17,18 hematoendothelial progenitors (Kdr, Tal1, and Etv2),¹⁹ and Gata1+ blood progenitors (Figures 3B, 3C, 4A, 4B, and S4A–S4E, Video S5-2; Table S1). An anterior primitive streak population (Cer1, Sox17, Foxa2, and Lefty1)^1,20–22 was positioned in the distal domain of the primitive streak region and extended into the axial midline of embryos 1/2 (Figures 3B, 3C, 4A, 4B, and S4A–S4C; Video S3; Table S1). Notably, this population includes both anterior primitive cells as well as anterior primitive streak derivatives in seqFISH embryos, including cell populations labeled DE precursors and axial mesendoderm in Figure 1H. Therefore, the cell population labeled anterior primitive streak in the extended scRNA mouse gastrulation atlas⁸ likely contains anterior primitive streak cells and anterior primitive streak derivatives.

E7.25 and E7.5 seqFISH embryos also contained ectodermal cells (Ptn and Cxcl12), ExE mesodermal cells (Foxf1 and Bmp4),^23,24 and primordial germ cells (Figures 4A and S4A–S4E; Table S1; Video S3). Primitive endoderm-derived populations, including visceral endoderm and ExE endoderm, were identified on the exteriors of egg cylinders (Figure 4A). Proximal and distal ExE ectodermal cells, assigned based on their spatial orientation relative to the ExE-Em junction (Figure 2C), were found in extraembryonic regions (Figures 4A and S4A–S4C, Video S3). Rostral ectoderm and caudal epiblast (Cdx1 and Nkx1-2) cells were present by E7.5 (Figures 4A and S4A–S4C; Video S3). Notably, the higher proportion of distal ExE ectoderm cells in E7.25 seqFISH embryos, compared to scRNA-seq samples, is likely due to the removal of the “sticky” ectoplacental cone prior to sequencing scRNA samples (Figure 4B). Finally, comparison of imputed gene expression patterns in early-stage seqFISH embryos with whole-mount in situ hybridization images from TS10/11 embryos in the eMouse Atlas²⁵ highlights the accuracy and reliability of imputed gene expression patterns (Figures S4F and S4G).

Previously, we assigned ~ 30 cell-type labels to E8.5 seqFISH embryos¹¹ using a gastrulation atlas covering stages E6.5–E8.5³. In this study, aligning E8.5 seqFISH embryos with the expanded gastrulation atlas⁸ (E6.5–E9.5) substantially increased the number of annotated cell types at E8.5 to 55 (Figures 4B and S2B). E8.5 seqFISH embryos contained first and second heart field cardiomyocytes (Nkx2-5, Isl1),²⁶ as well as endocardial cells #1 and #2 (Hand2, Vwf, Gata5, Pecam1, and Sox17),^27,28 cardiopharyngeal progenitors (Smarcd3),²⁹ and epicardium (Wt1 and Tcf21),^30,31 which localize to distinctive heart regions in embryos 5–7 (Figures S5A–S5F; Table S1; Video S4). The allantois endothelium (Dlk1 and Bambi)^3,8 was positioned in the posterior regions of embryos 5 and 7, adjacent to the allantois (Tbx4 and Hoxa10)^32,33 cell population in embryo 7 (Figures S5A–S5F; Video S4). Embryo proper endothelium #1 and #2, which was depleted in cardiac tissues, contributed to intersomitic and cranial vessels and extended along the trunk (Figures S5D–S5F; Video S4).

Somitic tissues, including the sclerotome (Pax1),³⁴ dermomyotome (Pax3),³⁵ and endotome, were also observed in the trunk, while somitic and presomitic mesoderm (Tbx6, T, Hes7, and Dll1/3)³⁶ localized to the posterior regions, corresponding to the differentiation front, where cells in the presomitic mesoderm transition from a proliferative state to form somites (Figures S5G–S5I; Video S4). The gut tube was composed of pharyngeal endoderm (Nkx2-3 and Pitx1),³⁷ thyroid primordium, foregut (Foxa1 and Gata6),^38,39 and hindgut (Cdx2 and Hoxa7),^39,40 distributed in distinct patterns along the anterior-to-posterior axes of E8.5 embryos, while notochord (Noto and Nog)⁴¹ cells were located dorsally to the gut tube in embryos 5 and 7 (Figures S6A–S6C; Table S1; Video S5). Additionally, various neural tube and ectodermal cell types were identified in their expected anatomical locations, including spinal cord progenitors (Hoxb8),⁴² the optic vesicle (Lhx2, Six3, and Otx2),^43–45 placodal ectoderm (Six3 and Pitx1), and the otic placode (Figures S6D–S6I; Video S5).

Notably, our spatiotemporal analysis provided additional insight into the annotations of certain cell types in the scRNA atlas, including ExE ectoderm and lateral plate mesoderm (LPM) (Figure S3B). Spatially distinct ExE ectodermal cell types emerge during mouse embryogenesis to form functionally specialized placental lineages. Our spatiotemporal atlas allowed us to identify distal ExE ectoderm cells, expressing Ascl2 and Gata2, located in the ectoplacental cone region at E7.25, which likely include trophoblast stem/progenitor cells⁴⁶ (Figures 4A and S4A–S4D; Table S1; Video S5). In contrast, proximal ExE ectoderm cells aligned with chorion progenitors, marked by Elf5 and Perp, found adjacent to the epiblast at E7.25 and in the chorion layer at E7.5 (Figures 4A and S4A–S4E; Video S3).

LPM gives rise to an array of critical cell types during murine embryogenesis, including those that form the cardiovascular system, such as the heart and blood vessels.⁴⁷ We relabeled a subset of Hand1- and Foxf1-expressing LPM^24,48 cells “ExE mesoderm and Anterior LPM,” as they localized to extraembryonic tissues at E7.25/E7.5 and were also present in the medial trunk of seqFISH embryos at E8.5, highlighting transcriptional similarities between ExE mesoderm and developing intraembryonic medial LPM (Figures S4A–S4C and S5A–S5C; Table S1). In the scRNA atlas, the relabeled “ExE mesoderm and Anterior LPM” cell population predominated at the earlier stages (E7.0–E8.25) and contained cells originally labeled LPM, mesenchyme, and nascent mesoderm. Moreover, a distinctive subset of LPM cells, marked by Cdx4⁴⁹ and Pitx1, was localized to the posterior of embryos 5–7 and identified in the allantoic bud of embryo 4, prompting relabeling as “ExE mesoderm and Posterior LPM” (Figures S5A–S5C; Video S4). These “ExE mesoderm and Posterior LPM” cells predominated in the scRNA datasets at later stages (E7.75–E9.25) and contained cells originally labeled LPM and allantois.

Altogether, our integrated spatiotemporal atlas provides a high-resolution framework for investigating the spatial and transcriptomic organization of embryonic cell populations, serving as a valuable resource for the research community. To support cell-type assessment, refinement, and characterization, all code and metadata are readily accessible (https://github.com/ltgharland/Spatiotemporal-Atlas-of-Mouse-Gastrulation). Table S1 details marker gene information, while Videos S3–S5 illustrate the spatial localization of cell types across seqFISH embryos, alongside imputed gene expression patterns and high-resolution clustering. We anticipate these resources will encourage continued exploration and refinement, ensuring that the spatiotemporal atlas remains a dynamic and valuable tool for developmental biology research.

Spatiotemporal analysis reveals dynamic gene expression changes along the AP and DV axes

During embryogenesis, gene expression patterns coordinate the establishment of embryonic axes, which orchestrate spatiotemporal signals for organ development and delineate the body plan. To investigate gene expression changes along these axes, we assigned normalized AP and DV coordinates to seqFISH embryos (Figures 5A and S7A–S7G). Next, using our integrated atlas, we imputed AP and DV values into the scRNA reference atlas (Figures 5B and S7H). Cells from E9.25-E9.5 scRNA samples, isolated from the anterior, medial, or posterior regions of sub-dissected embryos, were accurately assigned imputed AP values consistent with their known anatomical origins⁸ (Figure 5C, see Figure 3A for sub-dissection). This approach provided finer spatial resolution compared to sub-dissection alone, as evidenced by AP gradients within medial and posterior sections (Figure 5C). Moreover, it enabled the estimation of AP and DV axes in earlier-stage embryos that had not been sub-dissected (pooled, Figure 5C). The average AP and DV positions of all cell types in seqFISH embryos are shown in Figure S7I.

Figure 5. Assigning and imputing AP positional identity across spatial and suspension single-cell transcriptomic datasets during mouse gastrulation.

(A) Spatial maps of E7.25, E7.5, and E8.5 embryos, with cells colored by normalized AP values. Gray cells represent ExE regions or low-quality cells filtered during QC. Orientation abbreviations: D, distal; V, ventral; R, right; L, left; A, anterior; P, posterior; PR, proximal.

(B) Joint projection of seqFISH and extended gastrulation atlas (scRNA) cells, colored by normalized AP values and imputed normalized AP values, respectively (red-blue).

(C) Cells from anatomical sub-dissections of the extended gastrulation atlas (scRNA), colored by imputed AP values. Anterior, medial, and posterior sections are from E9.25–9.5 embryos.

(D) Binned imputed expression of Hoxa and Hoxb genes along the AP axis in E7.25, E7.5, and E8.5 seqFISH embryos. Curves represent locally weighted LOESS (locally estimated scatter plot smoothing) regression fits across AP bins; shading indicates 95% confidence intervals around the fitted mean. See also Figure S7.

Incorporating spatial coordinate information into the integrated spatiotemporal atlas provides the developmental biology community with a powerful tool to explore gene expression dynamics along embryonic axes in cell types and tissues of interest at specific time points. For instance, we analyzed the imputed expression patterns of Hox genes (key developmental regulators with well-established AP-biased expression) along seqFISH embryos from different embryonic stages. Predicted Hox expression patterns (Figure 5D) align with previous studies,⁵⁰ highlighting the utility of our approach.

Next, we investigated gene expression changes in a tissue context along the primitive streak region of embryo 2, focusing on three key populations—primitive streak cells, nascent mesoderm, and the adjacent endoderm layer—spanning positions 0–50 along the AP axis (Figures 6A and 6B). This region was selected because mesodermal cells migrating through different AP regions along the primitive streak are known to acquire distinct fates.^51–53 Previous studies have explored bulk RNA-expression patterns in embryonic slices along the AP axis.^54,55 However, single-cell spatially resolved experiments during this initial process of cell fate diversification have not been performed.

Figure 6. Regionalized gene expression along the AP axis highlights rapid fate allocation in the primitive streak during gastrulation.

(A) Spatial plot highlighting primitive streak (PS), nascent mesoderm (NM), and endoderm (END) cells in embryo 2 (red-blue, AP 0–50), with positions indicated by dotted black lines.

(B) Ridge plots showing density distribution of major cell types along the AP axis in seqFISH embryo 2. ExE, extraembryonic; Visc., visceral; PS, primitive streak; APS, anterior primitive streak; NM, nascent mesoderm; Epi., epiblast.

(C) Heatmap of smoothed, scaled predicted gene expression profiles (using Tradeseq) along the AP axis in primitive streak, nascent mesoderm, and endoderm cells, ordered by AP values and grouped by cell type (END = endoderm, PS = primitive streak, NM = nascent mesoderm). Significant associations identified by Tradeseq⁷⁴ (p < 0.01, mean log fold-change [logFC] > 0.25). Colored rectangles to the right of the heatmap indicate gene clusters that were identified using hierarchical clustering. Enriched GO terms were identified using Metascape.⁷⁵

(D) Visualization of imputed normalized log expression counts of selected genes, in the primitive streak region indicated in (c). See also Table S2.

Our analysis of imputed gene expression reveals that numerous genes are dynamically expressed along the AP axis in a germ layer-specific manner within the primitive streak region of embryo 2 (Figure 6C; Table S2). Genes associated with the transport of vitamins and organic anions were enriched in posterior endodermal (Posterior END) cells, while genes related to in utero embryonic development, such as Tead1, Junb/d, and Gjb3, were expressed in the anterior endoderm (Anterior END) (Figure 6C). Primitive streak cells expressed pluripotency genes, including Nanog and Pou5f1, that were downregulated in the nascent mesoderm, which displayed limited expression changes along the AP axis. By contrast, the nascent mesoderm expressed genes in the anterior related to paraxial mesoderm development and Notch signaling, including Tbx6, Mesp2, and Dll1/3, while posterior nascent mesoderm was enriched for genes associated with Wnt signaling and vasculature development including Etv2, Hand1, Foxf1, Isl1, Kdr, Msx1/2, Tbx3, and Wnt2 (Figures 6C and 6D).

These findings revealed the nature of rapidly established transcriptional programs along the AP axis in nascent mesodermal cells, distinguishing between posterior ExE mesoderm and anterior paraxial mesoderm, in line with fate biases observed in previous fate mapping studies within the primitive streak. To validate the accuracy of these predicted gene expression pattern changes from our imputed gene expression counts in seqFISH embryo 2 (Figure 6C), we explored the expression patterns of the same gene set in the tissue sections of E7.25 embryos profiled via GEO-seq (geographical position sequencing).⁵⁴ This comparison confirms the reliability of our predictions in capturing spatial gene expression changes (Figure S7J). Altogether, these findings highlight the power of our integrated spatiotemporal atlas to explore gene expression dynamics at single-cell resolution across embryonic axes.

Projecting models of mouse gastrulation into in vivo spatial and temporal contexts

Determining which aspects of embryonic development in vitro models faithfully recapitulate is crucial for accurately interpreting experimental outcomes and improving differentiation protocols in stem cell biology. A key strength of our spatio-temporal atlas is its ability to integrate and project external single-cell transcriptional datasets into an in vivo spatiotemporal context. To facilitate this, we have developed a bioinformatics pipeline that enables researchers to explore their datasets within the spatial framework of mouse development. As a proof of concept, we projected scRNA-seq datasets from gastruloids, embryonic stem cell-derived models of murine gastrulation, into our atlas (Figure 7A).

Figure 7. A bioinformatics pipeline to project additional scRNA-seq datasets into a spatial and temporal context during murine embryogenesis.

(A) Schematic of the bioinformatics pipeline used to project additional scRNA-seq datasets from day 4 (96 h)–7 (168 h) gastruloids, into the spatiotemporal atlas of murine gastrulation and early organogenesis.

(B) Spatial maps showing gastruloid cells projected onto the seqFISH embryos in gray, with gastruloid cells colored according to their spatiotemporal cell type annotations.

(C) Gastruloid cells projected onto the spatiotemporal atlas, with each panel colored by different factors. Top row: stage; gastruloid annotations from Rossi et al., 2021; refined cell type annotations from the spatiotemporal atlas (current study). Bottom row: spatial cluster quality control (QC); imputed anterior-posterior (AP); imputed dorsal-ventral (DV) values. Please see STAR Methods section for further information about spatial QC.

(D and E): Ridge plots displaying the density distributions of selected cell types from different germ layers (D) or somitic cell types (E) along the AP axis in the scRNA atlas (red) and gastruloids (blue). Density difference plots are shown below, Δ(g-s), which highlight the regions of increased (red) or decreased (blue) gastruloid density relative to the scRNA atlas in (D). A Kolmogorov-Smirnov (KS) test was performed to assess differences in density distributions. To quantify cell type composition similarities along the AP axes per germ layer, binned Jaccard indices (mean ± standard deviation, determined by bootstrapping) are displayed above the ridge plots in 10 AP bins.

(F) Heatmap of smoothed, scaled gene expression profiles along the imputed AP axis for the somitic cell types shown in (E) for gastruloids (left) and the scRNA-seq atlas (right). Significant associations were identified using Tradeseq association test (p < 0.01, mean log fold-change > 0.6). See also Figure S7 and Table S2.

Gastruloids replicate key aspects of the AP axis development in the absence of extraembryonic tissues.^56–60 To assess their transcriptional and spatial alignment with in vivo development, we projected scRNA-seq datasets from day 4 (96 h) to 7 (168 h) gastruloids, grown using an enhanced induction protocol that promotes cardiovascular development¹³ (Figure 7A). In line with previous findings,^13,61 our projections highlight the formation of first and second heart field cardiovascular progenitors and endocardial cells as well as hematoendothelial progenitors and YS endothelial cells (Figures 7B and 7C). Beyond cardiovascular progenitors, gastruloids also developed an array of mesodermal, ectodermal, and endodermal cell types (Figures 7B and 7C). These included a substantial amount of paraxial mesoderm-derived tissues, such as anterior and posterior somites, and other key cell types typically formed during mouse embryogenesis between E6.5 and E9.5 (Figures 7B and 7C).

Since the spatiotemporal atlas is constructed from seqFISH cells collected from embryonic tissue sections at E7.25, E7.5, and E8.5, without full coverage of the medio-lateral axis, it represents only a subset of the cell populations captured in the scRNA-seq atlas, which includes whole embryos and yolk sacs spanning E6.5 to E9.5 (Figure 3A). Certain regions of the spatiotemporal atlas are therefore dominated by scRNA-seq cells alone, lacking seqFISH reference points (see red cells in poor joint clusters, Figure S3A). To make users aware of this, our projection pipeline includes a spatial QC (quality control) step that flags query cells aligning with regions where seqFISH data are absent or underrepresented (see STAR Methods, spatial QC in the bioinformatics pipeline, for details). For example, a subset of gastruloid cells, including YS endothelial cells and SHF cardiomyocytes, align with scRNA-seq-dominated regions and therefore fail spatial QC as they lack seqFISH reference points, making their spatial positioning less reliable (Figure 7C, “spatial cluster QC”).

Next, we assigned imputed AP and DV values to the gastruloid cells (Figure 7C). Comparing cell densities of gastruloid versus scRNA-seq cells along imputed AP axes, stratified by germ layer, revealed that this cardiovascular gastruloid model does not fully capture the complete range of AP development in mesodermal and ectodermal germ layers (Figure 7D). More specifically, a subset of the most posterior mesodermal cell types, including posterior LPM and allantois (Figures 7C and S7K), with the most anterior ectodermal cell types, were underrepresented, highlighting a need for further protocol optimization to better induce these tissues. Nonetheless, these cardiovascular gastruloids successfully generated a diverse array of somitic cell types (Figures 7C–7E).

Taking advantage of the true single-cell resolution of our approach, we assessed the predicted AP distribution of somitic cell types and gene expression patterns in gastruloids, induced using the modified cardiovascular protocol, and compared them to in vivo somitogenesis using our scRNA-seq datasets (Figures 7E and 7F). Previous TOMO-seq (tomography sequencing) studies,⁵⁷ which assess gene expression patterns in tissue slices but not at single-cell resolution, have revealed that the dynamic transcriptional patterns along the AP axis of individual 120-h gastruloids, generated using standard somitogenesis-inducing protocols, appear to align with those observed in E8.5 mouse embryos. Our analysis similarly revealed that cardiovascular gastruloids generate the full spectrum of somitic cell types typically found along the AP axis in developing embryos (Figures 7C–7E). Moreover, a suite of genes exhibited similar expression gradients along imputed AP axes in both gastruloids and embryonic somitic tissues (Figure 7F; Table S2). Many genes showed expression patterns consistent with those previously observed in TOMO-seq-analyzed gastruloids,⁵⁷ thus (i) validating our approach, (ii) bench-marking the effectiveness of gastruloids in modeling key aspects of early development, and (iii) introducing a broadly relevant strategy for assessing the cellular output of in vitro models at true single-cell spatiotemporal resolution.

Discussion

Here, we present an integrated spatiotemporal atlas of mouse gastrulation and early organogenesis, introducing broadly applicable computational strategies to merge spatial and suspension transcriptomic datasets, leveraging the strengths of both technologies. By extracting per-cell coordinates along the AP and DV axes from spatial transcriptomic data, we accurately predicted the AP and DV positions of cells profiled using dissociation-based single-cell technologies. Additionally, we enriched the spatial transcriptomics dataset by imputing whole-transcriptome gene expression at the single-cell level, creating a comprehensive transcriptional resource for studying mouse embryogenesis within a spatiotemporal framework. To ensure broad accessibility, we developed a user-friendly web portal (http://shiny.maths.usyd.edu.au/SpatiotemporalMouseAtlas/) for providing researchers with an intuitive platform to explore, visualize, and analyze these complex datasets and derive new hypotheses to advance our understanding of fundamental biological processes as well as inform new stem cell differentiation strategies for disease modeling and cellular therapy applications.

Our analysis of gene expression changes in the primitive streak of a late-streak stage embryo underscores the utility of assigning axial coordinates. Previous studies that examined genome-wide expression along the AP axis in mini-pools derived from tissue slices of mouse gastrulas at similar stages identified gene expression domains associated with mesodermal fate patterning.^54,55 The single-cell resolution of our dataset reveals that these significant transcriptional changes along the AP axis predominantly occur in migratory nascent mesodermal wings rather than in cells immediately exiting the primitive streak. This precise spatial resolution offers fresh insights into the dynamic shifts in gene expression during gastrulation and highlights the potential of our atlas to explore axial patterning in an unbiased manner, providing a powerful tool for understanding gene regulation during embryogenesis.

A recent study⁶² employed whole-mount transcriptomics using cycleHCR (hybridization chain reaction) to measure the expression of 254 genes in a mid-streak stage mouse gastrula (E6.5) in a three-dimensional (3D) context. Consistent with our findings for our earliest time point, this single snapshot study identified nine transcriptionally distinct cell clusters, including epiblast, primitive streak, nascent mesoderm, ExE mesoderm, visceral endoderm, and extraembryonic ectoderm, mapped precisely to their spatial locations. In addition, Qiu et al. presented spatial transcriptomic datasets of whole E9.5 and E11.5 mouse embryos, accompanied by a computational framework for analyzing and exploring 3D spatiotemporal data.⁶³ While our study focuses on earlier developmental stages (E7.25, E7.5, and E8.5), integrating these complementary datasets in future computational efforts will extend the temporal window of investigation, offering a more complete spatiotemporal roadmap of early embryogenesis. Moreover, as additional spatial transcriptomic datasets emerge, particularly those covering key stages of mouse gastrulation that only measure hundreds of genes, our computational methods can be applied to align these datasets with our spatiotemporal atlas, enabling transcriptome-wide gene expression imputation and refined cell type characterization.

Mapping cardiovascular gastruloids¹³ onto the spatiotemporal atlas using our data integration approach provides a qualitative benchmark of in vitro murine models against the normal spatiotemporal context of mouse development. Our analysis reveals that while these gastruloids effectively recapitulate key aspects of embryogenesis, including anterior-posterior patterning of somitic tissues, they lack specific posterior LPM cell types. Additional datasets from alternative murine culture systems or in vivo perturbations, such as drug-treated, mutant, embryoid body,⁶⁴ or chimera cells, can similarly be mapped and benchmarked using the atlas, although special care needs to be taken to avoid biases due to the mapping process.⁶⁵ Moreover, our group and others are actively developing computational pipelines to cross-examine the early developmental processes across species.^66,67 Given the sparsity of human spatial reference data during gastrulation, these methodologies will be especially valuable in extending benchmarking practices to emerging human pluripotent stem cell-derived models.^68–70

Looking ahead, advancing methods to integrate comprehensive spatiotemporal transcriptional atlases with time-resolved 3D embryo models, such as those generated through light-sheet imaging,⁷¹ will pave the way for creating 4D models that predict transcriptome-wide gene expression changes throughout development. Such future approaches will enable detailed exploration of gene expression dynamics across differentiating cellular lineages, providing essential boundary conditions for performing trajectory inference with suspension datasets and offering deeper insights into cell-cell communication during developmental processes. This work will likely contribute to creating a “virtual tissue” of mouse gastrulation and early organogenesis, offering an unprecedented tool for modeling and understanding embryonic development at a new level of precision.

Limitations of the study

Despite the advances we describe in generating this atlas, some further avenues must be explored. While our atlas contains spatial coordinates and whole-transcriptome level gene expression, a consistent “average” coordinate system is yet to be built over both spatial and temporal axes. Further work could leverage recent methodological advances in mapping cells between spatial and temporal coordinates that harness optimal transport, such as moscot⁷² or DeST-OT (developmental spatiotemporal optimal transport).⁷³ However, using these methods would only be applicable for spatially resolved cells rather than leveraging the entire set of cells in the spatiotemporal mouse gastrulation atlas. Moreover, it is important to consider that for any abundance-based downstream analysis, cells in the spatial data may be subject to enrichment or depletion biases due to sample selection. Here, we ameliorated this potential issue by comparing cell densities between gastruloids and scRNA-seq-resolved cells rather than with spatial transcriptomic-resolved cells, though it may be that further biases remain across the entire spatio-temporal atlas, in which case further effort is required to determine the effective matched controls for comparison against a reference atlas.⁶⁵ Finally, while we provide refined AP and DV coordinate predictions for scRNA-seq-resolved cells using spatial integration, these annotations are based on computational inference. Without further direct experimental validation, their spatial accuracy remains putative and should be interpreted with appropriate caution.

It is important to note that our seqFISH datasets do not fully capture the complete medio-lateral extent of each embryo. Consequently, the observed cell types represent only a subset of the total cellular composition at the developmental stages examined. Moreover, while imputed gene expression patterns—particularly in earlier stage embryos such as E7.25—show strong concordance with established in vivo expression domains, discrepancies are more apparent at later stages. These may arise from technical challenges in spatial integration and imputation, embryo-to-embryo variability, or inherent limitations in spatial transcriptomic resolution and reconstruction. These factors should be considered when interpreting individual gene expression patterns within the atlas.

Resource Availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Shila Ghazanfar (shila.ghazanfar@sydney.edu.au).

Materials availability

Reagents generated in this study are available upon request from the lead contact, and other materials in this study are commercially available.

Star★Methods

Key Resources Table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Anti-E-cadherin (clone 36)	BD Biosciences	Ref. 610181; RRID: AB_397580
Anti-N-cadherin (13A9)	Cell Signaling Technology	Ref. 14215; RRID: AB_2798427
Anti-mouse IgG secondary  antibody (oligo-conjugated)	Custom	See Lohoff et al.11
Anti-pan-cadherin	Abcam	Ref: ab22744; RRID: AB_447300
Anti-B-catenin (15B8)	Abcam	Ref: ab6301; RRID: AB_305406
Chemicals, peptides, and recombinant proteins
4% Paraformaldehyde (PFA)	Thermo Scientific	Ref. 28908
Acryloyl-X SE	Thermo Fisher	Ref: A20770
Alexa Fluor 647	Thermo Fisher	Ref: A20006
Poly-D-Lysine	Sigma	Ref: P6407
Proteinase K	NEB	Ref: P8107S
SDS
Salmon Sperm DNA	Thermo Fisher	Ref: AM9680
Uracil-specific excision reagent	NEB	Ref: N5505S
DAPI	Sigma	Ref: D8417
Dextran Sulfate	Sigma	Ref: D8906
Formamide	Sigma	Ref: F9027
Bovine Serum Albumin (BSA)	Thermo Fisher	Ref: AM2616
APS	Sigma	Ref: A3078
TEMED	Sigma	Ref: T7024
Tris-HCl	Invitrogen	Ref: AM9856
EDTA	Invitrogen	Ref. 15575020
NaCl	Sigma	Ref: S5150
Trolox	Sigma	Ref. 238813
D-glucose	Sigma	Ref: G7528
Catalase	Sigma	Ref: C3155
Glucose oxidase	Sigma	Ref: G2133
Ethylene carbonate	Sigma	Ref: E26258
Dextran sulfate (EC buffer)	Sigma	Ref: D4911
Acryoloyl-X succinimidyl ester	Thermo Fisher	Ref: A20770
M2 medium	Sigma Aldrich	Ref. 7167
PBS	Invitrogen	Ref: AM9624
Rnase-free sucrose	Sigma Aldrich	Ref. 84097
Tissue base mold	Sakura	Ref. 4162
OCT compound	Sakura	Ref. 4583
Isopropanol	VWR	Ref. 20842
Critical commercial assays
QIAquick PCR Purification Kit	Qiagen	Ref. 28104
Deposited data
seqFISH Images	This paper	https://zenodo.org/records/13977985.
Processed seqFISH data	This paper	http://shiny.maths.usyd.edu.au/SpatiotemporalMouseAtlas/
Code repository	This paper	https://github.com/ltgharland/Spatiotemporal-Atlas-of-Mouse-Gastrulation
Experimental models: Organisms/strains
C57BL/6J Mice	Charles River
Software and algorithms
Napari	Python	(https://zenodo.org/records/8115575)
U-Net	Ronneberger et al.76	http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net.
StarDist	Schmidt et al.77
CellPose	Stringer et al.78	www.github.com/mouseland/cellpose
Ilastik V1.4.0	Berg et al.	https://www.ilastik.org/
scran V1.26.2
princurve	Hastie & Stuetzle79
Seurat V4.2.0	Hao et al.14	https://github.com/satijalab/seurat
StabMAP V0.1.8	Ghazanfar et al.15	https://marionilab.github.io/StabMap/index.html
batchelor V1.12.3	Haghverdi et al.16	https://bioconductor.org/packages/release/bioc/html/batchelor.html
Tradeseq V1.10.0	Van den Berge et al.74	https://www.bioconductor.org/packages/release/bioc/html/tradeSeq.html
ComplexHeatmap V2.15.3	Gu et al.80	http://www.bioconductor.org/packages/devel/bioc/html/ComplexHeatmap.html
Metascape	Zhou et al.75	https://metascape.org/gp/index.html
ImageJ
scater V1.24.0	Bioconductor	https://www.bioconductor.org/packages/release/bioc/html/scater.html
SingleCellExperiment V 1.20.0	Bioconductor	https://www.bioconductor.org/packages/release/bioc/html/SingleCellExperiment.Html
data.table v 1.14.8	CRAN	https://cran.r-project.org/web/packages/data.table/index.html
dplyr v 1.1.0	CRAN	https://cran.r-project.org/web/packages/dplyr/index.html
ggplot2 V3.4.1	CRAN	https://cran.r-project.org/web/packages/ggplot2/index.html
ggridges V0.5.4	CRAN	https://cran.r-project.org/web/packages/ggridges/index.html
R	N/A	https://www.r-project.org

Experimental Model and Study Participant Details

Mice and tissue preparation

Experiments were performed in accordance with EU guidelines for the care and use of laboratory animals, and under the authority of appropriate UK governmental legislation. Eight-to 12-week-old wildtype C57BL/6J (Charles River) were used and all mice used in this project were housed under a 12-h light/12-h dark cycle, with constant access to food and water. No sex selection of the used embryos was performed. Tissue sections were prepared and sectioned ready for seqFISH imaging as previously described,¹¹ further details are provided in the method details section below.

Method Details

Library and probe design

We used the same library selection as previously described,¹¹ using a total of 351 seqFISH barcoded genes (Table S1) and 36 additional non-barcoded sequential smFISH imaging. We used the same primary probe design, readout probe design, primary probe library construction, readout probe synthesis, encoding strategy, and coverslip functionalization as described¹¹ descibred briefly in the next sections.

Library and probe design – Gene panel selection

To construct a spatial transcriptomic gene panel tailored to mouse gastrulation, we selected marker genes that discriminated between cell types in a single-cell RNA-seq atlas (Pijuan-Sala et al., 2019). Differentially expressed genes were identified using the find-Markers() function from the scran R package with pval.type = "any" and a minimum log fold change threshold of 0.5. This analysis was repeated at each stage from E6.5 to E8.5 in 0.25-day intervals, excluding cell types with fewer than 10 cells per stage. Genes were filtered out if the upper quartile of normalized expression for any cell type exceeded a threshold of 20 counts, to prevent optical crowding during imaging. The top five discriminative genes per cell type and stage were selected based on the Top column from findMarkers(), which reflects the minimal gene set needed to distinguish the cell type from all others. Genes were aggregated across all stages to form a comprehensive panel. To further refine the panel, we ensured that the total estimated number of transcripts per cell type remained below 300, calculated by exponentiating log-normalized expression values. This constraint was informed by prior benchmarks suggesting 200–300 transcripts per cell as an upper limit for reliable detection without optical saturation. Genes of particular interest (e.g., transcription factors) were added manually, followed by iterative re-filtering to manage total transcript load. The final panel comprised 387 genes: 351 for barcoded detection by seqFISH and 36 for non-barcoded detection by sequential smFISH (Table S1).

Library and probe design – Primary probe design

We designed 30-nt gene-specific primary probes targeting the coding sequences of the 387 selected genes using the mm10 mouse genome and UCSC annotations. Probes were required to have 45–65% GC content and avoid homopolymers of ≥5 identical bases. For each gene, 28–48 probes were designed, with genes yielding fewer than 28 usable probes excluded from the panel. All probe candidates were aligned against the mouse transcriptome using BLAST to remove off-target sequences with >15-nt matches to un-intended transcripts. To prevent cross-hybridization within the probe set, a local BLAST database of all probe sequences was built. Probes predicted to have more than seven off-target hits against other library probes were iteratively removed until all probes passed specificity criteria.

Library and probe design – Readout probe design

Readout probes of 15 nucleotide length were designed as reported in Shah et al.⁸¹ In brief, readout sequences were randomly generated to have a GC content of 40–60% and were screened by BLAST to ensure no homology to the mouse transcriptome. To reduce cross-hybridization between readout probes, all pairs of sequences with ≥10 consecutive identical bases were eliminated. The reverse complements of these validated sequences were included in the primary probes.

Library and probe design – Primary probe library construction

The final library of 15,989 primary probes (targeting 387 genes) was synthesized as an oligo pool by Twist Bioscience. Each probe consisted of a 30-nt mRNA-targeting region, four gene-specific readout sequences (15-nt each) separated by 2-nt spacers, and universal primer sequences for amplification. Two formats were used.

• (1)seqFISH probes (barcoded): 5′–(primer 1)–(readout 1)–(readout 2)–(probe)–(readout 3)–(readout 4)–(primer 2)–3′
• (2)smFISH probes (non-barcoded): identical readout sequence used four times: 5′–(primer 1)–(readout 1)–(readout 1)–(probe)–(readout 1)–(readout 1)–(primer 2)–3′

The probe pool was amplified by limited-cycle PCR to produce a template for in vitro transcription (NEB, E2040S), followed by reverse transcription using uracil-containing primers (Thermo Fisher, EP7051). Forward primers were cleaved enzymatically using uracil-specific excision reagent (NEB, N5505S), and single-stranded DNA was alkaline hydrolyzed with 1 M NaOH at 65°C for 15 min. After neutralization with 1 M acetic acid, probes were purified via ethanol precipitation and phenol–chloroform extraction. Probes were resuspended in hybridization buffer (40% formamide, 2× SSC, 10% dextran sulfate) at 40 nM per probe and stored at −20°C.

Library and probe design – Readout probe synthesis

Fifteen-nucleotide readout probes were synthesized by Integrated DNA Technologies (IDT) and conjugated to fluorophores (Alexa Fluor 488, Cy3B, Alexa Fluor 647). Probes were stored at −20°C in nuclease-free water.

Library and probe design – Encoding strategy

We implemented a 12-pseudocolor encoding scheme, distributing pseudocolors evenly across three fluorescent channels (Alexa Fluor 488, Cy3B, Alexa Fluor 647). The 12 pseudocolors were cycled over four imaging rounds to generate 20,736 unique barcodes. An additional round of pseudocolor imaging was performed to obtain error-correctable barcodes, yielding robust gene identification across the 351 barcoded targets.

Library and probe design – Coverslip functionalization

Glass coverslips (Thermo Scientific, 3421) were washed in nuclease-free water and immersed in 100% ethanol, air-dried, and plasma-cleaned for 5 min (Harrick Plasma, PDC-001, high setting). Coverslips were incubated in 1% bind-silane (GE, 17-1330-13) in 10% acidic ethanol (pH 3.5) for 1 h at room temperature, rinsed three times in ethanol, and heat dried at >90°C for 30 min. They were then incubated with 100 μg/mL poly-D-lysine (Sigma, P6407) in water for at least 1 h, rinsed in nuclease-free water, and air-dried. Coverslips were stored at 4°C and used within one week.

seqFISH microscopy

Three tissue sections from two experimental blocks, containing four embryos (2 from E7.25 and 2 from E7.5), were imaged as previously described,^10,81 which is described in brief in the following sections. For each FOV, snapshots were acquired with 4 μm z steps for six z slices. Serial hybridization and imaging were repeated for 29 rounds.

seqFISH microscopy – Tissue preparation and hybridization

Embryos were dissected from uteri, washed in M2 medium (Sigma Aldrich, M7167), and fixed in 4% paraformaldehyde (PFA; Thermo Scientific, 28908) in 1× PBS (Invitrogen, AM9624) for 30 min at room temperature. Fixed embryos were washed and incubated in 30% RNase-free sucrose (Sigma Aldrich, 84097) in 1× PBS at 4°C until they sank. Each embryo was then embedded sagittally in optimal cutting temperature (OCT) compound (Sakura, 4583) using tissue base molds (Sakura, 4162), snap-frozen in dry ice–isopropanol (VWR, 20842), and stored at −80°C. Tissue sections (20 μm) were cut using a cryostat, collected on functionalized coverslips (see STAR Methods: Library and probe design – Coverslip functionalization), and stored at −80°C.

Tissue sections were post-fixed in 4% PFA in 1× PBS for 15 min at room temperature, permeabilized with 70% ethanol for 1 h, and cleared with 8% SDS in 1× PBS for 20 min. Sections were washed with 70% ethanol, air-dried, and blocked for at least 2 h at room temperature in blocking buffer (1× PBS, 0.25% Triton X-100, 10 mg/mL BSA [Thermo Fisher, AM2616], and 0.5 mg/mL salmon sperm DNA [Thermo Fisher, AM9680]) in a humid chamber. Tissue sections were incubated for 2 h at room temperature with primary antibodies diluted 1:200 in blocking buffer: anti-pan-cadherin (Abcam, ab22744), anti-N-cadherin (Cell Signaling, 14215), anti-β-catenin (Abcam, ab6301), and anti-E-cadherin (BD Biosciences, 610181). After three washes with PBS containing 0.1% Triton X-100 (PBS-T), sections were incubated with a CCTTACACCAACCCT-conjugated secondary antibody (anti-mouse IgG) diluted 1:500 in blocking buffer for ≥2 h at room temperature. Samples were again washed three times in PBS-T and post-fixed in 4% PFA in 1× PBS for 15 min, followed by three 10-min washes in 2× SSC (Thermo Fisher, 15557036).

Hybridization was performed in primary probe hybridization buffer (40% formamide [Sigma, F9027], 2× SSC, 10% dextran sulfate [Sigma, D8906]) containing ~2.5 nM of each primary probe, 1 nM Eef2 probe set A and B, and 1 μM locked nucleic acid oligo-d(T)30 (Qiagen). Samples were incubated in a humidified chamber at 37°C for 24–36 h. They were washed with 40% formamide wash buffer (40% formamide, 0.1% Triton X-100 in 2× SSC) for 30 min at 37°C and rinsed three times with 2× SSC. Sections were then hybridized with 200 nM tertiary probe (/5Acryd/AG GGT TGG TGT AAG GTT TAC CTG GCG TTG CGA CGA CTA A) in EC buffer (10% ethylene carbonate [Sigma, E26258], 10% dextran sulfate [Sigma, D4911], and 4× SSC) for ≥2 h. Samples were washed for 5 min in 10% formamide wash buffer (10% formamide, 0.1% Triton X-100 in 2× SSC), followed by two 5-min washes in 2× SSC. For gel embedding, samples were treated with 0.1 mg/mL Acryloyl-X SE (Thermo Fisher, A20770) in PBS for 30 min, rinsed three times in 2× SSC, and post-fixed in 4% PFA in PBS for 15 min, followed by three washes in 2x SSC. Sections were incubated for 30 min in 4% acryl-amide/bis-acrylamide (1:19 ratio) in 2× SSC, followed by gelation with degassed 4% hyrdogel solution containing 0.05% ammonium persulfate (APS; Sigma, A3078) and 0.05% TEMED (Sigma, T7024). The sample was sandwiched by a GelSlick-coated slide (Lonza, 50640) in a nitrogen chamber for 30 min at room temperature and then at 37°C for ≥3 h. After polymerization, the hydrogel was separated from the coverslip and rinsed with 2× SSC. Embedded tissue was cleared at 37°C for 3 h in digestion buffer (1:100 Proteinase K [NEB, P8107S], 50 mM Tris-HCl pH 8.0 [Invitrogen, AM9856], 1 mM EDTA [Invitrogen, 15575020], 0.5% Triton X-100, 1% SDS, and 500 mM NaCl [Sigma, S5150]). Sections were rinsed with 2× SSC and treated with 0.1 mg/mL label-X for 45 min at 37°C for chemical labeling. Samples were then re-embedded in a second 4% hydrogel (2.5-h gelation), trimmed, and mounted into custom flow cells for imaging.

seqFISH microscopy – Imaging

Three tissue sections from two experimental blocks, containing four embryos, were imaged. Flow cells were connected to an automated fluidics system. Prior to hybridization rounds, samples were stained with 10 μg/mL DAPI (Sigma, D8417) in 4× SSC to identify fields of view (FOVs). All imaging was performed in antibleaching buffer (50 mM Tris-HCl pH 8.0, 300 mM NaCl, 2× SSC, 3 mM Trolox [Sigma, 238813], 0.8% D-glucose [Sigma, G7528], 1:100 catalase [Sigma, C3155], and 0.5 mg/mL glucose oxidase [Sigma, G2133]). Sixteen hybridization rounds were imaged for the decoding of the barcoded mRNA seqFISH probes followed by a repeat of the first hybridization. Then, 12 serial hybridization rounds were imaged for 36 non-barcoded sequential smFISH probes, followed by 1 hybridization to visualize the cell segmentation staining using Cy3B conjugated to/5AmMC6/TTAGTCGTCGCAACG. Hybridization rounds were automated using a deep-well 96-well plate system. Each hybridization buffer contained three 15-nt readout probes (Alexa Fluor 647, Cy3B, Alexa Fluor 488; 50 nM each) in EC buffer. The cell segmentation round used a single Cy3B-labeled readout (/5AmMC6/TTAGTCGTCGCAACG). Each hybridization step lasted 25 min at room temperature in the dark, followed by washing with 10% formamide wash buffer, rinsing in 4× SSC, and DAPI staining (1.5 min, 10 μg/mL in 4× SSC). The flow cell was then filled with antibleaching buffer, and all selected FOVs were imaged. Imaging was performed using a Leica DMi8 microscope equipped with a Yokogawa CSU-W1 spinning disk, an Andor Zyla 4.2 Plus sCMOS camera, a Leica 63 ×1.40-NA oil objective, a motorized stage (ASI MS2000), and Semrock filters. For each FOV, images were acquired using 4 μm z-steps across six z-planes. After each round, readout probes were stripped using 55% formamide wash buffer (55% formamide, 0.1% Triton X-100 in 2× SSC) for 4 min, followed by a 4× SSC rinse. Serial hybridization and imaging were repeated for 29 rounds in total. RNA integrity was verified using colocalization of interleaved Eef2 probe sets visualized with distinct fluorophores. Integration of the automated fluidics and image acquisition system was managed using custom μManager scripts.

Imaging processing and registration

Raw image data were processed as described previously.¹¹ Briefly, effects of chromatic aberration, tissue background and auto-fluorescence were removed, non-uniform background was corrected and background signal subtracted following the same approach as described previously.¹¹ For each round of hybridization, Images were registered to the DAPI images of the first hybridization for each FOV using the same two-dimensional phase correlation algorithm as previously described.¹¹

Image processing and registration – Chromatic correction, background subtraction, and alignment

Raw imaging data were processed to correct for chromatic aberration, remove background signal, and register images across hybridization rounds. To correct for chromatic aberration between fluorescence channels, 0.1 μm TetraSpeck microspheres (Thermo Scientific, T7279) were imaged, and geometric transforms were computed to align all fluorescence channels. Tissue background and autofluorescence were removed by dividing background with the fluorescence images. A flat field correction was applied by dividing the normalized background illumination with each of the fluorescence images while preserving the intensity profile of the fluorescent points. Background signal was then subtracted using ImageJ rolling ball background subtraction algorithm using a radius of 3 pixels and despeckle algorithm to remove any hot pixels. For each round of hybridization, the 405 nm channel capturing DAPI signal was used for registration. Within each field of view (FOV), all DAPI images from subsequent hybridization rounds were registered to the DAPI image from the first round using a two-dimensional phase correlation algorithm.

Transcript detection and smFISH processing

We called individual transcript barcodes using the dot matching algorithm with the same parameters as described in Lohoff et al., with more details in Shah et al.⁸¹ In brief, detected fluorescent spots across all hybridizations and channels were matched using a 2.45-pixel search radius to identify symmetric nearest neighbors. Spot sets corresponding to a unique barcode were directly assigned to the on-target barcode list. For ambiguous matches, combinations were filtered by minimizing residual spatial distance, and assigned if a single on-target barcode with a Hamming distance of 1 remained. If ambiguity persisted, the point was excluded. This matching process was repeated using each hybridization round as the seed, and only barcodes consistently detected in at least three out of four rounds were retained as valid gene calls. For the 36 genes that we probed using smFISH, we found that assignment of an optimal light intensity threshold to separate background noise from hybridized mRNA was particularly challenging due to each gene’s expression only being measured over a single round of hybridization. To address this problem, we extracted a continuous measure of gene expression per cell, taken to be the 95th percentile of the pixel intensities across all pixels within each cell’s segmented area, where a high value corresponds to some evidence of detection of expression of each gene in each cell. We did not use the smFISH-captured genes for further high-dimensional analysis.

Cell segmentation

Since we visually inspected images associated with the cell membrane channels in the first hybridization and found there was very low signal intensity associated with the cell membranes, we opted for an alternative cell segmentation strategy to what was performed previously.¹¹ We describe the steps of this segmentation strategy below.

Human-in-the-loop cell segmentation

In this situation, we used a human-in-the-loop cell segmentation strategy, where we performed manual annotation with napari (https://zenodo.org/records/8115575) of single points for each DAPI-stained nucleus for each E7.25 and E7.5 image across all z slices. Then, for each z slice and FOV, we performed a seeded watershed algorithm to extend the manually annotated points into solid segments covering the nucleus of each cell. The watershed used nuclear boundary predictions from a U-Net⁷⁶ as heightmap and foreground predictions from the same U-Net as mask. This network was initially trained on the nucleus segmentation dataset⁸² and then retrained several times on the segmentation results obtained via the watershed from manual seed points. Note that we tried using StarDist⁷⁷ and CellPose⁷⁸ for the nuclear segmentation, but found the segmentation quality of their pretrained models insufficient. To further extend nuclei to full cell segments, we trained a feature based classifier to segment the full cell and cell boundaries using ilastik⁸³ pixel classification on a few cells using the DAPI channel and the total set of detected transcripts as input channels. We then used a watershed with nuclear segmentations as seeds and ilastik prediction as heightmap and mask to obtain the cellular segmentation.

3D segments for early embryos

Our data comprises optical sections captured 4 μm apart within 20 μm thick physical sections. These optical sections were merged into 3D reconstructions without requiring additional registration across separate physical sections (see Videos S1 and S2 which display the optical sections). Since the z slices were 4um apart, a distance that could conceivably cover a single cell over multiple z slices, we performed a matching algorithm to identify matching 2D segments assigned to a single 3D segment. We applied this matching algorithm to all seqFISH data, including the previously published E8.5 data. The matching algorithm worked by querying consecutive z slices from z slice 2 and so on, identifying for each segment the next segment with the largest overlapping area, and calling the pair of segments a match if the Jaccard index between the pixels was 0.5 or above. Finally, for each matched segment, we calculated the new 3D cell centroid as the centroid of the pixel locations across the x, y, and z planes.

Data processing - Filtering and quality control

Following cell segmentation, we assigned each transcript belonging to the barcoded seqFISH library to the corresponding cell in which the segment belonged. We then filtered cells to retain only those that i) had at least three detected genes, and ii) had between 5 and 1,000 total counts. To further check for cells that may be low quality, we performed unsupervised clustering of the log-transformed gene counts, and identified clusters with particularly low total counts to be removed from further analysis.

Data processing - Normalisation

To normalise the gene expression data, we first needed to account for the systematic effect where we observed fewer counts with higher z-slice values. To do this, we first extracted the size factors for embryo 1 and z-slice 2 using the calculation embedded within the ‘logNormCounts()’ function in the scran package, ignoring the Xist gene in the calculation of size factors. We then normalised all of the seqFISH-resolved cells using the ‘logNormCounts()’ function in scran, where size factors were quantile normalised to match the distribution of those of embryo 1 z-slice 2. In doing so, we extracted normalised logcounts for the cells, which we used for subsequent analyses.

Data processing - Defining AP- and DV-axes

To ascertain anterior-posterior (AP) axis positioning of the seqFISH-resolved embryos, using anatomical landmarks, we performed manual annotation of the AP axis, as well as manual annotation of the extraembryonic region and the embryo proper for the E7.25-E7.5 samples. Then we used the ‘princurve’ package⁷⁹ to extract the relative positions along the AP axis as well as the perpendicular distances to the AP axis curve to determine the positions along the dorsal-ventral (DV) axis. To enable consistent comparison between each of the seqFISH-resolved embryos, we used z-scaling followed by reordering to start on the anterior part of the AP axis. These scaled AP values were then used for further downstream analysis.

scRNA-seq analyses - Quality control filtering after integration and clustering

ChatGPT (GPT-4, OpenAI; accessed via ChatGPT Plus in 2024–2025) was used to assist in writing and optimizing R scripts for scRNA-seq analyses. All code generated using this tool was reviewed, edited, and validated by the authors. The full analysis code is available in the accompanying repository. Initially, we applied batch correction using the rPCA method in Seurat v4¹⁴ to log-normalized counts matrices from E7.25, E7.5 and E8.5 seqFISH embryos. rPCA integration was performed across different embryos (e.g., embryo 1 and embryo 2) within each embryonic stage (e.g., E7.25). To perform rPCA integration, principal components (PCs) were generated by running ScaleData and RunPCA using all seqFISH genes set as features. Integration was then carried out using FindIntegrationAnchors (with k.anchor = 5, scale = FALSE, and reduction = ‘rpca’), followed by IntegrateData (with features.to. integrate = all features (genes) and k.weight = 100). Finally, clustering was performed after applying ScaleData and RunPCA on the batch-corrected matrices. We used FindNeighbors(dims = 1:30) and FindClusters(resolution = 1) to identify clusters. As a quality control measure, we excluded clusters with abnormal RNA/feature counts, as well as a cluster from the most distal region of the ecto-placental cone in E7.25 embryos (see Figures S1B and S1D red clusters) prior to further analyses described below.

scRNA-seq analyses - integration, clustering and UMAP for E7.25/E7.5 seqFISH Embryos

We applied multiple rounds of batch correction using the rPCA method in Seurat v4¹⁴ to log-normalized counts from E7.25 and E7.5 seqFISH embryos. Cells that passed the previously described quality control criteria were included in this analysis. In round 1, rPCA integration was performed across different embryos (e.g., embryo 1 and embryo 2) within each embryonic stage (e.g., E7.25). First, principal components (PCs) were generated by running ScaleData and RunPCA on all seqFISH probes. Integration was then carried out using FindIntegrationAnchors (with k.anchor = 5, scale = FALSE, and reduction = ‘rpca’), followed by IntegrateData (with features. to.integrate = all features (genes) and k.weight = 100). In round 2, batch-corrected count matrices from round 1 were re-integrated using rPCA, with k.weight = 200 and k.anchor = 5. Finally, clustering was performed on the batch-corrected matrices after applying ScaleData and RunPCA. We used FindNeighbors(dims = 1:30) and FindClusters(resolution = 1.2) to identify clusters, followed by dimensionality reduction using UMAP, generated via RunUMAP(dims = 1:30). A second round of clustering was performed on a cluster that was initially annotated ‘Anterior Primitive Streak’ using FindNeighbors(dims = 1:30) and FindClusters(resolution = 1).

This unsupervised sub-clustering revealed three transcriptionally and spatially distinct subpopulations within the original APS cluster: axial mesendoderm, definitive endoderm precursors, and a mixed cluster comprising primitive streak and axial mesendoderm cells.

scRNA-seq analyses - integration, clustering and UMAP for E8.5 seqFISH Embryos

For the E8.5 seqFISH embryos, a similar procedure was followed. After quality control, batch-corrected count matrices from within-stage integration were processed with ScaleData and RunPCA. Clustering was then performed using FindNeighbors(dims = 1:30) and FindClusters(resolution = 1), and clusters were visualized via UMAP, generated using RunUMAP(dims = 1:30).

scRNA-seq analyses - Marker gene identification

To identify marker genes for the clusters from the E7.25/E7.5 and E8.5 seqFISH embryos, we applied FindAllMarkers(slot = "data"), filtering for marker genes with an average log2 fold change (avg_log2FC) greater than log2(1.5).

scRNA-seq analyses - principal components from the extended gastrulation scRNA atlas

We used log-normalized counts from a downsampled version (10,000 cells per embryonic stage from E6.5 to E9.5) of the mouse gastrulation atlas from Imaz-Rosshandler et al., 2024, to perform three rounds of rPCA integration using Seurat v4.¹⁴ In round 1, rPCA integration was carried out across sequencing batches within each embryonic stage (e.g., E6.5) using 2,000 highly variable features identified with FindVariableFeatures. We applied FindIntegrationAnchors (k.anchor = 5, scale = FALSE), followed by IntegrateData (features.to.integrate = all features (genes), k.weight = 100). In round 2, batch-corrected count matrices from round 1 were integrated across all embryonic stages within each atlas version (original vs. extended). In round 3, integration was performed across both the original and extended atlas versions. For rounds 2 and 3, rPCA integration was executed with k.weight = 200 and k. anchor = 5, using PCs that were identified using a combined list of VariableFeatures (5716 genes) identified from each embryonic stage during round 1 of rPCA integration.

scRNA-seq analyses - principal components from the seqFISH datasets

We applied multiple rounds of batch correction using the rPCA method in Seurat v4¹⁴ to log-normalized counts from E7.25, E7.5 and E8.5 seqFISH embryos. Cells that passed the previously described quality control criteria were included in this analysis. In round 1, rPCA integration was performed across different embryos (e.g., embryo 1 and embryo 2) within each embryonic stage (e.g., E7.25). First, principal components (PCs) were generated by running ScaleData and RunPCA on all seqFISH probes. Integration was then carried out using FindIntegrationAnchors (with k.anchor = 5, scale = FALSE, and reduction = ‘rpca’), followed by IntegrateData (with features.to.integrate = all features (genes) and k.weight = 100). In round 2, batch-corrected count matrices from round 1 were re-integrated using rPCA, with k.weight = 200 and k.anchor = 5. Finally, principal components were calculated on the batch-corrected matrices after applying ScaleData and RunPCA.

scRNA-seq analyses - StabMAP and reducedMNN to align seqFISH and scRNA datasets

To align all seqFISH cells that passed quality control with the downsampled scRNA cells from the extended gastrulation atlas, we applied StabMAP¹⁵ followed by reducedMNN.¹⁶ For StabMAP, we used the top 30 principal components calculated for both the scRNA and seqFISH datasets, as described above, with projectALL = TRUE. The resulting StabMAP embedding was then reweighted to ensure equal contribution from both datasets using reWeightEmbedding(). To correct for any remaining technical differences between the seqFISH and scRNA datasets, we applied reducedMNN_batchFactor() with k = 10, setting the batch factor to either seqFISH or scRNA as appropriate.

scRNA-seq analyses - UMAP and joint clustering for spatiotemporal atlas

A low-dimensional UMAP was generated to visualize both scRNA and seqFISH cells within the same space, using runUMAP applied to the batch-corrected StabMAP embedding. Joint clustering and subsequent sub-clustering were performed on this embedding with FindNeighbors (dims = 1:60) and FindClusters (resolution = 1 for clustering; resolution = 3 for sub-clustering). As a quality control step, clusters containing more than 98% or fewer than 2% of either seqFISH or scRNA cells were labeled as ‘poor joint clusters.’

scRNA-seq analyses - Label transfer across scRNA and seqFISH datasets

We applied K-nearest neighbors (K = 5) to classify the seqFISH cells based on cell type annotations from the extended mouse atlas, incorporating both extended and original atlas labels, as well as embryonic stage and subdissection labels. Additionally, we used the same K-nearest neighbors approach (K = 5) to classify the scRNA cells according to the cell types assigned to the seqFISH cells.

scRNA-seq analyses - Gene expression imputation for seqFISH

We leveraged the scRNA atlas data, which provides whole transcriptome counts, to impute full gene expression profiles for the seqFISH-resolved cells. For each seqFISH-resolved cell, we identified the K-nearest neighbors (K = 5) and calculated the mean expression vector across all genes from the extended mouse atlas dataset.

scRNA-seq analyses - Gene expression and AP/DV coordinate imputation for seqFISH and scRNA

We utilized the seqFISH atlas data, which includes AP and DV axial coordinates, to impute these coordinates in the scRNA cells. To achieve this, we identified the K-nearest neighbors (K = 5) for each scRNA cell and calculated the mean expression vector across the AP or DV coordinates from the seqFISH dataset.

scRNA-seq analyses - Refining cell type annotations using seqFISH data

After obtaining cell type annotations for the seqFISH cells, we visualized their spatial distribution to further distinguish subtypes based on both spatial positioning and gene expression. This analysis was focused on the endothelium, lateral plate mesoderm, and ExE ectoderm. We then back-mapped the scRNA-seq extended mouse atlas to correlate the newly refined cell types with the broader atlas, allowing us to manually annotate subclusters with more specific cell type labels (e.g., proximal and distal ExE ectoderm, anterior and posterior LPM, embryo proper endothelium 1 and 2, and endocardium 1 and 2) based on additional spatial localization information.

Manual curation of cell type annotations

Cell type annotations within the spatiotemporal atlas were initially assigned using a semi-automated pipeline that incorporated high-resolution clustering and label transfer from the extended gastrulation atlases. While many annotations were reliably transferred via exploration of physical localization and marker gene expression patterns, additional manual curation was performed to refine specific populations, particularly embryo proper endothelium, lateral plate mesoderm (LPM), and extraembryonic ectoderm (ExE ectoderm). Some annotations may still require further refinement, and we provide the full dataset, clustering data, and transferred labels as an open-access resource to enable researchers to make additional improvements based on their expertise.

To enhance transparency and support further evaluation, Table S1 provides detailed marker genes for each cell type, while Videos S3–S5 illustrate the spatial distribution of these annotated cell types across the seven seqFISH embryos, along with gene expression patterns for the related cell types. Moreover, to support ongoing refinement by the community, we encourage users to provide feed-back or suggestions for improved annotations to the lead contact: Shila Ghazanfar (shila.ghazanfar@sydney.edu.au).

Manual curation - Updating ExE ectoderm and LPM annotations

To improve the resolution of populations initially labeled as ExE ectoderm and lateral plate mesoderm (LPM) following the first round of majority voting, we refined annotations by back-referencing and assigning majority-vote seqFISH cell type labels. These labels provided crucial spatial localization information, allowing us to better resolve these populations. Subsequently, the labels were consolidated into ‘ExE ectoderm proximal’, ‘ExE ectoderm distal’, ‘ExE mesoderm and anterior LPM’, or ‘ExE mesoderm and posterior LPM’, in alignment with their distinct spatial positions as indicated by seqFISH labels.

Manual curation - Reclassification of embryo proper endothelium

Annotations for embryo proper endothelium were refined by integrating spatial information from seqFISH embryos and clustering data. Cells initially labeled embryo proper endothelium through majority voting were reassigned to more specific subtypes, including endocardium #1/#2, Allantois endothelium, and Embryo proper endothelium #1/#2.

Manual curation - Refining poorly labeled populations using spatial information

To improve annotation accuracy, we refined a subset of primitive streak cells based on their sub-cluster identities and spatial positioning, reassigning them as rostral ectoderm. Similarly, a sub-cluster of epiblast was relabeled ExE endoderm, and another sub-cluster initially labeled haematoendothelial progenitors were reclassified as ExE mesoderm, better reflecting anatomical localization. Cells previously assigned as parietal endoderm were reassigned to visceral endoderm, while a group of anterior primitive streak cells were relabeled notochord.

Manual curation - Resolving ambiguous labels

As the initial annotation process relied on majority voting, two populations required further refinement to resolve ambiguous labels. Cells initially labeled ‘Dorsal spinal cord progenitors, Neural tube’ were relabeled as ‘Dorsal spinal cord progenitors’, while cells assigned ‘Non-neural ectoderm, Surface ectoderm’ were renamed ‘Surface ectoderm’ to reflect their spatial identity more accurately.

Overall, these refinements aimed to ensure that cell type annotations align more accurately with the spatial and transcriptomic landscape captured in our dataset. By combining automated label transfer with expert-driven curation, we improve annotation specificity while maintaining flexibility for further refinement by the research community. All updated annotations, code, and metadata are publicly available to ensure transparency and reproducibility.

Marker gene identification and expression visualization

To identify marker genes for the refined cell types, we used the FindAllMarkers function on seqFISH cells with the imputed gene expression data (slot = "imputed gene expression") and a curated list of established marker genes from Imaz Rosshandler et al. (2024). This analysis was performed on subsets of cell types (Figures S2–S4), with markers filtered for those exhibiting an average log2 fold change (avg_log2FC) greater than log2(1.5). The original gene expression patterns for these marker genes in both seqFISH and scRNA cells are visualized in dot plots (Figures S2–S4).

Differential gene expression analyses along AP axes in the primitive streak region and somitic tissues in the gastruloids and scRNA cells

To identify genes with altered expression over the AP axes in the primitive streak region of embryo 2, or the gastruloids somitic tissues, the AP coordinates were provided as pseusotime values to Tradeseq⁷⁴ (1.10.0; Van den Berge et al., 2020). Generalised additive models were fit (fitGAM, nknots = 6, cellWeights = 1) to genes that were highly variable [top 2000 highly variable genes were identified using VariableFeatures function in Seurat (4.2.0)] among the cells that were part of each cell type. The associationTest function was then used to identify which of these genes had altered expression over the AP axes. ComplexHeatmap⁸⁰ (2.15.3; Gu et al., 2016) was used to visualise the expression patterns from the GAMs with the highest 300 waldStat scores (p-value < 0.01 and mean-LogFC>2). Metascape⁷⁵ (Zhou et al., 2019) was used to identify Gene Ontology (GO) terms that were enriched for clusters of genes.

Bioinformatics pipeline to project in vitro models into a spatiotemporal framework

To project additional scRNA-seq datasets into the spatiotemporal atlas, we used StabMAP¹⁵ and reducedMNN.¹⁶ Principal components were calculated for the gastruloid cells after performing rPCA integration using Seurat v4¹⁴ as described above in the ‘Calculating Principal Components’ sections. Briefly, multiple rounds of rPCA integration were performed, first across samples within each gastruloid timepoint and then across timepoints. Next, to align all gastruloid cells with the spatiotemporal atlas cells, we applied StabMAP followed by reducedMNN. For StabMAP, we used the top 30 principal components calculated for the scRNA, seqFISH and query gastruloid datasets, as described above, with projectALL = TRUE. The resulting StabMAP embedding was then reweighted to ensure equal contribution from all three datasets using reWeightEmbedding(). To correct for any remaining technical differences between the seqFISH, scRNA and gastruloid datasets, we applied reducedMNN_batchFactor() with k = 10, setting the batch factor to either seqFISH, scRNA or gastruloid as appropriate. Cell type labels, and imputed AP and DV values for the gastruloid cells were assigned as described in the ‘Label Transfer across scRNA and seqFISH Datasets’ and ‘Gene Expression and AP/DV Co-ordinate Imputation for seqFISH and scRNA Cells’ sections above.

Spatial QC in the bioinformatics pipeline

To construct the spatiotemporal atlas, we integrated seqFISH and scRNA-seq datasets into a shared low-dimensional embedding and performed high-resolution clustering (see UMAP Generation and Joint Clustering for spatiotemporal Atlas sub-section in scRNA-seq Analyses section). Poor joint clusters were identified as clusters where >98% of cells originated from only one dataset (see Figure S3A). scRNA-seq cells in these clusters lacked sufficient neighboring seqFISH cells, making spatial imputation (e.g., AP or DV values) unreliable. This discrepancy arises because the scRNA dataset includes tissues absent from seqFISH slices (e.g., mature yolk sac endothelium and mesothelium). When projecting additional query datasets, such as gastruloids, onto the spatiotemporal atlas, poor joint cluster label information is transferred to the query cells. Query cells receiving the poor joint cluster label = TRUE are identified as those cells failing spatial QC (Figure 4C), as their spatial information is unreliable due to alignment with scRNA-seq cells that lack spatial reference points in seqFISH. For instance, in Figure 4, some gastruloid cardiomyocytes and yolk sac endothelium fail spatial QC because their scRNA-seq counterparts lack seqFISH equivalents.

Quantification and Statistical Analysis

Statistical analysis of cell type distribution along the AP axis

To quantitatively assess differences in cell type distribution between query cell and scRNA cells along the anteroposterior (AP) axis, we implemented two complementary statistical tests into the bioinformatics pipeline.

• (1)Kolmogorov-Smirnov (KS) Test: To determine whether the distribution of cells along the AP axis differs significantly between query and scRNA-seq cells, we applied a KS test. This test provides a statistical measure of divergence in AP positioning between the two datasets.
• (2)Jaccard Index Analysis: To further assess similarity in cell type composition, we calculated Jaccard indices across discrete bins spanning the AP axis. This analysis identifies regions long the AP axis where cell type composition is similar or divergent in query versus scRNA datasets.

Additional Resources

Data can be interactively explored at this link: http://shiny.maths.usyd.edu.au/SpatiotemporalMouseAtlas/.

Processed data can be downloaded using the links provided on the front page of the interactive exploration Shiny app above.

Supplementary Material

Supplemental Information

Supplemental information can be found online at https://doi.org/10.1016/j.celrep.2025.116047.

Highlights.

• Integrated spatiotemporal atlas of mouse embryogenesis from E6.5 to E9.5

• Resolves 80+ refined cell types across germ layers and embryonic stages

• Spatial gene expression across the anterior-posterior and dorsal-ventral axes

• Enables projection of additional datasets for comparative developmental analysis

In brief.

Harland et al. present a spatiotemporal transcriptomic atlas of early mouse development, integrating spatial and single-cell transcriptional data from gastrulation through organogenesis. This resource explores axial patterning across space and time and provides a bioinformatic pipeline to benchmark in vitro models against in vivo development.

Data and code availability

• Data can be interactively explored at this link: http://shiny.maths.usyd.edu.au/SpatiotemporalMouseAtlas/.

• Processed data can be downloaded using the links provided on the front page of the interactive exploration Shiny app above.

• DAPI images, molecules, and segmentation boundaries can be downloaded from Zenodo at https://zenodo.org/records/13977985.

• All analyses were performed in R (version 4.2.1). Scripts for analysis and figure panels in this manuscript are available at https://github.com/ltgharland/Spatiotemporal-Atlas-of-Mouse-Gastrulation.