Single-cell profiling of an in vitro model of human interneuron development reveals temporal dynamics of cell type production and maturation

Summary

GABAergic interneurons are essential for neural circuit function and their loss or dysfunction is implicated in human neuropsychiatric disease. In vitro methods for interneuron generation hold promise for studying human cellular and functional properties and ultimately therapeutic cell replacement. We describe here a protocol for generating cortical interneurons from hESCs and analyze the properties and maturation timecourse of cell types using single-cell RNAseq. We find that the cell types produced mimic in vivo temporal patterns of neuron and glial production, with immature progenitors and neurons observed early and mature cortical neurons and glial cell types produced late. By comparing the transcriptomes of immature interneurons to more mature neurons, we identified genes important for human interneuron differentiation. Many of these genes were previously implicated in neurodevelopmental and neuropsychiatric disorders.

eTOC

Close et al. describe the types of cells present in cultures of human interneurons generated from embryonic stem cells by characterizing their gene expression. Computational methods were used to identify genes that are expressed as these cells mature.

Introduction

GABAergic neurons are a diverse population with crucial processing roles (Tremblay et al., 2016). These cells have been implicated in human disorders, including autism, schizophrenia and epilepsy (Marin, 2012). Mammalian inhibitory interneurons originate from the embryonic ganglionic eminences, with the majority produced in the medial ganglionic eminence (MGE) (Bandler et al., 2016). These cells either migrate tangentially to the pallium and integrate into the cortical circuitry, or remain in the ventral telencephalon and become striatal interneurons (Marin et al., 2000). MGE-derived cortical interneurons differentiate into either Parvalbumin (PVALB) or Somatostatin (SST)-expressing cells (Butt et al., 2005; Cobos et al., 2006; Xu et al., 2004).

Mouse studies have carefully detailed the timing, subregional origin, and molecular events required for the production and survival of PVALB⁺ and SST⁺ expressing interneurons (Bandler et al., 2016). Multiple subtypes of SST⁺ interneurons exist in adult cortex, but it remains unknown whether specification is a result of intrinsic or extrinsic influences (Ma et al., 2006; Tasic et al., 2016; Zeisel et al., 2015).

Production of GABAergic interneurons from human embryonic stem cells (hESCs) makes therapeutic transplant possible, and multiple protocols to create MGE-like cells from hESCs exist (Southwell et al., 2014). The MGE produces cortical and striatal interneurons as well as oligodendrocytes, but the proportion of specific cell types generated in vitro is not known because these studies have characterized gene expression at the population level (Kessaris et al., 2006). It is also not clear how human interneuron identity and development in culture differs from primary tissue.

Here we describe the gene expression changes that occur as MGE-like cells are generated from hESCs. We performed extensive transcriptional profiling of subpopulation and single-cells using RNAseq at multiple timepoints to characterize the diversity and maturation status of these cells. We compared the transcriptomes of interneurons generated using this protocol to in vivo-derived human fetal interneurons using principal component analysis (PCA) to place them in temporal and developmental context. Our single-cell RNAseq (scRNA-seq) analysis revealed 41 distinct populations of progenitors, neurons and glia over the course of the differentiation protocol. We identified a set of genes associated with SST⁺ interneuron maturation. This study is the first to characterize the diversity and dynamic transcriptomic changes of human MGE progenitors and neurons generated in vitro at single-cell resolution, and the first transcriptomic comparison between human fetal interneurons and hESC-derived human interneurons.

Results

In vitro generation of MGE progenitors and neurons

To characterize human interneuron development, we developed a protocol to generate MGE-like cells in vitro based on previous methods (Maroof et al., 2013). Telencephalic neural induction of hESCs was initiated with dual SMAD inhibition, combined with a small-molecule inhibitor of the WNT pathway (Figure 1A)(Chambers et al., 2009). On day 10 (D10), cells were ventralized using sonic hedgehog (Shh) and purmorphamine for 8 days (orange bar). On D24, the ventralized cells were seeded at low density to induce neuronal differentiation (green bar). Immunofluorescence at D10 indicated widespread SOX2 expression, and MKI67-positive proliferating cells, indicating that these cultures were correctly specified as neuronal progenitors (Figure 1B). We assessed high (100 ng/ml Shh and 1 μM purmorphomine) and low (50 ng/ml Shh and 0.5 μM purmorphamine) Shh treatment conditions to determine the effects on MGE subtype specification (Xu et al., 2010). With high levels of Shh signaling factors, 74.5 ± 9.4 % of cells were NKX2-1⁺ compared with 65.2 ± 6.2% in cultures treated with low levels of Shh, though this difference was not statistically significant (Figure 1C). Telencephalic specification was equivalent between conditions, as FOXG1 expression levels were 86 ± 2.8% and 85 ± 1.0% in high and low Shh conditions, respectively (Figure 1F). No NKX2-1⁺, FOXG1⁻ cells were observed in our cultures and there was no difference in terms of the percentage of SST⁺ cells or intrinsic factors that would bias these cells to either a dorsal or ventral MGE identity between Shh levels (data not shown) (Flames et al., 2007). However, in 30% of experiments treated with higher levels of Shh agonists, cultures detached from the plate after D24. Since the conditions were otherwise comparable, cultures treated with low levels of Shh signaling factors were used.

Figure 1. MGE-like progenitors and neurons are generated in vitro from hESCs.

(A) Summary of hESC differentiation procedure. (B) D10 immunostaining for SOX2 (red), and MKI67 (green) (C) FOXG1 (red) and NKX2-1(green) at D24. (D,E) GAD67 (green) and SST (red), at D54 and D100, respectively. (F) Percentage of FOXG1 and NKX2-1 expressing cells for 100 ng/ml rmShh + 1 μM purmorphamine (dark green bars) and 50 ng/ml Shh + 0.5 μM purmporphamine (light green bars). 86 ± 2.8% of cells expressed FOXG1 in high Shh and 85 ±1.0% in low Shh cultures. NKX2-1 was expressed in 74.5 ± 9.4% and 65.2 ± 6.2% of cells in high and low Shh-treated cultures, respectively. (G) Quantification of interneuron markers at D54 (dark blue bars): GAD67, 77.2 ± 4.5%; SST, 46.3 ± 6.1%; CALB2, 16.4 ± 2.9%; NR2F2, 37.3 ± 9.6%; 7.2 ± 3.3%. D100 (light blue bars): GAD67, 47.2 ± 7.5%; SST, 24.9 ± 4.5%; CALB2, 12.3 ± 1.7%; NR2F2, 23.4 ± 3.0%; TH, 5.3 ± 1.7%. (H) Percentage SST cells expressing marker at D54 (dark orange bars): CALB2, 21±4.7% ; NR2F2,35.6 ± 9.9%; TH, 5.8 ± 2.1%. D100 (light orange bars): CALB2, 10.3 ± 2.5% ; NR2F2, 29.9 ± 4.9% ; TH, 8.21 ± 3.41%. (I) SOX2 (red) and MKI67 (green) at D10 in DCX-citrine cells. (J) FOXG1 (red) and NKX2-1 (green) expression at D24 in DCX-citrine cells. (K) GAD67 (red) and Citrine (green). (L) MKI67 (red) and Citrine(green). (M) The percentage of citrine-positive cells observed during FACs analysis at each timepoint: D19, 8.7 ± 1.7%; D24, 22.1 ± 1.1% ; D54, 80.7 ± 1.2%; D100, 70.6 ± 6.0% ; D125, 40.8 ± 6.2%. (N) D24 quantification of NKX2-1 (76.2 ± 2.9%, dark grey bar) and FOXG1 (85.7 ± 0.3%, light grey bar) in DCX-citrine (O) Quantification of MAP2 (89 ± 2.3%, dark blue bar), GAD67 (80.5 ± 6.8 %, blue bar), and SST (32.6 ± 3.5%, light blue bar) in DCX-citrine cells at D54. Scale bars = 100 μm. Error bars = s.e.m.

At D54, a majority of cells were GAD67⁺ (77.2 ± 4.5%, Figure 1D, G, dark blue bar). We did not observe any PVALB⁺, CHAT⁺ or VGLUT1⁺ cells in these cultures. However, SST⁺ GAD67⁺ were present (Figure 1D, E). By D100 the percentage of GAD67⁺ cells was reduced (47.2 ± 7.5%, Figure 1G, light blue bar). A similar reduction in SST cell percentages was observed between D54 (46.3 ± 6.1%, dark blue bar) and D100 (24.9 ± 4.5%, light blue bar). We quantified the expression of SST subtype markers CALB2, NR2F2 and TH (Figure 1H, S1)(DeFelipe et al., 2013; Tasic et al., 2016). CALB2⁺ and NR2F2⁺ cell percentages decreased over time. At D54, 21.0 ± 4.7% (dark orange bar) of SST cells expressed CALB2, but only 10.3 ± 2.5% (light orange bar) by D100. SST and NR2F2 double-positive cell numbers were reduced between day 54 (35.6 ± 9.9%) and D100 (29.9 ± 4.9%). Only few SST⁺ cells were also TH⁺ at both D54 (5.8 ± 2.1%) and D100 (8.2 ± 3.4%).

We constructed an hESC line with Citrine (Cit) fused to the endogenous copy of the DCX gene in the H1 parent line to allow us to profile neurons (Cit⁺) and progenitors (Cit⁻) (Yao, 2017). A majority of cells were FOXG1 (76.2 ± 2.9%) and NKX2-1⁺ (85.7± 0.3%) by day 24 (Figure 1J, N), indicating that the DCX-citrine line differentiated comparably to the parent line (Figure 1N; compare with Figure 1F). By D54, we observed many GAD67⁺Cit⁺ cells (Figure 1K), but no overlap between MKI67 expression and Cit (Figure 1L). FACs analysis indicated that the percentage of Cit⁺ cells in these cultures increased between D19 and D54, declining thereafter (Figure 1M). At D54, 89 ± 2.3% of Cit⁺ cells were MAP2⁺, indicating Cit was confined to post-mitotic neurons (Figure 1O). In addition, 80.5 ± 6.8 % of Cit⁺ cells were GAD67⁺, and 32.6 ± 3.5% were SST⁺ (Figure 1O), indicating the DCX-Cit line will allow us to enrich for GABAergic cells for transcriptome profiling.

To determine the physiological properties of our in vitro-derived interneurons, we performed patch clamp electrophysiology between D65 and D75 (Figure S2). Most patched cells exhibited electrophysiological properties of immature neurons, with relatively depolarized membrane potentials and high membrane resistance. As a result of the relatively high membrane resistances of these cells, it was difficult to obtain stable recordings. However, by holding cells at −65 mV, we elicited action potentials with current injection (Figure S2D). Spontaneous inhibitory post-synaptic currents in voltage clamp mode were abolished with application of picrotoxin, suggesting that these events were mediated by GABAergic cells (Figure S2E). These data suggest that the in vitro-derived inhibitory interneurons are indeed GABAergic.

RNAseq analysis confirms in vitro generation of MGE-like cells

Cells differentiated for 24, 54 and 100 days were purified by fluorescence activated cell sorting (FACS) and analyzed by RNA-Seq for validation. Analysis of key genes revealed that the in vitro generation of MGE-like cells replicates the temporal progression of gene expression observed in vivo. For descriptive purposes, we considered log₁₀(TPM+1) values of <1 to be low expression, 1 – 2 to be moderate, and values above 2 to be high levels of expression. Forebrain patterning genes (OTX2, FOXG1, LHX2, SIX3, SOX2) were highly expressed in progenitors, with OTX2 and FOXG1 expressed in Cit⁺ cells at moderate levels (Figure 2A)(Acampora et al., 1995; Dou et al., 1999; Ferri et al., 2013; Gestri et al., 2005; Kelly and Moon, 1995; Matsuo et al., 1995; Monuki et al., 2001). PAX6 was expressed at moderate levels in Cit⁻ cells at D24, but low levels thereafter (Briscoe et al., 1999). Pallial markers (EMX1) or midbrain/hindbrain markers (PAX2, IRX3, EN2 and GBX2) were expressed at low levels or absent. These results suggest this in vitro protocol correctly specified cells as ventral telencephalic progenitors.

Figure 2. In vitro and in vivo subpopulations are comparable.

10K cell Citrine⁻ and Citrine⁺ populations were sorted at D24, D54, D100, and RNAseq was performed. N=3 experiments/timepoint. (A) Telencephalic atterning genes: OTX2, FOXG1, LHX2, SIX3, SOX2, PAX6 and EMX2. Midbrain/hindbrain markers: PAX2, IRX3, EN2 and GBX2. MGE transcription factors: NKX2-1, ASCL1, DLX1, DLX5, LHX8, LHX6, ZEB2, SOX6, abARX, and SATB1. (B) Migration genes: DCX, CXCR7, CXCR4, and ERBB4. Interneuron maturation genes: GAD1, SLC32A1, SST, NPY, CALB2, GRIA2, and GRIA4, CCK, RELN, NOS1, PVALB, and VIP. Values are expressed in Log10 (TPM+1). (C) Representative flow plot of mid-gestation human cortical cells stained for PAX6 and SOX2. Gated sub-populations are named and frequency averaged across four donors (96 to 115 dpc) is shown in parentheses. (D) Hierarchical clustering of marker genes for distinct cortical cell types generated from 100-cell sub-populations from the four cortical specimens as in panel B. (E) Principal component analysis (PCA) of Citrine⁺ and Citrine⁻ 10,000-cell populations from indicated stages of differentiation and 100-cell populations of primary cortical interneurons (P9-green dots). Primary cortical interneurons are most similar to D54 Citrine⁺ cells (neurons) in PCA space. For Citrine⁺ vs Citrine ⁻ gene expression levels, *P<0.05, **P<0.01, ***P<0.001, Mann-Whitney (BH FDR correction).

Transcription factors important for MGE specification were expressed at moderate to high levels (NKX2-1, ASCL1, DLX1, DLX5, LHX8 and LHX6)(Anderson et al., 1997a; Anderson et al., 1997b; Casarosa et al., 1999; Eisenstat et al., 1999; Flandin et al., 2011; Hansen et al., 2013). ZEB2, SOX6, ARX, and SATB1, were expressed at moderate levels from D54 on (Batista-Brito et al., 2009; Close et al., 2012; Colombo et al., 2007; McKinsey et al., 2013). The expression pattern of these transcription factors is reminiscent of the temporal progression of gene expression for MGE-like progenitors and interneurons.

Interneuron migration genes DCX, CXCR4, CXCR7, and ERBB4 were expressed at moderate or high levels in postmitotic cells, with CXCR4 and CXCR7 also expressed in Cit⁻ cells (Figure 2B)(Anton et al., 2004; Friocourt et al., 2007; Vogt et al., 2014; Wang et al., 2011). SST was expressed in both Cit-negative and -positive populations at D54 and 100. DCX expression is highest in immature neurons and downregulated in the most mature and functional cells during development (Brown et al., 2003; Couillard-Despres et al., 2005). SST expression in Cit⁻ cells represents a small number of SST neurons that downregulates DCX and Citrine, as 3.3 ± 1% and 9.6 ± 2.9% of SST-expressing cells were Cit⁻ at D54 and D100, respectively (Figure S3). SST subtype markers, NPY and CALB2 were expressed in Cit⁺ cells at D54 and D100, but CCK, RELN and NOS1 expression were low. AMPA receptors (GRIA2, GRIA4) were expressed in Cit⁺ cells at D54 and D100 (Akgul and McBain, 2016). There was low or no expression of PVALB, VIP or HTR3A, indicating that these interneuron subtypes were not produced. Raw expression levels from pooled progenitors and neurons (TPM), as well as p-values (Mann-Whitney, BH FDR correction) indicating whether expression differences between progenitors and neurons are significant, are provided in Table S1. Taken together, these data indicate that the current protocol generates MGE-like progenitors and interneurons that express the appropriate genes for specification, migration and function of GABAergic interneurons.

To compare in vitro-derived neurons to those generated in vivo, we analyzed the gene expression of cells derived from four human fetal neocortical samples at approximately 100 days post-conception (dpc). FACS was used to sort fixed cortical progenitors (SOX2^hi) and neurons (SOX^lo/−PAX6⁺) for gene expression profiling (Thomsen et al., 2016). 100-cell pools from seven reproducibly-observed populations were evaluated by RNAseq (Figure 2C). Populations expressing high levels of SOX2 were composed of neuronal and glial precursors, including a mixture of radial glia and intermediate precursors (P4: EOMES, PAX6, HOPX, VIM), astrocyte precursors (P11: AQP4), and oligodendrocyte precursors (P10: OLIG1, PDGFRα) (Figure 2D). Fetal projection neurons were SOX2⁻ P6-P7. Only two populations expressed detectable levels of GAD1: P8 and P9. The P8 population likely consists of LGE-derived interneurons, express expressed PAX6 and SP8. The P9 population expressed the highest levels of GAD1, and was the only population to show substantial expression of SST or LHX6, indicating this population is enriched for cortical interneurons. We performed principal component analysis (PCA) with the top 1% of genes based on maximum pairwise fold change across the six in vitro-derived subpopulations and mapped the in vivo-derived interneurons onto this PCA space (Figure 2E). Replicate in vitro populations show reproducible patterning, with the first PC separating progenitor (Cit⁻) and neuronal (Cit⁺) populations at each time point, and the second PC corresponding to the time point of collection. This analysis reveals that interneurons from around 100 dpc human fetal cortex most closely resemble D54 interneurons generated in vitro.

ScRNA-seq analysis of in vitro-derived human interneurons

We performed scRNA-seq on hESC-derived MGE-like cells at D24, D54, D100 and D125 to characterize transcriptomic changes with single cell resolution. Progenitors (Cit⁻) or neurons (Cit⁺), were dissociated, sorted and prepared for scRNA-seq (Picelli et al., 2014)(Figure S4). To determine the maturation landscape and relationship between cells, we performed PCA based the top 1% of differentially expressed genes across all populations based on maximum pairwise fold change (Figure 3A). PCA plots using 10% or 100% of expressed genes were also generated, and result in separation of progenitors and neurons as well as cells in various states of maturation (Figure S5). The first two PCs correspond to Cit expression status and days in culture, allowing us to define the temporal order of these cells by differentiation stage (PC 1) and pseudotime (relative maturity based on gene expression; PC 2), respectively. Progenitors and neurons largely grouped in gene expression space by Cit status; however, some Cit⁻ cells mapped with neurons and vice versa, and a small number of cells (~8.3%) mapped to an intermediate differentiation stage. Similarly, neurons and progenitors both grouped predominately based on number of days in culture. Of note, not much separation was detected between D100 and D125 cells, which could be an indication that transcriptional changes do not underlie late-stage differentiation, or that cells do not mature in vitro after D100.

Figure 3. Single cells are distinguishable by transcriptomic signatures.

(A) PCA of single cells based on high variance genes separates neurons from progenitors along PC1 (Differentiation stage), and separates cells based on their maturation along PC2 (Pseudotime). Top genes loading on PCs 1 and 2 are shown as heat maps along the top and right edges of the PCA plot, respectively. Each box shows the average expression of a particular gene for PCs 1 or 2 for all cells within 1 of 40 evenly spaced bins in PC space. For both D and E, white represents minimal log-transformed expression while red represents maximal expression of a particular gene. P < 10⁻¹⁰⁰ for all genes (B) Example genes showing extensive changes in expression with pseudotime, differentiation stage, or both. Grey dots indicate a cell with FPKM=0.

Expression of many genes was highly correlated with differentiation status and pseudotime. For example, VIM, PDPN, and S1PR1 were expressed predominantly in progenitors, while interneuron genes such as SST, GAD1, and STMN2 gradually increased with differentiation status (Figure 3A, top panel. P < 10⁻¹⁰⁰ for all genes described in this figure, see Table S2). Similarly, several genes showed significantly increased (BCAN, PTN, and C1orf6 (CROC4)) or decreased (FAM60A, CRABP2, and DLK1) expression with pseudotime (Figure 3A, right panel), likely representing a general program of development. Gene expression was often restricted by both differentiation status and pseudotime. In contrast to ZFP36L1, which was expressed in all progenitors (Cit⁻) populations, COL2A1 and AQP4 were confined to early- and late-stage progenitors, respectively (Figure 3B, left panels). Similarly, while DCX expression increased with differentiation stage, LHX8 was maximally expressed in early (e.g. newly generated) pseudotime neurons, whereas SST expression was highest in late (e.g. more mature) pseudotime neurons (Figure 3B, right panels). To verify these results, we compared gene expression changes correlated with pseudotime for our in vitro cells to gene expression patterns observed in the developing non-human primate and human brains (Figure S6)(Bakken et al., 2016). We find that genes positively or negatively correlated with pseudotime in vitro are expressed in similar patterns in vivo. These results provide insight into the dynamics of shifting cell type populations emerging during interneuron development, and their underlying transcriptional changes.

Weighted gene co-expression network analysis (WGCNA) was performed to identify modules of co-expressed genes, and the genes were then used to cluster cell types, as described in previously (Figure 4A) (Tasic et al., 2016; Thomsen et al., 2016). We observed 41 cell clusters—8 at D24, 11 at D54, 13 at D100 and 10 at D125—which showed clear separation between progenitor (Cit⁻) and neuron (Cit⁺) cells at each time point. Cit⁻ clusters had higher VIM and SOX2 expression, whereas Cit⁺ clusters expressed more DCX and GAD1 (Padj < 0.001, FDR for Cit⁺ vs. Cit⁻).

Figure 4. Single cell transcriptome analysis reveals progenitor and neuron diversity dynamics.

(A) Iterative WGCNA reveals distinct transcriptomic groups at each time point. Gene expression levels represent centered cluster averages of log2(tpm+1) values. (B-G) Immunofluorescence was used to visualize cell types present in some of the clusters identified. B): ASCL1(red) and NKX2-1(green)-positive cells at D24; C) SP8 (red) and CITRINE (green)-positive cells at D24; D) OLIG2 (red) and ASCL1 (green)-positive cells at D54; E) CORT (red) and CITRINE (green)-positive cells at D54; F) HOPX (red) and CRYAB (green)-positive cells at D100; G) RBFOX3 (red) and CITRINE (green)-positive cells at D100. Arrowheads indicate double-positive cells, Scale bar = 50 μm.

More Cit⁻ clusters expressed PAX6 at D24 than later timepoints. DLL1, DLL3, HES6 and ASCL1 tended to be co-expressed in at least one Cit⁻ cluster at each timepoint, perhaps indicating the emergence of intermediate progenitors. After D54, Cit⁻ clusters expressed higher levels of MGE migration genes (i.e. ERBB4, CXCR4), and from D100 through D125, Cit⁻ clusters expressed HOPX, AQP4 and S100B (Padj < 0.001 for Cit⁻ cells at D100 and D125). Cit⁺ populations at D24 expressed DLX2, DLX5 and DLX6, indicating that these clusters were maturing interneurons (Padj <0.001, D24 Cit⁻ vs. Cit⁺). D24 neurons did not express high levels of functional maturity genes, such as GRIA2, GRIA4, SST and RBFOX3, though these genes were expressed after D54 (Padj <0.001 D24 Cit⁺ vs. D54). Clusters that co-expressed OLIG1, OLIG2 and PCDH15 emerged after D100, indicating that oligodendrocyte precursors were present at late timepoints. For a detailed heatmap of all genes distinguishing D54 clusters, see figure S7. These data suggest that the cell types produced shift over time in terms of identity and maturation status with a temporal progression like that observed in vivo.

We performed immunofluorescence to validate the clusters we observed. At D24, ASCL1 and NKX2-1 co-localized in a subset of cells, as expected for the ipMGE.24.DLL1 cluster (Figure 4B). We also observed expression of SP8 in a subset of Cit⁺ neurons, which made up the nLGE clusters at D24 (Figure 4C). At D54, co-localization of OLIG2 and ASCL1 was observed in presumptive ipMGE cells (Figure 4D), while CORT expression was present in Cit⁺ nCTX cells (Figure 4E). One Cit⁻ cluster expressed CRYAB at D100, while six were positive for HOPX. These genes were expressed in mutually exclusive clusters by gene expression (Figure 4A) and antibody staining (Figure 4F). RBFOX3 was expressed in a limited number of nCTX clusters, and we observed its expression via immunofluorescence in a subset of Cit+ cells (Figure 4G).

If the cell types produced in vitro mimic those produced in vivo, we would expect unventralized telencephalic (PAX6⁺) and mitotic MGE progenitors (NKX2-1⁺) early, followed by immature MGE-like neurons and intermediate progenitors (DLL1⁺, ASCL1⁺, CCND2⁺), with functionally mature interneurons, and glia produced at the latest timepoints. Therefore, we assigned putative cell type identities to each cluster produced based on important developmental regulators (Figure 5A). We observed three categories of Cit⁻ cells: PAX6⁺ clusters, which were deemed telencephalic progenitors (pTEL); NKX2-1⁺ clusters; and clusters negative for both. NKX2-1⁺ progenitors were further divided into ASCL1⁻ MGE progenitors (pMGE) and ASCL1⁺ intermediate MGE progenitors (ipMGE). NKX2-1 and PAX6 negative cells were further divided into ASCL1⁻ progenitors (p) and ASCL1⁺ intermediate progenitors (ip). Four main categories of Cit⁺ clusters emerged: GAD1⁺,PAX6⁺, SP8⁺ clusters were deemed LGE neurons (nLGE); ERBB4⁻,GAD1⁺ clusters were called MGE neurons (nMGE); ERBB4⁺, GAD1⁺ clusters were categorized as cortical interneurons (nCTX); while clusters that expressed OLIG1 and PDGFRα were classified as oligodendrocyte precursors (oMGE).

Figure 5. Gene expression differences within cell type groups.

(A) Assignment to cell type groups based on gene expression: Citrine⁻ cells were assessed based on Z-scores for canonical transcription factors. Z > 0 = positive expression. PAX6⁺/NKX2-1⁻ cells were designated pTEL (light blue), NKX2-1⁺/PAX6⁻ cells were either designated as pMGE (ASCL1⁻, dark blue) or ipMGE (ASCL1⁺, pink), NKX2-1⁻/PAX6⁻ cells were designated as progenitors (p, green) or intermediate progenitors (ip, ASCL1⁺, periwinkle). Citrine⁺ cells were assigned identity based on gene expression. PAX6⁺/SP8⁺ cells were designated nLGE (yellow), ERBB4⁻ neurons were designated nMGE (light orange), SST⁺/ERBB4⁺ cells were designated nCTX (dark orange), and OLIG1/OLIG2/PDGFRα+ cells were designated oMGE (maroon). B) Cell type groups identified and classified using WGCNA emerged in distinct temporal patterns. Citrine⁻ cell types: pTEL (light blue), pMGE (dark blue), ipMGE (pink), p (green), and ip (periwinkle). Citrine⁺ cell types: nLGE (yellow), nMGE (light orange), nCTX (dark orange) and oMGE (maroon). Large circle: >75 cells, Medium circle: 26–75 cells, Small circle: 1–25 cells. Violin plots show gene expression distribution in log2 (tpm+1) for each cluster, scaled relative to the maximal value. (C) Temporal transcriptomic changes within Citrine⁻ cell types. (D) Temporal transcriptomic changes within Citrine⁺ cell types.

A few overall trends emerged from these analyses (Figure 5B). PAX6⁺ progenitors (pTEL) were observed from D24 to D100, with peak production at D24. The only progenitor type observed at every timepoint was the pMGE, NKX2-1⁺ progenitor, which constituted 37% of all Cit⁻ cells collected and profiled. ASCL1⁺ intermediate MGE progenitors (ipMGE), were observed only during the initial 54 days of culture, whereas PAX6⁻NKX2-1⁻ progenitors (p, ip) were only observed from D100 through D125, possibly indicating a switch from neurogenesis to gliogenesis. In addition to the expression of ASCL1, clusters classified as intermediate progenitors also tended toward higher expression of DLL, DLL3 and CCND2 (Figure 4A). Previous work has indicated that CCND2 expression indicates intermediate progenitor status, and may bias interneuron progenitors toward the PVALB fate (Glickstein et al., 2007; Petros et al., 2015).

Cit⁺ cells included both neurons and oligodendrocyte precursors. LGE-like neurons (nLGE), which expressed SP8, and only constituted 6% of the Cit⁺ total. As with the pTEL PAX6⁺ populations, this may suggest that at the earliest timepoint, D24, progenitors and the neurons that arise from them are incompletely ventralized. ERBB4⁻ neurons (nMGE) constituted 28% of the Cit⁺ population, and were observed at every timepoint. These neurons could be potential striatal neurons, as many clusters lacked ZEB2 or LHX6 expression, but they may also be immature cortical interneurons yet to initiate the genetic program necessary for tangential migration. The majority of Cit⁺ cells were GAD1⁺ERBB4⁺SST⁺ cortical interneurons (nCTX), constituting 64% of all Cit⁺ cells. We did not observe significant PVALB RNA expression in any cluster. Lastly, oligodendrocyte precursors (oMGE) were observed at and after D100. These trends mimic in vivo patterns of cell type production.

We also observed temporal changes within cluster types (Figure 5C, D). At D24, pTEL progenitor clusters expressed slightly higher levels of MKI67 and GPC3, indicating these early populations are mitotically active (Filmus, 2001; Scholzen and Gerdes, 2000). By D54 and D100, pTEL progenitor clusters express NFIA and HOPX, consistent with a transition to astrogliogenesis and/or outer radial glial cell production (Deneen et al., 2006; Pollen et al., 2015; Shu et al., 2003; Thomsen et al., 2016). After D54, pMGE clusters also express higher levels of HOPX, in addition to AQP4, S100B, ANXA1, and S100A10 compared to earlier clusters. These genes have been implicated in astrocyte differentiation and/or function, another indication that progenitors in vitro make the transition to gliogenesis at late stages, as occurs in vivo (Donato et al., 2013; Eberhard et al., 1994; Nielsen et al., 1997).

LHX8 is expressed early during MGE-derived interneuron production, and its expression is reduced in cortical interneurons (Flandin et al., 2011; Fragkouli et al., 2005; Fragkouli et al., 2009). Indeed, we observed that the nMGE clusters from earlier timepoints expressed LHX8, but there was little or no LHX8 expression at later timepoints, or in nCTX neurons (Figure 5D, Padj <0.001). There was a split between CORT⁻ and CORT⁺ nCTX clusters that indicated the latter represented a more mature population. CORT⁻ cells expressed more CXCR4 and NKX2-1, which are repressed as cortical interneurons mature. The CORT⁺ cells expressed higher levels of SST, GRIA4, and RBFOX3 (see Figure 4A, Padj <0.001 for all genes described, CORT⁻ vs. CORT⁺ clusters). These gene expression trends indicate that differentiation and maturation of in vitro-derived interneurons and oligodendrocytes occurs side by side in vitro.

Gene expression changes during SST cell differentiation

To determine genes associated with specific phases of SST cell maturation, we performed differential expression (DEX) analysis of SST-expressing clusters using the R package limma (Ritchie et al., 2015). At D54, four clusters of SST+ cells were present, two of which were CORT⁺ (Figure 4,5). Cortistatin (CORT), a neuropeptide expressed by a subset of cortical interneurons, was previously implicated in regulation of slow-wave sleep (de Lecea et al., 1997). The CORT⁺ clusters (nCTX.54.CORT, nCTX.54.CORT/NEFL) appeared more mature than the CORT⁻ clusters (nCTX.54.SST, nCTX.54.SST/GRIA2), due to CXCR4 expression in the latter, and increased expression of GRIA2, SATB1 and MEF2C in the former (Figure 4A, 5C). We looked for differentially expressed genes across these clusters to determine the signatures of SST cell maturation. We identified 2304 DEX genes (FDR < 0.05), with the majority (90%) upregulated in the CORT⁺ populations (see Table S3 for complete list of differentially expressed genes). We plotted the top 50 DEX genes ranked by test significance (Figure 6A), and the 29 genes enriched in maturing SST cells includes a number of genes implicated in migration (DAB1, CNTNAP2, PHLDA1), synaptogenesis (SERPINI1, NTM, OPCML) and axon outgrowth (CNTN1, DCC), as well as three immunoglobin superfamily genes (CNTN1, OPCML and NTM). GABA receptors, ion channels and proteins that directly regulate ion channel function are also prominent features (RIC3, GABBR2, GABRG2, SCN2A, KCNC2). RAB3B is a small GTP-binding protein known to modulate synaptic release, but it has not been implicated in interneuron differentiation or function to date (Schluter et al., 2006). These results emphasize the fact that transcriptomic changes occurring during neuronal differentiation can be observed in vitro using single-cell and computational methods, providing a means to identify processes that affect brain development and function in specific ways.

Figure 6. Differentially expressed genes during SST neuron differentiation.

(A) Analysis of genes differentially expressed between D54 SST clusters reveal the top 21 genes highly enriched in the more immature CORT⁻ clusters, and the top 29 genes enriched in CORT⁺ clusters are shown. B) Mouse brain expression of select genes found to be differentially regulated during SST cell differentiation.

The expression of many of these genes showed expected patterns in the Allen Developing Mouse Brain Atlas (http://developingmouse.brain-map.org/). At E13.5 and E15.5, SST RNA is expressed in a pattern indicative of interneurons in active tangential migration (Figure 6Ba,b). By E18.5 and P4, SST is expressed sparsely in the cortex (Figure 6Bc,d). At P56 (Figure 6Be), SST expression is most prevalent in layers IV-VI. CORT expression was not observed until P4, with a sparse expression pattern reminiscent of interneurons at P56 (Figure 6Bi, j). At P4, RAB3B is present in the intermediate zone and olfactory bulb (Figure 6Bn), and in a pattern resembling SST expression at P56 (Figure 6Bo). This data validates our observations, and suggests that some of these genes are expressed in a pattern indicating they have a role in interneuron maturation and/or function.

Many of the differentially expressed genes have an indirect link to disease endophenotypes. For instance, RIC3 is a chaperone protein affecting maturation and expression of α7 nicotinic acetylcholine receptors (Halevi et al., 2003; Williams et al., 2005), agonists of which have shown promise in treating schizophrenia (Olincy et al., 2006; Olincy and Stevens, 2007). Two IgLON immunoglobin superfamily members implicated in synaptogenesis and neuritogenesis were also enriched in CORT⁺ cells: OPCML and NTM (Hashimoto et al., 2009; Sanz et al., 2015; Yamada et al., 2007). Loss of these genes is linked with autism spectrum disorder and developmental delay (Minhas et al., 2013). The presence of these disease genes at a specific stage of interneuron differentiation and maturation suggests these diseases may involve specific neurodevelopmental processes that occur late in brain development.

Discussion

Here we provide a detailed transcriptomic analysis of in vitro-derived human interneurons and progenitors using scRNA-seq and computational approaches. The temporal order of cell production in vitro mimicked that observed in vivo, and we determined that ~100 dpc fetal interneurons matched best with D54 hESC-derived interneurons. We found that many genes were previously implicated in neuronal migration and function are upregulated during SST cell differentiation, and that mutations in some of these genes were implicated in human neurodevelopmental and neuropsychiatric disorders. This dataset could be mined to examine additional transitional gene expression cascades, allowing for a refinement of culture techniques to produce specific cell types, or for discovery of additional developmental processes.

We determined that the majority of cells produced using the current protocol were MGE-like progenitors or cortical interneurons, but we found that additional cell types were present in culture, including unventralized PAX6⁺ progenitors and LGE-like neurons. The cell types we observed may be unique to the conditions used, but single-cell resolution will provide insight into the development of effective protocols for generating specific populations, especially for therapeutic transplant purposes in which it is important to have an understanding of the cells present. Notably, there were no PVALB-expressing interneurons present in our cultures at any time point. Previous work has suggested that this population is sensitive to extrinsic factors such as activity and connectivity, and requires a protracted period of maturation (Cellerino et al., 1992; Nicholas et al., 2013). Likely we did not observe these cells because there were no excitatory neurons present in culture to provide input. Future studies connecting the transcriptomic identity of human MGE-like progenitors, with fate acquired after transplant into mouse cortex, will likely be required to determine which conditions are optimal for production of this cell type, as they are unlikely to reach maturity in vitro.

We also observed a marked reduction in the percentage of cells expressing GAD67 and Citrine between D54 and 100, likely because GAD67 and DCX are predominately expressed by postmitotic cells, and mitotic cells outnumber differentiated neurons over time. Reductions in GAD67⁺ cells have also been observed in vivo, and it is possible in vitro-derived cells undergo a similar process of programmed cell death (Southwell et al., 2012). It is also possible that some interneuron subtypes require excitatory input for survival and maturation, rendering monocultures suboptimal for survival and function of GABAergic neurons (Close et al., 2012; De Marco Garcia et al., 2011). These factors must be taken into account, as a clear understanding of the molecular mechanisms of subtype specification relies on the ability to produce these cells in the first place.

We made a number of important observations with the current conditions. First, single-cell analysis over the entire 125 day period of culture revealed that cell-type production proceeded as expected based on in vivo studies, with the production of unventralized telencephalic progenitors early, followed by mitotic NKX2-1⁺ MGE progenitors, ASCL1⁺NKX2-1⁺ intermediate progenitors, and glia. Interneurons to immature to be classified as cortical interneurons appeared earliest, followed by ERBB4⁺SST⁺ cortical interneurons, and finally immature oligodendrocyte precursors. We mapped expression data from GAD1⁺ migrating interneurons from D100 dpc fetal human cortex onto the PCA space generated from the in vitro data and found that they most closely resembled neurons obtained from D54 in vitro. Between D100 and D125 in vitro, the rate of transcriptomic changes appeared to slow. A slow-down in gene expression dynamics has been observed in previous studies of human cortical development (Bakken et al., 2016; Colantuoni et al., 2011). It is possible that our in vitro observations reflect similar changes occurring at equivalent ages in vivo, or that some extrinsic element necessary for continued development of these cells is missing from our protocol, as previous work has suggested that development is comparatively slow in vitro (Sadegh and Macklis, 2014).

We observed important transition states by analyzing single-cell data. Specifically, we found that SST⁺ interneurons were present at various levels of maturity at D54. SST cell death has been observed in mood disorders and dysfunctional SST interneurons have recently been shown to exacerbate cortical excitotoxicity in neurodegenerative disorders (Lin and Sibille, 2013; Zhang et al., 2016). Our analysis determined that several genes important for migration, synaptogenesis, axon outgrowth and ion channel function were differentially expressed by more mature interneurons. None of the differentially expressed genes were previously shown to play a specific role in SST interneuron differentiation, though a number of them play a role in interneuron migration or differentiation in general. Disabled-1 (DAB1), has been implicated in interneuron migration, though it may not be required for this process (Hammond et al., 2006; Pla et al., 2006). Deletion of contactin-associated protein-like 2 (CNTNAP2) disrupts interneuron migration, results in a loss of interneurons, and mutations in CNTNAP2 are associated with epilepsy and autism in humans (Penagarikano et al., 2011). Calcium/calmodulin-dependent 3′,5′ cyclic nucleotide phosphodiesterase 1A (PDE1A) was recently identified as a marker of a SST neuron subtype (Zeisel et al., 2015). The potassium voltage-gated channel subfamily C member 2 (KCNC2) has been characterized in interneuron function, and may underlie some of the unique intrinsic physiology of these cells (Chow et al., 1999). PHLDA1 was identified as a downstream target of aristalless (ARX), which is itself crucial for interneuron development and migration (Friocourt and Parnavelas, 2011). A specific role for these genes in SST cells will require careful genetic dissection, as past work has shown that genes expressed in multiple subtypes do not necessarily perform similar functions at equivalent stages of maturation, and may depend on combinations of factors within each subtype as they diverge (Batista-Brito et al., 2009; Close et al., 2012). Their precise role will likely need to be determined in a non-human model in order to uncover any specific effects of deletion or overexpression in SST-expressing interneurons. But the fact that mutations in many of these genes have been described in neuropsychiatric and developmental disorders makes them intriguing targets for study.

STAR Methods

Contact for Reagent and Resource Sharing

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact Jennie Close (jenniec@alleninstitute.org).

Experimental model and subject details

Human ESC Differentiation

Undifferentiated human ESCs (H1, WiCell) were maintained in mTESR (Stemcell Technologies) on matrigel (BD Biosciences)-coated plates. We did not authenticate this line, though we did conduct pathogen testing Colonies were dissociated with Dispase solution (Thermo-Fischer) and replated onto matrigel-coated plates at ~15% confluence. Induction of neural progenitors was adapted from Maroof et. al. 2013. On D0, hESCs were dissociated into a single cell suspension with Accutase (Life Technologies) and plated onto matrigel-coated 24-well plates at 3 × 10⁵ cells/cm² in mTeSR media with 2 μM thiasovivin (StemRd). Neural/telencephalic induction was initiated on D1 by replacing medium with NIMX (all reagents from Thermo Fisher Scientific unless indicated), a dual SMAD inhibition medium consisting of DMEM/F12, 1X N-2 supplement, 1X B-27 supplement, 2mM Glutamax Supplement, 0.1 mM MEM Non-Essential Amino Acids Solution, 0.11 mM 2-mercaptoethanol, 0.05% (v/v) Bovine Albumin Fraction V Solution, Penicillin-Streptomycin, 100 nM LDN193189 (Reagents Direct), 10 μM SB431542 (StemRD), 2 μM XAV939 (TOCRIS bioscience). Medium was changed daily. On D5, the medium was replaced with 75% NIMX media, 25% N2 medium (DMEM/F12 (1:1), N-2 supplement, 25% (w/v) dextrose (Sigma Aldrich), 55 μM 2-Mercaptoethanol, Penicillin-Streptomycin). On D7, cells were fed with a 50/50 NIMX/N2 medium, on D9, cells were fed with 25% NIMX medium. On D10, cells were dissociated into a single cell suspension with Accutase and plated onto matrigel-coated 24-well plates in 25% NIMX supplemented with 2 μM thiasovivin at a density of 1.3 × 10⁶/cm². On D11 the medium was replaced with N2/B27 medium (N2 medium (above) + B-27 supplement) containing Purmorphamine (Stemgent) and Sonic Hedgehog (R&D Systems). On D19-23, cells were fed daily with N2/B27 medium without Shh and purmorphamine.

On D24 of neural induction, cultured cells were dissociated with accutase to allow clumps of 10–50 cells to remain and plated at a density of 8 × 10⁵ cells/cm² onto poly-D-lysine (Millipore) and laminin (Invitrogen)-coated 35mm and 24-well plates in N2/B27 + 2 μM thiasovivin.

On D25, medium was replaced with NBND medium for neuronal differentiation (Neurobasal medium, 0.5× N-2 supplement, 0.5× B-27 supplement, 2 mM Glutamax supplement, 0.025% Bovine Albumin Fraction V Solution, 1X Penicillin-Streptomycin, 55 μM 2-Mercaptoethanol, 0.2 mM Sodium-L-Ascorbate (Sigma Aldrich), 0.1 mM Adenosine 3′,5′-cyclic monophosphate (Sigma Aldrich), 5 ng/mL NT-3 (R&D Systems), 5 ng/mL BDNF (R&D Systems), 5 ng/mL GDNF (R&D Systems). Half the medium was replaced every 48 h for the duration of the experiment.

For RNAseq analysis, samples were taken from at least 3 separate experiments that passed our quality control metrics. Our quality control metrics included: 1) a minimum percentage (70%) of FOXG1 and NKX2-1 positive cells by immunofluorescence at D24; 2) a minimum percentage (10%) of Citrine+ cells by FACs analysis at D24; 3) a minimum (30%) percentage of SST+ cells by immunofluorescence at D54.

Human fetal tissue

Human fetal tissue was staged, collected and dissected by the Laboratory of Developmental Biology (BDRL), utilizing University of Washington IRB protocol 41557. All human tissue was donated with written and informed consent and in compliance with the Health and Human Services Code of Federal Regulations (CFR) 45 Part 46.101–124 (subparts B, D, E). Four human fetal samples were utilized, and these were staged at 113 dpc (unknown gender), 115 dpc (male), 101 dpc (unknown gender) and 96 dpc (unknown gender).

Method Details

Immunocytochemistry

Cultures were fixed in 4% (w/v) paraformaldehyde (Electron Microscopy Sciences) in 1× PBS for 15 min at room temperature (RT), and washed twice with PBS. Fixed cells were incubated for 1 hr at RT in blocking solution (PBS with 2% (v/v) goat serum (Vector Laboratories), 0.1% (v/v) Triton X-100 (Sigma Aldrich)). Immunostaining was performed overnight at 4°C with primary antibodies diluted in PBS and 0.1% (v/v) Triton X-100. Cells were rinsed three times with PBS and incubated for 1 hr at RT with secondary antibodies, protected from light. Fixed cultures were rinsed once with DAPI in PBS and coverslipped with Fluormount-G (SouthernBiotech). Primary antibodies used: FOXG1 (Abcam, ab18259), TH (Millipore, AB5986); CALB2 (Millipore, MAB1568), CORT (LifeSpan BioSciences, LS-C143717), GAD67 (Millipore, MAB5406), GFP (Nacali Tesque, 04404-84, or Millipore, 06-896), MKI67 (BD Pharmingen, 550609), NR2F2 (R&D Systems, PP-H7147-00 or Abcam, ab64849); SOX2 (Millipore, AB5603); SST (Millipore, MAB354); NKX2-1 (Progen, 16108). Secondary antibodies used were species-specific Alexa Fluor dyes (ThermoFisher Scientific).

Image Analysis

Images were captured using a Nikon Eclipse Ti fluorescent microscope with NIS-Elements AR software and edited with Adobe Photoshop CS5. Image J was used to manually count cells. Three images were taken for analysis per experimental condition. Cells were counted at random by extracting the image’s channels; 200+ DAPI cells were counted per image.

Fetal brain tissue processing

Cortical samples were identified by morphology and then dissociated into single cell suspensions. Tissue was minced (approx. 0.25 – 0.5 mL total volume) with #5 forceps (Fine Science Tools, Foster City CA) in Ca²⁺- and Mg²⁺-free HBSS (Thermo Fisher) and digested with 2 mL trypsin solution for 20 min at 37 °C (Ca²⁺- and Mg²⁺-free HBSS, 10 mM HEPES, 2 mM MgCl₂, 0.25 mg/mL bovine pancreatic trypsin (EMD Millipore), 10 μg/mL DNase I (Roche), 100 nM TTX (Tocris), 20 μM DNQX (Tocris), and 50 μM DL-AP5 (Tocris), pH 7.6). Digestion was quenched with 6 mL of ice-cold Quenching Buffer (440 mL Leibovitz L-15 medium, 50 mL water, 5 mL 1M HEPES pH 7.3–7.4, 5 mL 100× Pen-Strep, 20 mg/mL bovine serum albumin (Sigma), 100 μg/mL trypsin inhibitor (Sigma), 10 μg/mL DNase I, 100 nM TTX, 20 μM DNQX, and 50 μM DL-AP5. Samples were resuspended with 1 mL of quenching buffer and triturated on ice with a P1000 pipette set to 1 mL, using 25 gentle cycles, and then diluted in 30 mL in Staining Medium (440 mL Leibovitz L-15 medium, 50 mL water, 5 mL 1M HEPES pH 7.3–7.4, 5 mL 100× Pen-Strep, 20 mL 77.7 mM EDTA pH 8.0 [prepared from Na₂H₂EDTA], 1 g bovine serum albumin, 100 nM TTX, 20 μM DNQX, and 50 μM DL-AP5), filtered, pelleted (220 × g, 10 min, 4°C), resuspended in 5 mL staining medium, and counted on a hemocytometer (typically ~20–40 M live cells isolated per cortical piece at ~50% viability). Cells were fixed (4% PFA in PBS, 15 min on ice), rinsed twice in staining buffer (SB, PBS, 0.2% w/v molecular biology grade BSA (Gemini Bio-Products), 0.25% v/v RNasin Plus (Promega)), and kept frozen in this buffer at −80°C at 2 million cells/1.5 mL until use. For sor ting specific populations and retrieving RNA from fixed fetal cortical cells, we followed a previously published method for staining and sorting these cells, followed by reversing cross-linking and RNA purification prior to SmartSeq2 RNAseq (Thomsen et al. 2016).

Single-cell sorting

hESC-derived cultures were dissociated with Accutase (ThermoFisher) at 37°C with trituration until nearly all clumps had been dissociated. Single-cell suspensions were filtered with a 40 μm cell strainer, washed in PBS with 1% FBS, stained with 0.5–1 μg/mL DAPI and sorted on a FACSAria II SORP (Becton Dickinson) directly into PCR strip tubes or plates held in chilled aluminum blocks. Doublets and dead cells were excluded based on forward scatter, side scatter and DAPI fluorescence. A 130 μm nozzle was used with the sort mode set to single cell. Accuracy of single-cell sorts was confirmed by sorting DAPI-stained fixed cells onto a dry well of a 96-well plate and analyzing by fluorescence microscopy.

Single-cell transcriptomics

Both FRISCR purified mRNA and live single cells were prepared for single-cell transcriptomics using SmartSeq2 (Picelli et al., 2013). After reverse transcription with ProtoScript II (New England Biolabs) and template switching, we amplified cDNA with KAPA HotStart HIFI 2 × ReadyMix (Kapa Biosystems) using 19–22 cycles. We purified PCR products using Ampure XP beads (Beckman Coulter), and quantified cDNA using a High Sensitivity DNA Chip on a Bioanalyzer 2100 (Agilent). We used 1 ng of cDNA to generate RNA-Seq libraries using the Nextera XT library prep system (Illumina) and sequencing of human cortical cells occurred on the Illumina HiSeq using 50 base paired-end reads.

RNA preparation from cell populations

For sub-populations, cells were dissociated as above and sorted based on positive or negative Cit fluorescence. Total RNA was harvested from populations of 10,000 cells collected and stored in RLT Lysis Buffer (Qiagen) and 2-Mercaptoethanol (Sigma Aldrich). RNA was extracted (Qiagen RNeasy Micro Kit) and validated with a RNA Pico Chip on a Bioanalyzer 2100 (Agilent). cDNA was amplified using SmartSeq2 on total RNA, and libraries constructed by tagmentation with as with single cell libraries. To characterize sub-populations of fixed and stained primary human cortical cells, FRISCR was carried out on 100-cell sample just as with single cells, however three fewer PCR cycles were used.

Electrophysiology

Patch clamp recordings were performed on human ESC-derived neurons of the DCX-Cit reporter line. At 65–75 days post-induction the culture dishes were transferred to the stage of an upright microscope equipped with differential interference contrast optics and perfused at 3–4 mL per minute with artificial cerebrospinal fluid (aCSF) bubbled with carbogen gas (95% O2/5% CO2). The aCSF was composed of: 121 mM NaCl, 2.5 mM KCl, 1.25 mM NaH₂PO4, 26 mM NaHCO₃, 12.5 mM glucose, 5 mM HEPES, 2 mM CaCl₂.2H₂O 2, and 2 mM MgSO₄.7H₂O with pH of 7.35 and osmolality of 305 mMol/Kg. Cells were patched based on healthy appearance, shape of somata, and Cit fluorescence. Borosilicate glass electrodes had a tip resistance of 4–6 MOhms when filled with the internal pipette solution composed of: 130 mM K-Gluconate, 4 mM KCl, 10 mM HEPES, 0.3 mM EGTA, 10 mM Na₂-phosphocreatine, 4 mM Mg-ATP, and 0.3 mM Na₂-GTP adjusted to pH 7.35 and osmolality of 285–290 mMol/Kg. Recordings were obtained with a Multiclamp 700B amplifier and Digidata 1550 digitizer using pClamp10 data acquisition software (Molecular Devices). Action potential firing behavior was examined in current clamp mode by injecting a series of 1 sec step current waveforms (from hyperpolarizing to depolarizing) in 5 pA increments until the full dynamic range of the neuron was probed. To standardize measurements for these experiments, a hyperpolarizing bias current was applied as needed to adjust baseline membrane potential to approximately −65 mV. Spontaneous synaptic events were recorded in voltage clamp mode with the holding potential set to −50 mV to facilitate distinction of excitatory (inward) versus inhibitory (outward) events. The GABA-A receptor antagonist picrotoxin was bath applied at 50 uM in the aCSF to block sIPSCs.

Quantification and Statistical Analysis

Clustering method

Cells with over 200,000 mapped mRNA reads were clustered iteratively based on WGCNA gene modules as described previously (Tasic et al., 2016; Thomsen et al. 2016). Due to strong temporal gene expression variation across time points, we clustered independently for cells at each age group. Within each age group, we assessed whether gene modules were represented by one batch. Modules that showed significant batch effect, or corresponded to cell cycle states, ribosome rRNAs, ribosomal proteins, immediate early response genes, mitochondrial genes, RNA binding, and mRNA quality (as assessed by mRNA mapping percentage of the single cells) were removed from downstream analysis, as our aim was to identify cell types independent of these factors. The cells were then clustered based on eigengenes from remaining gene modules using hierarchical clustering with “Ward’s” method. The number of clusters were selected dynamically so that every pair of clusters can be distinguished by at least 5 differentially expressed genes (2-fold change, adjusted Pvalue < 0.01) whose sum of –log₁₀(adjusted Pvalue) is greater than 40 (each gene’s contribution capped at 20). We produced cluster heatmaps at each age group by selecting the top 20 differentially expressed genes in both directions between every pair of clusters. Top markers (Figure 4, 5) were selected manually from the top differentially expressed genes between clusters such that every pair of clusters differ by at least one marker. For comparison of gene expression between clusters, we used the R “limma” package to calculate raw and adjusted P values based on average expression of a given gene within each cluster (log₂(TPM+1)).

Differentially expressed genes in SST+ clusters

313 cells in D54 SST⁺ clusters were selected for differential expression analysis. The R package “limma” was used to test each gene for variable expression between clusters, and 2304 genes were statistically significant (Benjamini-Hochberg corrected P < 0.05). For the top 50 most significantly variable genes, the expression z-score across the four clusters was calculated and plotted as a heatmap using the R package “pheatmap”.

Comparison of primary cortical cell and in vitro interneuron populations

Seven populations of primary cortical cells were sorted based on SOX2 and PAX6 antibody staining, with n≥3 unique specimens and n = 9–16 technical replicates per population. Six populations of 100 in vitro cells were collected based on Cit fluorescence and stage (D26, D54, and D100), with N = 3–11 replicates per population. In vivo and in vitro cells were compared as described below.

Predominant cell types in each primary cortical cell population were identified using known marker genes, and visualized by taking the average expression across population replicates, clustering these data using marker genes, and plotting the results as a heat map. Gate P09 populations, which showed highest expression of SST interneuron markers, were then directly compared with in vitro interneuron populations. First, data from both sets of populations were log transformed by taking Log₂(FPKM+1), and absent genes (FPKM>1 in >5% of in vitro populations) were excluded from the analysis. Next, standard principal component analysis (PCA) was performed (using the prcomp R function) on the top 1% most differential genes based on maximum pairwise fold change across 6 groups (DCX⁺ D24, DCX⁻ D24, DCX⁺ D54, DCX⁻ D54, DCX⁺ D100, DCX⁻ D100). Finally, we mapped populations from gate P09 into PCA space to estimate age of in vitro cells, and plotted the results. We note that this analysis is robust to choice of threshold for differentially expressed genes.

Defining pseudotime and differentiation stage in single cells

Data from each cell was log transformed by taking Log2(FPKM+1), and absent genes (FPKM>1 in >5% of in vitro cells) were excluded from the analysis. Next, standard principal component analysis (PCA) was performed (using the prcomp R function) on the top 1% most differential genes based on maximum pairwise fold change across eight groups (DCX⁺ D24, DCX⁻ D24, DCX⁺ D54, DCX⁻ D54, DCX⁺ D100, DCX⁻ D100, DCX⁺ D125, DCX⁻ D125). Differentiation stage and pseudotime were defined as the first two principal coordinates, respectively.

Gene expression was visualized using two strategies. First, gene expression was mapped directly onto the MDS plot, with grey points indicating no expression and with a white-to-red vector of expression indicating low (nonzero) to high expression. Second, pseudotime and differentiation stage were summarized by binning cells into 40 equally spaced groups and plotting the average metric value for all cells in that bin as one row of a heatmap. Top genes loading on each PC were plotted in this manner. The Pearson correlation between every gene and differentiation stage (across all, early, and late cells) and pseudotime (in all, high diff stage, and low diff stage cells) are included in Table S2.

Data and Software Availability

Raw and normalized RNA-Seq data from hESC-derived cells can be obtained at NCBI GEO accession super series ID number is GSE93802, with the single-cell data under id number GSE93593, and the subpopulation data under id number GSE93801. Raw and normalized RNA-Seq data from human fetal cortical cells can be obtained at dbGaP (phs001276.v1.p1).

Highlights.

• We characterized cell types present during in vitro production of human interneurons.

• Genes expressed during somatostatin interneuron maturation were identified.

• Computational methods allowed us to describe temporal changes