Conserved properties of dentate gyrus neurogenesis across postnatal development revealed by single-cell RNA sequencing

The dentate gyrus of the hippocampus is a brain region in which neurogenesis persists into adulthood; however, the relationship between developmental and adult dentate gyrus neurogenesis has not been examined in detail. Here we used single-cell RNA sequencing to reveal the molecular dynamics and diversity of dentate gyrus cell types in perinatal, juvenile, and adult mice. We found distinct quiescent and proliferating progenitor cell types, linked by transient intermediate states to neuroblast stages and fully mature granule cells. We observed shifts in the molecular identity of quiescent and proliferating radial glia and granule cells during the postnatal period that were then maintained through adult stages. In contrast, intermediate progenitor cells, neuroblasts, and immature granule cells were nearly indistinguishable at all ages. These findings demonstrate the fundamental similarity of postnatal and adult neurogenesis in the hippocampus and pinpoint the early postnatal transformation of radial glia from embryonic progenitors to adult quiescent stem cells. Using single-cell RNA-seq, the authors show that early developmental neurogenesis in the dentate gyrus of the hippocampus is largely conserved in the adult, but with a perinatal transformation of stem cells to an adult type.

T he dentate gyrus, part of the hippocampus, is involved in learning, episodic memory formation, and spatial coding. It receives unidirectional input from the entorhinal cortex through the perforant path and projects to the CA3, forming the first step of the hippocampal trisynaptic circuit. The dentate gyrus is composed of three layers: the cell-sparse molecular layer and the hilus, separated by a dense layer of granule neurons whose dendrites extend into the overlying molecular layer 1 .
The majority of granule cells are born perinatally, with substantial numbers of cells generated until around postnatal day (P) 10 ( Fig. 1a) 2,3 . Radial glia are neuronal stem cells during embryonic development 4,5 and are still abundant in the first postnatal weeks. Radial glia-like cells differentiate into proliferating progenitors (neuronal intermediate progenitor cells or nIPCs) 6 and exit the cell cycle to become migrating neuroblasts, then settle in the granule cell layer and mature around postnatal week 3 into fully functional excitatory granule cells.
In adult mammals, radial glia-like cells are neuronal stem cells (NSCs) in the subventricular zone 7 and continue to generate granule cells from the dentate gyrus subgranular zone [8][9][10] . Thus, a small number of progenitor cells in the subgranular zone continue to generate new granule cells throughout life, albeit at low rates [11][12][13][14][15][16] . The process is believed to contribute to learning and memory formation, and it is sensitive to stress and antidepressants, as well as to enrichment and exercise [17][18][19][20] . Upon differentiation in the adult, newborn neuroblasts go through a several-weeks-long process of maturation, acquiring mature membrane properties after about 3-4 weeks 21 . While adult NSCs and early postnatal radial glia-like cells show many similarities, they act in different environments, perhaps requiring different regulatory mechanisms to maintain their neurogenic potential 22 . However, it is not currently known whether NSCs are simply radial glia-like cells that persist or whether there are any molecular properties that make adult NSCs unique. Similarly, it is not currently known whether early postnatal and adult neurogeneses proceed through identical intermediate states or whether the adult process is different beyond timing 23 from the initial postnatal waves.
Recent work using single-cell RNA-seq and targeting the neurogenic populations has suggested that neurogenesis forms a continuum of molecular states 24,25 . However, we suspected that the limited number of cells in those studies, and the targeted sampling using specific markers, might have masked the presence of distinct cell types. We therefore performed larger-scale single-cell RNA-seq and unbiased sampling to delineate the cellular states along the granule cell lineage. We examined perinatal and juvenile development, as well as adult mice, and we used two complementary platforms to rule out batch effects and obtain both wide and deep sampling of cells and time points. We unraveled a branching lineage including all major neural cell types: astrocytes, oligodendrocytes, granule cells, and hippocampal pyramidal neurons or mossy cells. We observed a clear sequence of distinct steps that constitute granule cell development from quiescent stem or progenitor cells to mature neurons, and we validated these findings using in situ hybridization. Comparing cells across time from embryonic day (E) 16.5 to P132, we demonstrate a perinatal transformation of neurogenesis in the dentate gyrus, from an embryonic to a postnatal configuration. During the second postnatal week, radial glia shifted sharply from an embryonic to an adult state, and during the same time, cycling intermediate precursor cells shifted from a less to a more neurogenic state. A similar stepwise maturation of the granule cells was observed around the third postnatal week, establishing a distinct postadolescent granule cell identity. In contrast, neuroblasts and immature granule cells were indistinguishable at all ages.
cell RNA-seq (10x Genomics Chromium; see Methods). To rule out batch effects, we include a set of experiments (Dataset B) using an orthogonal single-cell RNA-seq technology (valve-based microfluidics chips; Fluidigm C1), sampled at 17 developmental timepoints, with 2,303 cells. Based on these initial findings, we acquired a third, larger, dataset (Dataset C, 10x Chromium), extending to earlier and later timepoints, comprising 24,185 cells from E16.5, P0, P5, P18, P19, P23, P120, and P132. We analyzed the datasets separately because they were obtained using different technologies (A and C vs. B) and kit versions (A vs. C). All datasets are summarized in Supplementary Fig. 1. We provide an online browsable resource (http://linnarssonlab.org/dentate) where gene expression can be visualized in individual cells and clusters of all three datasets.
Focusing first on Dataset A (Fig. 1a), graph-based clustering analysis (Markov clustering 26 of the mutual k nearest neighbor graph; see Methods) revealed 22 distinct types of cells (Fig. 1b), which we arranged in four major categories (Fig. 1c): vascular (endothelial, pericytes, and vascular leptomeningeal cells), immune (microglia and perivascular macrophages), glial, and neuronal cell types. Visualizing these cells using t-distributed stochastic neighbor embedding (t-SNE 27 ; Fig. 1d) revealed the entire architecture of the postnatal developing dentate gyrus in detail (Fig. 1e). At center, dominating the structure, was the granule cell lineage, starting with radial glia-like cells (RGLs), via nIPCs, two neuroblast stages (Neuroblast 1 and 2), immature granule cells, and mature granule cells. RGLs were closely related to, but clearly distinct from, astrocytes. Oligodendrocytes also formed a three-stage lineage (oligodendrocyte precursor cells, newly formed oligodendrocytes, and mature oligodendrocytes 28 Fig. 3). nIPCs were divided by their position along the transition between RGL and NB, as indicated by the line. Numbers of cells as in a. f, Distribution of the percentage of NB-like or RGL-like nIPCs positive for genes associated with cell cycle, neurogenesis, or progenitor cells.
that in addition to the more broadly expressed Reln, Cajal-Retzius cells highly specifically expressed homeobox domain genes such as Lhx1 and Lhx5 ( Supplementary Fig. 2) 29 , as well as cell death regulators Trp73 (encoding p73) and Diablo.

Radial glia-like cells are different from astrocytes and nIPCs.
We first sought to identify the key cell types of early neurogenesis, i.e., RGLs, nIPCs, neuroblasts, and granule cells (Fig. 2). RGLs were identified by their close similarity to astrocytes (Gfap, Hes5, and Sox9) and distinct expression of recently described neuronal stem cell marker Lpar1 30 (also expressed in oligodendrocytes; Fig. 3c). They almost completely lacked expression of the aquaporin-encoding gene Aqp4, characteristic of mature astrocytes. They expressed both Nes (nestin) and Prom1 (also known as CD133), commonly used stem cell markers, although both genes were more highly expressed in endothelial cells (Fig. 3c). The complete absence of expression of cell-cycle genes such as Cdk1, Top2a, and Aurkb showed that these cells are quiescent, in contrast to the actively dividing nIPCs. nIPCs were an actively cycling progenitor population, with a strong cell cycle machinery gene signature (for example, Cdk1, Aurkb, Top2a, and many more). In addition, these cells had taken on an early neuronal fate, expressing the known neurogenic transcription factor-encoding genes Neurog2, Eomes, and Neurod4. This is consistent with previous findings in the developing neocortex, where Eomes (also known as Tbr2) is not expressed by RGL cells; only when entering a proliferating intermediate progenitor state is Tbr2 switched on 31 . In the t-SNE plot, which arranges cells based on their overall transcriptional similarity, nIPCs were located at the root of a trajectory leading to mature granule cells (Fig. 1d).
We further found two neuroblast populations, here defined as the first step of differentiation after nIPCs exit the cell cycle. The first population, Neuroblast 1, retained expression of Eomes (shared with nIPCs), Tac2, and Calb2 (which encodes calretinin), while the second, Neuroblast 2, expressed Gal, and both populations shared expression of Sox11 and doublecortin (Dcx). Both neuroblast stages also expressed Igfbpl1, a marker that we found to be more specific than the widely used Dcx and Sox11 markers (Fig. 2). Together, our findings identify these cells as two sequential steps in the maturation of migrating neuroblasts 32 .    To identify possible transitions between these cell types, we reanalyzed astrocytes and the early neurogenesis clusters (RGL, nIPC, and Neuroblasts 1 and 2) using a graph representation that shows pairs of mutual k nearest neighboring cells in gene expression space (Fig. 2a). In this representation, cells are linked by graph edges when they are each other's nearest neighbors in the highdimensional gene expression space, suggesting possible transitions in expression space. Conversely, the absence of edges suggests disallowed paths through expression space. Graph edges confirmed the order of the trajectory from RGLs, via nIPCs to Neuroblast 1 and then Neuroblast 2. Assuming parsimony in gene expression, it also confirmed that neurogenesis proceeds essentially linearly through a sequence of steps and that the two neuroblast populations are sequential, not parallel. It also demonstrated the close relationship yet clear distinction between astrocytes and RGLs.
With dentate gyrus neurogenesis dissected into several molecularly distinct steps, we performed an unbiased search for celltype-specific gene expression. Focusing on transcription factors, we rediscovered genes encoding previously known key regulators Hes5 (RGLs and astrocytes), Neurog2 (nIPCs), Neurod4 and Eomes (Neuroblast 1), and Sox11 (Neuroblast 2). In addition, using the wide scope of the current study, we identified several highly specific markers and present an expression sequence of single and combinatorial genes (Figs. 2a-c and 3c, Supplementary Fig. 3, and Supplementary Tables 1 and 2), marking the progression from RGL via nIPC to Neuroblasts 1 and 2. For instance, we found the transcription-factor-encoding Tfap2c specifically expressed in quiescent RGLs and nIPCs and propose it as a marker in the context of early neurogenesis. Vnn1, Rhcg, and Wnt8b are further examples of genes more narrowly expressed in RGL cells than commonly used markers.
A fundamental outstanding question in dentate gyrus neurogenesis concerns the distinction between quiescent and active progenitors or stem cells. Is quiescence a distinct transcriptional state, or is it just the absence of cell division? What is the nature of the transition from quiescence to active division, and how is it related  . Nodes represent cells (total n = 3,822 cells); edges connect mutual nearest neighbors. b, Heatmap of genes with differential expression between the maturing cell types, sorted by their peak cluster; colors as in Fig. 1c. c, t-SNE (as in Fig. 1d), stained by contribution from each sampling timepoint. P24* and P35* are estimated to be composed of ~85% of P24 or P35 cells, while the remaining 15% are from the P16 and P12 samples, respectively, due to sex-based sample-separation inaccuracy (see Methods). d, t-SNE (as in Fig. 1d), stained by expression of four markers: Prox1 (pan-GC), Fxyd7 (immature GC-enriched), and Plk5 and Ntng1 (mature GC-enriched). e, RNAScope multiplex in situ hybridizations of dentate gyrus granule layer (as in Fig. 3a,b) from mice aged P10, P16, P24, and P37; representative images from hybridizations on 6 sections in 2 mice per timepoint. Scale bars, 50 μ m. f, Schematic summarizing maturation of granule cells, with immature GCs enriched in blue shaded area (Fxyd7 + ) and mature GCs enriched in yellow shaded area (Ntng1 + Plk5 + ; see Fig. 3b).
to neuronal commitment? As mentioned, nIPC was the only cluster of cells that was proliferating, but this cluster also already expressed neurogenic transcription factors. Both nIPCs and neuroblasts were depleted with age ( Fig. 2b), whereas RGLs were maintained. But how could RGLs be maintained in the absence of cell division? We speculated that a small population of proliferating RGLs might have been clustered with nIPCs, which would not be unexpected since cell division involves the activation of large numbers of genes, which would dominate and drive clustering. To further dissect the proliferating nIPCs, we removed all genes expressed specifically in nIPCs (which would include cell-cycle genes), but not those shared with either RGLs or neuroblasts. We then asked whether individual nIPC cells were now more similar to neuroblasts or to RGLs. After recomputing the k nearest neighbor graph, we found that most nIPCs were now intermingled among the earliest neuroblasts ( Fig. 2e and Supplementary Fig. 4), consistent with their commitment to neuronal differentiation. However, notably, a small number of cells remained close to RGLs in gene expression (expressing, for example, Hes5, Ascl1, Lpar1, and Nes) and did not express most neurogenic transcription-factor-encoding genes (Neurod1, Tbr1, Eomes, Neurod2, and Neurod4), but were still actively proliferating (many genes, including Cdk1, Top2a, Cenpe, Aurkb, Mcm2-8, and Mki67). We confirmed this observation by splitting nIPCs into RGL-like and neuroblast-like subsets ( Supplementary Fig. 4) and measuring the enrichment of progenitor-, cell-cycle-, and neuroblast-specific genes (Fig. 2e,f). This analysis suggests the existence of a small pop-ulation of dividing RGLs that had not entered the neurogenic program. Thus RGLs were mostly quiescent, but could be maintained by occasional proliferation without differentiation (or, alternatively, by occasional asymmetric cell division). This is also consistent with the observation in the subventricular zone that most stem cells are quiescent (91.4% based on the hGFAP:GFP reporter 33 ).
To validate genes expressed specifically in each of the steps of early neurogenesis (Fig. 2) and locate cell types in the tissue, we performed multiplexed RNA staining of the dentate gyrus (Fig. 3a). Ascl1 + Tfap2c + Cdk1 -RGLs were almost entirely confined to the subgranular zone at all ages, as were the proliferating Ascl1 + Tfap2c + Cdk1 + nIPCs (Fig. 3b). Neuroblasts 1 (Igfbpl1 + Eomes + ) or 2 (Igfbpl1 + Eomes -) accumulated just above the subgranular zone, spanning almost the entire thickness of the granule cell layer at P10, but were transient and replaced by mature granule cells (Plk5 + Eomes -) at the later ages (Fig. 3b,d). These findings confirm the existence of distinct RGL, nIPC, and neuroblast cell types and reveal the spatial and temporal dynamics of these cells (Fig. 3d). Further, we showed that the separate or combinatorial expression of these genes labeled the cell types more specifically than previously used markers such as Gfap, Hes5, or Sox2 (shared with astrocytes), Prom1 and Nes (shared with vascular endothelial cells), or Lpar1 (also expressed in oligodendrocytes) (Fig. 3c).

Rapid maturation of granule cells around postnatal week 3.
Noting the rapid maturation of neuroblasts into mature granule   cells between P16 and P24 (compare Igfbpl1 and Plk5 stainings in Fig. 3b), we next focused on the later stages of neurogenesis. As granule cells matured, they transitioned from the still glia-like neuroblasts (Fig. 1d,e) into an immature granule cell state and finally to mature granule cells (Fig. 4a). This transition was characterized by a shift of gene expression most pronounced between the neuroblast to immature stage but also noticeable from the immature to mature stage ( Fig. 4b and Supplementary Tables 3 and 4). Comparing developmental timepoints, we noted that mature granule cells appeared rather abruptly sometime between P16 and P24 (Fig. 4c). This finding was confirmed in tissue by staining for Fxyd7 (neuroblasts), Prox1 (all stages), and Ntng1 (mature granule cells; Fig. 4d-f).
Although a small number of Ntng1-staining cells were observed before P16, their numbers greatly increased at P24 and filled most of the granule cell layer at P37. We took precautions to eliminate batch effects, including analyzing different ages in the same cell preparation and droplet emulsions. However, we were still concerned that our findings of an abrupt shift from immature to mature granule cells around postnatal week 3 might have been spurious. To address this concern, and to extend the previous findings to a higher temporal resolution, we performed a second set of experiments using an orthogonal singlecell RNA-seq technology (Dataset B, Fluidigm C1). We sampled 2,303 single cells at 17 developmental timepoints extending into adulthood, from P8 to P68 ( Supplementary Fig. 1). We performed clustering and identified cell types independently in this separate dataset, which confirmed the findings previously obtained by droplet microfluidics (Supplementary Fig. 5). We confirmed the existence of distinct astrocyte, RGL, nIPC, and neuroblast cell types, as well as the distinction between immature and mature granule cells. At P50 and P63, we enriched for neuronal stem cells by exploiting the hGFAP:GFP reporter mouse 34 . We observed that GFP + cells did not necessarily express endogenous Gfap and included not only astrocytes and RGLs but also perivascular macrophages (PVMs), oligodendrocyte precursor cells, oligodendrocytes, and neuroblasts ( Supplementary Fig. 5). This demonstrated not only that a broader expression in several cell types needs to be considered to avoid confounding results when using hGFAP:GFP to study adult neurogenesis but also the power of single-cell RNA-seq to dissect and resolve the heterogeneity of complex samples.
Taking advantage of the denser temporal sampling, we not only confirmed the rapid maturation of granule cells around the third postnatal week, but also found that it occurs in a very short time interval (Supplementary Fig. 1). Immature granule cells virtually disappeared after P18, whereas mature granule cells appeared abruptly at P20 (which also agreed with our in situ validation;  Thus, our results so far show that the dentate gyrus granule cell lineage proceeds through a sequence of distinct cellular stages, with characteristic and stereotypical spatial distribution from the subgranular zone to the outer granule cell layer. The early cell types-RGLs, nIPCs, and neuroblasts-are clearly distinct states. The presence of a small fraction of dividing RGLs is consistent with self-renewal of the RGL pool, whereas the active cell division of all nIPCs, and their decline over time, support the notion that these cells are a transient amplifying cell type. As granule cells mature, they transition around the third postnatal week from an immature to a mature state, associated with a shift in gene expression involving large numbers of neuronal-related genes. Together, these findings refine and extend previous literature, and we provide extensive lists of cell-type-specific gene expression ( Fig. 4b and Supplementary Fig. 6).

Adult neurogenesis resembles early postnatal development.
A second fundamental question in postnatal neurogenesis concerns the relationship between embryonic development and adult neurogenesis. To what extent are these processes related? Is adult neurogenesis fundamentally different, relying on a distinct type of adult stem cell that differentiates through distinct cellular states to become mature granule cells, or are they one and the same, with developmental stem cells persisting into adulthood and reiteration of the developmental differentiation program? Notably, since these questions are about similarities, not differences, they cannot be adequately addressed using markers alone (i.e., the fact that a marker is shared does not rule out other molecular differences). In contrast, single-cell RNA-seq strongly limits the possibility of unobserved differences, and thus can reveal similarities with much greater confidence.
To address these questions, we generated a third, larger, dataset (Dataset C; 10x Chromium), spanning perinatal (E16.5-P5), juvenile (P18-P23), and adult (P120-P132) animals ( Fig. 5a-d and Supplementary Figs. 1 and 7). Notably, this dataset extended the range of ages sampled to both embryonic and true adult stages. We again confirmed the previous findings: RGLs, nIPCs, and neuroblasts were distinct cell states conserved between juvenile and adult mice, and granule cells matured abruptly around P20. RGLs were also confirmed to be clearly distinct from juvenile (and adult) astrocytes (Fig. 5e), particularly expressing transcription-factor-encoding genes Sox4 and Ascl1, as well as Thrsp (also known as SPOT14), a regulator of adult neurogenesis via lipid metabolism 35,36 that is expressed in the quiescent radial and nonradial cells that give rise to the neuronal lineage in adult mice.
Overall, the t-SNE visualization revealed a similar process to that of Datasets A and B. We found all the same non-neuronal clusters and GABAergic neurons and the same trajectory from RGLs to nIPCs, to neuroblasts, to immature granule cells and then to mature granule neurons. These clusters also largely expressed the same enriched markers as previously observed (Fig. 5d). However, because the dissection at the earlier timepoints necessarily included some hippocampus proper, we now also observed immature pyramidal neurons (Fig. 5a). Notably, these cells branched from the granule cell lineage only at the late postmitotic neuroblast stage, indicating a shared differentiation trajectory.
Furthermore, the extension to perinatal time points (E16.5, P0, and P5) revealed the existence of a perinatal-specific early neurogenesis program (Fig. 5b,c). Perinatal radial glia (RG) were distinct from juvenile and adult RGLs, with greater expression of Sox4 and Sox11 (which are jointly required to generate the hippocampus 37 ), Vim at the younger ages, and an increased expression of Notch2 and Padi2 in the juvenile and adult (Fig. 5d). Both cell types were largely quiescent as indicated by the absence of reliable cell-cycle genes, including Top2a and Cdk1.
We also found perinatal-specific cycling radial glia (RG-cycling; Fig. 5d), which differed from nIPCs in their lower expression of genes encoding neurogenic transcription factors such as Eomes, Elavl2, Elavl4, Neurog2, Neurod4, Sox4, and Sox11. Although RG-cycling were only rarely observed after P5, they would be analogous to the cycling radial glia-like nIPCs (Fig. 2e,f) found in juvenile and adult animals, likely representing RGLs occasionally entering the cell cycle without differentiating. Thus, the stage of initiation of neurogenesis (RG and RG-cycling) showed a clear shift in the second postnatal week from an embryonic form (RG, RG-cycling) to a juvenile and adult form (RGL and cycling radial glia-like nIPCs).
In contrast, nIPCs spanned the full range of ages, from E16.5 to P132 (Fig. 5d), as did neuroblasts and immature granule cells. That is, adult forms of these cell types did not generate separate clusters from those of the perinatal or juvenile stages, in contrast to the clear distinction between perinatal and juvenile forms of radial glia. This was true both for cells sampled without selection and for hGFAP:GFP + cells sorted by fluorescence-activated cell sorting (FACS, Supplementary Fig. 1g; compare also the result from Dataset B in Supplementary Fig. 5).
To quantify this observation (Fig. 6), we performed pairwise comparisons of perinatal versus juvenile and adult types (where we would expect significant differences for RG only), as well as juvenile versus adult (where we would expect minor differences). Comparing perinatal RG to juvenile RGLs (Fig. 6a), we found many genes that were both statistically significantly differentially expressed (red circles) and showed greater than twofold differences (outside the dashed lines), including Thrsp in (2.48-fold, P value 1.7 × 10 -10 ) the juvenile and Vim (2.79-fold, P value 7.7 × 10 -17 ) in the perinatal. In contrast, comparing juvenile to adult RGLs (Fig. 6d), we found few genes that were significantly differentially expressed (e.g., Aldoc, q value 4.2 × 10 -11 ) and only three (Pla2g7, Sparcl1, and Aldoc) showed greater than twofold change. Thus, RG cells, which initiate neurogenesis, switch their molecular identity after P5 and then maintain it until at least P132.
In contrast, both nIPCs and neuroblasts at all ages were near indistinguishable. nIPCs differed by only two genes between perinatal and juvenile (expressed at least twofold; Fig. 6b) and by three between juvenile and adult (Fig. 6e). In neuroblasts, not a single gene showed twofold difference between perinatal and juvenile or juvenile and adult mice (Fig. 6c,f). We conclude that although the initiation stage of neurogenesis differs between juvenile and adult compared to perinatal and embryonic, the differentiation trajectory of cells committed to neuronal fate is conserved from the nIPC stage onwards.
One unique feature was observed during development that was absent in the adult. Immature granule cells were found only in young animals, while in adults these were almost absent (Supplementary Figs. 1 and 5d). This was validated in situ, where detection of intermediate marker Fxyd7 was strongly decreased, but not absent, in P24 and P37 (Fig. 4e). This does not necessarily indicate that adult neurogenesis skips a step, but more likely that the great abundance of mature granule cells (which accumulate by age) and the decreasing rate of neurogenesis at this age makes transient intermediate stages more difficult to observe.

Discussion
We have described the perinatal development and persistent adult neurogenesis of the dentate gyrus using single-cell RNA-seq. Our data strongly supports a unified process of early postnatal and adult neurogenesis, with a set of clearly defined, rather than continuous, cell types and transitions. Both the cell types and transitions were invariant over time, but with changes in the number of cells of each type as well as their spatial distribution (Figs. 2 and 5). The long-lived stem cell of the dentate gyrus is the RGL, as supported by clonal analysis of Nes + RGLs 38 . We found that RGLs were predominantly quiescent, even during development (Fig. 2), and that in that state they were marked by a very small number of highly enriched markers, which included Rhcg, Vnn1, and Lpar1.
After RGLs entered the cell cycle and the nIPC state, a large number of neurogenic transcription-factor-encoding genes, including Neurod1, Neurod2, Neurod4, and Eomes were induced, after which the cells likely would proceed irreversibly to a neuronal fate. However, a small proportion of cycling cells retained features of RGL and did not show robust activation of the neurogenic factors ( Supplementary Fig. 4). Thus the pivotal event in dentate neurogenesis is the fate choice occurring as an RGL begins to divide.
Once nIPCs have committed to a neuronal fate, they either continue to divide or exit the cell cycle as neuroblasts. Data from the SVZ suggests that the average nIPC divides about four times 33 , resulting in an approximately eightfold amplification from RGLs to the neuroblast stage. Since we observed a decrease in actively cycling nIPCs with age-likely explained by a reduced amplification rate-depletion of neuroblasts should be expected. At the same time, the gradual reduction of nIPCs and neuroblasts (Fig. 5) confirms that they must be transient cellular states, while the abundances of RGLs and astrocytes remained roughly stable in adult animals. We thus speculate that once neurogenic transcription factors are induced, cells proceed unidirectionally and irreversibly through nIPC, Neuroblasts 1 and 2, and immature and mature granule cell stages.
Extending the analysis to late embryonic and early postnatal stages (perinatal, E16.5-P5), we found that the differentiation lineage was unchanged from the nIPC stage through neuroblast and immature granule cell stages, which were indistinguishable between E16.5 and P132. This reinforces the fundamental importance of the induction of neurogenesis, which once induced will proceed through a stereotyped differentiation trajectory. In contrast, on and before P5, including in the embryo at E16.5, the initiation stage of neurogenesis differed markedly compared with that in juveniles and adults. The quiescent radial glia differed, with a perfect separation at the second postnatal week: no perinatal RGs were found after P5, and no juvenile or adult RGLs were found before P18 (Fig. 5d). These findings agree with and extend previous findings 39 , which describe a shift in the expression of key stem cell marker genes Sox2, Nes, and Fabp7 between P7 and P14, indicating a change in the identity of radial glia.
Resequencing at close intervals during postnatal development (Dataset B) revealed another sharp transition during development at P20, when mature granule cells were formed rapidly (a similarly sharp transition was observed for mossy cells). A parsimonious interpretation is that before P20, permissiveness for maturation of granule cells is absent, resulting in the accumulation of an immature granule cell type; these cells mature all at once when the environment becomes permissive. A shift in the neurogenic profile around P20 is consistent with early studies by Altman and colleagues, in which the transition of developmental to adult neurogenesis was seen to occur between P10 and P30 in the rat 2 . In Nestin-eGFP mice, a strong drop in GFP + neuronal progenitors was reported between P14 and P28 40 .
Major efforts have been invested into isolating and studying populations of stem and progenitor cells, using a wide range of animal models and marker genes 41 . We showed that several of these marker genes are not specific to quiescent or cycling progenitors, greatly restricting their utility for such purposes. For instance, Gfap, Hes5, and Sox2 expression were shared with astrocytes, while markers like Prom1, Lpar1, and Nes were highly coexpressed by entirely different cell types (endothelial cells and oligodendrocytes; Fig. 3c), and even Reln expression by Cajal-Retzius cells is not exclusive to this population ( Supplementary Fig. 2).
The wide scope and unbiased nature of our study allowed us to suggest an array of sequential or combinatorial markers that are more narrowly expressed in the relevant cell types and stages of early and late neurogenesis. One such marker gene is Tfap2c, highly specific to RGL and nIPC in our dataset. Tfap2c (also known as Tcfap2c and AP2ɣ) encodes a transcription factor that plays a key role in early mammalian extraembryonic development and organogenesis. A murine knockout for this gene is embryonic lethal as early as E7.5 42,43 . Recently, Tfap2c was shown to be involved directly in hippocampal neurogenesis and suggested to regulate the key neurogenic genes [44][45][46] . Our data suggest that Tfap2c is an exclusive marker for the hippocampus stem cell niche, and we speculate that one role could be to ensure a certain level of proliferation 47,48 , while at the same time determining the fate of subgranular zone RGLs by prespecification. This would be similar to its function in the developing cortex where it specifies onset of a Tbr2-NeuroD neurogenic lineage program specifically generating certain layer II/III neurons 45 .
Our findings establish the essential unity of perinatal, juvenile, and adult neurogenesis in the mammalian hippocampus ( Supplementary Fig. 8), but with sharp changes at around the second (RGLs) and the third (granule cells) postnatal weeks; demonstrate that neurogenesis proceeds through distinct discrete states; show that the intermediate progenitor and neuroblast stages are molecularly conserved from E16.5 to P132; provide markers with high specificity for each state; show their spatial distribution during development; and suggest a simple model for hippocampal neurogenesis driven by the choice RGLs make to differentiate, or not, as they divide.

Animals.
We used male and female wild type CD-1 mice (Charles River) and C57Bl/6 mice and hGFAP-GFP reporter mice 49 on embryonic day E16.5 and between postnatal days P0-P132. Details on the mice used per experiment are found in Supplementary Tables 5 and 6. All experimental procedures followed the guidelines and recommendations of Swedish animal protection legislation and were approved by the local ethical committee for experiments on laboratory animals (Stockholms Norra Djurförsöksetiska nämnd, Sweden).

Single cell dissociation.
Dissection and single-cell dissociation were carried out as previously described 28 . Briefly, mice were killed by an overdose of isoflurane, followed by transcardial perfusion through the left ventricle with artificial cerebrospinal fluid (aCSF, in mM: 87 NaCl, 2.5 KCl, 1.25 NaH 2 PO 4 , 26 NaHCO 3 , 75 sucrose, 20 glucose, and varying concentrations of CaCl 2 and MgSO 4 ). The concentrations of CaCl 2 and MgSO 4 were adapted according to sampling age: 2 mM each for mice aged < P18, 1 and 7 mM CaCl 2 and MgSO 4 , respectively, for mice aged ≥ P18. Crucially, aCSF was equilibrated in 95% O 2 5% CO 2 before use and cells were kept on ice or at 4 °C at all steps except enzymatic digestion. The brain was removed, 300-μ m vibratome sections were collected, and the dentate gyrus was microdissected. Single-cell suspensions were prepared using the Papain kit (Worthington) with 25-35 min enzymatic digestion (depending on age), followed by manual trituration using a BSA-coated p1000 pipette or fire-polished Pasteur pipettes.
FAC sorting. hGFAP-GFP mice were killed for analysis and single cell suspensions prepared as described above. On a BD FACSAria II, GFP + cells were sorted into oxygenated aCSF at 4 °C, inspected for viability, counted, and loaded on the Fluidigm C1 or 10x Chromium systems, as described below.
Fluidigm C1. For processing on Fluidigm C1, cells were resuspended in aCSF with 5% DNase, counted, and diluted to 500-800 cells/μ L for cell load at 4 °C on a medium-sized chip. All capture sites were imaged and manually checked for the presence of a single, healthy cell; other wells were excluded from library preparation. cDNA synthesis, with unique molecular identifiers (UMIs) and 5′ sequencing library preparation by tagmentation followed by multiplexing, was carried out using the C1-STRT protocol, as described before 50 . An overview of Fluidigm C1 experiments is found in Supplementary Table 6. 10x Genomics Chromium. Dataset A sampling was carried out on two experimental days (Supplementary Table 5), pooling samples from male and female mice in a single experiment each day to further reduce the risk of batch effects. The dentate gyrus of male (day 1, P35; day 2, P24) and female (day 1, P12; day 2, P16) mice were dissected and dissociated separately. Suspensions were pooled at equal ratios after visual inspection and counting. Dataset C was sampled as indicated in Supplementary Table 5. We then resuspended 300-1,000 cells/μ L in aCSF and added to the RT mix to aim for sampling of around 3,500-5,000 cells. All downstream cDNA synthesis (13 PCR cycles), library preparation, and sequencing were carried out as instructed by the manufacturer (10x Genomics Chromium Single Cell Kit Version 1 (Dataset A) or Version 2 (Dataset C)). Data analysis on 10X Chromium Datasets A and C, clustering, and visualization. 10x Chromium Dataset A (Dataset C) data files were loaded and merged to one dataset. For more uniform representation of timepoints, Dataset C data files were downsampled as indicated in Supplementary Table 5 ('Cells_analysed'). Valid cells were defined as having > 600 (C:> 800) total genes, between 800 and 20,000 (C: 1,000-30,000) molecules, and a ratio of molecules to genes > 1.2. This resulted in 6,090 (C: 25,546) cells. Next, we removed 46 (C: 732) doublet cells based on coexpression (> 1 molecule) of any pair of the following marker genes: Stmn2 (neurons), Mog (oligodendrocytes), Aldoc (astrocytes), C1qc (microglia), and Cldn5 (endothelial). In Dataset A, we here allocated cells to their sampling age; see separate section below. In Dataset C, we here normalized the number of total molecules per cell to 5,000, rounding to integers. We then performed basic feature selection by removing genes with low or wide expression (less than 20 detected molecules over all cells or expressed in more than 60% of the cells). Next, we plotted coefficient of variation versus mean (log(CV) vs. log(mean)), fit a linear function, and ranked the genes according to distance from the line to select the top 5,000 differentially expressed genes. To exclude an effect of sex or stress genes on clustering, we removed genes previously found to be involved in either (Ehd2, Espl1, Jarid1d, Pnpla4, Rps4y1, Xist, Tsix, Eif2s3y, Ddx3y, Uty,   Kdm5d, Rpl26, Gstp1, Rpl35a, Erh, Slc25a5, Pgk1, Eno1, Tubb2a, Emc4, and Scg5). Before clustering, the order of cells and genes was permuted, and in Dataset A the total number of molecules was normalized to 10,000 per cell. We then calculated PCA projections with 100 components on the log 2 centered normalized data and calculated the correlation matrix between cells on the PCA coordinates. On a correlation-based mutual k nearest neighbor (KNN) graph (Dataset A, k = 20; Dataset B, k = 30), we then calculated the Jaccard distance between cells (fraction of mutual neighbors). Jaccard distances were transformed to edges connecting cells in a graph object (Matlab 'graph' function) and the layout was calculated (Matlab 'force' option). Cliques (cells connected by edges) of less than 10 cells were removed. This Jaccard KNN graph was clustered using MCL (http://micans.org/ mcl/), with MCL parameter 1.25 (C: 1.4), minimum cluster size 5, and maximum cluster size 1,000 (C: 3,000). This resulted in 40 (C: 46) clusters, which we plotted on KNN using the force layout. We next calculated a marker table matrix of enriched genes, using top 20 markers power 0, 0.5, and 1. Manual inspection of this marker table was necessary to merge mainly granule cell clusters (oversplit by MCL), split the cluster of mixed PVM and VLMC, and remove one cluster of suspected doublets, resulting in 22 (C: 25) final clusters. We updated our marker table matrix based on the final clusters, using top 20 markers power 0, 0.5, and 1, resulting in 675 (C: 756) genes. The same gene selection was used to calculate a dendrogram using linkage (Ward's, correlation), correlation matrix (each averaged over clusters), and t-SNE projection of all cells with the following parameters: Dataset A: initial PCA, 80; perplexity, 50; epsilon, 100; correlation distance, 1,000 iterations; Dataset C (fast implementation BH-tSNE): initial PCA, 50; perplexity, 100; theta, 0.2.

Data analysis on Fluidigm C1 (Dataset B), clustering, and visualization.
For C1 data analysis we used exactly the same analysis pipeline described above, with the following modifications: (i) minimum total molecules set to 1,500; (ii) for doublets removal we added C1ql3 as an additional neuronal marker, and the threshold for exclusion was set more stringently to > 0 detected molecules; (iii) for KNN, k = 10; and (iv) perplexity in t-SNE was set to 60. The number of clusters after MCL was 39; these were manually inspected and partly merged, resulting in 17 final clusters.
Here, we started out with 4,425 valid cells and removed 719 suspected doublets, as well as cells belonging to small cliques, resulting in 2,706 cells. Clusters suspected to be low quality (lack of specific marker genes) were removed, resulting in 2,303 cells in the final dataset.
Dataset A: allocation of sampling age to cell. Dataset A sampling was carried out on two experimental days, each day pooling samples from male and female mice in a single experiment to reduce the risk of batch effects. Thus, the dentate gyrus of male (day 1, P35; day 2, P24) and female (day 1, P12; day 2, P16) mice were dissected and dissociated separately, but pooled at equal ratios to a single Chromium well. To separate male and female cells (and thus timepoints), the sampling timepoint (P12, P16, P24, or P35) was assigned to each cell according to sample name (10 × 43_1 or 10 × 46_1) and expression of sex-specific genes: Ddx3y, Uty, and Eif2s3y for male-derived and Xist and Tsix for female-derived cells. We thus allocated each cell to a female (P12, P16) or male (P24, P37) source (i.e., if the sum of female genes was higher than that of the male genes, the cell was defined as female and vice versa). For zero expression of either gene sets, the sex was labeled unknown, but was most likely to be male (P24 or P37), as the expression of female genes (mainly Xist) was detected in 85% of cells coming from similar samples of known female-only origin (data not shown). In these cases, we labeled the cells P24* or P37*, and expect that about 15% of cells classified were false-positives.

Analysis of early neurogenesis and granule cell maturation.
To analyze early neurogenesis (Fig. 2) and granule cell maturation (Fig. 4), we generated separate mutual KNN graphs and lists of peak cluster enriched genes. For this, we loaded data only from relevant cell clusters (early neurogenesis: astrocytes, RGL, nIPC, NB1, and NB2; GC maturation: NB2, immature granules, and mature granules). Low or widely expressed genes (expressed in less than 20 cells or in more than 60% of all cells) were filtered, and the 5,000 most differentially expressed genes were selected by plotting log(CV) vs. log(mean) (as above). All cells were normalized to 10,000 molecules. PCA projections were calculated with 100 components, and correlation based on the PCA coordinates was calculated. A mutual KNN graph (k = 40) was calculated based on correlation as distance, and weights in the mKNN graph were replaced by the Jaccard distance (fraction of shared neighbors). Cliques (connected components) of fewer than 10 cells were removed. The graph was calculated (Matlab force option), and original clusters were visualized on the new layout. For analysis of early neurogenesis without nIPC-specific genes, the same pipeline was used, with the additional removal of genes significantly specific to nIPCs by rank-sum test (FDR 5%) after feature selection.
To identify enriched genes per peak cluster, we loaded data from relevant cell clusters (as above). For each transition identified by the KNN (for example, astrocytes to RGL), we then calculated significantly expressed genes using ranksum tests and fold enrichment as the log mean average in the population divided by the log mean average in all groups. Significantly expressed genes were defined as those with q < 0.01 in one of the transitions, fold enrichment > 1.5, and followed the requirement that fold enrichment in the first or last population be < 0.1 (early neurogenesis) or < 0.4 (GC maturation).
Pairwise cluster comparison. Prior to comparison, cells were normalized to 5,000 molecules. To remove nonrelevant genes potentially adding noise and false positives in later testing, rough-feature (gene) selection was performed, testing clusters of interest against all other clusters. Here, only genes passing FDR 20% after t test on log 2 (x + 1)-transformed data were selected. Next, the top 3,000 differential genes were selected based on log(CV) vs. log(mean) (as above). Finally, we performed gene-by-gene rank-sum testing for each pairwise group, recording q values (FDR corrected P values) and log 2 (x)-fold change. The number of cells included in each cluster by age is listed in Supplementary Table 7.
RNAscope. CD-1 mice (Charles River) were killed with an overdose of isoflurane and perfused through the left ventricle with PBS. Brains were immediately dissected out, embedded in OCT on dry ice and stored at -80 °C. We collected 10-μ m cryostat sections covering the anterior-posterior axis of the dentate gyrus. The sections were either fixed in 4% PFA for 5 min, rinsed in PBS, and stored at -80 °C until staining, or immediately frozen at -80 °C and fixed after thawing. RNAScope hybridizations were carried out according to the manufacturer's instructions, using the RNAscope Multiplex Fluorescent (Advanced Cell Diagnostics) for fresh frozen sections. Briefly, thawed sections were dehydrated in sequential incubations with ethanol, followed by 30  Immunohistochemistry. CD-1 mice (Charles River) were killed with an overdose of isoflurane and perfused through the left ventricle with PBS, followed by 4% PFA. Brains were dissected out, postfixed in 4% PFA for 16 h, cryoprotected in 10% and 30% sucrose, embedded in OCT, and frozen and stored at -80 °C. We cut 16-μ m sections on a cryostat, covering the anterior-posterior axis of the dentate gyrus, and frozen at -80 °C until staining. Before staining, thawed sections were treated with 1× Target Retrieval solution (Dako), according to the manufacturer's instructions. Primary antibodies (goat anti-GFP (Rockland 600-101-215, 1:1,000), rabbit anti-GFAP (Dako Z-0334, 1:500), rabbit anti-GFP (Molecular Probes A-11122, 1:1,000), goat anti-AldolaseC (Santa Cruz Biotech. N-14 SC-12065, 1:500), and goat anti-PDGFRA (R&D AF1062, 1:200)) were incubated O/N, at 4 °C, in PBS with 5% normal goat or donkey serum, 3% BSA, 0.5% sodium azide and 0.3% Triton X-100, followed by washing and incubation with appropriate secondary antibodies (donkey anti-goat Alexa Fluor 488, donkey anti-rabbit Alexa Fluor 555, donkey anti-rabbit Alexa Fluor 488, donkey antigoat Alexa Fluor 555; all Molecular Probes, 1:1,000), overnight at 4 °C in PBS, counterstained with Hoechst (Molecular Probes), and mounted with ProLong Gold (Molecular Probes).

Statistics and reproducibility.
No statistical methods were used to predetermine sample sizes; our sample sizes were determined iteratively. The study did not include any intervention, so no randomization was performed and data collection and analysis were not performed blind to the conditions of the experiments. The conclusions drawn from scRNA-seq data presented in this study are based on three independent experimental series performed on three different scRNA-seq protocols. We provide details of the number of experiments contributing to each dataset in Supplementary Tables 5 and 6. To identify genes enriched to specific cell types, we performed two-sided Wilcoxon rank-sum tests (no underlying distribution assumed), with correction for multiple testing (FDR), as detailed in the Method sections "Analysis of early neurogenesis and granule cell maturation" and "Pairwise cluster comparison. " For validation of marker genes identified in scRNAseq, in situ hybridizations (see "RNAScope") were carried out on 6 sections each of 2 mice per timepoint. Similar results were repeatedly obtained and representative images are presented. Immunohistochemistry was carried out on sections from two animals, with similar results. Further details on experimental design are found in the Life Sciences Reporting Summary.
Life Sciences Reporting Summary. Further information on experimental design is available in the Life Sciences Reporting Summary.
Data and code availability. The raw datasets generated in the current study are available in the NCBI Gene Expression Omnibus (GEO) repository, accession number GSE95753. We provide an online browsable resource of single-cell expression data for all three datasets at http://linnarssonlab.org/dentate/. The code used to perform analyses in this paper is available on GitHub at https://github.com/ linnarsson-lab/DG_paper_public. nature research | life sciences reporting summary June 2017

Statistical parameters
For all figures and tables that use statistical methods, confirm that the following items are present in relevant figure legends (or in the Methods section if additional space is needed).

n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement (animals, litters, cultures, etc.) A description of how samples were collected, noting whether measurements were taken from distinct samples or whether the same sample was measured repeatedly A statement indicating how many times each experiment was replicated The statistical test(s) used and whether they are one-or two-sided (note: only common tests should be described solely by name; more complex techniques should be described in the Methods section) A description of any assumptions or corrections, such as an adjustment for multiple comparisons The test results (e.g. P values) given as exact values whenever possible and with confidence intervals noted A clear description of statistics including central tendency (e.g. median, mean) and variation (e.g. standard deviation, interquartile range)

Clearly defined error bars
See the web collection on statistics for biologists for further resources and guidance.

Software
Policy information about availability of computer code

Software
Describe the software used to analyze the data in this study.
scRNA-seq analyses were performed using Cell Ranger v1.1, and v1.3 and MATLAB 2016B. RNAscope and Immunofluorescence imaging was carried out using the NIS-Elements AR (v 4.40) software and analyzed in Fiji/ImageJ (v2.0.0) For manuscripts utilizing custom algorithms or software that are central to the paper but not yet described in the published literature, software must be made available to editors and reviewers upon request. We strongly encourage code deposition in a community repository (e.g. GitHub). Nature Methods guidance for providing algorithms and software for publication provides further information on this topic.

Materials and reagents
Policy information about availability of materials 8. Materials availability Indicate whether there are restrictions on availability of unique materials or if these materials are only available for distribution by a for-profit company.
No restrictions apply 9. Antibodies Describe the antibodies used and how they were validated for use in the system under study (i.e. assay and species).
Commercial antibodies against well-known antigens were used, and validated by the manufacturer: