2.1 Sample collection and preprocessing

This study is approved and supervised by the ethic committee of Shanghai Tongren Hospital. Five fresh human foetal tooth germ tissues were collected at different embryonic developmental stages (2, 2 and 1 samples at 17, 20 and 24 PCW, respectively). Informed consent was obtained from each donor before the termination of pregnancy. All prenatal examinations showed that the foetus had no developmental abnormalities. The method of pregnancy termination was medically induced. The parents of the foetus are healthy, with no family history of facial and dental developmental abnormalities. Each sample was processed in sequenced separately. Details of sample pre-process was provided in Data S1.

2.2 scRNA and ST

Library construction, sequencing and data filtration were described in the Supplementary methods in Data S1. Both libraries of scRNA and ST were sequenced with paired-end 150 bp sequencing (PE150) by NovaSeq 6000 platform. We retained non-doublet cells with number of detected genes >500 and <2500, and mitochondria gene percentage <5%. Library size normalization was performed with NormalizeData function in Seurat¹⁶ to obtain the normalized count. Specifically, the global-scaling normalization method ‘LogNormalize’ normalized the gene expression measurements for each cell by the total expression, multiplied by a scaling factor (10,000 by default), and transformed by Seurat: SCTransform function. Cell phase was defined by Seurat: CellCycleScoring function based on cell cycle gene list.¹⁷ Spatial transcriptome data were pre-processed by Seurat in a similar way. We observed a median of 1378 genes and 250,835 reads per spot (1892 spots in total), where 95.9% barcodes and 97.0% UMIs were valid.

2.3 Cluster analysis

For scRNA data, we applied Louvain community detection-based cluster analysis implemented in Seurat: FindClusters function with number of principle component = 20 and resolution = 0.5. We applied Seurat: FindMarkers Wilcoxon test to find top 15 markers for each cluster, and defined them according to existing knowledge on dental cell types. Epithelium and mesenchyme cell were further extracted and underwent a second round of clustering, with the same parameters.

2.4 Pseudotime analysis

For epithelium and mesenchyme (excluding isolated follicle and MSC clusters) cell groups, we applied monocle3:learn_graph and order_cells function¹⁸ to learn the pseudotime trajectory. We then applied monocle3: graph_test function on top 5000 variable genes detected by Seurat: FindVariableGenes function to find all genes that significantly altered along the trajectory. Genes with FDR-adjusted p < 0.01 were further used for module detection by monocle3: find_gene_modules function. For the ease of interpretation, in epithelium cell group we separately applied module detection in each branch, and only retained gene modules that showed up-regulation or down-regulation along each of the three branches.

2.5 Transcription factor regulation network analysis

We applied SCENIC¹⁹ to construct transcription factors (TF)-target genes regulation network, separately for epithelium and mesenchyme. TF-motif annotation was downloaded from SCENIC website, which was corresponded to hg38 refseq gene annotation and contained a 10 kb window of each gene. We preserved the target genes pertaining to each transcription factor's extended network and subsequently subjected the module genes identified through pseudotime analysis to a Fisher test, thereby determining whether these genes exhibited enrichment within any of the transcription factor regulation networks. TFs with FDR-adjusted p < 0.01 and odds ratio >2 were retained for each module. We further extracted the overall expression of a TF and all its targets (so-called regulon) of each cell to verify that this expression value had a similar expression pattern as module genes. For epithelial SCENIC analysis result, we removed TF which was expressed in <25% of cells from the corresponding branch. For mesenchyme SCENIC result, we removed TF expressed in <25% of cells in the pulp-odontoblast trajectory.

2.6 Gene Ontology analysis

We used ClusterProfiler R package²⁰ to apply Gene Ontology-Biological Process (GO-BP) analysis by a hypergeometry test. The background gene list was set as all genes with GO-BP annotations. Only terms with >5 and <500 genes were considered. We applied simplify function to remove highly similar terms. Only terms with FDR-adjusted p < 0.01 were retained. The odds ratio and p-value were reported in the word cloud plot.

2.7 The nichenetr analysis

We used nichenetr R package²¹ to apply signalling network analysis as described previously.¹² We separately considered (1) up-regulation modules of each of the three epithelium branches; (2) immature (first and second wave) and mature (fourth wave) of mesenchyme differentiation as target gene list, and the final signalling network were combined in each of these two categories. Ligands were prioritized by area under curve (AUC) and Pearson correlation coefficient calculated by nichenetr.

2.8 Spatial transcriptome analysis

After defining cell clusters on scRNA data, we projected it to spatial transcriptome data by Seurat: FindTransferAnchors and TransferData functions to estimate the cell type probability of each spot. We assign cell types with the highest probability to each spot. We used misty R package²² to estimate the spatial relationship among highlighted cell types, TFs and expression levels of ligands and target genes. We chose the ‘intra’ mode in misty analysis and trained a linear model to predict each of the target values (cell probability or gene expression) by adjacent predicting values. Spatial adjacency was quantified by importance score in the prediction model.

3.1 Overview of cell types in human embryonic tooth

We collected tooth germ tissues from aborted foetuses spanning 17–24 PCW (Figure 1A) for scRNA and ST. In scRNA, we obtained expression data for 11,218 single cells of nine general cell types that passed quality control (Figures 1B–D and S1). The cell groups were classified based on gene panel (Table S1) from existing cell atlas of mouse tooth,^{7, 8} adolescent tooth⁹ and histology knowledge (Figures 1D and S1). Cells with epithelial origin were defined based on the expression of dental epithelial markers KRT5, KRT14, FXYD3 and COL17A1 (Figures 1D and S1). Mesenchyme origin cells were characterized by their expression of EMILIN1 and MSX1. A subgroup of mesenchyme origin cells expressing SFRP2 was defined as apical pulp-like cells. Lastly, we defined mitotic cells, pericytes, endothelium, neural crest cells (NCC), macrophages and lymphocytes based on their specific markers: CENPF, RGS5, CD34, FOXD3 + SOX10, C1QA and PTPRC + CD52, respectively (Table S1). Across different gestational weeks, we observed differences in the proportion of epithelium and pericytes, but no differences in cell cycle distribution (Figure 1C). We conducted a differential expression analysis between developmental time points for pericytes, and found minimal differences with biological significance. This result indicated that pericytes did not show significant differentiation between the two time points, thus we mainly focused on dental epithelium and mesenchyme clusters.

ST was applied to tooth slices from three donors: a slice from 24 PCW embryo containing both the tooth germ of the primary tooth and the tooth germ of the permanent tooth, a slice from a 17 PCW primary tooth and a slice from a 20 PCW primary tooth (Figure 1E). The proportion of each cell cluster of scRNA in each spatial plot was evaluated using the anchor-based integration method implemented in Seurat v3 (Figures 1F and S2). As expected, we observed that the epithelium incompletely surrounded the pulp tissues, which consisted mainly of mesenchyme cells, scattered pericytes and endothelium in the 24-week sample. AP-like cells were only observed in the permanent tooth germ of 24-week sample. We also found an isolated NCC group surrounded by endothelium (Figure 1F). This group was adjacent to the permanent tooth germ and likely represented the process of neuron outgrowth. The spatial distribution of cell markers at the transcription and protein levels, such as KRT15, COL17A1also supported these spatial characteristics (Figure 1G,H). Taken together, we successfully constructed a cell type atlas of the human embryonic tooth germ that covers various developmental timepoints and spatial structures.

3.2 Branched trajectory of dental epithelium differentiation

Next, we conducted an in-depth analysis of the cell constitution and developmental process of epithelial cell groups. By applying cluster analysis and pseudotime analysis, we partitioned all the epithelial origin cells into three branches (Figure 2A). The root branch expressed GAS6 and IGFBP5, and consisted of two continuous cell clusters: outer enamel epithelium expressing KRT15 and SPARCL1 (OEE, Figure 2B), and inner enamel epithelium expressing CRABP1 and HOPX (IEE). Most of the 17w epithelial cells were grouped into a root branch, supporting that root branch was the origin of epithelial development (Figure S3). The root branch ended at a cluster of ALCAM-positive cells that also expressed CD24 and HOPX, which were associated with epithelial stem cells. The ALCAM+ cell then gave rise to two branches. The first one was the SI branch consisting of SI progenitors expressing KRT17 and TAGLN and stratum intermedium expressing CLDN10, which was further validated by immunofluorescence (Figure S3). The second was the ameloblast branch expressing ASCL5 and AMELX. Lastly, we found a small subcluster of isolated cells near ameloblast expressing SHH, WNT10A, and exclusively WNT10B, which we defined as enamel knot (EK). Twenty-four PCW tooth germ contained moderately more ameloblasts and fewer ALCAM+ SC than 17 PCW tooth germ, and the overall cell proportions were similar (Figure 2C). These results suggested that the development of SI and ameloblast have a divergent trajectory, even though they share the same set of OEE-IEE lineage as progenitor. This result resembled previous identification of a dental progenitor that could give rise to both SI and ameloblast after injury,²³ and suggested that the ALCAM+ population in our study might have similar roles as dental progenitors.

We further compared our human embryonic data with mouse data meta-analysed by Hermans et al.¹³ using anchored-based integration. Figure 2D shows profound cross-species differences in expression profile, highlighting the necessity to study human tissues in dental developmental biology. Many human cell types mostly resembled dental epithelium progenitor (DEP), and were dissimilar to their corresponding cell types in mouse molars and incisors except for the IEE. We also found no similar cell types to any of the mouse ameloblast subtypes (pre-ameloblast, secretary ameloblast and mature ameloblast). In fact, human ameloblast and enamel knot were mostly similar to DEP in mouse. There were also considerable differences in cell marker expression. In humans, KRT17 (Figure 2E) was highly expressed in multiple epithelia, especially the SI lineage, whereas in mice, there were no notable expression patterns. On the other hand, although most human epithelia were similar to mouse DEP, the marker SFRP5 (Figure 2E) exhibited minimal expression in human dental epithelium.

Finally, we analysed the biological processes underlying these branches. Using a threshold of FDR-adjusted p < 0.01, we found 622, 1397 and 842 genes whose expression levels significantly altered along pseudotime in the root, SI and ameloblast branch, respectively. The gene module detection algorithm implemented in monocle3¹⁸ to find genes that were consistently up- or down-regulated in each branch (Tables S2–S4). SCENIC¹⁹ was then used to identify transcription factors (TF) that regulate these genes. TFs like NR1D1, DLX3, RUNX2, SOX11 and SMAD9 were predicted to regulate a large number of genes showing alterations in each branch (Figure 2F and Tables S5–S7). For example, among 18 target genes of DLX3, DLX14 were down-regulated in the root branch (Odds ratio, OR = 102, Fisher test p = 6.45 × 10⁻¹⁸). Consistently, the overall expression levels of DLX3 and its targets (so-called regulon¹⁹) were highly expressed in the primitive OEE, and gradually diminished in the root branch (Figure 2G). We highlighted regulons of several other TFs showing consistent expression patterns in SI and ameloblast branches, such as SOX11 and SMAD9 that were activated during SI and ameloblast development (Figure 2H).

3.3 Spatial patterns of dental epithelium development

We linked single-cell signals to the spatial transcriptome. Figures 3A,B and S4 show that OEE is present on the outer surface of the permanent tooth germ. ALCAM+ SC are found deeper, inside the OEE layer. In the permanent tooth germ, ameloblasts are few in number and located deep, near the stem cell layers. But in the primary tooth germ, they are abundant and closely packed. Ameloblasts are located at the boundary of pulp and epithelial tissues. The SI progenitors are located in the outer layer of ameloblasts. These patterns match the cell markers' spatial expression levels (Figure S5). We analysed the cell proportions in each spatial spot of all tissues (Figure 3C). We found IEE-OEE mainly in permanent tooth germ and stratum intermedium mostly in 24 PCW tooth germ. Specifically, in 20-week sample, we found that enamel knot cells were mainly distributed in the cusp region (Figure S4). This finding suggests that they might be secondary enamel knots.

Misty²² elucidated the spatial adjacency of different epithelial cell subtypes and key TFs. Figure 3D shows that when analysing epithelial cells, OEE and ALCAM+ SC had a close distribution with each other, while ameloblasts, enamel knots and stratum intermedium aggregated together. When including all cell types together, only mesenchyme showed close relationship with multiple epithelial cells. When analysing key TFs involved in epithelial differentiation (i.e., those TFs highlighted in Figure 2), we also observed that the expression levels of various key epithelial TFs was significantly associated to adjacent mesenchyme proportion (Figure 3D). In Conclusion, we inferred that the signal from mesenchyme cells had an important role in regulating dental epithelial differentiation. This process involved DLX3, which is known to be implicated in bone, hair and tooth development. Another example is RUNX2, a key factor in osteoblast differentiation and skeletal morphogenesis, which was found to be notably modulated by mesenchymal signals. In fact, the intercellular communication between mesenchymal and epithelial cells appears to employ a complex signalling cascade, ensuring the temporal and spatial control of these molecules during the various stages of tooth formation.

3.4 Involvement of LHX9, TCF7 and SP7 in dental pulp development

Similar to the analysis of dental epithelium, we further deciphered the spatiotemporal dynamics of mesenchyme cell groups. We identified seven subtypes of mesenchyme origin (Figures 4A–C and S6). Apical pulps (AP) were defined by the expression of SFRP1, SOSTDC1 and SMOC2. We defined the distal pulp according to the expression of SOX9, which contained two subtypes: one group expressed TNC, DKK3 and HEY1 (TNC+ DP) and the other expressed FRZB, FGF3 and TWIST2 (TWIST2+ DP). We further defined two odontoblast groups by SALL1 expression, and distinguished them by FGF3-TWIST2 expression (FGF3+ OD) and DKK3-FBN2 expression (FBN2+ OD). Lastly, we found two clusters of follicle cell, one expressed IGFBP5 + SPON1 + FOXF1 (follicle 1) and the other expressed GDF10 and COL12A1 (follicle 2). Compared with 24-PCW tooth germ, the 17-PCW tooth germ had significantly larger proportions of follicle cells, and significantly smaller proportion of distal pulps (Figure 4C).

By integrating our data with the previous dental cell atlas from Hermans et al.¹³ (Figure 4D), we found that human embryonic dental mesenchyme was generally similar to postnatal cells: the follicle, odontoblast, and distal pulp subgroups were all closest to the same cluster in postnatal data. The only exception was apical pulp, which was moderately distinct from the postnatal apical pulp. Nonetheless, when comparing human dental cells with mouse dental cells, we found substantial differences between the two species. In fact, only TNC+ DP resembled the corresponding DP cluster in mouse data, while all other cell types were close to mouse apical pulp.

We further applied pseudotime analysis to reconstruct the developmental trajectory of the continuous cell clusters from follicle 1 to FBN2+ OD (Figure S7), and identified four waves of gene expression patterns along this trajectory (Table S8). As shown in Figure 4E, the first and second waves corresponded to progenitor cell differentiation, and the third and fourth waves corresponded to pulp cell development. The second wave peaked at the apical pulp population, where the pulp and follicle populations merged, resembling the segregation of the mouse dental follicle and dental papilla lineages discovered by Jing et al.⁶ The first wave included HES1 and TWIST1, and was enriched in cell fate determination (GO p-adjust = 2.28 × 10⁻³) and other processes related to cell cycle (Figure 4F). SCENIC¹⁹ analysis revealed that TFs regulating the first wave genes were immediate genes like FOS and JUN family like FOSB (Fisher test p = 3.60 × 10⁻¹⁴; Figure 4E and Table S9). The second wave of genes including PAX9, SCUBE3 and SPARC, was found to be enriched in extracellular matrix organization (GO p-adjust = 1.23 × 10⁻¹⁸) and formation of primary germ layer (GO p-adjust = 8.38 × 10⁻⁹). They were also enriched in the predicted targets of LHX9, TCF7 and SOX18 (Fisher test p < 10⁻¹⁰; Figure 4F and Table S9). The third wave consisted of genes involved in cytoplasmic translation (GO p-adjust = 7.66 × 10⁻²⁹), without significant enrichment in specific TF regulon. Finally, the fourth wave genes, including DKK2, VCAN and THY1, were involved in the regulation of odontogenesis (GO p-adjust = 1.68 × 10⁻⁴), Wnt signalling pathway (GO p-adjust = 0.0001) and ossification (GO p-adjust = 3.18 × 10⁻⁶). These genes are mainly regulated by TFs such as DLX5, DLX3, SP7 and MSX2 (Fisher test p < 1.78 × 10⁻⁵; Figure 4F and Table S9). The overall expression levels of these highlighted TFs were also in line with the four wave patterns (Figure 4G).

3.5 Pulp cell subtypes exhibited a distinct spatial distribution

We applied the Seurat anchor-based integration algorithm¹⁶ to map single-cell data to the spatial transcriptome (Figure 5A). In the 17 PCW primary tooth germ, we observed that follicle, distal pulp, apical pulp, and epithelium layers distributed from distal to proximal end, in line with their order in the developmental trajectory (Figure 4A). In the 24 PCW primary tooth germ, we further discovered that FBN2+ OD was at the outer layer of pulp cells, which consisted of TWIST2+ DP and FGF3+ OD. One exception was the sparsely distributed TNC+ DP in permanent tooth germ without significant spatial structure (Figures 5A and S8). We plotted the expression of its marker TNC and HEY1 and also found no matching patterns (Figure 5B). We inferred that TNC+ DP was loosely distributed in all regions of pulp tissues. As a comparison, the density as well as the marker expression of other mesenchymal cell subtypes (Figures 5B, S8 and S9) were all aggregated to specific regions. The temporal distribution (Figure 5C) showed that TNC+ DP and two follicle and subgroups were mainly found in immature permanent tooth germ (p < 10⁻¹⁰), whereas odontoblast and pulp cells were mainly found in mature primary tooth germ (p < 10⁻¹⁰).

We also applied misty²² to quantify the spatial adjacency among cell subtypes and key TFs (Figure 5D). Within mesenchyme cell groups, we found a close relationship between two subgroups of follicle cells (misty adjacency score >2.5) as well as FGF3+ OD and FBN2+ OD (misty adjacency score >2). Apart from mesenchyme cell groups, only epithelium and mitotic cells had a close relationship with pulp cells (adjacency score >2). For TFs, pulp cells had close association with nearby expression values of multiple TFs that played important roles in mesenchyme differentiation, except for FOSB which took effect in early progenitors. Another spatial association was found between FBN2+ OD and MSX2 gene expression (adjacency score = 2.02).

3.6 Signalling network analysis of dental development

After annotating the refined cell subtypes of the embryonic tooth germ, we were able to study the intercellular signalling pathways that regulate dental development. We first applied nichenetr²¹ on scRNA data and misty²² on ST data to build a signalling network that regulates a branch-specific key genes in epithelial development (181 key genes included in nichenetr; Figure 6A). By integrating the prioritized upstream ligands of the both root branch, SI branch and ameloblast branch, it was found that signals from the mesenchyme (including follicle cells) and pericytes regulated the largest number of epithelial key genes (n = 142 and 92, respectively, Figure 6 and Table S10). The top credible ligands, including BMP5, SFRP2, JAG1 and HMGB2 (Figure 6B,C and Table S11), were identified. Among them, JAG1 was predicted to regulate the largest number (n = 91) of key genes in epithelial differentiation, including GJA1, NOTCH1, and so on. Two notable exceptions were GPNMB and SHH: they were the only prioritized ligands that showed the highest expression in ameloblasts (Figure S10) and had high potential to regulate SI and ameloblast branch development (AUC = 0.59 and 0.53, respectively), suggesting that they might act through autocrine or paracrine signalling. Despite their expression in epithelium, they were also found to be expressed in NCC (Figure S10).

Finally, we analysed the intercellular signalling networks that regulate dental mesenchyme development. For all 162 mesenchyme key genes (genes of the first, second and fourth waves of Figure 4D), 135 were regulated by ligands from pericytes, and 133 were regulated by ligands from epithelium (Figure 7). These included ligands such as BMP5, BMP7, DSC3, COL4A1 and so on. Since various lines of evidences predicted that COL4A1 was interacted with multiple mesenchymal trajectory genes, considering its known role in the basement membrane and angiogenesis, we hypothesized that COL4A1 might regulate the spatial distribution of mesenchyme cells. By applying misty, we found that COL4A1 was indeed spatially adjacent to the apical pulp (p < 2.2 × 10⁻¹⁶), supporting our hypothesis. We proposed that the high expression of COL4A1 represented the formation of the endothelial and vascular niche, which is vital for mesenchyme differentiation.²⁴ We also found that different ligands tended to regulate different stages of mesenchyme development. For example, BMP5 and BMP7 regulated all first-wave genes and over 70% of the second-wave genes that were included in the nichenetr analysis, whereas all predicted targets of APP (mainly expressed in endothelium) and TGFB1 (mainly expressed in macrophage) belonged to the fourth wave (Figure S11 and Table S12). In general, upstream signalling pathways of mesenchyme development were restricted to a few vital ligands and cell types, such as BMP signals from pericyte and epithelium. Finally, we showed that the spatial distribution of these signalling molecules was in accordance with their predicted functions, as detailed in Supplementary Materials in Data S1; Table S13; Figures S12 and S13.