Single‐cell analysis reveals dynamic changes of neural cells in developing human spinal cord

Abstract

During central nervous system development, neurogenesis and gliogenesis occur in an orderly manner to create precise neural circuitry. However, no systematic dataset of neural lineage development that covers both neurogenesis and gliogenesis for the human spinal cord is available. We here perform single‐cell RNA sequencing of human spinal cord cells during embryonic and fetal stages that cover neuron generation as well as astrocytes and oligodendrocyte differentiation. We also map the timeline of sensory neurogenesis and gliogenesis in the spinal cord. We further identify a group of EGFR‐expressing transitional glial cells with radial morphology at the onset of gliogenesis, which progressively acquires differentiated glial cell characteristics. These EGFR‐expressing transitional glial cells exhibited a unique position‐specific feature during spinal cord development. Cell crosstalk analysis using CellPhoneDB indicated that EGFR glial cells can persistently interact with other neural cells during development through Delta‐Notch and EGFR signaling. Together, our results reveal stage‐specific profiles and dynamics of neural cells during human spinal cord development.

Using single‐cell nuclei and single‐cell RNA‐seq, this study reveals a developmental timeline and uncovers genetic heterogeneity in neural cells during embryonic and fetal stages of human spinal cord development.

Introduction

The spinal cord is a highly organized tissue, composed of three distinct neural cell types: neurons, astrocytes, and oligodendrocytes. The vertebrate spinal cord transmits signals between the periphery and brain and vice versa. The sensory dorsal horns mediate exteroceptive signals, while the ventral cord is critical for the execution of locomotor behaviors (Bican et al, ²⁰¹³; Puelles & Martinez, ²⁰¹³; Leung & Shimeld, ²⁰¹⁹). Understanding spinal cord development will reveal fundamental principles of how the central nervous system is built. Although model systems have been studied to reveal this process, comparatively few investigations have examined the human spinal cord (Bradley et al, ²⁰¹⁹). In rodent spinal cord, proliferative progenitor cells originated from the neuroepithelium of the ventricular zone during neural induction. These cells, which are maintained by SOX2, can give rise to all neurons and glial cells in the spinal cord (Graham et al, ²⁰⁰³; Ellis et al, ²⁰⁰⁴; Pevny & Nicolis, ²⁰¹⁰). The antiparallel signaling gradients formed by sonic hedgehog (SHH) and bone morphogenic proteins regulate the transcriptional network in progenitor cells, ultimately determining neuronal subtype identity within a Cartesian‐like coordinate system (Caspary & Anderson, 2003; Zagorski et al, ²⁰¹⁷). Along with this spatial information, glial progenitors expressing the transcription factors SOX9 and NFIA are born in a precise order from the same progenitor cell zone following specification of neuronal progenitors (Miller & Gauthier, 2007; Kang et al, ²⁰¹²). During spinal cord development, radial glial cells served as both progenitors and a migration guidance scaffold for neural cells. Radial glial cells are also important for regulating axon outgrowth and pathfinding processes during gliogenesis (Puche & Shipley, 2001; Brusco et al, ²⁰⁰⁹; Barry et al, ²⁰¹³). The differentiation and lineage relationships of radial glial cells are regulated by genetic factors, cell–cell interaction, and microenvironmental factors (McDermott et al, ²⁰⁰⁵).

Single‐cell RNA sequencing has become a powerful tool to dissect tissue development at single‐cell resolution. Several single‐cell RNA sequencing studies of spinal cord development have been performed in model systems; however, as most of these studies focused on neurogenesis, there is still a lack of single‐cell transcriptomics information for glial development, especially for astrocytes. Glia are a central component of the nervous system. Recent studies point to an increasingly broad spectrum of roles for glial cells both during development and in the mature central nervous system (Freeman & Rowitch, 2013). Single‐cell analysis would provide more information on the diversity and development of human glial cells. Furthermore, compared with model vertebrates, the human embryo experiences a long gestation period and extended duration of neurogenesis and gliogenesis. This prolonged development is advantageous for analyzing neural cell subtypes and their transitional states during maturation.

Accordingly, we performed both high‐throughput single‐cell nuclei RNA sequencing (snRNA‐seq) and single‐cell cellular RNA‐seq (scRNA‐seq) of the developing human spinal cord. Altogether, we analyzed 0.8 million cells from human embryonic spinal cords covering gestational week (GW) 7‐ GW 23. Cells were clustered according to their transcriptional state and subsequently visualized using Uniform Manifold Approximation and Projection (UMAP) plots, in which related cell types were placed in proximity to one another (Cao et al, ²⁰¹⁹). Our findings describe development from an early stage of neuron generation to the differentiation of glial cells. Moreover, we identified a type of epidermal growth factor (EGFR)‐expressing transitional glial cell that arises from neuronal progenitor cells during the late stage of neurogenesis and thereafter shifts progressively toward the astrocyte and oligodendrocyte lineages. Immunostaining showed that EGFR‐expressing transitional glial cells adopted a markedly position‐specific distribution during spinal cord development.

Results

snRNA‐seq and scRNA‐seq of the developing human spinal cord

We obtained spinal cord samples from GW7 to GW23 (scRNA‐seq: GW7, 8, 9, 10, 11, 12, 13, 15, 16, 17, 19, 20, 22, 23; snRNA‐seq: GW7, 8, 10, 11, 14, 18, 20, 21, and 23). GW was measured from the first day of the woman’s last menstrual period. Samples were processed for snRNA‐seq and/or scRNA‐seq (Fig 1A and Table EV1). According to the region from which cells were collected, data were designated as cervical (C), thoracic (T), lumbar (L), or whole spinal cord (S). After applying quality filters 827,265 cells were retained for further analysis (see Materials and Methods, Appendix Figs S1A, and S2A and B). Data were projected onto two dimensions via UMAP, and major cell types were identified based on differentially expressed genes (DEGs) (Dataset EV1) and gene ontology (GO) analyses. Eight major classes were identified neurons, astrocytes, oligodendrocytes, ependymal cells (including roof and floor plate cells), neural progenitor cells (NPCs), microglia, meningeal cells, and vascular cells (Fig 1B and Dataset EV2). A discrepancy in neuronal cell types was found between the two sets of sequencing results (Appendix Fig S1B, Dataset EV2). Neurons were undetectable after GW13 in the scRNA‐seq dataset, whereas neuronal clusters were consistently detected in the snRNA‐seq dataset. This may be because mature neurons were more vulnerable to the cell separation and enzyme digestion processes of scRNA‐seq preparation. To rule out the possibility that cell preparation affected the transcriptional characteristics of cells, batch correction and integration analyses were performed using datasets from similar gestational stages (Appendix Fig S3A–D). Integration results showed no obvious differences in cluster distribution of the GW7 dataset. Although neuronal cell types were missing from scRNA‐seq data with increasing GW, numbers of other neural cell types (astrocytes, oligodendrocytes, and NPCs) in datasets correlated well with gestational stages. These results indicate that the transcriptional profiles of cells were unaffected by scRNA‐seq preparation. Hereafter, our analysis concentrated on snRNA‐seq data, with scRNA‐seq data used as a contrasting dataset for glial cell types. To assess for potential region‐specific effects, snRNA‐seq datasets of cervical, thoracic, and lumbar regions from the same spinal cord sample were compared at three gestational ages (GW10, GW14, and GW20). When data of different regions from matched samples were integrated, we found no significant difference in cluster distributions or proportions among segments (Appendix Fig S4A). Moreover, clusters from different segments shared similar DEGs (Dataset EV1), with the exception of HOX family genes, which are related to the polarization of axial patterning (Appendix Figs S4B and S5A–D).

Figure 1. Integrated snRNA‐seq data reveal transcriptomic features and changes in the proportions of neural cells in early and mid‐gestation human spinal cord.

1. A schematic diagram of the single‐cell sequencing process.
2. Visualization of major classes of spinal cord cells by integrated UMAP plots of cell clusters. Cell types are labeled and circled with dotted lines (top). Violin plots show the most distinct and commonly used marker genes for each cell cluster (bottom). Cell‐type annotations are listed on the right and shown in varying colors.
3. Three‐dimensional visualization of neural subtype developmental trajectories for individual samples from GW7, GW10, GW14, GW18, and GW23. Stage‐dependent differentiation trends changed with development. The red arrow indicates the proliferative neural progenitor cell cluster (MKI67, PAX6, NES, and SOX2). Neural cell types are named and circled with dotted lines. Gray represents non‐neural cell types (e.g., vascular and meningeal).
4. Immunofluorescence staining of the spinal cord at GW7, GW8, GW10, GW11, and GW23 for RBFOX3 (red), OLIG2 (green), and GFAP (white). The appearance of astrocytes and oligodendrocytes was observed from GW8 onward. Scale bars: 100 μm.

Glial differentiation commenced around GW8 and ventricular zone specification occurred around GW10

NPCs are capable of proliferating and differentiating into neuronal and glial cell types (Alaynick et al, ²⁰¹¹; Hagey et al, ²⁰¹⁸; Ma et al, ²⁰¹⁸). To investigate differentiation trends of NPCs during development, differentiation trajectories within individual samples were examined (at GW7, GW10, GW14, GW18, and GW23) (Fig 1C). Proliferative NPCs were defined by expression of the proliferation‐related transcription factor MKI67 and neural progenitor markers NES and SOX2, as well as the absence of expression of genes characteristic of matured neural cells. At GW7, proliferative NPCs were associated with neuronal classes. Expression of glial progenitor‐specification genes SOX9 and NFIA was observed in GW7 samples (Appendix Fig S6A–C). We did not observe expression of differentiation markers for astrocytes (GFAP and ALDH1L1) or oligodendrocytes (PDGFRA and SOX10) in GW7 spinal cord glial cell subtypes (Appendix Fig S7A–C). Glial differentiation emerged as early as GW8, when upregulated expression of GFAP, ALDH1L1, PDGFRA, and SOX10 was observed in glial cells (Appendix Fig S7D–F). Expression of HOPX and SLC1A3 (GLAST) was observed in GW8 progenitor cell clusters, which indicated the appearance of radial glia (Penisson et al, ²⁰¹⁹) (Appendix Fig S7F). NPC clusters lost their connection with neuronal groups after GW14, while links with astrocyte and oligodendrocyte clusters remained. After GW22, most dividing NPCs tended to be linked with the oligodendrocyte group (Fig 1C). Expression of myelination genes also occurred at an earlier developmental stage (GW13) in human (Appendix Fig S8) compared with mouse (1 day prior to birth) (Foran & Peterson, 1992). Immunofluorescence analysis also showed that RBFOX3‐positive neuronal cells were the primary cell type in GW7 spinal cord. GFAP‐positive astrocytes and OLIG2/PDGFRA‐positive oligodendrocytes were first observed at GW8. In GW8 samples, GFAP‐positive astrocytes were localized in the roof plate area of the spinal cord, while OLIG2‐positive oligodendrocyte progenitors (OPCs) were observed in the area ventral to the ventricular zone. After GW10, both GFAP‐ and OLIG2‐positive glial cells were found in the ventral spinal cord. From GW10 onward, GFAP‐positive astrocytes showed a radial glia morphology, which could assist glial cells migration. The spinal cord achieved its mature morphology around GW23 (Fig 1D). Collectively, these results reflect the dynamic production of the glial cell population during development.

Proliferative NPCs reside in the ventricular zone (later becomes the ependymal zone) of the developing spinal cord—an area that changed dramatically during embryonic development (Appendix Fig S9A–C). After scanning a set of adult human ependymal‐specific genes (Ghazale et al, ²⁰¹⁹), we detected expression of primary cilia formation related genes, CCDC114 and NEK5 starting from GW10. Expression of the cilia‐related gene DNAAF1 was also detected at this time. Moreover, the adult ependymal zone‐specific gene, ODF3B, was observed at GW21 (Appendix Fig S9B). SHH is responsible for regulating neurogenesis, especially motor neurons specification, and is expressed by the floor plate at the beginning of spinal cord development (Ericson et al, ¹⁹⁹⁶; Lee & Pfaff, ²⁰⁰¹). Our data indicate that after initiation of cilia‐related genes expression within the ependymal cluster at GW10, expression of SHH in this cluster sharply declined from GW10 to GW14. Meanwhile, motor neurons (expressing MNX1) began to express SHH for further maturation (Appendix Fig S9B). These results indicate that NPCs of the ventricular zone developed into ciliated ependymal cells during the fetal stage.

Neuronal development and cross‐species comparison

To gain insight into the neuronal differentiation process, subcluster analysis of neuronal lineages based on NPCs and neuronal cells from GW7 to GW11 was performed (Appendix Fig S10A and B). Neuronal progenitors were identified by expression of basic helix–loop–helix and paired box transcription factors such as HES6, DLL1, and PAX3. Excitatory (Exc), inhibitory (Inh), and motor (MN) neuron populations were distinguished by expression of SLC17A6, GAD1, and MNX1, respectively. Neuronal classes could be further subdivided (e.g., DI3, DI4, DI5) (Appendix Fig S10A). Heatmap visualization revealed the top 10 DEGs among clusters (Fig 2A, Dataset EV2). Selected GO enrichment analyses of differential expression data indicated that neuronal progenitors were enriched for translation, initiation, and membrane‐targeting proteins, whereas immature dorsal neurons were enriched for GO terms related to neuron differentiation and synaptic signaling (Fig 2B). We observed that mitotic neuronal precursor clusters strongly expressed the dorsal neural progenitor gene PAX3, while expression of the ventral gene NKX6‐2 was less detectable in this population (Appendix Fig S10C and D). Immunostaining confirmed that the dorsal side of the ventricular zone was filled with neuronal progenitors after GW7 (Fig 2C). Across all integrating time points, there were 50,533 cells, including proliferative NPC (pro‐NPC), glial progenitor (GPC), neuronal restricted progenitor (NRP), immature dorsal neuron (IDN), and differentiated neuronal classes. Next, we determined the relative proportion of each cell population at each developmental time points to estimate developmental trends (Fig 2D). The proportion of neuronal progenitors declined with increasing gestational age (21.5% at GW7 compared with 3.0% at GW11). Significant increases in the proportions of DI4 and DI5 cells were observed from GW7 to GW11 (0.4% at GW7 compared with 19.2% at GW11; GW7, 0.1% at GW7 compared with 24.2% at GW11, respectively). Although there was a suggestion of ventral neurons development at GW7, our data indicated that dorsal neuron development is the main event during this stage (GW7–GW11). Analysis of GW10 (Fig 2E) also showed that the neuronal differentiation path was enriched for dorsal neuron transcriptional regulators, such as POU4F1 and TLX3. Ventral neuron and MN clusters were located at considerable distances from the neuronal progenitor class in the trajectory plot, hence reflecting a more mature state. These data indicated that dorsal neurons generation peaks around GW10, while ventral neuronal generation occurs earlier.

Figure 2. Integrated trajectory and expression feature analysis of neuronal cell developmental trajectory in GW7, GW8, GW10, and GW11 samples.

1. UMAP plot of cells integrated from datasets of GW7, GW8, GW10, and GW11 samples (GW7S‐NC, GW8S‐02NC, GW10C‐01NC, GW10T‐01NC, GW10L‐01NC, GW10C‐02NC, GW10T‐02NC, GW10L‐02NC, GW11C‐NC, GW11T‐NC, GW11L‐NC) showing neuronal differentiation tendency. Cell types are labeled (left). Heatmap showing differentially expressed genes (DEGs) in neuronal restricted progenitors (NRP) and immature dorsal neurons (IDN) (right).
2. Visualization of selected top gene ontology (GO) enrichment terms for NRP and IDN clusters based on DEG analysis.
3. Immunofluorescence staining of the spinal cord from GW7, GW8, and GW10 for a dorsal marker (PAX7, green), neuronal precursor marker (ASCL1, red), and nuclei marker (DAPI, blue). Images show the neuronal differentiation process on the dorsal side of the spinal cord. Scale bar: 100 μm.
4. Stacked proportional changes of neural progenitor cells (NPC) and neuronal cell classes from GW7–GW11. Cell types were identified as: proliferative neural progenitors (Pro‐NPC), glial restricted progenitors (GRP), IDN, DI4 interneurons (DI4), DI5 interneurons (DI5), motor neuron (MN), DI1 and DI2 interneurons (DI1/2), DI3 interneurons (DI3), DI6 interneurons (DI6), and V0 and V1 interneurons (V0/1).
5. Three‐dimensional developmental trajectories for snRNA‐seq samples from the GW10 cervical sample (top view in the upper right corner). Cells are classified by differentiation stage according to expression of specific markers (neuronal progenitor: PAX3 and PAX7; DI5 interneuron: POU4F1 and TLX3; V3 interneuron: SIM1; and motor neuron: MNX1). Cells in the neuronal development trajectory are colored by gene expression. Distances from progenitor cells to differentiated neurons represent the differing maturation status between dorsal neurons (DI4, DI5) and ventral neurons (MN, V3). Cells with no detectable expression for a given gene were omitted from the plot.

Cross‐species analysis can reveal the homology and heterogeneity of neural cells between species (Yu et al, ²⁰¹⁴; Hodge et al, ²⁰¹⁹). Thus, we next aligned our human neurogenesis trajectory results (GW7–GW11) with mouse spinal cord single‐cell neurogenesis transcriptomics data (E9.5–E13.5) (Delile et al, ²⁰¹⁹). Alignment based on expression covariation showed that the neural composition was largely conserved between species, confirming the reliability of our data (Fig EV1A, B and E). Our alignment results showed that human neuronal progenitor subtypes (NPC‐Neuronal NOTCH1 and NPC‐Neuronal HES6) were most similar to mouse PD4/5 subtypes (NPC‐Neuronal ASCL1, NPC‐Neuronal CENPA, and NPC‐Neuronal TFAP2B) (Fig EV1C). Furthermore, GFAP expression was broadly detected in humans but not mouse (Fig EV1D), although we noticed the appearance of astrocyte progenitor cluster in mouse dataset from E10.5 (Fig EV2H). Gene families with the most divergent expression (> 5‐fold change) showed a change in patterning across cell types. These gene families included extracellular matrix elements, neurotransmitter receptors, ion channels, and fatty acid transporters (Fig EV1F and G, and Table EV2). Our data indicated that between human and mouse, intermediary DI3 interneuron, and ipsilaterally projecting V2a interneuron classes exhibited more conserved cellular properties compared with most of the remaining classes (Fig EV2A–D and F). Furthermore, five human neuronal subclasses lacked homologous mouse equivalents: Exc‐DI5 PIEZO2, Exc‐DI5 ROBO, Inh‐DI4 ROBO, Inh‐VI ONECUT3, and MN COL6A2 (Fig EV2E). This may reflect either a difference among species or development phases. For example, because expression of PIEZO2 plays a role in the mechanotransduction channel for proprioception, the Exc‐DI5 PIEZO2 subcluster in humans may represent a refined mechanosensation function (Florez‐Paz et al, ²⁰¹⁶) (Fig EV2G). Similarly, in early motor neuron development classes, a subcluster of human MN COL6A2 may indicate an MN subcluster with promoted axon growth as COL6A2 deficiency has been related to severe motor axon outgrowth defects (Ramanoudjame et al, ²⁰¹⁵). Furthermore, we noticed several genes (such as CHD5, EGFR, GAD1, and NEUROG2) exhibited different expression patterns between human and mouse (Appendix Fig S11). These genes are related to several important aspects of spinal cord development, including neuronal gene activation, synaptic vesicle transporters, and extracellular matrix structure.

Figure EV1. Alignment analysis of cell‐type conservation and divergent expression between human and mouse.

1. UMAP visualization of human (n = 50,533 nuclei) and mouse (n = 26,039 cells) neuronal and neural progenitor cell (NPC) clusters after alignment.
2. Radar chart showing similar distribution of neuronal subtype proportions among the two species during the corresponding developmental period.
3. Human and mouse cell‐type homologies for NPCs predicted from shared cluster membership. Gray shade corresponds to the minimum proportion of co‐clustering between species. Columns show human clusters and rows show mouse clusters. Homologous clusters are labeled based on human and mouse cluster membership. Inferred cell‐type homologies highlighted in red boxes. Orange boxes indicate subclasses of PD4/5‐related neuronally restricted progenitors in human and mouse data.
4. Violin plots showing GFAP expression in human and mouse integrated data. Blue boxes indicate astrocyte subclasses in human and mouse.
5. Comparison of expression levels of orthologous genes between species for NPC‐Ependymal and EXC‐DI4 types.
6. Gene families exhibiting the most divergent expression patterns included neurotransmitter receptors, extracellular matrix elements, ion channels, and fatty acid transporters.
7. Expression of ionotropic glutamate receptors differed between homologous cell types. Scores are listed on the right.

Figure EV2. Alignment analysis of cell‐type conservation and divergent expression between human and mouse.

• A, B
  Integrated UMAP plots of NPCs and neuronal cell types from developing human (GW7–GW11) and mouse (E9.5–E13.5) spinal cord data. Red line show cell subtypes base on marker gene expression.
• C, D
  Human and mouse cell‐type homologies for inhibitory and excitatory neuron types, predicted based on shared cluster membership. Inferred cell‐type homologies highlighted in red boxes. Gray color corresponds to the minimum proportion of co‐clustering between species. Columns show human clusters, and rows show mouse clusters. Homologous clusters are labeled on the basis of human and mouse cluster membership.
• E
  Taxonomy of 22 neuronal and seven NPC‐homologous cell types and cell classes. Color icons represent different neural types. Arrows indicate one‐to‐one matches. Asterisks indicate subclasses lacking homologous types.
• F
  Similar functional gene families (n = 384 gene sets) discriminate inhibitory neuron types in human and mouse. Error bars correspond to the standard deviation of mean MetaNeighbour AUROC scores across ten subsamples of cells.
• G
  Violin plots showing expression of PIEZO2 in human and mouse integrated data (left). Red box indicates human Exc‐DI5 PIEZO2 subclass.
• H
  Time split UMAP plots of mouse data show the gradual appearance of an astrocyte progenitor cluster.

Glial cell development and region‐specific EGFR‐expressing glial cells

To evaluate the diversity of neural cells and to dissect their development in human spinal cord, we integrated datasets of neural cells from GW7 to GW23 (select sample data: GW7S NC, GW8S NC, GW10T_01 NC, GW11T NC, GW14T NC, GW18T NC, GW20T NC, GW21T NC, and GW23S NC). Integration resulted in faithful recapitulation of the cellular transcriptomic diversity of neural cell types (Fig 3A and C, Dataset EV2). Doublets and glial cells that lacked corresponding cell types between snRNA‐seq and scRNA‐seq data were excluded from further analysis. We calculated the proportion of each neural class at each stage. The proportion of the neuronal population increased dramatically before GW10 (39% at GW7 compared with 79.3% at GW10). Meanwhile, the astrocyte population (Clusters 0, 3, 7, and 14 including the progenitor population at early development stages) reached its highest proportion at GW20 (33.5%), and the oligodendrocyte population (Clusters 4, 6, 8, and 11) increased from GW8 onward to reach 51% at GW23 (Fig 3B). These findings are consistent with lineage and immunofluorescence analyses from individual samples at different developmental stages (Fig 1D). Integrated analysis revealed mature oligodendrocytes expressing NKX2‐2 from GW10, while expression of the myelination‐related genes MBP and MOG was observed in oligodendrocytes from GW11. The astrocyte population clustered into four groups (Clusters 0, 3, 7, and 14). While expression of GFAP, AQP4, and FGFR3 was generally found in all astrocyte subtypes, a distinct gene expression pattern was detected within each subtype. For example, expression of the axon navigation gene ROBO3 was observed in Cluster 3, while cluster 7 expressed the flagellum‐related gene LRRIQ1 and dorsal marker PAX3 (Appendix Fig S12A and B). GO enrichment analysis of DEGs showed that cluster 0 was enriched for genes related to neural precursor cell proliferation and glial cell differentiation, Cluster 3 was enriched for axonogenesis and neuron migration, Cluster 7 was enriched for cilium assembly and cilium organization, and Cluster 14 was enriched for glial cell development and astrocyte differentiation (Fig 3D).

Figure 3. Neural cell‐type integration analysis and characterization.

1. UMAP plot of snRNA‐seq data from 64,381 combined neural cells from GW7 to GW23. Cell cluster phenotypes are labeled and the red arrow indicates the EGFR‐expressing glial cluster (left). Selected UMAP plots from different development stages (GW7, GW11, GW18, and GW23) represent the neural development trend (right). Gray represents non‐specific cell types.
2. The stacked proportions of neural cell classes showing changes in stage‐dependent differentiation trends with developmental stage.
3. Violin plot showing the expression of selected genes. EGFR⁺ Cluster 14 is circled with a red dotted line.
4. Visualization of selected top gene ontology (GO) enrichment terms of astrocyte clusters based on DEG analysis.

Notably, we observed unique expression patterns in Cluster 14. For instance, EGFR and the neural development regulator genes DLL1 and HES6 were uniquely expressed in Cluster 14. Moreover, Cluster 14 exhibited expression of radial glial genes HOPX and SLC1A3. In addition, NESTIN and MIAT were expressed in this EGFR‐positive cluster, which indicated a possible progenitor property of these cells (Fig 3C).

We noticed Clusters 14 and 0 were composed of cells from all gestational stages (GW7–GW23) while Clusters 3 and 7 were comprised of cells that only appeared after GW11 (Fig 3A). This difference in cellular composition suggested that Clusters 14 and 0 contained progenitor cells that arose before glial differentiation. Investigation of Clusters 14 and 0 from GW7 by GO enrichment analyses of DEGs showed that Cluster 14 was enriched for terms associated with spinal cord development and neuronal fate commitment (Fig 4A and B). Cluster 0 was enriched with glial fate commitment genes (such as DAAM2, FAM189A2, and ID4) and exhibited a differentiation‐arrest state at GW7 (Fig 4A and B), consistent with our results showing no glial differentiation at the GW7 stage. Thus, our results indicate that at GW7, Cluster 14 cells are neuronal progenitors, while Cluster 0 represents resting glial progenitors.

Figure 4. Characterization of GW7 progenitor cell populations and EGFR‐expressing transitional glial cells.

1. UMAP plot of Cluster 0 and Cluster 14 populations from the GW7 sample.
2. Heatmap of differentially expressed genes (DEGs) in Clusters 0 and 14 at GW7. On the right of the heatmap are the top‐twelve DEGs. Selected top gene ontology (GO) terms related to the corresponding DEGs (right).
3. Single‐cell resolution trajectory of EGFR⁺ transitional glial cells (Cluster 14) was constructed from sample (GW11T‐NC, GW14T‐NC, GW18T‐NC, GW20T‐NC, GW21T‐NC) during the transitional period using Monocle. Linear arrangement of the transition trajectory with developmental stages represents the transition process of a regulatory cell type. The arrow indicates the developmental trend.
4. Violin plots produced by VlnPlot function of Seurat package representing changes in expression of EGFR, DLL1, MIAT, and FAM189A2 in cluster 14 with development based on snRNA counts data.
5. Overview of selected statistically significant interactions between neural cell types according to cell–cell communication analysis from statistical test of CellPhoneDB v.2.0. Size indicates P value, and color indicates the means of the average expression level of the receptor–ligand pairs. The EGFR⁺ glial cell type exhibited the most prominent interactions with Delta‐Notch and EGFR signaling within the network.

Further analysis indicated that cells from Cluster 14 gradually acquired an astrocyte expression profile after GW11. To illustrate this fate‐transformation process, we construct a single‐cell trajectory showing the cell‐state transition process. Cells from late neurogenesis to the gliogenesis stage of male samples (GW11T‐NC, GW14T‐NC, GW18T‐NC, GW20T‐NC, GW21T‐NC) formed a continuous linear trajectory according to developmental stage, representing the transition process of EGFR⁺ glial cells (Fig 4C). During this process, EGFR expression increased after GW11. A high level of HES6 expression was constantly found within the EGFR⁺ glial cell cluster. DLL1 expression decreased after GW18 and subsequently remained relatively low. Meanwhile, a decrease in MIAT expression, which regulates neural progenitor differentiation and survival of newborn neurons (Aprea et al, ²⁰¹³; Wang et al, ²⁰¹⁷), was observed after GW14. Expression of astrocyte‐related genes FAM189A2, GFAP, and AQP4 continually increased from the late neurogenesis stage (GW11) (Figs 4D and EV3A). We further analyzed the lineage relationship of EGFR‐expressing cells. By examining the developmental trajectories of individual sample data, we ensured that the spatiotemporal information of all trajectory‐relevant factors was uniform. The results of trajectory analyses were consistent with integrated analyses showing that EGFR⁺ cell clusters (co‐expressing EGFR, DLL1, and HES6) were mainly related to the astrocyte lineage (Fig EV3B–D). Next, we predicted the intercellular communication network between neural cell types based on potential receptor–ligand pair interactions (Efremova et al, ²⁰²⁰). This analysis indicated interactions between EGFR⁺ cells and other glial cell types (namely astrocytes, ependymal, NPCs, OPCs, and oligodendrocytes) through Delta‐Notch ligand–receptor and EGFR signaling pathways (Fig 4E). The Notch signaling pathway plays critical roles in diverse developmental and physiological processes (Artavanis‐Tsakonas et al, ¹⁹⁹⁹). Our analysis showed that Notch1 was expressed in all astrocyte populations (Clusters 0, 3, 7, and 14) and OPCs (Cluster 17) while DLL1 expression was enriched in the EGFR⁺ cluster (Cluster 14) (Fig 3C).

Figure EV3. Altered gene expression levels and differentiation trajectory of EGFR⁺ glial cells at different gestational time points.

• A–D
  (A) Violin plots produced by VlnPlot function of Seurat package representing expression levels of HES6, NES, GFAP, and AQP4 in the EGFR⁺ glial cluster across developmental process based on snRNA counts data. The EGFR⁺ glial cluster cell type (DLL1 ⁺ EGFR ⁺ HES6 ⁺) could be found in GW9T (B), GW16T (C), and GW20T (D) scRNA‐seq datasets (left, red circle). Dot plots showing expression of DLL1 and EGFR were predominantly found within the EGFR⁺ glial cell type (middle). Pseudotime analysis suggesting that EGFR⁺ glial cells serve as neuronal progenitors in neuronal differentiation trajectory during neurogenesis stage and jion astrocyte differentiation trajectories in gliogenesis stage (right).

To further investigate EGFR‐expressing cells, we performed immunostaining of the spinal cord samples from GW7 to GW23. We found that the abundance of EGFR⁺ cells was low at GW7 but increased dramatically around GW11. Furthermore, the morphology of these cells changed over time. We also noticed that the distribution of EGFR⁺ cells showed position‐specific features (Fig 5A and B). There was weak expression of EGFR on the ventral side of the ventricular zone at GW7. However, a large population of EGFR⁺ cells was observed both the dorsal and ventral sides of the ventricular zone at GW11. Dorsal EGFR⁺ cells had a typical radial glial morphology, including a single thin process extending from the ventricular zone to the pial surface that could be double labeled with GFAP (Fig 5D). The nuclei of these cells, which were initially located in the dorsal ventricular zone and accompanied by SOX9 expression, exhibited only low OLIG2 expression (Fig EV4A–F). Meanwhile, colocalization of EGFR and OLIG2 was observed in the ventral EGFR⁺ cell population (Figs 6A and B, and EV4C), consistent with our sequencing data. In GW11 scRNA‐seq data, the EGFR ⁺ cluster showed a distinct expression pattern; namely, most EGFR ⁺ PAX3 ⁺ cells did not express OLIG2, whereas OLIG2 expression was detected in PAX3‐negative EGFR⁺ cells (Fig 5C). According to immunostaining of samples from GW17‐GW23, dorsal EGFR⁺ cells lost their radial morphology and ventral EGFR⁺ cells nearly disappeared. An increasing number of dorsal EGFR⁺ cells appeared to translocate their cell bodies to the pial surface, and most were concentrated within the dorsal horn area (Fig 5B). Co‐expression of EGFR and the astrocyte lineage marker SOX9 or oligodendrocyte lineage markers OLIG2 was observed outside the ventricular zone, which indicates that EGFR⁺ glia were a transitional state that could differentiate into both astrocytes and oligodendrocytes (Fig 6C and D). Notably, we found difference in EGFR expression between human and mouse spinal cord in transcriptomics data (Appendix Fig S11) and immunostaining results (Appendix Fig S13).

Figure 5. Expression profile of EGFR‐expressing glia during spinal cord development.

1. Schematic diagram showing the spatial and temporal dynamics of EGFR‐expressing cells during spinal cord development. The solid green line represents EGFR‐expressing radial glial cells, while the green dotted line represents EGFR‐expressing transitional glial cells. The blue line represents the gray matter boundary.
2. Immunofluorescence staining of spinal cord sections at GW7, GW11, GW17, and GW23 for EGFR (green) and nuclei (DAPI, blue). EGFR expression on the ventral side was observed in GW7 and GW11 spinal cord. EGFR expression on the dorsal side was observed in GW11, GW17, and GW23 spinal cord. Dorsal EGFR expression was mainly concentrated within the dorsal horn area. Scale bar: 100 μm. (Top) Higher‐magnification view of boxes at bottom. White dashed lines indicate dorsal horn borders.
3. UMAP plot of GW11 sample. Dot plot shows the expression of featured genes in the EGFR‐expressing progenitor cluster mapped on the UMAP plot. Expression of PAX3 and OLIG2 displays a complementary pattern, while PDGFRA was expressed at a low level within the EGFR⁺ glial cluster.
4. Representative immunofluorescence staining of a GW11 spinal cord section. Most dorsal EGFR‐expressing cells at GW11 co‐expressed GFAP and SOX9, but not OLIG2. Scale bars: 100 μm.

Figure EV4. Characterization of EGFR‐expressing cells in GW11 human spinal cord.

• A–E
  Co‐staining of EGFR with GFAP (A) SOX9 (B) OLIG2 (C) IBA1 (D) and NOTCH1 (E). Higher‐magnification view of boxes at the bottom. The SOX9 staining image is the same as the immunofluorescence image shown in Fig 6A.
• F
  Immunofluorescence staining of GW11 spinal cord with glial lineage marker SOX2 (green) and neuronal lineage marker NEUN (red). Higher‐magnification view of boxes at the bottom. Scale bars: 100 μm.

Source data are available online for this figure.

Figure 6. Comparison of EGFR‐expressing cell features between GW11 and GW17 spinal cord.

• A–D
  GW11 EGFR‐expressing cells exhibited radial glia morphology and most co‐expressed SOX9 (A) but not OLIG2 (B). However, EGFR co‐expression with SOX9 or OLIG2 was found in GW17 EGFR‐expressing cells (C, D). White dashed lines marked dorsal horn borders. Higher‐magnification view of boxes on left. Scale bars: 100 μm.

Gene expression analysis showed the expression of several astrocyte lineage genes in the EGFR⁺ glial cluster, such as SOX9, FGFR3, GFAP, and AQP4 (Fig 3C). We also observed a high level of OLIG2 expression in the EGFR⁺ glial cluster, whereas the expression of other OPC genes (such as PDGFRA) was present only at extremely low levels (Fig 5C). EGFR ⁺ OLIG2 ⁺ cells might be in the early stage of OPC development, consistent with a recent report of human cortical OPC development (Huang et al, ²⁰²⁰). Our results indicated that the EGFR⁺ glial cluster represents a heterogeneous glial population. Our scRNA‐seq result also showed EGFR and reported ventral astrocyte subtype identity genes (PAX6, SLIT1) (Hochstim et al, ²⁰⁰⁸) are differentially expressed among the different astrocyte subsets (Appendix Fig S14). This was consistent with our immunostaining results that EGFR expression are mainly located in dorsal area of human spinal cord.

A recent study by Fu et al reported the heterogeneity of EGFR⁺ cells in the developing human cerebral cortex. They divided cortex EGFR⁺ cells into twelve subtypes: cells not in G0 of the cell cycle (Cycling cells), intermediate progenitor cell 1 (IPC1), intermediate progenitor cell 2 (IPC2), immature inhibitory neurons (IN), primitive oligodendrocyte progenitor cells (PriOPC), oligodendrocyte progenitor cells (OPC), HOPX ⁺ SPARCL1 ⁺ glial progenitor cells (OAPC), astrocyte progenitor cell 1 (APC1), astrocyte ependymal cell (AstroEpen), and endothelial cells (Endo) (Fu et al, ²⁰²¹). To further characterize EGFR‐expressing glial cell populations in the spinal cord, we next constructed a reference dataset based on cortical EGFR⁺ cells of the brain. Next, we selected spinal cord EGFR⁺ cell data from samples of four different gestation stage (GW11T, GW16T, GW20T, and GW23S) and projected our data onto the reference dataset (Fig EV5A and B). This analysis offers a way to investigate features of the spinal cord EGFR⁺ population. We showed that spinal cord EGFR⁺ cells mapped well to most of the cortical EGFR⁺ clusters, with the exception of IPC1 and APC1. The main projected subtypes from spinal cord EGFR⁺ cells were cycling cells, OAPC, PriOPC, IPC2, and OPC (Fig EV5A–C). We also noticed that the composition of spinal cord EGFR⁺ cell subtypes changed over time; specifically, the proportion of EGFR⁺ IPC2 and EGFR⁺ PriOPC subtypes decreased with gestational age (IPC2: 14.4% at GW11 compared with 0.1% at GW16; PriOPC: 37.1% at GW11 compared with 27.3% at GW20), whereas the proportion of EGFR⁺ OPC subtypes increased from 1.7% at GW11 to 40.3% at GW23 (Fig EV5D and E). These results indicate that EGFR⁺ cells in cord represent a heterogeneous glial population whose progenitor characteristics change during gestation.

Figure EV5. Characterization of spinal cord EGFR‐expressing cells based on brain cortex data.

1. UMAP plot of spinal cord EGFR⁺ dataset mapped alongside human brain cortex EGFR⁺ dataset (Fu et al, ²⁰²¹). Cycling cells, cells not in G0 of the cell cycle; IPC2, intermediate progenitor cell 2; IN, immature inhibitory neuron; PriOPC, primitive oligodendrocyte progenitor cell; OPC, oligodendrocyte progenitor cell; OAPC, HOPX ⁺ SPARCL1 ⁺ glial progenitor cell; AstroEpen, astrocyte ependymal cell; Endo, endothelial cell.
2. Separate UMAP plots showing the subtype distribution of the reference brain cortex EGFR⁺ dataset (left) and spinal cord EGFR⁺ dataset (right).
3. Heatmap showing differentially expressed genes (DEGs) of spinal cord EGFR⁺ subtypes.
4. Time split UMAP plots of the spinal cord EGFR⁺ dataset shows the subtype distribution EGFR⁺ cells at GW11, GW16, GW20, and GW23.
5. Histogram showing that the composition of spinal cord EGFR⁺ subtypes changes with developmental stage.

Taken together, our findings indicated that a group of EGFR‐expressing transitional glial cells arise from neuronal progenitors and gradually transform through intermediate stages into astrocytes and oligodendrocyte lineages. These cells can regulate gliogenesis through receptor–ligand interactions. Notably, our results indicated that EGFR‐expressing transitional glial cells are distributed in a position‐specific manner, with dorsal EGFR⁺ cells mainly concentrated within the dorsal horn area during development.

Discussion

Although neuronal diversity and specification have been studied in the embryonic spinal cord of mouse and zebrafish (Rosenberg et al, ²⁰¹⁸; Delile et al, ²⁰¹⁹; Wan et al, ²⁰¹⁹), neural lineage development of human spinal cord has not been previously explored. Here, we performed large‐scale single‐cell transcriptome analyses of the developing human spinal cord from embryonic to fetal stages. We incorporated snRNA‐seq to ensure the capture of neurons, as well as scRNA‐seq to achieve a larger gene expression dataset and enrich for glial cell populations. Our single‐cell resolution data revealed detailed neuronal and glial dynamic properties, allowing us to explore the characteristics of neural subsets and predict intercellular interactions between cell types during human spinal cord development. We also investigated spatiotemporal variations in protein expression levels of specific cell types by immunofluorescence analysis.

Our finding identified transcriptional characteristics and revealed the composition of neural populations at various stages. Changes in the proportions of both general neural cell types and dividing NPC subsets showed that in the human spinal cord, increases in neuronal cells cease before GW14. The development of astrocytes starts at GW8 and then gradually stops before GW23; in contrast, oligodendrocyte development was detected from GW8 onwards and continued to increase with advancing fetal age. Further analysis of NPCs during the neuronal development phase shows that GW7 neuronal progenitors contained a group of progenitors specific for ventral neuron development, while the majority of cells exhibited gene expression profiles corresponding to commitment of progenitors to a dorsal sensory neuron fate. Moreover, comparison of human and mouse single‐cell transcriptomics data enabled analyses of the conservation and divergence of cell patterns across species. Alignment based on expression covariation showed that the neuronal composition of the spinal cord is largely conserved between species.

Apparent differences in the morphology, proportion, and diversity of glial cells found between human and other species are thought to be associated with the evolution of nervous system complexity (Freeman & Rowitch, 2013; Majo et al, ²⁰²⁰). We showed a difference in the timing of glial development between human and mouse. From a neurogenesis perspective, human DI4/5 interneurons are born between GW7 and GW11, corresponding to E10.5–E12.5 in mouse (Lu et al, ²⁰¹⁵). Human astrocytes and oligodendrocytes differentiated as early as GW8, whereas in mouse the earliest appearance of GFAP occurs at E16 (Barry & McDermott, 2005) and the first wave of OPC generation commences around E12.5 (Tilborg et al, ²⁰¹⁸). Our results showed that the human spinal cord undergoes a more extended period of gliogenesis during fetal development compared with than rodents. This extended period is one possible explanation for the greater number of glial cells in human. The increased complexity of the human nervous system also requires enhanced glial cell diversity. Our data reveal the transcriptional heterogeneity of the oligodendrocyte lineage relative to its maturation state, while oligodendrocytes from the same maturation stage were a transcriptomically homogeneous. In contrast, four clusters exhibiting stable transcriptomic features were found within astrocyte subsets after GW11, potentially representing regional specificity or functional diversity among astrocytes, although this needs further exploration.

In addition, we found a few human neuronal subclasses that lacked mouse homologues. Such subclasses, such as MN COL6A2 and Exc‐DI5 PIEZO2, might have a more refined role during circuit establishment. Studies in model systems indicate that motor neurons are the earliest active neuronal class in the spinal cord; accordingly, the heterologous subcluster MN COL6A2 was unlikely to be caused by a difference occurring in the development phase. Instead, this MN COL6A2 subtype may represent a group of motor neurons with promoted axon growth. Severe defects in motor neuron axon outgrowth occur in COL6A2‐deficient embryos, but could be rescued by injection with COL6A2 RNA (Ramanoudjame et al, ²⁰¹⁵). Another neuronal subtype that caught our attention was Exc‐DI5 PIEZO2. Excitatory DI5 neurons, which migrate dorsally after derivation, are involved in pain, temperature, itch, and touch sensation. PIEZO2 expression, which plays a role in the mechanotransduction channel for proprioception (Zheng et al, ²⁰¹⁹), may be involved in the formation of sensory circuit in spinal cord. Our results also showed that the expression of EGFR is different between human and mouse spinal cord during development. Thus, differences between cell subtypes and gene expression might represent the diverse requirements of spinal cord development between species, which require further investigation. Additional studies are required to further investigate these species difference.

EGFR is essential for growth and differentiation in various developing systems, including the nervous system (Caric et al, ²⁰⁰¹; Xian & Zhou, ²⁰⁰⁴; Chen et al, ²⁰¹⁶; Romano & Bucci, ²⁰²⁰). Here, we found that EGFR‐expressing transitional glial cells were present and exhibited a radial morphology at the onset of gliogenesis. These EGFR‐expressing glial cells act as glial progenitors and progressively acquire either astrocyte or pre‐oligodendrocyte lineage properties, consistent with previous reports showing that astrocytes and oligodendrocytes can arise from EGFR‐expressing radial glia (Burrows et al, ¹⁹⁹⁷; Aguirre et al, ²⁰⁰⁷; Hayakawa‐Yano et al, ²⁰⁰⁷; Galvez‐Contreras et al, ²⁰¹³). Using annotation analysis based on a recent brain cortex dataset, we characterized EGFR⁺ cells in the spinal cord. Changes in features of EGFR⁺ cells corresponded to the stages of spinal cord development.

Cellular communication, which has a critical role during development, helps generate distinct cell types among initially similar neighboring cells by defining adjacent regions in developing tissues (Sprinzak et al, ²⁰¹⁰). Our intercellular communication analysis suggested that EGFR‐expressing glial cells interact with several glial cell types through Delta‐Notch and EGFR signaling pathways. Gene expression analysis revealed Notch1 expression was found in the astrocyte group (including EGFR‐expressing glia) and OPCs, while DLL1 expression was only enriched in the EGFR‐expressing glial cluster. Notch1/DLL1 signaling allows communication both within the same cell (cis‐interactions) and between neighboring cells (trans‐interactions) (Sprinzak et al, ²⁰¹⁰). During development, EGFR‐expressing glial cells might experience both cis‐interactions and trans‐interactions of Notch signaling within the surrounding microenvironment. It was previously reported that the interplay of Notch and EGFR signaling pathways could affect the balance between numbers of NSCs and NPCs in the subventricular zone (Aguirre et al, ²⁰¹⁰). Taken together, our finding suggests that EGFR‐expressing glial cells may play an important role in the fate decision processes of spinal cord cells.

Interestingly, our study indicated a markedly position‐specific distribution of EGFR‐expressing glial cells in the developing human spinal cord. Dorsal EGFR‐expressing glial cells were mainly concentrated within the dorsal horn gray matter, which has not been previously reported. We showed that EGFR‐expressing cells exhibited a radial glial morphology in the dorsal area at GW11, which indicated that EGFR⁺ cells may be involved in the arrangement of dorsal neural cells localization. This result is consistent with previous studies demonstrating that EGFR expression levels affect the timing of migration and settling patterns of cells (Burrows et al, ¹⁹⁹⁷). Furthermore, these position‐specific distributions could potentially be associated with different cellular functions among glial subtypes. EGFR signaling supports astrocyte survival and, thus, prevents neurodegeneration in the cortex (Wagner et al, ²⁰⁰⁶). Finally, the spatiotemporal arrangement of EGFR‐expressing cells during development suggests that they may play a role in establishing accurate axonal projections between dorsal root ganglia and the spinal cord by maintaining appropriate axonal branching levels. EGFR expression is required for proper axon innervation during skin development by limiting neurite outgrowth and branching in dorsal root ganglia neurons (Maklad et al, ²⁰⁰⁹). Hence, further study is warranted to determine whether EGFR‐expressing glial cells are functionally involved in spinal cord development, especially in sensory circuits.

Overall, the work presented here has outlined the landscape of neural development in the human spinal cord, covering the period from generation of neurons to the differentiation of astrocytes and oligodendrocytes. Transcriptome characteristics of neural cell types present at various stages of development were defined at single‐cell resolution. Our study provides fundamental insights into development of the human spinal cord that we expect to be a resource for further understanding of human spinal cord development. Moreover, our finding may inform the realization of cell therapy approaches to treat spinal cord diseases.

Materials and Methods

Study subjects

Thirty‐four aborted fetuses were collected (26 were used for sequencing and eight were used for immunofluorescence staining) from 33 donors (twin fetuses at GW17 were collected from one donor). Three fetuses were collected at GW9, GW10, GW11, GW13, GW17, and GW20; two fetuses were collected at GW7, GW12, GW15, GW22, and GW23; and one fetus was collected at GW8, GW14, GW16, GW18, GW19, and GW21. Gestational age was measured in weeks from the first day of the woman’s last menstrual cycle to the sample collection date.

Tissue sample collection and preparation

Fetal spinal cords were collected whole or separated into regions (cervical, thoracic, and lumbar) and placed in ice‐cold Hanks’ balanced salt solution medium. For cell preparation, tissues were digested using a papain dissociation system according to the manufacturer’s protocol (Worthington Biochemical Corporation, Lakewood, NJ, USA). Spinal cords were roughly cut into small pieces and then incubated in papain solution containing DNase I at 37°C for 40 min to 1 h with occasional pipetting. Albumin‐ovomucoid inhibitor solution was used to terminate the enzymatic reaction. Subsequently, the cell suspension was further pipetted to generate a single‐cell suspension, which was centrifuged at 300 g for 5 min to obtain a cell pellet. The supernatant was carefully removed, the cell pellet resuspended in 500 μl of Hanks’ balanced salt solution medium, and the sample was placed on ice. Quality control was assayed by measuring cell viability and rate of clumping before the sample was loaded onto a chromium single‐cell controller (10 × Genomics, Pleasanton, CA, USA). Samples with > 80% viability were used for sequencing. 3’‐Gene expression libraries were prepared using V2 Chromium Single Cell Reagent Kits in accordance with the manufacturer’s instructions (10 × Genomics). Quality control of cDNA and prepared libraries was performed using a 4200 TapeStation System (Agilent, Santa Clara, CA, USA). Libraries were sequenced using a HiSeq 4000 system (Illumina, San Diego, CA, USA).

For nuclei preparation, we adapted the protocol of Sathyamurthy et al Tissue samples were homogenized using a glass Dounce tissue grinder (D8938; Sigma‐Aldrich, St. Louis, MO, USA) in 1 ml of cold Nuclei EZ lysis buffer (EZ PREP NUC‐101; Sigma‐Aldrich) on ice. Each sample was ground 20 times with pestle A, followed by 20 times with pestle B. The lysate was then diluted with 3 ml of cold sucrose buffer and centrifuged at 3,200 g for 10 min. The supernatant was discarded, and the sediment was carefully resuspended with 3 ml of cold sucrose buffer. Next, the suspension was gently overlaid on 12 ml of density buffer in a 50‐mL conical tube and centrifuged at 3,200 g for 20 min. The supernatant was aspirated and discarded. Nuclei deposited on the tube wall were collected with 3 ml of cold phosphate‐buffered saline (PBS) and spun at 3,200 g for 10 min. Nuclei were then resuspended in PBS for sequencing. 3’‐Gene expression libraries were prepared using V3 Chromium Single Cell Reagent Kits in accordance with the manufacturer’s instructions (10 × Genomics). Quality control of cDNA and prepared libraries was performed using a 4200 TapeStation System. Libraries are sequenced using the HiSeq 4000 system.

Immunofluorescence staining of spinal cord sections

Human embryonic spinal cords were dissected, post‐fixed overnight in 4% paraformaldehyde, and subsequently cryoprotected in 30% sucrose for 48 h. Spinal cords were embedded in optimal cutting temperature medium (Thermo Fisher Scientific, Waltham, MA, USA) and cryosectioned into 20‐μm sections using a Leica CM1950 cryostat (Wetzlar, Germany). Sections were treated with blocking buffer containing 10% normal donkey serum and 0.3% Triton X‐100 in PBS. The following primary antibodies, which were diluted in blocking buffer, were used: mouse anti‐GFAP (1:500; MAB360; Millipore, Burlington, MA, USA), rabbit anti‐SOX2 (1:500; ab97959; Abcam, Cambridge, UK), mouse anti‐PAX7 (1:500; ab218472, Abcam), goat anti‐OLIG2 (1:500; AF2418; R&D Systems, Minneapolis, MN, USA), mouse anti‐NeuN (1:100; MAB377, Millipore), rabbit anti‐NeuN (1:500; ab177487, Abcam), rabbit anti‐PIEZO2 (1:100; HPA031975; Sigma–Aldrich), rabbit anti‐LRRC50 (1 μg/ml; PA5‐59463, Thermo Fisher Scientific), mouse anti‐ASCL1 (10 μg/ml; H00000429; Abnova), rabbit anti‐PDGFRA (PA5‐16571; 2 μg/ml, Thermo Fisher Scientific), mouse anti‐SOX9 (1:1,000, MA5‐17177; Thermo Fisher Scientific), chicken anti‐DRGX (1:500; PA5‐72713, Thermo Fisher Scientific), rabbit anti‐DLL1 (1:1,000; ab10554, Abcam), mouse anti‐LHX3 (1:300; ab2135805, Abcam), mouse anti‐EGFR antibody (1:200; ab30, Abcam), and rabbit anti‐FOXP2 (1:1,000; ab16046, Abcam). Alexa Fluor 488, Alexa Fluor 561, or Alexa Fluor 647 fluorophore‐conjugated secondary antibodies (Life Technologies, Carlsbad, CA) were used to visualize antibody binding. Cell nuclei were stained using a fluorescence mounting medium containing DAPI (ZLI‐9557; ZSGB, Beijing, China). Imaging was performed using a Leica TCS SP8 laser‐scanning confocal microscope.

Data analysis

Raw reads mapping and quality control

Raw files were processed with Cell Ranger 3.1.0 (10× Genomics) using default mapping options. Reads were mapped to the human GRCh38 genome, and a gene expression matrix was obtained for each nuclei and cellular sample.

For nuclei samples, quality control was performed with Seurat 3.0 (https://satijalab.org/seurat/). We first trimmed the original gene expression matrix of each sample before subsequent analysis; only cells that featured more than 200 genes and only genes that were detected in at least 10 single cells were included. To reduce noise from low‐quality cells, we next filtered cells according to the number of unique genes and the number of RNA molecules.

For cellular samples, we used the same series of quality controls as described above for nuclei samples, with the exception of additional criteria for the percentage of mitochondrial genes.

A common strategy employed to generate violin plots of single‐cell transcriptomic data is to remove cell outliers for the number of genes, number of unique molecular identifiers, and percentage of mitochondrial counts. Quality was validated if a significant correlation between the number of molecules detected and the number of unique genes within each cell was obtained. We also manually identified clusters that co‐expressed neuronal and non‐neuronal markers as doublets.

Cluster analysis of cell types and identification of DEG among clusters

After quality control, gene expression measurements were normalized by a global‐scaling normalization method, “LogNormalize”. This approach identified approximately 2,000 highly variable genes for cell‐type cluster analysis. To minimize extensive technical noise from any single gene, principal component analysis (PCA) was performed on these selected genes as a dimensional reduction technique after linear transformation (“scaling” of Seurat). Subsequently, significant PCs were identified by the “JackStraw” method and “Elbow plot” to drive graph‐based clustering analysis. The “FindNeighbors” function accepted previously defined principal components to construct a KNN graph based on the Euclidean distance in PCA space. Finally, the “FindClusters” function produced cell‐type clusters at 0.5 resolution. To clearly visualize and further explore cell clusters in low dimensions, we used non‐linear dimensional reduction (UMAP) plots. We applied the “FindAllMarkers” and “FindMarkers” functions to clusters to obtain DEGs; expression levels of these genes and several hundreds of classical marker genes for specific cell types, are illustrated in violin plots and feature plots for each cluster in all samples. All these steps enabled us to finally identify cell types present in each cluster. A series of R packages were also used to assist with the analysis process, such as ggplot2, Matrix, grid, and tidyverse.

Integration analysis of specific cell types

To better understand the development of specific cell types in the spinal cord, we isolated clusters belonging to a certain cell type in samples at different time points and integrated them into a new gene expression matrix. After removing the “batch effect” between gene sets from distinct samples using the “anchors identify” step and “IntegrateData” function, we repeated the above cluster analysis and marker gene searching process. Furthermore, we examined all the integrated cells at different stages of development separately so that we could identify the moment of emergence, transition, or vanishing of certain cell types during development. Trajectory analysis was performed for different cell types in each sample or in an integrated dataset after removing batch effects. To understand cell‐state transitions in spinal cord development, we performed trajectory analysis using the Monocle (V 3.0, http://cole‐trapnell‐lab.github.io/monocle‐release/) pseudotime reconstruction function without manually defined start point. For either a single sample or integrated “batch‐corrected” datasets, we extracted a gene expression matrix for all cells and directly set up a new Monocle object. Next, all cells in the object were mapped onto a pseudotime trajectory, to decipher transitions between cell types. When the trajectory analysis is applied to more specific subtypes, gender‐related genes showed some influence of results. To avoid gender impact, we choose male samples from GW11T‐NC, GW14T‐NC, GW18T‐NC, GW20T‐NC, and GW21T‐NC and mapped them onto a pseudotime trajectory, to decipher transitions between cell types. Moreover, as with a marker gene plot in clustering analysis, we visualized how expression of individual genes varied along the trajectory using the “plot_cells” function of Monocle. We applied the “FindAllMarkers” function to obtain DEGs. To identify cell‐type markers conserved across development, we performed the function, “FindConservedMarkers”, of the Seurat package.

MetaNeighbour analysis

To identify correlation of different gene families in each major cell class between human and mouse, we performed separate MetaNeighbour analyses of each species. Data were divided into two artificial experiments by selecting random groups of equal size up to a maximum of 10 cells per cluster for each experiment. Next, MetaNeighbour was ran separately for clusters from each major class (Exc, Inh, and NPC) using the R function run_MetaNeighbour. Mean area under the receiver operating characteristics (AUROC) scores for each gene set were then calculated by averaging the reported AUROC scores for a gene set across all clusters. This process was repeated for 10 divisions of the human and mouse data into random experimental groups.

Estimation of cell‐type homology

Analysis was performed using the Seurat “FindIntegrationAnchors” and “IntegrateData” functions. To specify the neighbor search space, 1–30 dimensions from the canonical correlation analysis (CCA) were used. We defined homologous cell types by clustering the aligned embedding output from Seurat and identifying human and mouse samples that co‐clustered. Clusters of cells were identified by a shared nearest neighbor modularity optimization‐based clustering algorithm. First, k‐nearest neighbors were calculated and a shared nearest neighbor graph was constructed. Next, the modularity function was optimized to identify clusters. For each pair of clusters, the overlap was defined as the sum of the minimum proportion of samples in each cluster that overlapped within each aligned cluster. This approach identified pairs of human and mouse clusters that consistently co‐clustered within one or more aligned clusters. Cluster overlaps varied from 0 to 1 and were visualized as a heat map with human clusters in columns and mouse clusters in rows.

Quantification of expression divergence

For each pair of homologous cell types, the average expression of 13,161 orthologous genes was calculated as normalized counts data with a scale factor of 1e4. The “FindMarkers” function was used to calculate differential expression between human and mouse for all genes and homologous cell types. Thresholds to identify differences were large (> 5‐fold), moderate (2‐ to 10‐fold), and small (< 2‐fold). A heat map was generated showing expression differences across cell types, and hierarchical clustering using Ward’s method was applied to group genes with similar patterns of expression change. Finally, changes in divergent expression patterns were calculated for functional gene families downloaded from the HUGO Gene Nomenclature Committee at https://www.genenames.org/download/statistics‐and‐files/. The enrichment analysis was performed using the Disease Ontology Semantic and Enrichment hypergeometric model.

Cell–cell interaction analysis

To enable a systematic analysis of intercellular interactions, cell–cell communication was predicted based on the receptor–ligand pairs using CellPhoneDB (https://www.cellphonedb.org). Only receptors and ligands expressed in at least 10% of the cells in a given cluster were further analyzed. Pairwise comparisons between all cell types were performed, and only those with P < 0.05 were used for subsequent prediction of cell–cell communication.

Mapping and annotation of scRNA‐seq datasets

We constructed a reference dataset based on Tang’s brain EGFR⁺ cells and projected our selected EGFR⁺ cells of four samples (GW11T, GW16T, GW20T, GW23S) onto it. Using “Mapping and Annotating” functions in Seurat 4.0, we annotated our query EGFR⁺ cells and presented them in reference UMAPs. We also calculated the proportion of cell types for each sample and indicated their difference in a stacked bar plot.

Study approval

Collection and analysis of human embryonic tissue were approved by the Reproductive Study Ethics Committee of Nanjing Drum Tower Hospital, Nanjing Medical University (Nanjing, China) (2018‐223‐01). Informed consent was provided by all donors, and aborted fetuses were obtained after legal pregnancy termination at the Nanjing Drum Tower Hospital.

Author contributions

JWD, ZFX, and YLH conceived the project and designed the experiments. QZ, YMY, and ZFX prepared the manuscript. QZ and YHF performed RNA‐seq. XMW and MHS performed bioinformatic analyses. QZ, ZFX, SFH, YNZ, BC, and MHS analyzed the data. PPJ, JH, GFZ, QZ, and YHF prepared the samples. QZ, SFH, YHF, and BX performed immunofluorescence and imaging. All authors discussed, edited, and proofread the manuscript. These authors contributed equally: QZ, XMW, YHF, and PPJ.

Conflict of interest

The authors declare that they have no conflict of interest.

Supporting information

Appendix

Expanded View Figures PDF

Source Data for Expanded View and Appendix

Table EV1

Table EV2

Dataset EV1

Dataset EV2

EMBO reports (2021) 22: e52728.

Data availability

Single‐cell RNA‐seq data generated for this study have been deposited in the Gene Expression Omnibus (GEO) under the accession number GSE136719. Raw image files and custom programs are available from the corresponding authors upon reasonable request.