Single-cell RNA-sequencing of mammalian brain development: insights and future directions

Early in development, secreted morphogens define spatial domains within the brain that express distinct transcription factor combinations. These transcription factors, in turn, specify thousands of transcriptionally distinct neural cell types in the adult brain (reviewed by Silbereis et al., 2016). However, this cellular complexity is likely exceeded several-fold by the number of distinct transient cell states that exist during development. Understanding brain development thus requires extraordinarily sensitive and scalable methods.

Single-cell RNA-sequencing (scRNA-seq) has emerged as the predominant technique for discovering the transcriptional cell type composition of complex tissues. Advances in miniaturization and multiplexing have reduced costs and increased throughput, enabling the generation of gene expression profiles of many thousands of cells per experiment (Svensson et al., 2018). Following the first scRNA-seq experiment reported on mouse blastomeres (Tang et al., 2009), the most studied tissue has been the brain, with over 250 publications to date. Here, we review studies of mammalian brain development, focusing on post-gastrulation events. Using a database of scRNA-seq publications (Svensson et al., 2020), catalogs of scRNA-seq datasets [NeMo and the University of California, Santa Cruz (UCSC) Cell browser] and PubMed, we manually identified 89 studies of developing mammalian brain and nervous system that used primary tissues (Table 1). To illustrate the scope of the article, we summarize mouse and human studies by anatomical regions and developmental time points. We found that, in recent years, scRNA-seq has been scaled to whole mouse embryos and brain samples across the full timeline of mouse prenatal brain development (Fig. 1). Together with advanced spatial techniques, a full 3D cellular map of mouse brain development seems within reach. However, owing to limitations of human tissue collection and experimentation, studies on human brain development remain less comprehensive (Fig. 2). We highlight important findings in the field so far and their implications for experimental models. Finally, we discuss remaining gaps in the field and suggest future research directions and technologies necessary to understand human brain development.

scRNA-seq studies that include data from multiple regions and time points provide the most complete view of cell-type heterogeneity during mammalian brain development. For example, one atlas comprised almost 300,000 sequenced cells from all major regions of the mouse brain during the entire timespan of prenatal brain development [embryonic day (E)7-E18] (La Manno et al., 2021). Eight main cell classes (Fig. 3) and nearly 800 clusters were identified in total, covering the onset and timeline of gastrulation, the emergence of the neural crest and neural tube, neurogenesis, early gliogenesis, and the branching of the meningeal lineage from the neural crest or mesenchyme. This atlas revealed that, although 87 radial glia subtypes were identified, most neuronal diversity was generated in postmitotic neuroblasts (171 subtypes) and maturing neurons (306 subtypes). The number of cell types identified depends on the protocol used (Mereu et al., 2020), the number of cells and depth sequenced (Svensson et al., 2020), as well as clustering methods (Kiselev et al., 2019). Comprehensive atlases are, therefore, the most accurate resources to compare cell heterogeneity between brain regions and time points. Such datasets are also the most accurate way to identify markers that can distinguish cell types globally across the brain or entire body (Box 1). Below, we explore cellular heterogeneity in the mammalian brain revealed by scRNA-seq and single-cell multi-omics.

A study of first-trimester human telencephalon found nine clusters of neuroepithelial and radial glial cells. The gradual transition of these cells from early to more mature progenitors was defined by gene expression gradients (Eze et al., 2021). Three major classes of radial glia were identified in humans: ventricular, outer and truncated radial glia, appearing in early-to-late development (Nowakowski et al., 2017). Cell cycle and differentiation genes were mostly responsible for the transcriptional variation seen in outer radial glia (Pollen et al., 2015). Furthermore, regional transcription factors [e.g. HOX genes (hindbrain); PAX7 (midbrain); GBX2 (thalamus); FOXG1 (cerebral cortex)] were expressed as early as Carnegie stage (CS) 13, but cell type-defining genes were mostly conserved across different brain regions (Eze et al., 2021). Combined scRNA-seq and single-cell assay for transposase-accessible chromatin-sequencing (scATAC-seq) data showed that chromatin activity of a set of genes in cycling cells precedes lineage-specific gene expression, suggesting that progenitors entering the cell cycle were primed towards specific lineages (Trevino et al., 2021). RNA-velocity analysis of the early human brain (CS 12-13) indicated examples of radial glia giving rise to neurons, suggestive of direct neurogenesis (Eze et al., 2021). Intermediate progenitor populations appeared at ∼CS 18, implying the beginning of stereotypical (indirect) neuronal differentiation (Eze et al., 2021). However, the expression of TBR2 (also known as EOMES), which is the basis for intermediate progenitor definition, was also found in postmitotic neurons, such as Cajal-Retzius cells (Eze et al., 2021). In addition, a novel mesenchymal progenitor was described that expressed LUM and ALX1, potentially supporting telencephalic development before the radial glial scaffold emerges (Eze et al., 2021). These cells may be the early mesenchymal progenitors destined to form the meninges, as recently described (La Manno et al., 2021).

In the mouse, forebrain excitatory neurons were a clear transcriptional outgroup distinct from the other neuronal lineages (La Manno et al., 2021), supporting data from adolescent animals (Zeisel et al., 2018). In humans, by mid-gestation there were populations of migrating, differentiating and maturing excitatory neurons (Zhong et al., 2018; Polioudakis et al., 2019). There were also area- and layer-specific pyramidal cell subtypes in the human cortex, which emerged during peak neurogenesis. These results differ from early models of a cortical ‘column’ consisting of similar cell types and lamination patterns irrespective of topographical location in the cortical sheet (Nowakowski et al., 2017; Zhong et al., 2018; Bhaduri et al., 2020). Multi-ome data defined a developmental trajectory for glutamatergic neurons, revealing a continuous progression of transcription factor motif activities associated with neuronal specification and migration (Trevino et al., 2021). Furthermore, a comparison of scRNA-seq and scATAC-seq profiles of human excitatory lineage cells revealed chromatin state signatures of cortical arealization already in intermediate progenitor populations, preceding that of areal transcriptomic signatures in excitatory neurons (Ziffra et al., 2021).

Mouse and human forebrain inhibitory neurons develop from the ganglionic eminences and migrate tangentially into the neocortex. The precise origins and development of these cells have been the subject of intense study (Lim et al., 2018). Precursors located in the mouse caudal, lateral and medial ganglionic eminences clearly expressed distinct transcription factors that bias their fates upon differentiation (Mayer et al., 2018; Mi et al., 2018), but whether the full set of mature interneuron types is prespecified in the eminences remains unclear. During parvalbumin and somatostatin interneuron development, combined scRNA-seq and scATAC-seq revealed a predominantly shared chromatin architecture of general interneuron identity from E13 to postnatal day (P)2, followed by the expression of specific transcription factors that initiated terminal differentiation and cell type diversification (Allaway et al., 2021). There were also conflicting reports concerning the existence of interneuron progenitors in the human neocortex (Zhong et al., 2018; Polioudakis et al., 2019). Recent findings indicate that most of the diversity of interneuron types is generated in the ganglionic eminences (Shi et al., 2021). However, genetic fate mapping revealed a shared origin for human excitatory and inhibitory cortical neurons (Delgado et al., 2022). Interestingly, the same was not true in mice (Bandler et al., 2022).

A common progenitor for astrocytes and oligodendrocytes, and two additional astrocyte precursor states, were found in humans (Zhong et al., 2018; Polioudakis et al., 2019). In mice, transitional intermediate states in astroglial and oligodendroglial lineages persist in postnatal gliogenesis and were altered in gliomagenesis (Weng et al., 2019). A population of late radial glial cells (glioblasts) was identified, that gradually lost proliferative capacity and switched to gliogenesis (La Manno et al., 2021). Oligodendrocytes were established as a transcriptionally heterogeneous lineage with twelve clusters identified in P21-P60 mice (Marques et al., 2016), whereas three subpopulations of early progenitor, late progenitor (myelinating) and mature oligodendrocytes were identified in humans (Perlman et al., 2020). Overall, oligodendrocytes were consistently found to be much less heterogeneous than neuronal or astrocyte lineages.

Microglia have a yolk sac origin (i.e. non-neural) and migrate into the human brain by gestational week (GW) 8 (Zhong et al., 2018). Their development involves three distinct temporal stages in synchrony with the homeostatic needs of the developing brain, and displays heterogeneity from embryonic to adult ages (Matcovitch-Natan et al., 2016; Li et al., 2019; Kracht et al., 2020). Interestingly, human microglia may be more heterogeneous than five other mammals surveyed (macaque, marmoset, sheep, mouse and hamster) (Geirsdottir et al., 2019).

Neurodevelopment is a highly dynamic process with precise spatiotemporal regulation of the transcriptome. To understand the sequence of events during development, and the origin of adult cell types, samples can be collected over time. One such study analyzed various brain regions from embryonic, fetal, adolescent and adult human brains using bulk and scRNA-seq and found a global temporal transcriptomic pattern (Li et al., 2018). This, and the molecular signatures of topographical gene expression and bilaterality in the developing human brain, were previously uncovered using bulk RNA-sequencing alone (Pletikos et al., 2014). During embryonic to mid-fetal stages, there was high transcriptomic variation between brain regions (Li et al., 2018). At late fetal stages, there was a sharp decrease in regional differences but an increase in genes related to neuronal maturation, synaptogenesis and myelination. High intra- and inter-regional variation resumed in childhood-adolescence. Alignment of prenatal and adult data revealed that putative embryonic and fetal excitatory neurons clustered near their adult counterparts, whereas fetal cells in the oligodendrocyte lineage did not, indicating a distinctly different maturation status (Li et al., 2018).

Focusing on a single region over time facilitates detailed sub-anatomical analysis of timing, patterning events and the effects of gene mutations. For example, in mice, it was shown that the neurogenic-to-gliogenic switch occurred between E12 and E16, that nearly all neuronal lineages passed through an early neuroblast state, and that the earliest astrocytes appeared at ∼E15 (La Manno et al., 2021). In mouse hypothalamus development, Nkx2-1 was identified as a potential positive regulator of ventral hypothalamic identity while also suppressing prethalamic identity, which was not predicted by the prosomeric model for forebrain organization (Puelles, 2009; Ferran et al., 2015), highlighting an unexpected developmental connection between the hypothalamus and prethalamus (Kim et al., 2020). Another study generated an atlas of cortical development in Fezf2 mutant mice (Di Bella et al., 2021). Fezf2 is a crucial transcription factor for forebrain development, the absence of which causes a loss of sub-cerebral projection neurons (reviewed by Molyneaux et al., 2007). At E13.5, postmitotic projection neurons were transcriptionally more like projection neurons of later time points. Thus, although Fezf2 was expressed in progenitors, its role in sub-cerebral projection neuron specification was primarily postmitotic (Di Bella et al., 2021). In addition, other specific regions that were studied over time include the mouse developing cerebellum (Carter et al., 2018), dentate gyrus (Hochgerner et al., 2018), as well as the human hippocampus (Zhong et al., 2020) and pituitary (Zhang et al., 2020a) (Table 1).

Computational tools, such as Monocle (Trapnell et al., 2014), URD (Farrell et al., 2018), RNA velocity (La Manno et al., 2018) and scVelo (Bergen et al., 2020), can be used to infer differentiation trajectories, model transcriptional dynamics and predict the direction of differentiation (reviewed by Saelens et al., 2019). This information can guide complementary experiments such as genetic lineage tracing, to prove the predicted lineages and clonal relationships (Wagner and Klein, 2020).

Evolutionary conservation can reveal the function of cell types and can be used to transfer knowledge about these functional properties from model organisms to humans. For example, little is known about the electrophysiological properties of human neurons, but such properties can be inferred from the corresponding mouse or non-human primate cell types. At the same time, non-conserved cell types or states may reveal human-specific traits and disorders. For example, one study comparing human cortical development with mice showed that the human cortex developed more slowly and certain human cell populations were absent in early mouse development (Eze et al., 2021).

A study using bulk and scRNA-seq compared human and macaque brain development and found numerous shared and divergent transcriptomic features. In the hippocampus, striatum and cerebellar cortex, differentially expressed genes were enriched in oligodendrocyte progenitors, medium spiny neurons, external granular layer neurons and microglia, indicating specific cell-type differences between human and macaque. In the prenatal dorsolateral prefrontal cortex, all human cell types had a close homologue in the macaque. However, there was a prolonged neurogenic period in humans with an enrichment for intermediate progenitor cells, suggesting that the duration of neurogenesis may differ between the two species (Zhu et al., 2018). These data highlight that distinct models for early human cortical development are necessary. Drawing correlations with genes known to be implicated in neuropsychiatric diseases, five heterotopic clusters specific to schizophrenia-associated genes were found. Three of them displayed exclusive species differences prenatally. One schizophrenia-associated gene, GRIA1 (an AMPA receptor subunit), was expressed in neocortical areas earlier in humans. Instead, autism spectrum disorder-associated genes (SHANK2 and SHANK3) were expressed earlier in macaque than human neocortex (Zhu et al., 2018). These findings are relevant for in vitro models of neurodevelopment and non-human primate models of neuropsychiatric disease, such as the SHANK3-deficient macaque model (Zhao et al., 2017).

Another study compared adult human, marmoset and mouse motor cortices and found that most cell types were homologous between the three species. Surprisingly, however, most cell-type-specific genes in each species were not conserved in the other species (Bakken et al., 2021). The same was observed in a study comparing interneurons in three primates (human, macaque and marmoset), mice and ferrets (Krienen et al., 2020). Evolutionary comparisons can, therefore, both reveal conserved genes likely to support the core function of each cell type, but also pinpoint divergent genes that have evolved recently in each species to serve species-specific adaptations.

Taken together, this suggests that interspecies differences can result from variations in cell-type diversity and abundance, more so than the transcriptional divergence between cell types. Interestingly, a new classification system based on an evolutionary definition of cell type was proposed (Arendt et al., 2016). The proposal can be tested with new tools, such as SAMap, which maps cell atlas manifolds across species and identifies homologous cell types (Tarashansky et al., 2021). We, therefore, stress the importance of studies on evolutionary conservation.

To fully understand brain development, data from all neuroanatomical regions and developmental time points is necessary. Regions of the nervous system and time points that have been studied in mice and humans are shown in Figs 1 and 2, respectively. Here, we clarify the remaining gaps in the field.

Whole mouse embryos were sequenced from E6 to E13 using scRNA-seq (Cao et al., 2019; Pijuan-Sala et al., 2019). In the central nervous system, the forebrain, midbrain, hindbrain, cerebellum and retina were sequenced almost throughout prenatal development, and the latter three during the first postnatal week as well. The spinal cord had almost no scRNA-seq data from the second half of prenatal development. The forebrain was studied the most, although with missing time points for subdivisions of the forebrain. Some neuroanatomical regions develop late, which may explain the absence of sequencing data on certain time points (e.g. the cerebellum develops from E10 through to adolescence) (Fig. 1; Table 1).

Peripheral nervous system (PNS) development was not as well covered across development. We found publications for neural crest-derived autonomic and sensory neurons, the enteric nervous system, the inner ear, sciatic nerve and dorsal root ganglia (Fig. 1; Table 1). More studies on dorsal root ganglia were found in adult mice (Usoskin et al., 2015; Li et al., 2016; Wang et al., 2021).

Although the sequencing coverage for mouse neurodevelopment was high, most regions and time points were only sequenced once (and a maximum of four times, e.g. E13-E15 cortex). Of 61 studies, 18 isolated specific cell types for sequencing [e.g. microglia from the whole brain (Hammond et al., 2019) and midbrain dopaminergic neurons (Tiklová et al., 2019)]. Furthermore, other than by inclusion in a larger dataset (La Manno et al., 2021), no study focused on analyzing all cell types in the developing hindbrain, thalamus or pons (Fig. 1; Table 1).

There were roughly half as many publications on human nervous system development compared with mice (Table 1). Several regions have been sequenced at almost all prenatal weeks up until birth: the cortex, pituitary gland and retina, and the spinal cord and cerebellum have also been covered extensively over prenatal time points (Fig. 2). Although it is not a trivial undertaking, data could be integrated across publications to create a full atlas of human cortical development. Comparing this to the atlas of mouse cortical development could yield interesting evolutionary findings and implications for experimental models. Notably, two studies covering wide age ranges and ten (Eze et al., 2021) or six (Zhu et al., 2018) different brain regions, mostly focused their analysis on the cortex, leaving avenues for interesting post hoc analyses.

Not many other regions were sequenced and analyzed in-depth, other than the hippocampus (Zhong et al., 2020) and ventral midbrain (La Manno et al., 2016). Only 5/30 studies isolated one specific cell type for sequencing (Table 1). We found no studies on PNS development in humans. The gaps in human studies are partly due to limitations on tissue acquisition (Box 2). Furthermore, the influence of other tissues on brain development are also important (Box 3).

The growth of scRNA-seq and the vast amount of data produced have raised a few considerations: how many cells and what sequencing depth is optimal to capture all cell types and distinguish them from each other can we define cell type diversity during brain development with similar datasets and methods as the adult brain, considering that rare and transient cell states vary between regions and time points during development; how do we create a common taxonomy of cell types (is a ‘glioblast’ in one study the same as ‘glial progenitor’ in another); can we bridge different sequencing techniques (single nucleus/cell sequencing, fresh/fixed tissue, chemistry A or B, multimodal sequencing assays, etc.)? To answer these questions and more, and to integrate single-cell data, international consortia, such as the Human Cell Atlas, have been formed. Its mission is to map every cell in the human body (a ‘Periodic Table’ of human cells) using single-cell and spatial genomics and computational techniques (Regev et al., 2017; Haniffa et al., 2021; Rozenblatt-Rosen et al., 2021). International consortia, together with integrative and interactive databases, and standardization efforts for data analysis (Luecken and Theis, 2019) are essential to combine findings and shape productive future directions.

In addition, although the most used single-cell method is scRNA-seq, methylation, chromatin accessibility, genetic variation and protein abundance can be measured using related techniques. Single-cell biology has entered a multi-omics age (Perkel, 2021), and key scATAC-seq work during mammalian development is already emerging (Preissl et al., 2018; Domcke et al., 2020; Ziffra et al., 2021). The combination of scRNA-seq and scATAC-seq is particularly powerful for developmental studies because it can simultaneously reveal transcriptional cell states and the gene regulatory elements – enhancers and promoters – that drive those states. Furthermore, cell-type-specific enhancer elements are powerful experimental tools for genetic targeting of cell types and states, such as by using viral vectors in vivo, in explants or in organoids (Mich et al., 2021).

Some important neurodevelopmental questions cannot be addressed by single-cell analyses alone. For example, scRNA-seq can reveal which cells express genes involved in brain patterning, but not their spatial location, as this information is lost upon tissue dissociation. Spatial-omics techniques are, therefore, under tremendous technical development and usage (Rao et al., 2021). For example, sci-Space allows the construction of spatially resolved single-cell atlases, recently mapping whole transcriptomes of 120,000 nuclei in whole E14.5 mouse embryos at single-cell resolution (Srivatsan et al., 2021). Comparing the transcriptome and location of cell types can also elucidate whether the transcriptome of a particular cell type is defined before or after migrating to its terminal niche. Data integration from scRNA-seq and spatial transcriptomics has been reviewed elsewhere (Longo et al., 2021).

One of the main goals in neuroscience is to understand human brain development and the mechanisms underlying neurodevelopmental disorders. A hurdle to achieving this – unlike other model organisms – is the difficulty of acquiring human samples at all developmental stages (Box 2) and the impossibility of in vivo experimentation. To a large extent, rodents and non-human primates serve as models of human biology. However, further development of advanced in vitro systems, such as ex vivo brain explant culturing and organoids, is essential to make progress in the field. As human brain development is slow compared with rodents, focusing on long-term culturing is necessary to investigate inaccessible time points. Single-cell and spatial-omics techniques can be used to verify the extent to which in vitro systems recapitulate normal development (Camp et al., 2015; Velasco et al., 2019; Bhaduri et al., 2020; Fleck et al., 2021), which can subsequently be used for live-tissue experiments, such as lineage tracing, live imaging, gene perturbations and other functional tests. Future integration of long-term time-lapse imaging with single-cell and spatial-omics could directly link cell behavior, lineage trajectories and molecular cell states. We believe data integration across organs is important to place brain development into the context of the whole body, as demonstrated by recent work on organogenesis and diverse fetal tissues (Cao et al., 2019, 2020; Pijuan-Sala et al., 2019; Lohoff et al., 2022). No model system is ideal for the study of all biological questions, but the integrated combination of existing models can eventually lead to a common atlas over space and time, with a coordinate framework of human brain development.