A dynamic gene regulatory code drives synaptic development of hippocampal granule cells

Abstract

Connecting neurons into functional circuits requires the formation, maturation, and plasticity of synapses. While advances have been made in identifying individual genes regulating synapse development, the molecular programs orchestrating their action during circuit integration of neurons remain poorly understood. Here, we take a multiomic approach to reconstruct gene regulatory networks (GRNs), comprising transcription factors (TFs), regulatory regions, and predicted target genes, in hippocampal granule cells (GCs). We find a dynamic gene regulatory code, with early and late postnatal GRNs regulating cell morphogenesis and synapse organization and plasticity, respectively. Our results predict sequential regulations, with early-active TFs delaying the activation of later GRNs and their putative synaptic targets. Using a loss-of-function approach, we identify Bcl6 as a regulator of pre- and postsynaptic structural maturation and synaptic transmission and Smad3 as a modulator of inhibitory synaptic transmission in GCs. Together, these findings highlight the networks of key TFs and target genes orchestrating GC synapse development.

A multiomic approach combined with loss-of-function analysis reveals the molecular program of granule cell synaptic development.

INTRODUCTION

Neurons are connected into neural circuits via synapses, highly organized and specialized cell-cell junctions. Proper assembly of neural circuits requires neurite outgrowth and guidance, recognition of appropriate cellular partners, and formation of synaptic contacts, followed by maturation and plasticity of synapses (1–3). The past decades have seen major progress in the identification of key genes regulating each of these steps in circuit assembly (4–9). How the functions of these individual genes are coordinated to orchestrate the integration of neurons into functional circuits remains unclear.

Gene regulatory networks (GRNs), consisting of transcription factors (TFs), target regulatory regions, and their target genes (TGs), coordinate the action of individual genes into molecular programs. Key insight has been derived from the analysis of GRNs in the nervous system development of Caenorhabditis elegans and Drosophila (10–13). These studies have postulated the terminal selector hypothesis, in which combinations of constantly expressed TFs define cell identity and key cellular properties such as connectivity. Work in the mouse nervous system has also provided evidence for individual TFs controlling both cellular identity and projection patterns by regulating the expression of axon guidance receptor genes (14, 15). Together, these studies imply a tight link between the transcriptional control of cell identity and the connectivity pattern. Less is known about the GRNs regulating the formation, maturation, and plasticity of synaptic contacts. This may be due to complicating factors such as the protracted period required for synapse development in mammalian nervous systems and the fact that trans-synaptic interactions between pre- and postsynaptic cells play a key role in determining synaptic properties (16).

In this study, we focus on hippocampal granule cells (GCs), a cell type with a well-characterized development in terms of morphology and connectivity. GCs represent the first relay in the hippocampal trisynaptic loop (Fig. 1A), receiving excitatory input from the entorhinal cortex on their dendrites and providing excitatory input to CA3 pyramidal neurons via their mossy fiber boutons (MFBs), and play an essential role in processing spatial representations and memory acquisition (17, 18). Several studies have examined the role of individual TFs in GC synapse development via loss-of-function approaches (19–23). Sox11 knockout (KO) severely disrupts the targeting of GC axons, the mossy fibers (MFs), in the CA3 region and alters MF-CA3 synapse function (24). Loss of Klf9 results in a transient increase in dendritic spine density and length in GCs (22). Mef2c KO increases dendritic spine density and synaptic transmission at entorhinal cortex-GC synapses (23). GCs in Bcl11b KO mice show a reduced density of dendritic spines and aberrant MF targeting (20). The synaptic organizer C1QL2 (complement component 1, Q subcomponent–like 2) was recently identified as a target gene of BCL11B (19), providing a first causal link between a TF and a synaptic effector in the development of structure and function of GC synapses. However, a comprehensive analysis of the gene regulatory code driving the synaptic development of GCs is lacking.

Fig. 1. Transcriptomic analysis reveals dynamic changes in synaptic gene expression in GCs.

(A) Scheme of the hippocampal formation and situation of GCs in the hippocampal trisynaptic circuit: GCs (red) in the DG receive excitatory inputs from the entorhinal cortex (EC; blue) and send axons to CA3. The inset shows postnatal GC development: NPCs are generated on P1; these will then give rise to NBs and iGCs, which extend dendrites to receive input from mossy cells and send axons toward the CA3 region. At P7, iGCs give rise to mGCs, which receive EC input and develop protrusions along their axon. At P14, mGCs present a complex dendritic arbor, increased dendritic spine density, and complex MFBs along their axon. At P28, mGCs reach peak dendritic spine density, while MFBs are refined, having fewer filopodia. (B) Hippocampi of Rbp4-Cre:Ai9 mice at P7, P10, P15, and P28. tdTomato (red) labels GCs in the outer part of the GC layer at each time point. Scale bars, 500 μm. (C) Time points used for scRNA-seq experiments; n = 1 mouse per time point, except P28 (n = 2). (D) PCA of highly variable genes (HVGs) separates cells by time point. (E) PCA colored by the expression of the GC marker gene calbindin 1 (Calb1). (F) Heatmap of the eight clusters obtained from the top 50 DEGs of each pairwise comparison between time points, highlighting, for each cluster, genes present in the SynGO database within the top 20 genes of each pairwise comparison. Cells are ordered by inferred pseudotime, and values represent GAM-fitted expression values for each gene. (G) Dot plot representing GO BP terms enriched in each cluster.

Here, we first use single-cell RNA sequencing (scRNA-seq) to map dynamics in synaptic gene expression in GCs as they integrate into the trisynaptic hippocampal circuit during postnatal development. We then use single-nucleus multiome (snMO) analysis, integrating transcriptome and chromatin accessibility data of the same cells, to reconstruct GRNs. This reveals a dynamic gene regulatory code consisting of early-postnatal GRNs controlling cell morphogenesis and late-postnatal GRNs regulating synapse organization and plasticity. We validate these observations using ribosome tagging, demonstrating that TF and target gene mRNAs are translated in GCs. To test the functional relevance of TFs in GC synapse development, we perform a loss-of-function analysis and identify Bcl6 as a regulator of dendritic spine density, MFB morphology, and synaptic transmission and Smad3 as a modulator of inhibitory synaptic transmission. Last, we construct a GRN model of postnatal GC development on the basis of our results, highlighting how early-postnatal repressor TFs target late-postnatal activators. Together, our findings uncover a dynamic gene regulatory code governing the synaptic development of GCs, provide insight in key TFs and target genes, and identify Bcl6 and Smad3 as regulators of specific aspects of GC synapse development.

RESULTS

Single-cell transcriptomic analysis reveals dynamic changes in synaptic gene expression in GCs

The structural integration of GCs into the hippocampal circuit is well characterized (Fig. 1A) (25–30). In mice, GC neurogenesis peaks around birth, as neural progenitor cells (NPCs) give rise to neuroblasts (NBs), which transition into immature GCs (iGCs) around postnatal day 3 (P3). At this stage, iGCs receive excitatory inputs from mossy cells located in the dentate gyrus (DG) hilus and begin extending axons toward the CA3 region. By P7, iGCs give rise to mature GCs (mGCs), receiving inputs from the entorhinal cortex in the medial molecular layer and displaying some axonal protrusions. By P14, mGCs exhibit an increased dendritic length and spine density, while MFBs in the CA3 region reach peak structural complexity, preceding synaptic remodeling. By P28, mGCs have achieved their final dendritic length and fully developed MFBs. In addition, GCs start receiving inhibitory input from different types of GABAergic neurons during this developmental period (31).

To map accompanying changes in gene expression, we performed scRNA-seq on GCs from Rbp4-Cre:Ai9 mice at key time points during synapse development (Fig. 1, B and C). The Rbp4-Cre line specifically expresses Cre recombinase in GCs (32, 33), resulting in tdTomato expression when crossed with Ai9 reporter mice (34). At each developmental time point, labeled GCs in Rbp4-Cre:Ai9 mice localize to the outer part of the GC layer, representing more mGCs (Fig. 1B). tdTomato-positive GCs were fluorescently sorted and processed for single-cell sequencing using Smart-seq2 to detect a high number of genes per cell (fig. S1, A to D, and Materials and Methods) (35). Alternating male and female samples were used to compensate for sex-related differences, with the exception of the P28 time point, for which both male and female samples were analyzed (fig. S1B). Differential expression analysis between the male and female P28 scRNA-seq samples showed only a small number (n = 18) of differentially expressed genes (DEGs; 0.05 adjusted P value, 0.5 |log fold change (LFC)|), indicating minimal sex-related transcriptional differences at this stage (fig. S1E). Principal components analysis (PCA) demonstrated a clear age-based separation of cells (Fig. 1D). Sorted cells expressed the postmitotic GC marker gene Calb1 (Fig. 1E), as well as other GC and neuronal markers (Prox1, Rfx3, Slc17a7, and Gad1; fig. S1, F and G) (36, 37).

Pairwise differential gene expression analysis [MAST (model-based analysis of single-cell transcriptomics) test, P value <0.05, |LFC| > 0.5] revealed significant gene expression changes in GCs across postnatal development, with prominent shifts at P7 and P15 (fig. S1H). Subsequent clustering of the top 50 DEGs per pairwise test (see Materials and Methods) showed eight clusters according to their gene expression profiles across pseudotime (Fig. 1F). We next examined synaptic gene expression patterns over pseudotime by analyzing DEGs annotated in the expert-curated SynGO database (38) within these clusters and linking them to the function of each cluster using Gene Ontology (GO) analysis [biological process (BP); P value of 0.05, q value of 0.05] (Fig. 1G). In the early clusters (clusters 1-2-3), we found genes related to axonogenesis and regulation of protein-containing complex assembly, like Gap43 and Ncam1, respectively, as well as Dpysl2 and Dpysl3, cytosolic phosphoproteins involved in growth cone collapse and other critical functions in developing neurons (Fig. 1, F and G) (39). Cluster 3 contained, among others, Efnb3, an axon guidance cue (Fig. 1, F and G) (40). The middle clusters (clusters 4-5-6) contained genes linked to vesicle-mediated transport (e.g., Apba2 and Ap2m1) and others linked to learning and memory. Last, the late clusters (clusters 7-8) contained genes peaking in expression around P15 to P28, which are involved in synaptic organization, such as the postsynaptic cell adhesion molecule Nlgn3, and synaptic plasticity, such as the secreted synaptic organizer Nptx1 (Fig. 1, F and G). Together, this analysis reveals the dynamics in synaptic gene expression during circuit integration of GCs, with an early-postnatal (P5 to P7) phase related to cell morphogenesis and a late-postnatal phase (P15 to P28) regulating synapse organization and plasticity.

snMO analysis reveals GRN dynamics during the postnatal development of GCs

To map the GRNs underlying the dynamic changes observed in synaptic gene expression during postnatal GC development (Fig. 1, F and G), we generated a 10x snMO dataset of P5, P10, P15, and P28 wild-type (WT) mouse hippocampus (Fig. 2, A and B). Combined transcriptome and chromatin accessibility from the same nuclei enables the reconstruction of enhancer-driven regulons (eRegulons) comprising TFs, their target regulatory regions, and target genes (41). After sequencing and quality controls, cell types were annotated, and GC lineage nuclei were extracted using marker genes (fig. S2, A to H, and Materials and Methods) (36, 37). NPCs, found in the subgranular zone of the DG, are the most immature cells, identified on the basis of Cdk6 and Top2a expression (fig. S2A). They are chronologically followed by type 1 NBs, expressing Elavl2 and Calb2, and type 2 NBs, which express Rgs6. Type 2 NBs give rise to iGCs, marked by Pcdh15 and Prox1 expression, and lastly, mGCs, the main excitatory neurons of the DG, expressing Meg3, Calb1, and Ntng1 (fig. S2A). Marker genes (fig. S2A) and similarity analyses using the cosine metric (fig. S2B) confirmed that the mGCs in this dataset align best with the GCs from our scRNA-seq dataset (Fig. 1), labeled by Rbp4-Cre:Ai9. Each age showed varying proportions in cell types, with a higher proportion of NPCs at P5 and more mGCs at P28 (fig. S2I). Uniform manifold approximation and projection (UMAP) based on transcriptome and epigenome profiles depicts the transition from NPCs to mGCs (fig. S2, J to L).

Fig. 2. snMO analysis reveals GRN dynamics during postnatal GC development.

(A) Methodology for snMO experiments. Hippocampi from C57Bl6 mice were homogenized, and nuclei were isolated and processed for both RNA-seq and ATAC-seq. From this dataset, the GC lineage was extracted using markers, and SCENIC+ was applied to infer eRegulons (TF – accessible target regions – target genes). Only mGCs are shown in the following panels. (B) Time points used for snMO experiments; n = 1 mouse per time point. (C) PCA based on eRegulon AUC scores of target genes separates mGC nuclei by time point. (D) Left panel: Heatmap of leading TF expression across time in snMO RNA-seq. Right panel: Heatmap of SCENIC+-filtered eRegulons’ (see Materials and Methods for filtering criteria) AUC scores in mGCs across time points. Cells are ordered by inferred pseudotime values. (E) Dot plot representing GO BP terms enriched in the target genes of representative eRegulons.

Focusing on mGCs, we analyzed snATAC-seq (single-nucleus assay for transposase-accessible chromatin using sequencing) data to track chromatin accessibility changes over time (fig. S3A). We focused on pairwise differentially accessible regions (DARs) across consecutive developmental stages, meaning that for each age, we compared its accessible regions with those of the subsequent age (Wilcoxon test, P_adj < 0.01 and LFC > 1). By using a simple approach and assigning individual peaks to the closest genes, we observed that DARs are associated with genes with distinct roles at different stages (fig. S3B). Between P5 and P10, DARs associated with genes related to cell fate commitment decrease in accessibility, while regions assigned with genes linked to axonogenesis or synapse organization increase in accessibility. From P15 to P28, the accessibility of regions near axonogenesis-related genes decreases, while accessibility for those near genes related to vesicle transport and signal transduction increases. These findings align with the timing of GC maturation and synapse development (Fig. 1A) (25–30).

We next used SCENIC+ (41) to infer GRNs. We applied SCENIC+ to the entire GC lineage (see Materials and Methods and fig. S3, C to E) and focused on eRegulons active in mGCs across postnatal development (Fig. 2A). PCA on SCENIC+ eRegulon enrichment scores in mGCs [AUC (area under the recovery curve) scores, calculated by AUCell (42)] reveals a clear temporal separation from P5 to P28 (Fig. 2C). eRegulon AUC scores represent the enrichment of their target genes within each cell’s expressed genes. A total of 36 eRegulons with distinct activity patterns was identified in mGCs (Fig. 2D and table S1), with leading TF expression following a similar pattern to eRegulon activity over time for activators and the opposite pattern for repressors (Fig. 2D and fig. S3F). eRegulons are defined by the name of their leading TF and a sign corresponding to their activator or repressor role [i.e., TF(+) or TF(−)], followed by the number of target genes they contain.

eRegulons grouped by activity patterns fell into three main clusters (hierarchical clustering, 0.85 cutoff value; fig. S3G). The first cluster (cluster 1, early; fig. S3G and Fig. 2D) was highly active at P5 and included Zbtb20(+) and Klf7(+). The second cluster (cluster 2, middle; fig. S3G and Fig. 2D) peaked at P10 to P15 and decreased thereafter, including Bcl11b(+), Foxo1(+), Bcl6(+), and Atf6(+). The third cluster (cluster 3, late; fig. S3G and Fig. 2D) peaked at P15 to P28 and included Smad3(+) and the repressors Bach2(−) and Nfib(−), among others.

We next explored whether eRegulon activity patterns during the synaptic development of mGCs correspond to target gene function using GO analysis (BP). Klf7(+) targets are associated with axonogenesis (Fig. 2E), consistent with Klf7’s role in promoting axon growth (table S1) (43). Bcl11b(+) regulates target genes involved in membrane potential and synapse organization (Fig. 2E), in agreement with the role of Bcl11b in the DG (table S1) (19). Smad3(+) target genes are important for synaptic plasticity and ion transport, aligning with Smad3’s role in long-term potentiation (LTP) in GCs (Fig. 2E and table S1) (44). Target genes of the NFIB repressor are involved in synaptic transmission and membrane potential (Fig. 2E), consistent with the fact that their expression is repressed at P5. Similarly, Bach2(−) target genes are related to plasticity and synapse organization. Of the 18 eRegulons without a known role in the DG (table S1), Bcl6(+) targets are related to protein localization to the cell periphery, and Atf6(+) controls genes involved in vesicle transport at the synapse (Fig. 2E). Together, these findings reveal a dynamic gene regulatory code, consisting of known and unknown GRNs, orchestrating the postnatal development of mGCs.

Predicted TFs and their targets are translated in developing GCs

We next determined whether the identified TFs and target genes are translated in GCs, given the complex relationship between the cellular transcriptome and the proteome (45). To this end, we used ribosome tagging to analyze transcripts of TFs and target genes recruited for translation and correlated these with inferences from our snMO analysis. We crossed the RiboTag mouse, which expresses the hemagglutinin (HA)–tagged ribosomal subunit Rpl22 in the presence of Cre recombinase (46), with the Rbp4-Cre line to capture ribosome-bound mRNA in GCs across postnatal stages P5 to P28 (Fig. 3, A and B). Following hippocampal dissociation, ribosome-bound mRNA was extracted and subjected to bulk RNA-seq after quality control (Fig. 3C and fig. S4, A to C). PCA showed a clear temporal separation between the samples from P5 to P28, with the P21 and P28 stages nearly indistinguishable (Fig. 3D). Differential gene expression analysis (pairwise analysis, adjusted P value <0.05, |LFC| > 0.5) again revealed prominent shifts at P7 and P15 (fig. S4D; compare to fig. S1H), with five clusters containing SynGO-annotated genes (fig. S4E) (38).

Fig. 3. Predicted TFs and their target genes (TGs) are translated in developing GCs.

(A) Hippocampi of Rbp4-Cre:RiboTag mice at P5, P7, P15, and P28. The HA tag (orange) labels the ribosomal subunit RPL22 expressed in GCs in the outer part of the GC layer at all time points. Scale bars, 500 μm. (B) Timepoints used for RiboTag experiments; n = 3 or 4 mice for P5 and P7 and n = 2 mice for P10 to P28. (C) Methodology for RiboTag experiments. Hippocampi from Rbp4-Cre:RiboTag mice were homogenized, and the homogenate was incubated with anti-HA magnetic beads for RPL22-HA pull-down. Ribosome-bound mRNA was separated from ribosomes, RNA was isolated, and libraries were prepared for bulk RNA-seq. (D) PCA shows the separation of biological replicates by time point. (E) Heatmap of SCENIC+ eRegulons leading TF (compare with Fig. 2D) expression in RiboTag samples across time points. The label on the left side of the heatmap indicates hierarchical clustering of the eRegulons (see fig. S3G). (F) Comparative plots of standardized expression across time for Klf7, Bcl11b, Bcl6, and Smad3 show TF expression in snMO (gene expression; light blue line), TF expression in RiboTag (light orange line), mean TG expression in snMO (gene expression; turquoise line), mean TG expression in RiboTag (orange line), and mean TG expression in RiboTag ±1 SD (cream shade around mean TG expression).

Using this dataset, we analyzed the expression patterns of ribosome-bound transcripts of TFs leading the SCENIC+-predicted eRegulons in mGCs (Fig. 2D). Ribosome-bound mRNA levels for these TFs largely correlated with their expression patterns in the snMO dataset (Fig. 3E and fig. S4F). For example, TFs such as Nfib or Klf7 peaked at P5 to P7 and decreased thereafter, while others, including Bcl6, Atf6, and Smad3, exhibited sharp increases after P15 (Fig. 3E and fig. S4F). Similarly, TFs with previously characterized roles in the development of GC connectivity, Bcl11b and Klf9, showed matching translatome profiles and expression patterns in the snMO dataset (Fig. 3E and fig. S4F).

We next assessed eRegulon target genes in the RiboTag dataset. By analyzing the mean expression of target genes over time, we observed a similar expression profile between the ribosome-bound transcript level and the expression pattern in the snMO dataset (fig. S4G). For example, ribosome-bound mRNA levels for Klf7(+), Bcl11b(+), Smad3(+), and Bcl6(+) target genes (Fig. 3F, orange lines) correlated with their expression pattern in the snMO dataset (Fig. 3F, turquoise lines, and fig. S4G). Thus, the ribosome-bound TF and target gene transcript dynamics show a strong correlation with the predicted eRegulon activity patterns, increasing confidence in the identified GRNs that regulate the postnatal development of GCs.

Loss of Bcl6 in GCs alters dendritic spine density and MFB morphology

To test the functional relevance of identified GRNs in the synaptic development of GCs, we focused on eRegulons from middle and late clusters (clusters 2 and 3; see Fig. 2D and fig. S3G), as their activity coincides with the timing of formation, refinement, and maturation of GC synapses. We selected two GRNs, Bcl6(+) and Smad3(+), on the basis of the roles of their predicted target genes and the unknown role of their leading TFs in GC synapse development. Focusing on these, we performed a loss-of-function analysis of the corresponding TFs in combination with sparse, bright labeling of GCs for morphological analysis.

We first investigated Bcl6(+), an eRegulon from cluster 2 (middle) whose activity peaks at P15 (Fig. 4A). Bcl6 itself has not been previously linked to synapse development and does not have a known role in the DG. In the cortex, BCL6 promotes neurogenesis via repression of Notch and Wnt pathways (47, 48). In the DG, Bcl6 is not detected in progenitors and iGCs but is expressed in mGCs (37). In Rbp4-Cre:Ai9 mice, we detected Bcl6 mRNA in GCs at P15, but not at P5, using single-molecule fluorescence in situ hybridization (fig. S5, A and B), consistent with Bcl6(+) eRegulon activity (Fig. 4A). Bcl6(+) includes 23 target genes encoding proteins with predominantly synaptic localization (Fig. 4B) that have been linked to hippocampal synapse development (49, 50). One of these predicted BCL6 target genes is Nptx1, encoding a secreted synaptic organizer (51) that is highly enriched at MF-CA3 synapses (6) and regulates the clustering of AMPA receptors (52). Nptx1 is controlled by a chromatin region that peaks in accessibility at P28 (Fig. 4, C and D, and fig. S5C).

Fig. 4. Loss of Bcl6 in GCs alters dendritic spine density and MFB morphology, affecting synaptic transmission.

(A) PCA based on eRegulon target gene cell enrichment scores (AUC), colored by Bcl6(+) target gene AUC. (B) GO cellular component (CC) terms enriched among Bcl6(+) targets. (C) Heatmap of mean normalized accessibility for the 24 predicted Bcl6(+) target regions across time points. (D) Coverage plot of the Bcl6-linked region near Nptx1 [chromosome 11 (chr11): 119550565 and 119551065], with arcs showing SCENIC+-predicted region-gene links. (E) AAV vectors used for sparse GC labeling: pAAV-U6-TRE-fDIO-mGFP-IRES-tTA (with scrambled or Bcl6 shRNA) and pAAV-TRE-FLP. IRES, internal ribosomal entry site. FRT (FLP recognition target) sites are indicated in yellow. (F) Experimental timeline: P0 DG injections of AAVs in C57Bl6J pups; brains collected at P28 for dendritic and MFB morphological analysis and at P27 to P35 for ephys recordings. (G) Representative images of dendritic segments from the medial molecular layer of the DG for control (gray) and Bcl6 KD (green) animals. (H) Bcl6 KD reduces spine density and mushroom spine density (n = 8 mice per group, 20-μm segments per animal; nested t tests, P = 0.0135 and P = 0.0220). (I) Representative images of MFBs in control and Bcl6 KD animals. (J) Bcl6 KD increases MFB filopodial number (n = 6 or 7 mice per group, six to eight MFBs per animal; P = 0.0209), with unchanged filopodial length. Box plots show the median, interquartile range, and range (each dot, one animal). (K) Representative mEPSC traces from GCs expressing control or Bcl6 shRNAs. (L) Bcl6 KD results in reduced mEPSC frequency (n = 43 or 44 neurons, six mice per group; P = 0.0189, Mann-Whitney test), with unchanged amplitudes. (M) Representative mIPSC traces from GCs expressing control or Bcl6 shRNAs. (N) Bcl6 KD results in reduced mIPSC frequency (n = 43 or 44 neurons, six mice per group; P = 0.0101, Mann-Whitney test), with unchanged amplitudes. Individual data points are shown, and bars represent the means ± SEM.

To explore Bcl6’s role in the synaptic development of GCs, we used short hairpin RNA (shRNA)–mediated knockdown (KD) of Bcl6, along with a flippase (FLPo) for Supernova-mediated sparse labeling (Fig. 4E) (53). This system relies on the leakage of the tetracycline-responsive element (TRE) promoter that induces tetracycline transactivator (tTA) expression, enhancing the signal in a sparse population of cells. Attempts to validate Bcl6 guide RNAs (gRNAs) for CRISPR-Cas9–mediated KO were unsuccessful because of a lack of specific BCL6 antibodies. We therefore resorted to previously validated Bcl6 shRNAs (47, 48), whose efficiency we confirmed in primary neurons (fig. S5, D and E). We injected adeno-associated viruses (AAVs) expressing either scrambled or Bcl6 shRNAs combined with FLPo into the DG of C57Bl/6J mice at P0 (Fig. 4F and fig. S5F). P0 AAV9 (AAV serotype 9) injections preferentially target postmitotic GCs (54), which will be the most mature GCs at P28 (fig. S5G). At P28, GC dendritic and axonal segments were imaged and reconstructed in 3D for morphological analysis (Fig. 4F). Bcl6 KD resulted in a 20% decrease in dendritic spine density compared to GCs expressing the scrambled shRNA (Fig. 4, G and H). This decrease was caused by a reduction in mature, mushroom spine density (Fig. 4H).

We next analyzed the MFB area, volume, and filopodial number and length. MFBs contain one or multiple filopodia that mediate feedforward inhibition to CA3 neurons via synapses onto inhibitory neurons (55). Bcl6 KD resulted in a 64% increase in the number of filopodia per MFB without affecting the filopodial length (Fig. 4, I and J). The MFB area and volume remained unchanged in Bcl6 KD animals (fig. S5H).

We then recorded spontaneous synaptic transmission in GCs using whole-cell voltage-clamp recordings in acute hippocampal slices from C57Bl6J mice (P27 to P35) injected with AAVs expressing Bcl6 or control shRNAs (Fig. 4F). The frequency of miniature excitatory postsynaptic currents (mEPSCs) in Bcl6 KD cells was reduced by 40% compared to controls (Mann-Whitney test, P value = 0.0189), while the mEPSC amplitude remained unchanged (Fig. 4, K and L). Likewise, the frequency of miniature inhibitory postsynaptic currents (mIPSCs) was reduced by 45% in Bcl6 KD cells compared to controls (Mann-Whitney test, P value = 0.0101), while the mIPSC amplitude remained unchanged (Fig. 4, M and N). Excitation/inhibition ratios in Bcl6 KD cells were similar compared to controls (Mann-Whitney tests) (fig. S5I). Together, these results show that the loss of Bcl6 in GCs alters dendritic spine density and MFB morphology and decreases the frequency of spontaneous excitatory and inhibitory synaptic transmission.

Loss of Smad3 in GCs alters inhibitory synaptic transmission

We next focused on the Smad3(+) eRegulon, a cluster 3 (late) eRegulon whose activity peaks at P28 (Fig. 5A). Consistent with this, we detected Smad3 mRNA using single-molecule fluorescence in situ hybridization in GCs of Rbp4-Cre:Ai9 mice at P15 and P28 but not at P5 (fig. S6, A and B). SMAD3 is predicted to regulate 224 genes linked to synaptic development (Fig. 2E) (56, 57) through 649 chromatin-accessible regions, which gradually increase in accessibility from P5 to P28 (Fig. 5B). One of these predicted SMAD3 target genes is Tanc1, a scaffold protein, highly expressed in the hippocampus, that interacts with the excitatory postsynaptic scaffold PSD-95 (postsynaptic density protein 95). Its deficiency causes reduced spine density in the CA3 region (56). Tanc1 is controlled by a chromatin region that peaks in accessibility at P28 (Fig. 5, B and C, and fig. S6C).

Fig. 5. Loss of Smad3 in GCs alters inhibitory synaptic transmission.

(A) PCA based on eRegulon target gene cell enrichment scores (AUC) colored by Smad3(+) AUC scores of target genes. (B) Heatmap showing mean normalized accessibility per time point of 649 predicted Smad3(+) target regions and labeling region-linked genes within the synaptic organization GO term. (C) Coverage plot for the Tanc1-associated region across time points in mGCs. The target region is located on chr2 between base pairs 59,537,947 and 59,538,447 (gray shade). Arcs represent SCENIC+-predicted region-gene links. (D) Plasmids pAAV-U6-TRE-fDIO-mScarlet-IRES-tTA containing control gRNA (lacZ) or Smad3 gRNAs and pAAV-TRE-DIO-FLP. FRT sites are indicated with yellow arrowheads; loxP sites (Cre recognition target) are indicated with green arrowheads. (E) Experimental design for Smad3 in vivo experiments. Rbp4-Cre:Rosa26-LSL-Cas9 P7 pups were injected in the DG with control or Smad3 gRNA AAVs and AAV-DIO-TRE-FLP for sparse labeling. Morphological analysis of GC dendrites (spine density and morphology) and MFB morphology (area, volume, filopodial number, and length) was performed at P28, and ephys was performed at P27 to P31. (F) Representative mEPSC traces from GCs expressing control or Smad3 gRNAs. (G) Analyzed mEPSC frequency and amplitude (number of neurons: n = 45 from n = 4 mice for control and n = 42 from n = 4 mice for Smad3 KO conditions; Mann-Whitney test). (H) Representative mIPSC traces from GCs expressing control or Smad3 gRNAs. (I) Analyzed mIPSC frequency and amplitude (number of neurons: n = 45 from n = 4 mice for control and n = 42 from n = 4 mice for Smad3 KO conditions; Mann-Whitney test, P value = 0.01). Individual data points are shown in (G) and (I), and bars represent the means ± SEM.

To assess whether the loss of Smad3 affects GC dendritic spine density or MFB morphology, we used CRISPR-Cas9–mediated KO. In vitro validation confirmed that Smad3 gRNAs effectively reduced SMAD3 protein levels in primary cortical cultures (fig. S6, D and E). We crossed Rbp4-Cre mice with Cre-dependent Cas9 mice [Rosa26-LSL-Cas9 (58)] for GC-specific Smad3 KO, combined with an AAV-DIO-FLPo to achieve Supernova-mediated sparse labeling (Fig. 5, D and E) (53). Here, we adapted the strategy used for Bcl6 KD experiments such that the random activation of the FLPo expression only occurs in Rbp4-Cre–positive cells. We injected AAVs expressing Smad3-targeting gRNAs or a nontargeting (lacZ) gRNA as a control (Fig. 5, D and E, and fig. S6F) into the DG of Rbp4-Cre:Rosa26-LSL-Cas9 mice at P7 and analyzed GC morphology at P28. The density and morphology of dendritic spines in Smad3 KO GCs did not differ from those of control GCs (fig. S6, G to I). In addition, the analysis of area, volume, and filopodial number and length of MFBs did not show significant changes following Smad3 loss (fig. S6, H and J), indicating that Smad3 KO in GCs does not affect dendritic spine density or MFB morphology.

The loss of Smad3 has previously been shown to abolish LTP in GCs via enhanced GABA_A (γ-aminobutyric acid type A) receptor neurotransmission (44). We recorded spontaneous synaptic transmission in GCs using whole-cell voltage-clamp recordings in acute hippocampal slices from Rbp4-Cre:Rosa26-LSL-Cas9 mice (P27 to P31) injected with AAVs expressing Smad3 or lacZ gRNAs (Fig. 5, D and E). The frequency and amplitude of mEPSCs in Smad3 KO cells did not differ from controls (Fig. 5, F and G). In contrast, the frequency of mIPSCs was increased in Smad3 KO cells compared to controls (Mann-Whitney test, P value = 0.01), while the mIPSC amplitude remained unchanged (Fig. 5, H and I). This led to an excitation/inhibition imbalance in Smad3 KO cells compared to controls (Mann-Whitney tests) (fig. S6K). Together, these results show that the loss of Smad3 in GCs increases spontaneous inhibitory synaptic transmission without affecting spontaneous excitatory synaptic transmission, dendritic spine density, or MFB morphology.

Early-active GRNs repress late TFs and predicted targets involved in synaptic organization

We next explored how the identified GRNs interact with each other. Most of the leading TFs target other TFs (fig. S7) belonging to the same or different clusters (see clustering analysis, fig. S3G). For example, cluster 1 TFs (Fig. 2D), such as the Nfi TFs (NFIB, NFIA, and NFIX), are highly interconnected but are also predicted to target later TFs, as seen for Rfx3(+), targeting Glis3 or Nr3c1, among others (fig. S7).

Focusing on predicted Bcl6(+) and Smad3(+) interactions (Fig. 6A) revealed an interesting link between early and late GRNs. Early-expressed repressor TFs like NFIB and BACH2 (P5; see Fig. 2D) target later TFs (P15 to P28), such as Bcl6 and Smad3 (Fig. 6A). This suggests that early GRNs could repress later ones, potentially delaying the activation of their downstream targets. Once active, late GRNs are also predicted to target each other, as seen with BCL6, KLF9, and FOXO1 targeting Bcl11b or KLF9 regulating Bcl11b, Nr3c1, and Smad3 (Fig. 6A and fig. S7). The binding of NFIB to Smad3 regulatory regions is confirmed by previous chromatin immunoprecipitation sequencing (ChIP-seq) studies in neuronal progenitors [18 of 33 studies in the ChIP-seq atlas (59), e.g., SRX7895561], and the binding of BACH2 to both Smad3 and Bcl6 regulatory regions was also shown in CD4+ T cells [SRX761324 and SRX761325 (59)]. In addition, early-expressed activator TFs like RFX3 are predicted to target later TFs, such as Bcl6, and in some cases, these predictions are bidirectional, suggesting feedback between early and late GRNs that would mutually regulate each other, e.g., Rfx3(+)-Klf9(+)-Bcl11b(+) (Fig. 6A). To link the GRNs to their putative roles in synapse development, we used SynGO to determine the size of each node on the basis of the percentage of eRegulon targets present in the database. Late-active eRegulons (clusters 2 and 3) show a higher proportion of SynGO-annotated synaptic genes than early-active eRegulons (cluster 1) (Fig. 6A and fig. S7), with Atf6(+) having the highest (40%) and Rfx3(−) the lowest (4.7%) proportion.

Fig. 6. Schematic representation of the GRN model centered on Bcl6(+) and Smad3(+) eRegulons.

(A) Simplified GRN model (see also fig. S7) centered on Bcl6(+) and Smad3(+) eRegulons and predicted interactions of leading TFs [incoming and outgoing; activating (black lines) and repressing (red lines)]. eRegulons are represented by circles, colored depending on the cluster the eRegulon belongs to (cluster 1-early, pink; cluster 2-middle, green; cluster 3-late, blue; see also Fig. 2D and fig. S3G). Early-expressed repressor TFs NFIB and BACH2 target later TFs Bcl6 and Smad3. (B) Simplified representation of BACH2, NFIB, BCL6, and SMAD3 interactions with their target genes Nptx1 and Tanc1. In both (A) and (B), the size of the circles for each eRegulon is determined on the basis of the percentage of SynGO-annotated synaptic TGs, normalized to the eRegulon with the highest proportion of SynGO-annotated synaptic genes, Atf6(+). eRegulons with leading TFs that act as both activators and repressors are represented by two circles, with the left circle representing the repressor and the right circle representing the activator. TFs labeled in gray also target themselves (e.g., Nfib), while those labeled in white do not (e.g., Bach2).

Early repression of late-expressed genes was also observed beyond TFs. For instance, the synaptic organizer–encoding gene Nptx1, a shared predicted target in Bcl6(+) and Smad3(+) eRegulons (see Figs. 4, C and D, and 5B), is also a predicted target of the Bach2(−) repressor (Fig. 6B). Similarly, Tanc1, a Smad3(+) target involved in synaptic organization (56) (Fig. 5B), is predicted to be repressed by Nfib(−). These predictions were also confirmed in previous ChIP-seq studies that linked BACH2-Nptx1 [SRX761325 (59)] and BACH2-Tanc1 [SRX761325 (59)] in CD4+ T cells, SMAD3-Nptx1 [SRX5251394 (59)] in embryoid bodies, and SMAD3-Tanc1 [7 of 33 studies in the ChIP-seq atlas (59), e.g., SRX130598 in neural progenitors].

These observations suggest a key role for early TFs in repressing late TFs and their target genes controlling synaptic organization and maturation. Cross-referencing the identified TFs with SFARI (60), a curated database of autism spectrum disorder (ASD) genes, showed that ~80% of the ASD-associated TFs in our dataset belong to the early cluster of GRNs [cluster 1; including Mef2c, Nfi family members, Zbtb20, and Rfx3 (fig. S8 and table S1)]. In contrast, no ASD-associated TFs were identified in the late cluster (fig. S8 and table S1). This suggests that the early TFs regulating GC synapse development are susceptible in ASD. Together, these findings highlight a sequential regulatory program in which early-active TFs act as gatekeepers, delaying the activation of later GRNs and their putative synaptic targets.

DISCUSSION

Here, we combine transcriptomics, epigenomics, and translatomics to reconstruct the GRNs regulating the synaptic development of GCs. Our multiomic analysis reveals distinct GRNs active at different stages of GC development and identifies the TFs Bcl6 and Smad3 as regulators of GC synapse development. In addition, our results suggest that early-active TFs repress GRNs that are active at later stages of GC development.

We find a dynamic gene regulatory code orchestrating the postnatal development of GCs as they integrate into the hippocampal circuit. GCs undergo prominent shifts in gene expression between early (P5 to P7) and late (P15 to P28) time points at both the transcriptomic and translatomic levels, with most genes up-regulated in the early phases and down-regulated in the late phases. Genes up-regulated in early stages are key for neurite development, while genes active at later time points control synaptic processes. GRN inference in mGCs identifies 36 eRegulons, which can be clustered into three main temporal groups [early (P5), middle (P10 to P15), and late (P15 to P28)]. Among the TFs with established roles in the DG, we find Bcl11b, Mef2c, and Klf9, all three of which are implicated in the development of GC connectivity [see refs. (19, 22, 23)]. Half of the TFs leading the eRegulons we identify have no known role in the DG (table S1).

We observe interesting parallels in the GRN structure of postnatally developing GCs (this study) and adult-born GCs (aGCs) (37). In the adult DG, radial glia–like cells continuously generate new neurons, which functionally integrate into the hippocampal circuit (61). These aGCs are more excitable but less precisely tuned than their embryonically born counterparts (62). SCENIC (42) analysis of scRNA-seq data from immature and mature aGCs identified nine regulons (37) that overlap with our findings, including Bcl6(+) and Foxo1(+) (table S1). Conversely, some eRegulons identified in our study are led by TFs previously reported to play a role in adult neurogenesis, such as Mef2a (63), Neurod2 (64), and Foxo1 (65). Together, these results suggest shared gene regulatory mechanisms between perinatal GC and aGC development.

In our functional analysis of the role of identified GRNs in the synaptic development of GCs, we characterize the BCL6 TF as a regulator of structural maturation of GC synapses. Previous work in the mouse cortex showed that Bcl6 promotes neurogenesis via repression of Notch targets (47), but a role in postnatal synapse development has not been described. Here, we identify BCL6 as a transcriptional activator targeting 23 genes that show similar expression patterns to Bcl6, further supporting an activator role for BCL6 in postnatal GCs. TFs can have dual repressor/activator roles, such as Fezf2 in corticospinal neurons (14), including TFs exerting their transcriptional regulation via ZiN/POZ (poxvirus zinc finger) domains like Bcl6 (66). We found that Bcl6 KD reduces the density of mature dendritic spines in GCs. Several of the predicted Blc6(+) target genes regulate dendritic spines. The KD of Pacsin1 in CA1 neurons reduces spine density (49), while overexpression of Arhgef25 increases spine size and the proportion of mature spines (50). Our observations suggest that Bcl6 is necessary for the maturation of dendritic spines in GCs, as mushroom spines are considered the mature type associated with stronger synapses critical for memory storage (67). In addition to the effects on dendritic spine density, Bcl6 KD increases filopodial density in MFBs, while other MFB features such as area and volume remain unaffected. MFB filopodial synapses target GABAergic cells in the stratum lucidum of CA3 (55), mediating feedforward inhibition essential for CA3 microcircuit function under normal and pathological conditions (5, 68). The MFB filopodial density peaks during development (P14) and decreases in adult mice (29), suggesting that the loss of Bcl6 impairs maturation or the refinement of filopodial density. Several synaptic genes affect the MFB filopodial number (5, 6, 68), and some of the predicted Bcl6(+) target genes could influence filopodial growth, such as the actin cytoskeleton organizer Pkp4 (69). Together, these findings suggest that BCL6 promotes the structural maturation of GC synapses through its interactions with specific synaptic target genes.

In contrast, the loss of Smad3 in GCs does not affect dendritic spine density or MFB morphology. These results were unexpected, considering that the Smad3(+) eRegulon contains 224 predicted target genes with roles in synapse organization and plasticity (51, 56). Further analysis revealed a selective increase in mIPSC frequency in Smad3 KO GCs. This increase could reflect an increase in the density of inhibitory input onto GCs or a change in inhibitory presynaptic properties. GCs in adult Smad3 KO mice show increased inhibitory, but not excitatory, neurotransmission (44), consistent with our analysis of the cell-autonomous loss of Smad3 in developing GCs. Together, our results suggest that the Bcl6(+) and Smad3(+) eRegulons control distinct aspects of GC synapse development.

Our analysis of interactions between GRNs predicts sequential regulations, with early-active TFs delaying the activation of later GRNs and their putative synaptic targets. This sequential regulation is exemplified by the interactions of the Bcl6(+) and Smad3(+) eRegulons. Both eRegulons are targeted by early-active repressor GRNs, such as Nfib(−) and Bach2(−). Nfib loss in the cortex induces severe loss of axonal projections and up-regulation of Notch signaling genes (70), suggesting a repressor role for this TF. Our GRN model (Fig. 6 and fig. S7) suggests that early-active Nfib(−) can repress Smad3(+), through shared target genes such as Tanc1. The Bcl6-Bach2 link has been previously studied in immune cells. Bcl6 is repressed by BACH2 in T cells (71), while BCL6 and BACH2 cooperate to repress other genes in B cells (72). BACH2 controls the selection of viable cells via the activation of the tumor suppressor protein p53 in the absence of BCL6, and their dysregulation is associated with high leukemic cell proliferation (73, 74). A similar antagonistic relationship between BACH2 and BCL6 may regulate GC synapse development. Our GRN model (Fig. 6 and fig. S7) suggests that early-active Bach2(−) can repress the Bcl6(+) and Smad3(+) eRegulons through shared target genes such as Nptx1 and Tanc1 (Fig. 6B). Further ChIP-seq studies in mGCs would be essential to corroborate the predicted regulatory interactions among TFs or between them and their targets. A recent study on the early postnatal development of cortical neurons identified a sequential regulation mechanism for Krüppel-like factor (KLF) TFs (75). A transcriptional shift from early KLF activators to late KLF repressors activates and then represses a set of shared target genes with cytoskeletal functions, including Dpysl3 that was also identified in the early clusters in our scRNA-seq dataset (75). Together with our findings in GCs, these observations suggest that the shift from axonogenesis to synapse formation involves coordinated repression of an active axonogenic transcriptional program and concurrent derepression of a synaptogenic program.

The terminal selector hypothesis proposed in Drosophila and C. elegans studies suggests that combinations of constantly expressed TFs define key cellular properties such as neurotransmitter identity and connectivity (10, 76, 77). Some of these selectors are required for synaptic differentiation via the regulation of key effectors such as synaptic organizers and ion channels, as well as other TFs (78, 79). However, the GRNs belonging to the middle and late clusters we identify here differ from terminal selectors in their postnatal onset of activity after the wiring pattern of GCs is established. Cell type–specific combinations of such GRNs could enable neurons to develop a large synaptic diversity without altering fundamental properties such as neurotransmitter profiles or wiring patterns. Notably, we find that ±80% of the early-active TFs we identify are associated with ASD, suggesting that the early GRNs regulating GC development are particularly susceptible in ASD. Whether this observation extends to other hippocampal neuron types remains to be determined.

In conclusion, our results uncover a dynamic gene regulatory code governing the synaptic development of GCs, provide insight into key TFs and target genes, and identify BCL6 and SMAD3 as regulators of specific aspects of GC synapse development. Our single-cell transcriptome, snMO, and GC translatome datasets provide an invaluable resource for the poorly mapped postnatal developmental period of hippocampal GCs. A systematic extension of our multiomic approach to other cell types during postnatal development will be critical to uncover the gene regulatory codes driving synapse formation, maturation, and plasticity in brain regions such as the hippocampus and cortex.

MATERIALS AND METHODS

Animals

All animal experiments were conducted according to the KU Leuven ethical guidelines and approved by the KU Leuven Ethical Committee for Animal Experimentation (ECD P164/2021). Mice were maintained in a specific pathogen–free facility under standard housing conditions with continuous access to food and water. WT C57BL/6J mice originated from Charles River. Rpb4-Cre (MGI:4367067) and Ai9 (JAX:007909) mouse strains were obtained from A. Takeoka (Neuro-Electronics Research Flanders, Leuven). RiboTag (JAX:029977) and H11-Cas9 line (JAX: 028239) were obtained from JAX. The Rosa26-LSL-Cas9 (JAX:024857) mouse line was obtained from J. Cools (VIB Center for Cancer Biology, Leuven). Genotypes were regularly checked by polymerase chain reaction (PCR) analysis. Rbp4-Cre heterozygous Cre-positive males were bred to WT, Ai9, RiboTag, or Rosa26-LSL-Cas9 females. For euthanasia, animals were either anesthetized with isoflurane and decapitated or injected with an irreversible dose of ketamine-xylazine.

scRNA-seq

Rbp4-Cre:Ai9 mice were euthanized at different developmental stages (P5, P7, P10, P15, and P28) after being anesthetized with isoflurane. The neuronal extraction protocol was adapted from Allen Brain Neuronal Isolation Protocol (80) and Tasic et al. (81). Briefly, immediately after extraction, the brain was immersed in ice-cold carbogenated artificial cerebrospinal fluid (table S2). The brain was sliced in a precooled brain matrix (AgnThos, 69-2165-1) to 500-μm sections, and the DG was dissected in a petri dish containing ice-cold artificial cerebrospinal fluid while carbogenating. Immediately after dissection, the DG pieces were transferred into an enzymatic solution containing Pronase Protease (VWR 537088, 2 mg/ml), diluted in digestion medium (table S2), and supplemented with 50 mM trehalose (Sigma-Aldrich, T0167). The enzymatic solution was placed in a 37°C water bath while carbogenating and incubated for 10 min (for P5, P7, and P10), 15 min (for P15), and 28 min (for P28), followed by the addition of trypsin (0.125%; Sigma-Aldrich, 59417C) and deoxyribonuclease (DNase I; 0.2 mg/ml; Thermo Fisher Scientific, 10649890) for 2 min (except for the P28 sample, which received no trypsin or DNase I). After digestion, the enzymatic solution was removed, and tissue pieces were washed with 10 ml of digestion medium (ice-cold and carbogenated). The tissue was triturated using Pasteur pipettes in 2 ml of Opti-MEM/trehalose solution (table S2) with polished tip openings of 500- and 250-μm diameters, layered on top of a 20% Percoll solution (table S2), and spun at 700g for 10 min at 4°C. When all solution and debris were removed, the cell pellet was resuspended in 200 to 500 μl of Opti-MEM/trehalose solution (table S2).

Single cells were isolated by FACS (fluorescence-activated cell sorting) into individual wells of 96-well plates (4titude, cat. no. 4TI-0960/C) containing 4 μl of cell lysis buffer [0.1% Triton X-100, Sigma-Aldrich; ribonuclease (RNase) inhibitor (1 U/μl), RNase OUT, Thermo Fisher Scientific; 2.5 μM oligo dT (25), IDT; 2.5 mM dNTP (deoxynucleotide triphosphate), Promega). Sorting was performed on a BD FACSAria Fusion using a 100-μm nozzle, at a sheath pressure of 20 psi (137.895 kPa), and in the single-cell sorting mode. To exclude dead cells, 4′,6-diamidino-2-phenylindole (DAPI; Invitrogen, 10184322) was added to the single-cell suspension to a final concentration of 5 ng/ml. FACS populations were chosen to select cells with low DAPI and high tdTomato fluorescence. The cells were sorted into wells with lysis buffer, and the plate was spun down at 2000g for 1 min before storage at −80°C.

Library preparations and sequencing for scRNA-seq

Smart-seq2 was performed using a previously published protocol (35). For the first-strand synthesis, the Smart-seq2 plates were thawed to room temperature (RT) for 1 min and subsequently spun down at 2000g for 1 min. RNA denaturation was carried out at 72°C for 10 min, and flash cooling on ice was performed for 5 min. The first-strand synthesis mix [1× SuperScript II first-strand synthesis buffer, Thermo Fisher Scientific; 5 mM dithiothreitol (DTT), Thermo Fisher Scientific; 100 U Oligo SuperScript II reverse transcriptase enzyme, Thermo Fisher Scientific; 10 U RNase OUT, Thermo Fisher Scientific; 1 M betaine, Sigma-Aldrich; 6 mM MgCl₂, Thermo Fisher Scientific; 1 μM template switch oligo (TSO), IDT Technologies] was added to the cell lysis mix to a total volume of 10 μM. The first-strand synthesis reaction was carried out with the following program: 42°C for 90 min; 10 cycles of 50°C for 2 min and 42°C for 2 min; 72°C for 15 min; 4°C indefinite hold. PCR amplification of the first-strand product was performed by adding 15 μl of the PCR mix to the first-strand product (1× KAPA HiFi HotStart Ready Mix, Roche; 0.1 μM IS PCR primer, IDT Technologies). The following PCR program was used to amplify the first-strand product: 98°C for 3 min; 22 cycles of 98°C for 20 s, 67°C for 15 s, and 72°C for 6 min; 72°C for 5 min; 4°C indefinite hold. Using the Hamilton STAR liquid handler (Hamilton), 20 μl of Ampure XP was added to each well and mixed. Standard Ampure XP purification was carried out as per the manufacturer’s recommendation, and the cDNA library was eluted in 22.5 μl of elution buffer. Twenty-two microliters of the eluted library was transferred to a fresh Echo Mass Spectrometry Qualified 384-well Polypropylene Microplate (Beckman Coulter). A Labcyte Echo 550 acoustic dispenser (Beckman Coulter) was used for the next steps. One microliter of Tn5 tagmentation mix (1× Tagment DNA buffer, Illumina; ATM mix, Illumina; cDNA library, 1 ng) for cDNA fragmentation. Tn5 tagmentation was performed with the following program: 55°C for 10 min; 4°C indefinite hold. The tagmentation reaction was stopped by quenching the reaction with 500 nl of neutralize tagment (NT) buffer for 5 min (Illumina). The i5 and i7 Illumina indexes (750 nl) were added to the plates. Last, 2 μl of NPM master mix was added to the plate, mixed well, and placed for index PCR amplification: 72°C for 3 min; 95°C for 30 s; 12 cycles of 95°C for 10 s, 55°C for 30 s, and 72°C for 30 s; 72°C for 5 min; 4°C indefinite hold.

The indexed PCR products were pooled in a single tube, 0.8× Ampure XP purification (Beckman Coulter) was carried out as per the manufacturer’s recommendation, and lastly, the sequencing library was eluted in 35.5 μl of elution buffer (Qiagen). Smart-seq2 libraries were sequenced on NextSeq 500 or NextSeq2000 (Illumina) sequencing platforms. Sequencing was done as per the protocol recommendations: paired-end reads of 61 base pairs (bp; read 1), 61 bp (read 2), 8 bp (index 1), and 8 bp (index 2). Sequencing was performed at a depth of 500,000 reads per cell.

Nucleus isolation and library preparation for 10x Genomics single-cell multiome

Nucleus preparation from mouse hippocampal tissue was done using the sucrose gradient protocol. Briefly, up to 50 mg of hippocampal tissue was homogenized using 1 ml of homogenization lysis buffer [1× salt-tris buffer: 146 mM NaCl, 10 mM tris (pH 7.5), 1 mM CaCl₂, and 21 mM MgCl₂; 0.03% (v/v) Tween-20, Sigma-Aldrich; 0.01% (v/v) NP-40, Sigma-Aldrich; 1% bovine serum albumin (BSA), Thermo Fisher Scientific; 25 mM KCl, Thermo Fisher Scientific; 250 mM sucrose, Sigma-Aldrich; 1 mM DTT, Thermo Fisher Scientific; 1× protease inhibitor, Sigma-Aldrich; RNase inhibitor, Promega]. The tissue was homogenized with 10× strokes with pestle A and 10× strokes with pestle B. The homogenized mixture was incubated in homogenization lysis buffer for 5 min. During the incubation, the homogenized sample was passed through 70-μm nylon mesh. The nuclei were pelleted by centrifugation at 500g for 5 min. The nucleus pellet was washed and resuspended with 520 μl of Wash Buffer 1 [1× salt-tris buffer; 1% BSA, Thermo Fisher Scientific; 25 mM KCl, Thermo Fisher Scientific; 250 mM sucrose, Sigma-Aldrich; 1 mM DTT, Thermo Fisher Scientific; 1× protease inhibitor, Sigma-Aldrich; RNase inhibitor (1 U/μl), Promega]. The nucleus suspension was then mixed with an equal volume of gradient medium [1 mM CaCl₂, Thermo Fisher Scientific; 50% (v/v) Optiprep, Sigma-Aldrich; 5 mM MgCl₂, Thermo Fisher Scientific; 10 mM tris (pH 7.5); Thermo Fisher Scientific; 75 mM sucrose, Sigma-Aldrich; 1 mM DTT, Thermo Fisher Scientific; 1× protease inhibitor, Sigma-Aldrich; RNase inhibitor (0.5 U/μl), Promega]. This was thoroughly mixed, and the mix was layered upon 948 μl of 29% density cushion [29% (v/v) Optiprep, Sigma-Aldrich; 72.5 mM KCl, Thermo Fisher Scientific; 14.5 mM MgCl₂, Thermo Fisher Scientific; 29 mM tris (pH 8), Thermo Fisher Scientific; 121 mM sucrose, Sigma-Aldrich; RNase inhibitor (0.25 U/μl), Promega]. The layered nucleus suspension was subjected to centrifugation at 9000g for 20 min at 4°C. After the centrifugation, the debris was removed from the top of the suspension, and if excess debris was present, this could be scooped out with a 1000-μl pipette tip. The supernatant was aspirated out gradually using a glass Pasteur pipette until about 75 μl of the supernatant was left. An equal volume of 0.2× lysis buffer [20 mM tris (pH 7.5), Thermo Fisher Scientific; 20 mM NaCl, Thermo Fisher Scientific; 6 mM MgCl₂, Thermo Fisher Scientific; 0.02% Tween 20, Sigma-Aldrich; 0.02% NP-40, Sigma-Aldrich; 0.002% digitonin, Promega; 2% BSA, Thermo Fisher Scientific; 2 mM DTT, Thermo Fisher Scientific; protector RNase inhibitor (2 U/μl), Sigma-Aldrich] was added, and the nuclei were resuspended and incubated for 2 min on wet ice. After the incubation, the lysis was quenched by adding 1 ml of Wash Buffer 2 [10 mM tris (pH 7.5), Thermo Fisher Scientific; 10 mM NaCl, Thermo Fisher Scientific; 3 mM MgCl₂, Thermo Fisher Scientific; 0.01% Tween 20, Sigma-Aldrich; 1% BSA, Thermo Fisher Scientific; 1 mM DTT, Thermo Fisher Scientific; protector RNase inhibitor (1 U/μl), Sigma-Aldrich], and the nucleus suspension was resuspended. The nuclei were pelleted by centrifugation at 500g for 5 min. The supernatant was discarded until about 20 μl of the supernatant was left in the Eppendorf. Removal was done carefully without disturbing the nucleus pellet. The nucleus pellet was resuspended with 20 μl of 2× Nuclei Buffer [2× Nuclei Buffer, 10x Genomics, PN-2000153; 1 mM DTT, Thermo Fisher Scientific; protector RNase inhibitor (1 U/μl), Sigma-Aldrich].

The nucleus count and the viability of the samples were accessed using a LUNA dual florescence cell counter (Logos Biosystems), and a targeted cell recovery of 10,000 cells was aimed for each of the samples. Library preparations for 10x Genomics single-cell multiome were performed using 10x Genomics Chromium NextGEM Single cell Multiome ATAC+Gene Expression chemistry (10x Genomics, Pleasanton, CA). Transposition of nuclei and the rest of the 10x Genomics Multiome library preparation were performed as per the protocol recommendations (Chromium Next GEM Single Cell Multiome ATAC + Gene Expression Reagent Kits User Guide; CG000338 Rev. G). At the different checkpoints, the library quality was accessed using Qubit (Thermo Fisher Scientific) and Bioanalyzer (Agilent).

10x libraries were sequenced either on an Illumina platform or on an MGISEQ-2000 sequencing platform (MGI Hong Kong SAR sequencing facility). Before being able to sequence on the MGI platforms, the 10x libraries need to undergo a conversion step using the MGIEasy Universal DNA Library Prep kit. Briefly, the final 10x sequencing libraries were circularized using the splint-ligation step, and the circularized libraries were converted to single-stranded DNA copies. DNA nanoballs were prepared from the circularized single-stranded DNA using rolling circle amplification. DNA libraries generated were then flown through the patterned flow cell of the MGISEQ-2000RS High-Throughput Sequencing kit. For a targeted sequencing saturation of 50 to 60%, sequencing was performed at a depth of 30,000 to 60,000 reads per cell for the RNA and ATAC libraries. For RNA libraries, the sequencing recipe of paired-end reads of 28 bp (read 1), 100 bp (read 2), 10 bp (index 1), and 10 bp (index 2) was used for sequencing. For ATAC libraries, the sequencing recipe of paired-end reads of 50 bp (read 1), 49 bp (read 2), 24 bp (index i5), and 8 bp (index i7) was used for sequencing.

RNAscope in situ hybridization

Rbp4-Cre:Ai9 mice were euthanized at different developmental stages (P5, P15, and P28), and brains were dissected and frozen in optimal cutting temperature (O.C.T.) compound (4583, Sakura Finetek). Cryosections (10 μm) were prepared on a CM3050 S cryostat (Leica), mounted on SuperFrost Ultra Plus adhesion slides (Thermo Fisher Scientific), and dried at RT. Sections were stained using the RNAscope Multiplex Fluorescent Detection Kit v2 (ACDbio, 323110) following the manufacturer’s protocols. Briefly, sections were treated for 10 min with a hydrogen peroxide solution and 12 min with protease IV, with 1× phosphate-buffered saline (PBS) washing in between. RNAscope probes were hybridized at 40°C for 2 hours, and subsequently, a horseradish peroxidase signal was developed using OPAL dyes. The Bcl6 probe was used in the C1 channel (+ Opal 520 1:2000), and the Smad3 probe was used in the C2 channel (+ Opal 520 1:2000), while a tdTomato probe was used in the C3 channel (+ Opal 570 1:2000) (table S3). Sections were counterstained with DAPI for 30 s and mounted with Mowiol. Images of tdTomato-positive regions in the DG were acquired on a Zeiss LSM900 using the 20× objective. Constant laser power and gain settings were used to image the same probe in sections of the three different ages. Images were quantified by measuring the probe mean intensity over area of a tdTomato-positive region of interest using ImageJ. Statistical analyses were performed using GraphPad Prism 9.

RiboTag sample preparation and bulk RNA-seq

Rbp4-Cre:RiboTag Cre mice were euthanized at different developmental stages (P5, P7, P10, P15, P21, and P28), and hippocampi were dissected in ice-cold Hanks’ balanced salt solution and frozen for subsequent use. Hippocampi from male and female mice were used at all time points, three or four animals were pooled per sample at P5 and P7, and two animals per sample were pooled at P10, P15, P21, and P28. Moreover, control samples from P10 animals negative for Cre recombinase were used to ensure the specificity of the pull-down.

The RiboTag sample preparation protocol was adapted from Sanz et al. (82). Briefly, hippocampi were thawed in prechilled homogenization buffer containing 50 mM tris-Cl (pH 7.5), 100 mM KCl, 12 mM MgCl₂, 1% NP-40, 1 mM DTT, RiboLock RNase inhibitor (200 U/ml; Thermo Fisher Scientific, EO0382), heparin (1 mg/ml; Sigma-Aldrich, H3393), cycloheximide (0.1 mg/ml; Sigma-Aldrich, C4859), and 1× EDTA-free protease inhibitors. Next, samples were placed in a Dounce homogenizer where they were dissociated using two sizes of glass pestles. The lysate was centrifuged at 10,000g for 10 min at 4°C, and the supernatant (named hereafter as input) was kept for further processing. Pierce Anti-HA Magnetic Beads (40 μl; Thermo Fisher Scientific, 13464229) were added to the input and incubated overnight at 4°C on a microtube rotator with gentle mixing. Next, samples were placed on a magnetic stand for at least 1 min to remove the supernatant (named hereafter as unbound) and washed 4× using high-salt buffer [50 mM tris-Cl (pH 7.5), 300 mM KCl, 12 mM MgCl₂, 1% NP-40, 0.5 mM DTT, and cycloheximide (0.1 mg/ml)]. After the last wash, 350 μl of RLT buffer (Qiagen RNeasy mini kit, 74104) supplemented with β-mercaptoethanol was used to elute RNA from the ribosomes. The mix was vortexed for 30 s, and tubes were placed back in the magnetic stand to collect the clear RLT buffer containing the RNA. Next, RNA was isolated using the Qiagen RNeasy mini kit with DNase treatment and following the manufacturer’s instructions.

RiboTag quality control: Bioanalyzer, quantitative RT-PCR, and Western blot

Bioanalyzer

The RNA quality and concentration was checked on a bioanalyzer instrument (Agilent Technologies) using High Sensitivity RNA ScreenTape (Agilent, 5067-5579). Only RNA samples with an RNA integrity number value higher than 7.5 were used.

Quantitative RT-PCR

For quality control using quantitative PCR (qPCR), 60 to 120 ng of RNA (P5, P10, and P15, n = 2 samples per condition) was retrotranscribed using the High-Capacity cDNA Reverse Transcription Kit (Thermo Fisher Scientific, 4368814). To determine the fold enrichment of respective marker genes in immunoprecipitated RNA compared with P10 whole hippocampal RNA lysates, the LightCycler 480 SYBR Green I Master kit (Roche) and comparative C_T (cycle threshold) method were used. Samples were considered specific if immunoprecipitated RNA exhibited correct de-enrichments or enrichments of respective marker genes and if the RNA of control samples did not show any selectivity for marker genes (see table S4 for primer sequences). For each assay, three technical replicates were performed, and the mean was calculated. The mRNA levels were normalized to Pgk1 mRNA (housekeeping gene) (table S4). Quantitative reverse transcription PCR (RT-PCR) assays were analyzed using GraphPad Prism 9.

Western blot

Input fractions of hippocampal samples were collected for n = 2 P5, P10, and P15 Rbp4-Cre:RiboTag and n = 2 P10 RiboTag mice (labeled as CTRL, Cre neg). A 2% volume of input was used for Western blotting. Briefly, samples were boiled for 5 min at 95°C, and samples were loaded on a 10% precast polyacrylamide gel (Bio-Rad, 4561036) and run at 180 V. Proteins were transferred to a 0.2-μm nitrocellulose membrane using semidry transfer (Bio-Rad) using the mixed-molecular-weight program. Membranes were blocked in 5% milk in TBS-T buffer [25 mM tris-base (pH 7.5), 300 mM NaCl, and 0.05% Tween 20] for 1 hour at RT. Primary and secondary antibodies were diluted in 5% milk in TBS-T. Primary antibodies were incubated overnight at 4°C, and secondary antibodies were incubated for 2 hours at RT (table S3). After primary and secondary antibody incubation, membranes were washed five times with TBS-T and once with 1× PBS before imaging on an ImageQuant 800 (Cytiva) using the SuperSignal West Femto PLUS Chemiluminescent Substrate (Thermo Fisher Scientific, 34577).

RiboTag library preparation and sequencing

Library preparation from total RNA extraction and sequencing was performed at the VIB Nucleomics Core. Per sample, an amount of 25 to 100 ng of total RNA was used as input. Using the Illumina TruSeq Stranded mRNA Sample Prep Kit (protocol version: 1000000040498 v00 October 2017), poly-A containing mRNA molecules was purified from the total RNA input using poly-T oligo–attached magnetic beads. In a reverse transcription reaction using random primers, RNA was converted into first-strand cDNA and subsequently converted into double-stranded cDNA in a second-strand cDNA synthesis reaction using DNA polymerase I and RNase H. The cDNA fragments were extended with a single “A” base to the 3′ ends of the blunt-ended cDNA fragments after which multiple indexing adapters were ligated, introducing different barcodes for each sample. Last, enrichment PCR was carried out to enrich those DNA fragments that have adapter molecules on both ends and to amplify the amount of DNA in the library. Sequence libraries of each sample were equimolarly pooled and sequenced on an Illumina NovaSeq 6000 [version 1.5, SE100 (100-8-8-0), 1% PhiX].

Cell lines

Human embryonic kidney (HEK) 293T-17 cells were obtained from American Type Culture Collection (cat. no. CRL-11268). HEK293T-17 cells were grown in Dulbecco’s modified Eagle’s medium (Invitrogen) supplemented with 10% fetal bovine serum (Invitrogen) and penicillin/streptomycin (Invitrogen).

Plasmids

shRNA sequences for Bcl6 KD and control were obtained from Tiberi et al. (47) (table S5). Hairpin sequences (fwd tcaagagagattctcagatccgtgtc; rev ctcttgaagattctcagatccgtgtc) were obtained from Gulmez Karaca et al. (83). gRNA sequences for Smad3 KO were generated using the CRISPick database with the mouse GRCm38 (National Center for Biotechnology Information RefSeq) chosen as the reference genome, CRISPRko as the mechanism, and the SpyoCas9 (Hsu 2013) transactivating CRISPR RNA (tracrRNA) as the enzyme. Two gRNA sequences with a target cut % below 30% were selected per target gene (table S5). For both gRNA and shRNA, forward and reverse oligos were designed, and Sap I (NEB) restriction site overhangs were added.

For the Supernova sparse labeling, pAAV-TRE-DIO-FLP (Addgene, 118027) and pAAV-TRE-FLP (Lab of Synapse Biology, from Addgene, 118027) were used for Smad3 KO and Bcl6 KD, respectively. shRNAs were cloned into the pAAV-U6-TRE-fDIO-MARCKS-EGFP-IRES-tTA plasmid (Lab of Synapse Biology, from Addgene, 118026). mGFP is tagged with a MARCKS (myristoylated alanine-rich protein kinase C substrate) sequence to be targeted to the membrane. gRNAs were cloned into the pAAV-U6-TRE-fDIO-MARCKS-mScarlet3-IRES-tTA plasmid (Lab of Synapse Biology, from Addgene, 118026, with mScarlet3 sequence from Addgene, 189753). pRep/Cap (AAV2/9; Addgene, 112865) and pΔF6 (Addgene, 112867) plasmids were used to generate AAVs.

AAV production

High-titer AAV production and purification were carried out as previously reported (84). Briefly, six plates of 80% confluent HEK cells were transfected with 20 μg of pDelta F6 plasmid, 10 μg of pRepCap 2/9 plasmid, and 10 μg of AAV genome plasmid per plate using polyethylenimine transfection. Three days later, cells were collected by scraping and centrifugation. AAV particles in the supernatant were precipitated with polyethylene glycol and added to cell lysates. Lysates were treated with benzonase nuclease to remove cellular DNA. Lysates were loaded onto an iodixanol gradient (60, 40, 25, and 15%) and ultracentrifuged for 1 hour 40 mins at 50,000 rpm at 12°C. The 40% fraction containing the purified AAV was carefully collected, desalted, and concentrated on a 4-ml Amicon column (Sigma-Aldrich). AAV purity was tested by silver staining using the ProteoSilver Silver Stain Kit (Sigma-Aldrich, PROTSIL1-1KT). The AAV titer was determined by qPCR to normalize titers between different viral vector batches using LightCycler 480 SYBR Green I Master (Roche, 04707516001). Primers for qPCR targeting the synapsin promoter were as follows: fwd TGATAGGGGATGCGCAATTTGG and rev GTGCAAGTGGGTTTTAGGACCA. Each AAV solution for KO/KD combined two different viruses, each carrying a different gRNA or shRNA, to enhance efficiency.

Stereotactic injections

For the Smad3 loss-of-function analysis [both morphology and electrophysiology (ephys)], Rbp4-Cre:Rosa26-LSL-Cas9 mouse littermates were injected at P7. P7 mice were anesthetized with 5% isoflurane and placed in a mouse stereotact (KOPF) on a hot plate kept at 37°C. During the rest of the procedure, 2% isofluorane was constantly administered. After disinfecting the mouse’s head, local anesthesia was administered by a subcutaneous injection with 100 μl of lidocaine (1% xylocaine). An incision was made on the skin to reveal the skull. One bilateral injection was performed; coordinates to target the DG at P7 were as follows: x: ±1.3, y: 1.4 from bregma, z: −2.1 from the surface of the skull. One hundred nanoliters of AAV mix containing AAV-TRE-DIO-FLP:AAV-gRNA-mScarlet in a 200:1 ratio was injected using a Nanoject III (Drummond) through a beveled capillary at 5 nl/s. After 2-min recovery, the capillary was pulled out at ~0.1 mm/5 s. The incision was stitched with surgical glue (Millpledge Veterinary). After 6 hours, their health was examined, and mice were injected with buprenorphine (0.1 mg/kg).

For the Bcl6 loss-of-function analysis, C57Bl6J littermates were injected at P0. P0 mice were anesthetized using hypothermia-induced anesthesia for 15 min and kept on an ice-cold surface during the procedure. The pups were immobilized manually, and 50 nl of AAV mix containing AAV-TRE-FLP:AAV-shRNA-mGFP in a 100:1 ratio for morphology experiments, or AAV-CAG-GFP:AAV-shRNA-mGFP in a 1:1 ratio for ephys experiments, was injected using a Nanoject III (Drummond) through a beveled capillary at 5 nl/s. After 30-s recovery, the capillary was pulled out at ~0.1 mm/5 s. Two injections per hemisphere were performed to target the DG at P0, the x and y coordinates were visually determined, and z: −1.6 from the surface of the skull was used.

Tissue processing and immunofluorescence

Mice were transcardially perfused with 10 ml of 4% sucrose in 1× phosphate buffer (PB) (0.1 M, pH 7.4) under nonreversible anesthesia. Brains were dissected and postfixed for 1 hour in paraformaldehyde, washed in PB, and sliced to a 80-μm thickness by a vibratome. Sections were blocked in 10% normal horse serum, 0.5 M glycine, and 0.5% Triton X-100 in PBS before incubation with antibodies in 5% normal horse serum and 0.5% Triton X-100 in PBS with extensive washing between primary and secondary antibodies (table S3). Sections were counterstained with DAPI before mounting on microscope slides with Mowiol. Imaging was performed on a Zeiss LSM900 using airyscan mode.

Neuronal cultures

Hippocampal or cortical cultures were obtained from dissected E18 H11-Cas9 mice or C57Bl6J WT embryos and plated on poly-d-lysine (Millipore)– and laminin (Invitrogen)–coated six-well plates. Neurons were maintained in Neurobasal medium (Invitrogen) supplemented with B27. Forty-eight to 72 hours after seeding, 10 mM glia inhibitor (5-fluoro-2-deoxyuridine; Sigma-Aldrich) was added to the medium. To achieve Bcl6 KD, WT neurons were infected at 1 day in vitro (DIV1) with high-titer AAVs expressing either shRNAs against the target gene or control shRNAs and collected at DIV15 for qPCR. To achieve Smad3 KO, H11-Cas9 neurons were infected at DIV1 with high-titer AAVs expressing either gRNAs against the target gene or control gRNA against lacZ, and the protein was collected at DIV15 for Western blotting.

Validation of shRNAs and CRISPR gRNAs (Western blotting and RT-PCR)

CRISPR KO validation

Primary neuronal cultures from H11-Cas9 mice were infected at DIV1 with 1 μl of a mix of AAV-U6-Smad3-gRNA-TRE-fDIO-MARCKS-mScarlet3-IRES-tTA (gRNA1 + gRNA2) (1:2000) or 1 μl of AAV-U6-lacZ-gRNA-TRE-fDIO-MARCKS-mScarlet3-IRES-tTA (1:2000). Neurons were lysed at DIV15 using radioimmunoprecipitation assay buffer (50 mM tris-HCl, 150 mM NaCl, 1% Triton X-100, 0.5% sodium deoxycholate, and 1% SDS). Fifty to 100 ng of protein lysate was loaded on a 10% precast polyacrylamide gel to evaluate KO efficiency between conditions (see table S4 for antibodies).

shRNA KD validation

Neuronal primary cultures from C57Bl6J mice were infected at DIV1 with 1 μl of a mix of AAV-U6-Bcl6-shRNA-TRE-fDIO-MARCKS-EGFP-IRES-tTA (shRNA1 + shRNA2) (1:2000) or 1 μl of AAV-U6-scrambled-shRNA-TRE-fDIO-MARCKS-EGFP-IRES-tTA (1:2000). Neurons were lysed at DIV15 using RLT buffer (Qiagen RNeasy mini kit, 74104) supplemented with β-mercaptoethanol following the manufacturer’s instructions. RNA was retrotranscribed using the High-Capacity cDNA Reverse Transcription Kit (Thermo Fisher Scientific, 4368814). To determine the fold enrichment of Bcl6 cDNA under KD versus control conditions, the LightCycler 480 SYBR Green I Master kit (Roche) and comparative C_T method were used. The mRNA levels were normalized to Rpl10 mRNA (housekeeping gene) (table S3). Statistical analysis was performed using GraphPad Prism 9.

Morphology analysis

Spine analysis

Neurolucida 360 and Neurolucida View Explorer were used to detect dendritic spines and perform spine analysis. Within Neurolucida 360, the backbone of a 20- to 30-μm dendritic segment was tracked. Three or four dendritic segments were reconstructed per animal. Furthermore, spines were classified using Neurolucida and manually confirmed in the three-dimensional (3D) environment on the basis of the fluorescent signal. Neurolucida View Explorer was used to obtain quantitative spine measurements. Spine type distribution and spine density were assessed for control and KO/KD conditions. Statistical analysis was done using GraphPad Prism 10.

MFB reconstructions

To reconstruct individual MFBs, aligned stacks were uploaded in Imaris (Bitplane). The number of filopodia was manually measured in 3D, and only filopodia with a minimum length of 1.5 μm were counted. The 3D reconstructions of the MFB core were obtained by initially segmenting the volume containing the entire MF terminal. The segmented volume was subsequently masked, and the MFB was reconstructed using the “New Surface” tool. The total surface area and volume of the MFB and the length of filopodia were determined for both control and KO/KD conditions through the “Statistics” function. Six to eight MFBs were reconstructed per animal. Statistical analysis was done using GraphPad Prism 10.

Whole-cell ephys

For whole-cell patch-clamp recordings, acute slices were prepared from P27 to P31 Rbp4-Cre:Rosa26-LSL-Cas9 animals injected with either lacZ gRNA (control) or Smad3 gRNAs (KO) or from P27 to P35 C57Bl6J animals injected with either scrambled shRNA (control) or Bcl6 shRNAs (KD) and an AAV-CAG-GFP for broad GC labeling. In short, animals were anesthetized using isoflurane, and after decapitation, the brain was quickly removed and transferred into an ice-cold cutting solution: 110 mM choline chloride, 26 mM NaHCO₃, 11.6 mM Na-ascorbate, 10 mM d-glucose, 7 mM MgCl₂, 3.1 mM Na-pyruvate, 2.5 mM KCl, 1.25 mM NaH₂PO₄, and 0.5 mM CaCl₂; 300 to 315 mosmol, pH adjusted to 7.4, with 5% CO₂/95% O₂. Parasagittal slices (250 μm) containing the hippocampus were made using a vibratome (VT1200, Leica Biosystems). Immediately after cutting, the slices were transferred to a 32° to 33°C cutting solution for 6 min to recover and subsequently stored in a holding solution: 126 mM NaCl, 26 mM NaHCO₃, 10 mM d-glucose, 6 mM MgSO₄, 3 mM KCl, 1 mM CaCl₂, and 1 mM NaH₂PO₄; 300 to 310 mM, pH adjusted to 7.4, with 5% CO₂/95% O₂ at RT. Slices were stored for ~1 hour before experiments. During experiments, brain slices were continuously perfused in a submerged chamber (Warner Instruments) at a rate of 3 to 4 ml/min with a recording solution: 127 mM NaCl, 2.5 mM KCl, 1.25 mM NaH₂PO₄, 25 mM NaHCO₃, 1 mM MgCl₂, 2 mM CaCl₂, and 25 mM d-glucose (pH 7.4 with 5% CO₂/ 95% O₂), including AP-5 [100 μM; (2R)-amino-5-phosphonovaleric acid] and TTX (1 μM; tetrodotoxin). All recordings were done between 31° and 33°C. We used borosilicate glass recording pipettes (resistance: 3.5 to 5 megohms; Sutter P-1000) filled with an internal solution: 126 mM CsMSF, 10 mM Hepes, 2.5 mM MgCl₂, 4 mM adenosine 5′-triphosphate, 0.4 mM guanosine 5′-triphosphate, 10 mM creatine phosphate, 0.6 mM EGTA, 5 mM QX-314, and biocytin (3 mg/ml). Whole-cell patch-clamp recordings of GCs in the DG were done using either a double EPC-10 amplifier under the control of Patchmaster v2 x 32 (HEKA Elektronik) or a Multiclamp 700B amplifier under the control of Clampex 10.7 (Axon Instruments). Currents were recorded at 20 kHz and low-pass filtered at 3 kHz when stored. Series resistance was compensated to 70 to 80% and monitored throughout the experiments. For each GC, both miniature excitatory synaptic inputs (mEPSCs; V_m = −90 mV) and miniature inhibitory synaptic inputs (V_m = 0 mV) were recorded. Miniature input was quantified using the Mini Analysis program (Synaptosoft). Statistical analysis was done using GraphPad Prism 10.

Read mapping

snMO read mapping

The generated raw fastq files were mapped to the Mus musculus (mm10) reference genome using the Cellranger Arc version 2.02 count function with intronic reads included by default.

RiboTag read mapping

The generated raw fastq files were mapped to the M. musculus (mm10) reference genome using STAR version 2.7.5b. Gene counts were then obtained using the featureCounts function from SAMtools version 1.10.

Smart-seq2 read mapping

Adapter trimming was performed using the fastp tool (table S6) (85). Reads were mapped to a custom M. musculus mm10 reference genome, which included the tdTomato gene, using the STAR tool (table S6). Quality control was applied to remove low-quality cells, filtering out cells with low mapping percentages or those lacking tdTomato gene detection (table S6).

Sequencing data quality control

SS2 analysis

Data from different samples were merged and subjected to quality control and filtering. Specifically, cells with fewer than 500 genes, more than 8000 genes, or more than 5% mitochondrial gene content were filtered out. In addition, genes expressed in fewer than six cells were removed, and all mitochondrial genes, along with a list of sex-related genes [as referenced in (37)] (table S7), were excluded from the resulting count matrix. This resulted in a total of 509 cells, with a median of 4276 genes detected per cell. The merged data were then normalized and log transformed using the “normalize_per_cell” and “log1p” functions in Scanpy version 1.9.1, respectively. Highly variable genes (HVGs) were identified using the “seurat_v3” method with n_top_genes = 4000, and the data were subsequently scaled. Last, PCA was performed, and the first two principal components (PC1 and PC2) were visualized.

RiboTag analysis

First, for quality control, genes with extreme or low expression values were filtered out. Specifically, we required each gene to have an expression count greater than 50 in at least three samples, and no gene could have an expression value exceeding 500,000. Genes that did not meet these criteria were excluded from further analysis. In addition, mitochondrial genes, sex-related genes [as referenced in (37)], and genes associated with the GO term “central nervous system myelination” (GO:0022010) were removed (table S7).

Next, to estimate factors of unwanted variation, the “RUVs” function from the RUVSeq package version 1.32.0 was applied with k = 2. Normalization for library size was performed, and the data were log₂ transformed using the “rlogTransformation” function from the DESeq2 version 1.38.3 package.

snMO snRNA-seq analysis

Gene expression data quality control was performed independently for each sample using the Scanpy tool version 1.9.1. For each sample, genes expressed in at least three cells were retained, as well as cells expressing a minimum of 200 genes and with a mitochondrial DNA content below 5%. Doublets were then identified and removed using the Scrublet tool (86), either by directly removing identified doublets or by excluding clusters with a high mean doublet score.

Merge samples GEX snMO

Before merging the high-quality cells from the four different samples, mitochondrial genes and a list of sex-related genes (table S7) (37) were removed. Next, normalization and variance stabilization (87) were applied to each sample independently using the “SCTransform” function from Seurat version 4.3.0 with the “glmGamPoi” method, regressing out the mitochondrial gene percentage. Samples were then merged, and batch effects were corrected using 30 dimensions with the Seurat and Harmony tool (88).

A total of 79,914 nuclei from four different time points were clustered using Louvain clustering (89) with a resolution of 0.9, resulting in 40 clusters. These clusters were manually annotated to cell types on the basis of marker gene expression (table S8).

For the focus of our analysis, cells belonging to the GC lineage—including NPCs, type 1 and 2 NBs, iGCs, and mGCs—were identified and extracted on the basis of GC lineage marker gene expression, resulting in a total of 22,912 high-quality snRNA-seq cells. These cells were further filtered on the basis of snATAC-seq quality (see the next section for further details), leading to 18,907 cells, with a median of 3499 unique molecular identifiers and 1796 expressed genes per nucleus.

The selected cells (GC lineage) were then normalized using the “SCTransform” function from Seurat version 4.3.0 (applied separately by age). Samples were then combined, and PCA was performed using all features. Next, nonlinear dimensionality reduction was applied and then visualized using the “RunUMAP” function from Seurat with the first 25 principal components used (number determined by inspecting the elbow plot).

Gene expression comparison snMO versus SS2

To focus on transcriptionally active genes, for each time point in the SS2 data, we selected the top 40% of the most highly expressed genes on the basis of their mean expression. The average expression profile of these selected genes was calculated for each cell type at the corresponding time point in the snMO snRNA-seq data. To quantify the similarity between gene expression profiles of SS2 and snMO, cosine similarity was computed using the cosine function from the SciPy package version 1.12.0.

snMO snATAC-seq analysis

Each snATAC-seq sample was processed separately using the pycisTopic tool version 2.0 [Python implementation of cisTopic (90)]. First, using cell type annotations for each cell from snRNA-seq analysis, we generated pseudobulk profiles per cell type in GC lineage and then per age only for mGCs. Next, peaks were called on each pseudobulk profile using MACS2 version 2.2.9.1 with the following parameters: --format BEDPE, --gsize mm, --qvalue 0.05, --keep-dup all, --shift 73, and --ext_size 146. To derive consensus peaks, we applied an iterative peak filtering approach (91) as implemented in pycisTopic version 2.0 using a peak_half_width of 250 bp and the mm10 blacklist version 2. This resulted in a total of 387,898 called peaks.

Quality control on consensus peaks was performed using Ensembl version 98 as the BioMart host for gene annotation. The pycisTopic “qc” function was applied separately to each sample. High-quality cells were retained on the basis of the following criteria: a minimum transcription start site (TSS) enrichment score of 5, a fraction of reads in peaks (FRIP) score higher than 0.28, and a minimum number of 1000 fragments. The different quality control criteria resulted in a total of 18,907 cells with high-quality snRNA-seq and snATAC-seq profiles. The 18,907 snATAC-seq nuclei yielded 257,570 accessible regions, with a mean of 4118 accessible regions and 8152 unique fragments per cell, a 53.13% FRIP, and a median TSS enrichment of 15.75.

For topic modeling, latent dirichlet allocation (92) using Gibbs sampling was performed, as implemented in pycisTopic, using Mallet version 202108 with 500 iterations across models of [30, 40, 45, 50, 60, 65, 70, 75, 80, 85, 90, 95, 100] topics. On the basis of pycisTopic model selection metrics, we selected the model with 80 topics.

Next, we binarized the topic-region distributions using both Otsu thresholding (93) and the ntop method, selecting the top 3000 regions. To impute dropouts, region-topic and topic-cell probabilities were multiplied with a scaling factor of 10⁶, followed by normalization using a scaling factor of 10⁴.

To identify DARs, highly variable regions were extracted using the default parameters in pycisTopic. Cell type–specific DARs were identified using the Wilcoxon test on the imputed accessibility matrix, with an adjusted P value threshold of 0.01 and an absolute LFC threshold of 1.

For comparisons across development, we identified DARs across different age groups of mGCs. This was done by comparing each age-specific group of mGCs against the remaining mGCs. In addition, we performed pairwise Wilcoxon tests to identify DARs between each pair of ages in mGCs.

For DAR heatmap visualization, Signac package version 1.9.0 in R version 4.2.2 was used. Chromatin accessibility data were first transformed by applying the term frequency-inverse document frequency method (94) using the “RunTFIDF” function on filtered high-quality cells. The average accessibility in mGCs of each age group was then plotted using ggplot2 version 3.4.1. Genes were linked to up- and down-regulated DARs with a simple approach by identifying the closest gene to each region using the “ClosestFeature” function from the Signac package. GO enrichment analysis was then performed on the resulting gene sets using the “enrichGO” function from the clusterProfiler package. The analysis used BP ontology, with P value and q value cutoffs of 0.05 and the Benjamini-Hochberg correction method. The top four enriched terms for each gene set were merged and visualized with a dot plot.

For the dimensionality reduction of the snATAC-seq data using UMAP, singular value decomposition was performed using the “RunSVD” function after identifying the top variable features (i.e., regions) with the “FindTopFeatures” function, applying a minimum cutoff of 20. Last, UMAP was computed using the first 30 dimensions derived from latent semantic indexing reduction, with the first dimension discarded because of its correlation with the sequencing depth. The snATAC-seq data were then visualized in a two-dimensional UMAP space.

Pseudotime inference SS2 GCs

Merged SS2 data was log normalized using the Seurat version 4.3.0 “NormalizeData” function. The top 2000 variable features of the merged data were identified using the “FindVariableFeatures” function, and the log-normalized data were scaled across all features using the “ScaleData” function. To infer the pseudotime order of GCs, regularized ordinal regression models were fitted using the “ordinalNetCV” function from ordinalNet package version 2.12 with the ordered age as the response using a cumulative link model (family = “cumulative”). The model was evaluated with nFolds set to 5 and nFoldsCV set to 10. The regularization parameter α was optimized using fivefold cross-validation, with the objective of minimizing the mean squared error. To determine genes most associated with pseudotime, nonzero coefficients were extracted, excluding any intercept terms. The prediction scores are generated by applying a linear combination of these selected gene expression values, with each gene weighted according to its corresponding coefficient.

Pairwise DEGs: SS2 and RiboTag

SS2

The normalized merged SS2 data were used with the “FindMarkers” function to perform pairwise differential gene expression using the MAST test to compare each age group with the following age group. For the P28 sample male versus female differential expression analysis, the MAST test was performed on the normalized gene expression matrix before sex- and stress-related gene exclusion (37). Genes were considered up-regulated or down-regulated if the adjusted P value was less than 0.05 and the absolute LFC exceeded 0.5. The numbers of up- and down-regulated genes between each pair of consecutive time points were visualized with a bar plot.

A generalized additive model (GAM) was fitted for each of the top 50 up- and down-regulated genes of each pairwise comparison (ranked by LFC) using the “gam” function from the mgcv package version 1.8.41, with inferred pseudotime as the predictor and the family set to Gaussian. To define gene clusters, the Euclidean distance matrix of the GAM-fitted values was computed, and hierarchical clustering was performed using the “hclust” function from stats (version 4.2.2, R) with Ward’s minimum variance method (“ward.D2”) as the linkage criterion, with a cutoff height of 28.

GO enrichment analysis for BPs was performed using the clusterProfiler version 4.6.2 package and the M. musculus organism database (org.Mm.eg.db), with P value and q value cutoffs set to 0.05. For visualization, the top 3 GO terms from each analysis were extracted, merged, and visualized in a dot plot. Gene expression patterns of the top 50 up- and down-regulated genes (table S9) from each pairwise comparison were visualized in a heatmap using the GAM-fitted expression values labeling only the top 20 genes per pairwise comparison that overlapped with genes from the SynGO database.

RiboTag

Pairwise differential expression analysis (comparing each age to the following age) was performed using the “DESeq” function from DESeq2 based on the negative binomial distribution and corrected for the estimated factors of unwanted variation (95, 96). The shrinkage of effect sizes was applied using the “lfcShrink” function (with the “apeglm” method). Genes were considered differentially expressed if the adjusted P value was less than 0.05 and the absolute LFC exceeded 0.5. The numbers of up- and down-regulated genes between each pair of consecutive time points were visualized with a bar plot.

For the clustering of DEGs, the top 50 up- and down-regulated DEGs per pairwise comparison (ranked by LFC; table S10) were selected and grouped on the basis of their expression profiles using the “degPatterns” function from the DEGreport package version 1.39.6 with nClusters = 6. Specifically, a distance matrix was created by calculating pairwise gene expression correlations using the mean expression per group, followed by the clustering of genes using the DIANA method.

GRN inference

To identify GRNs, the SCENIC+ algorithm was applied to the snMO data using the scenicplus Snakemake pipeline from the scenicplus package version 1.0a1 (41). Briefly, motif enrichment analysis was performed using pycisTarget version 1.0a2 (41, 97) on the region sets consisting of the identified DARs (see above) and binarized topic regions (from the Otsu and ntop methods) using the M. musculus species and a custom ranking database using the consensus peaks identified for the GC lineage, with a padding of 1000 bp. The motif similarity false discovery rate threshold was set to 0.001, with an orthologous identity threshold of 0 and motif-TF annotations based on the v10 clustering version (41).

pycisTarget analysis resulted in a set of cistrome modules consisting of TFs with their bound regions (regions that have a significant enrichment of the TF motif). Regions in the resulting cistromes were then filtered to retain only those that were accessible in any mGC cluster. Gene annotations were obtained using the archived BioMart host version 98 (September 2019). Last, TF-to-gene and region-to-gene importance scores were inferred using a gradient boosting machine and with a search space defined as the larger of either the nearest gene boundary or 150 kb and at least 1 kb upstream of the TSS or downstream of the gene’s end.

Last, eRegulons, in which TFs are linked to their target regions that are subsequently linked to their target genes, were constructed using a gene set enrichment analysis approach, and then cell enrichment scores of both eRegulon target gene and eRegulon target regions were calculated using the AUCell algorithm (97). SCENIC+-inferred eRegulons were filtered to retain only activators (+/+, positive correlation of both TF expression-target gene expression and target region accessibility-target gene expression) and repressors (−/+, negative correlation of TF expression-target gene expression and positive correlation of target region accessibility-target gene expression). For simplicity in this text, activators are referred to as TF(+) and repressors as TF(−) (rather than TF_+/+ and TF_−/+, respectively). In total, 66 eRegulons were obtained.

PCA was performed on normalized and scaled eRegulon enrichment scores per cell (AUC eRegulon matrix using the AUCell algorithm) using the Scanpy package. Nonlinear dimensionality reduction was then applied using UMAP with the first eight principal components. We observed a separation of cells by their cell type following GC lineage across development from progenitors to the mature state.

Pseudotime inference snMO mGCs

To infer pseudotime values for individual cells in snMO mGCs, we used eRegulon AUC scores, representing the enrichment of eRegulons per cell. The computed AUC scores were normalized per cell and log transformed using a pseudocount of 1. Subsequently, z-score normalization was applied to each eRegulon to center the activity scores to a mean of 0 and an SD of 1 across all cells. The standardized values were then used to perform PCA via the “prcomp” function from the stats package (version 4.2.2, R). An elbow plot was generated to visualize the proportion of variance explained by each principal component to determine the optimal number of principal components to retain for the diffusion map analysis. A diffusion map of mGCs was created using the “DiffusionMap” function from the destiny package version 3.12.0 with 12 principal components. mGCs were then clustered using the “kmeans” function (stats package, version 4.2.2, R) with 10 centers to obtain the trajectory starting cluster. The mGC maturation trajectory was subsequently inferred using the “slingshot” function from the slingshot package version 2.6.0, with the diffusion map as the reduced dimensional space.

GC lineage eRegulon selection

To select cell type–specific eRegulons across the GC lineage, two parallel selection methods were followed. eRegulons were first selected if their Spearman correlation value between the eRegulon activity score and the TF expression is higher than the mean correlation of all eRegulons. Next, eRegulons were also selected if they showed differential activity between two cell types [Wilcoxon rank sum test, adjusted P value smaller than 0.0025 (Bonferroni-corrected α), LFC higher than 0.2, and active in at least 25% of the cells in that cell type]. In addition, the eRegulon’s leading TF should also be differentially expressed in the same cell type [Wilcoxon rank sum test, adjusted P value smaller than 0.0025 (Bonferroni-adjusted), LFC higher than 0.5, and active in at least 25% of the cells in that cell type]. In total, 48 eRegulons were selected.

mGC eRegulon selection

To identify key eRegulons, two parallel approaches were used, focusing exclusively on cells annotated as mGCs. First, Spearman correlation was calculated between eRegulon activity scores (AUC values) and TF expression levels. eRegulons were selected if their correlation value exceeded the mean correlation across all eRegulons. Second, eRegulons were also chosen if they exhibited differential activity between two age groups based on a Wilcoxon rank sum test [adjusted P value <0.0033 (Bonferroni correction), LFC >0.2, and active in at least 25% of cells in that age group]. In addition, the corresponding TF had to be differentially expressed in the same age group [adjusted P value <0.0033 (Bonferroni correction), LFC >0.5, and expressed in at least 25% of cells]. This analysis resulted in a total of 36 eRegulons active in mGCs. To identify and visualize different eRegulon clusters in mGCs, a linkage matrix of eRegulons was obtained by performing hierarchical clustering using the “clustermap” function from the seaborn package version 0.13.2 with correlation as the metric for distance calculation. The dendrogram was then created using the hierarchy cluster module from scipy version 1.12.0, with a color threshold (cutoff distance) set at 0.85.

PCA was performed on eRegulon AUC scores (enrichment per cell) in mGCs. This analysis revealed a clear separation of cells by age group, with cells following a continuous trajectory from the early time point (P5) to the late time point (P28).

GO enrichment analysis for eRegulon target genes was performed for BP or cellular component (CC) using the clusterProfiler version 4.6.2 package and the M. musculus organism database (org.Mm.eg.db), with P value and q value cutoffs set to 0.05. For visualization, the top 7 terms were selected, merged, and displayed in a dot plot.

eRegulon target region accessibility was visualized using Signac version 1.9.0 in R version 4.2.2. Region accessibility was first transformed by applying the term frequency-inverse document frequency method using the “RunTFIDF” function on filtered high-quality cells. The average accessibility in mGCs for each age group was then plotted using ComplexHeatmap library version 2.14.0. Regions were then labeled by their predicted target genes by SCENIC+. The signal plot of observed chromatin accessibility over specific target regions was shown using Seurat’s “CoveragePlot” function with 1–kbp (kilo–base pair) (Nptx1) and 500-bp (Tanc1) and 5-kbp (Nptx1) and 75-kbp (Tanc1) extension upstream and downstream, respectively. SCENIC+-predicted target region-target gene links are also shown.

Correlation RiboTag GEX versus snMO GEX

To compare gene expression across postnatal developmental time points in both the multiome and RiboTag data, the expression patterns of TFs and their predicted target genes were visualized for a subset of selected eRegulons. Specifically, the mean gene expression of eRegulons’ leading TFs and target genes (normalized and log transformed) per age group was calculated for both RiboTag RNA-seq and multiome snRNA-seq. These values were then z-score normalized and visualized. In addition, the ±SD of target gene expression in RiboTag was calculated and visualized.

To assess the linear correlation in TF gene expression between RiboTag RNA-seq and multiome snRNA-seq, the Pearson correlation coefficient was calculated. Specifically, the mean gene expression per age was used for this calculation, and the Pearson correlation was computed using the stats module from the SciPy version 1.12.0 package. P values were determined using Monte Carlo resampling with a uniform distribution for random sampling. However, given the relatively small sample size, these P values may be subject to increased variability and potential inaccuracies and should be interpreted with caution. To quantify the linear correlation in eRegulon activity between RiboTag RNA-seq and multiome snRNA-seq, the Pearson correlation metric was also applied. Specifically, the Pearson correlation for each target gene was computed using the “corrwith” function from pandas version 1.5.0, and the mean correlation of all target genes was used to represent each eRegulon. The 95% confidence intervals for the Pearson correlation coefficients were estimated using the “bootstrap” function from the scipy.stats module version 1.12.0. A total of 1000 random paired resamples with replacement was generated to compute confidence intervals.

Supplementary Materials

The PDF file includes:

Figs. S1 to S8

Tables S1 to S5

Legends for tables S6 to S10

Other Supplementary Material for this manuscript includes the following:

Tables S6 to S10