To examine the diversity of primate inhibitory neurons (INs) during development, we dissected progenitor zones in the ventral telencephalon and migratory destinations in the cortex and basal nuclei of prenatal rhesus macaque brains. We focused on the lateral, medial and caudal ganglionic eminences (LGE, MGE and CGE, respectively) and ventromedial forebrain (VMF) regions including the septum, preoptic area and preoptic hypothalamus, where distinct configurations of transcription factors specify initial IN classes3,4,5,6. We also sampled migratory destinations in the cortex and basal nuclei, where local signals further influence the maturation and postmitotic refinement of neuron classes7. In total, we collected 71 samples across 9 specimens, spanning the onset of cortical neurogenesis, postconception day 40 (PCD40), to the conclusion, PCD100 (ref. 8; Fig. 1a and Extended Data Fig. 1). We performed single-cell RNA sequencing using the 10x Chromium Controller, incorporating recent data from PCD110 (ref. 9) and mouse studies from PCD13 to PCD21 (refs 10,11) as well as adult olfactory bulb (OB)12 (Supplementary Table 1). We applied stringent quality control, batch correction, dimensionality reduction, Leiden clustering and RNA velocity trajectory analysis to identify transcriptionally similar classes of progenitors and postmitotic INs among 109,112 macaque cells and 141,065 mouse cells, which were identified by expression of DLX and GAD genes (Methods).

Fig. 1: Transcriptional diversity of IN precursors in developing macaque and mouse telencephalon.
figure 1

a, Regions dissected for single-cell RNA sequencing, labelled on PCD80 macaque lateral, medial and frontal coronal section traces. Columns of stacked boxes represent samples from each individual, with each box representing a region dissected and sampled. PCD110 samples from Zhu et al.9. PFC, prefrontal cortex; POH, preoptic hypothalamus; POA, preoptic area. b, Model of inhibitory neurogenesis. c, d, UMAP projections coloured by progenitor state and initial class for mice (c) and macaques (d). Insets in cd show the dissection region from a, the scVelo dynamical RNA velocity shared latent time and age.

Macaque and mouse IN progenitors clustered mainly by cell cycle phase rather than spatial origin (Fig. 1b–d and Extended Data Figs. 13). Similarly, the most immature new-born neurons clustered by both class and differentiation stage, according to the RNA velocity latent time (Fig. 1b–d and Extended Data Figs. 13). From the Leiden clusters, we delineated 11 discrete initial classes of macaque postmitotic neurons, which resolved to 17 initial classes in mouse (Fig. 1c, d). We found that canonical marker genes for established progenitor territories exhibited significant correlations among the progenitors, suggesting that core transcriptional regulatory programmes that are present at or before the last cell division predict the identity of postmitotic neurons in the initial classes. For example, signatures reflecting spatial origin and subtypes in the CGE (NR2F1 and NR2F2), LGE (MEIS2, EBF1 and ISL1), MGE (MAF, LHX6 and CRABP1) and VMF (ZIC1 and ZIC2) were visible in G1 progenitors, and signatures related to four distinct LGE initial classes further emerged in dividing cells as correlations of the pan-LGE marker MEIS2 to PAX6, FOXP2, ISL1 or PENK (Extended Data Fig. 1d). These observations support a model in which we refer to differentiating postmitotic neuron clusters that are largely defined by germinal zone regional transcription factors as initial classes of new-born neurons. This terminology reflects the idea that a small number of discrete transcriptional classes are initially produced following a neuron’s final cell division and that each of these immature transcriptional states is later partitioned into one or many different mature classes by nurture and circumstance as neurons migrate and integrate into the circuitry (Fig. 1b)7. Although many studies have demonstrated that these transcriptionally defined classes result from shared lineage relationships13, we refrain from referring to classes as lineages in the absence of direct lineage tracing data.

Conserved and divergent initial classes

To construct a taxonomy of IN development, we sought to identify evolutionarily conserved cell classes and to link them to candidate adult populations14. We found most well-known initial class markers to be conserved between species (Fig. 2a and Extended Data Fig. 4). Gene expression signatures for each progenitor and postmitotic initial class in macaques correlated strongly with at least one comparable class in mice (Extended Data Fig. 4). Most classes had one-to-one relationships, although subclasses such as LGE_FOXP1/ISL1/NPY1R and LGE_MEIS2/PAX6/SCGN as well as a number of VMF classes were apparent in the mouse data but were undersampled in macaques. Because cell type correlation methods depend on clustering resolution in each species, we further examined homology at the level of individual cells. Mutual nearest-neighbour analysis showed that all telencephalic initial classes present in mice were also present in macaques (Fig. 2b). To infer the putative fates of the initial classes in the absence of lineage tracing, we compiled the most complete available data of adult mouse brains (Extended Data Fig. 5). We then computed the terminal class absorption probabilities for prenatal neurons using nearest-neighbour relationships and RNA velocity with equal weight in CellRank’s Markov chain model15. Our predicted mapping of postmitotic differentiation and partitioning of each initial class using transcriptional similarities recapitulated known lineage relationships and made a number of unexpected predictions that support unresolved linkages in the literature, as summarized in Fig. 2c, such as an NKX2-1+ MGE-derived LAMP5+ cortical chandelier population16,17,18 and a shared origin of amygdala intercalated cells (ITCs) and striatal eccentric spiny neurons19,20,21. The widespread distribution and diversity of derivatives from some initial classes such as MGE_LHX6/NPY also highlights the shared genetic programmes underlying the initial specification of populations that later diversify according to regional destinations, where terminal classes are commonly subdivided into many transcription types and morphotypes. Although our results suggest that the initial classes are largely uniform, minor axes of variation may already exist within classes, which could trigger downstream cascades that bias terminal fate partitioning. However, such variation cannot be identified here without knowing a cell’s fate a priori.

Fig. 2: Unified taxonomy of Euarchontogliran telencephalic IN specification.
figure 2

a, Heatmap of macaque initial class marker expression, scaled by column. Stacked barplots correspond to region of origin for cells in each class. The dendrogram represents complete linkage of the Pearson correlation distance of mean expression values. Stacked bar plots show the regional distribution of each class. b, Sankey diagram in which the thickness of the lines between the left and middle columns represents the number of mutual nearest-neighbour (MNN) cells shared between each class and that between the middle and right columns reflects the initial class (IC) identity of the 100 cells with the highest (CellRank) probabilities of being absorbed in each terminal class (TC). c, Summarized taxonomy of initial and terminal classes observed in macaques and mice. Forked lines represent subclasses that become apparent postmitotically. Initial classes of INs are organized by the presumptive birthplace based on the expression of regional marker genes and putative birthdates, presented in the manner of Lim et al.5. Inferred terminal fates are based on our gene expression and histology analysis and the literature, as denoted and discussed in detail in Supplementary Table 4. S/DWMIN, superficial/deep white matter inhibitory neuron; BN, basal nuclei. RMTW_ZIC1/RELN and VMF_TMEM163/OTP were not included because they are excitatory cortical and hypothalamic classes, respectively.

Recent comparative studies of adult primate, rodent and ferret telencephalon showed a primate-specific population of striatal INs that express the neuropeptide TAC3 (ref. 2), although the developmental origin of this evolutionarily novel population remains unclear. These striatal INs are important exceptions to the one-to-one conservation of initial classes between primates and rodents (Fig. 2b). Instead, mice have a single ancestral class of MGE_CRABP1/MAF neurons that shows strong homology to the MGE_CRABP1/MAF and MGE_CRABP1/TAC3 clusters in macaques (Extended Data Fig. 4c). We further examined the gene networks that define this primate-specific population (Fig. 3a, b and Extended Data Fig. 6). Using RNAscope, we quantified the co-expression of the dividing cell marker MKI67 and the initial class markers CRABP1, TAC3, MAF and LHX8 across the rostrocaudal expanse of the MGE and striatum at PCD65. Our results showed a bias of MGE_CRABP1/MAF neurons rostrally and MGE_CRABP1/TAC3 neurons caudally in the MGE progenitor zone (Fig. 3c, d). In addition, we detected a low fraction of cells co-expressing CRABP1, TAC3 and MKI67 that were displaced from the ventricle, suggesting that subventricular zone (SVZ) progenitors upregulate the programme for this novel initial class at or before their final cell division (Extended Data Fig. 7). In the striatum, both classes showed uniform distributions, which was also confirmed by RNAscope for STXBP6, ANGPT2 and RBP4 in two additional individuals (Extended Data Figs. 7 and 8). LHX8 expression was restricted to a subset of CRABP1+TAC3+ cells outside the MGE (Fig. 3d), highlighting early postmitotic specification of a TAC3/LHX8 subclass observed in adult marmosets2. Interestingly, although they were clearly distinct from primate MGE_CRABP1/TAC3 neurons, mouse cholinergic and pallidal neurons (VMF_CRABP1/LHX8) also expressed Zic1 and Lhx8 (Fig. 3a), hinting that a combination of transcriptional programmes used by neighbouring initial classes may define the novel TAC3 population. Differential expression and regulon analysis showed that the earliest molecular programmes that distinguish TAC3 INs involve distinct neuropeptides, acetylcholine receptors and immediate early gene networks (Extended Data Fig. 6 and Supplementary Table 6), suggesting that TAC3 neurons may receive signals from nearby cholinergic neurons. Notably, the primate-specific TAC3 population emerged as a distinct class as cells became postmitotic by PCD65. This occurred far earlier in development than the conserved PTHLH+, PVALB+ and TH+ terminal fates that ultimately arise from the related MGE_CRABP1/MAF class2,22. Lastly, we found that MGE_CRABP1 classes emerged in vitro as rare populations in human pluripotent stem cell-derived telencephalon organoids (Extended Data Fig. 6).

Fig. 3: Emergence of primate-specific MGE_CRABP1/TAC3 striatal INs.
figure 3

a, Dot plot of expression of striatal IN marker genes. b, Schematic summarizing properties distinguishing new-born MGE CRABP1+TAC3+ and CRABP1+MAF+ neurons, from markers given in Supplementary Table 6. c, Line plots showing the Rostro–Caudal distribution of classes of CRABP1+ cells. TAC3 (MAF) denotes that the MAF class is inferred by the lack of TAC3, as distinct positive markers for this class are not apparent until later in differentiation. Each point is the sum of all cells in at least five random fields of view in each section/region. Cells were counted from whole-section scans of RNAscope in situ hybridization on representative sections (full size shown in Extended Data Fig. 8). The solid (dotted) outlines of the GE region in the images represents MGE (LGE). One individual was used with four pairs of tandem sections interspersed with four single sections. d, Representative image of MGE_CRABP1/MAF (blue arrows), MGE_CRABP1/TAC3 (pink arrows), MGE_CRABP1/TAC3/LHX8 (yellow arrows) and VMF_CRABP1/LHX8 INs (LHX8+TAC3 cells) in the putamen from section 66 (Extended Data Fig. 8c). Scale bar 50 μm.

Reuse of OB neurons in primate cerebrum

We next analysed the initial classes of neurons detected within and probably derived from the LGE6. Two classes, LGE_FOXP2/TSHZ1 and LGE_MEIS2/PAX6, showed unexpected enrichment in the cortical frontal lobe in addition to the ventral telencephalon (Fig. 4a, b and Extended Data Fig. 9). LGE_MEIS2/PAX6 neurons express ETV1, SP8, MEIS2, SALL3, TSHZ1 and PAX6 during differentiation, all of which are markers of and are required for proper production of OB granule cells and dopaminergic TH+ periglomerular cells (PGCs)23,24. Indeed, the transcriptomes of cells in this class showed strong correlations to mouse adult-born granule cells (OB-GC_MEIS2/PAX6; Extended Data Fig. 4c). Similarly, trajectory analysis linked the mouse LGE_FOXP2/TSHZ1 class to OB-PGC_FOXP2/CALB1 PCGs of the OB, connecting each LGE class to distinct olfactory populations (Fig. 2b).

Fig. 4: Redistribution of LGE_MEIS2/PAX6 granule cells.
figure 4

a, Approximate ganglionic eminence transcription factor territories in PCD80 macaque brain, showing estimated section planes. b, Dot plot of expression of markers of LGE- and CGE-derived classes showing overlap of transcription factor domains. The expression values are scaled from 0–1 for each gene. The dot size represents the percentage of cells that express each gene. Stacked bar plots show the regional distribution of each class. NAc, nucleus accumbens. c–f, Immunohistochemistry for LGE class markers in macaque and human. c, Coronal–axial section of the A-dLGE and PFC at PCD80 with the inset highlighting DLX2+SP8+FOXP2 parenchymal chains. d, Arc–ACC SCGN+MEIS2+ cells (teal arrows) shown migrating within the boundary of dense TH+ axons at PCD120. e, Whole PCD120 coronal section from d showing the Arc–ACC and RMS wrapping around the striatum from the A-dLGE. Box marked ED13 corresponds to Extended Data Fig. 13f. f, Low-magnification sagittal image of a human postconception week 33 (PCW33) specimen showing large streams of MEIS2+SP8+ neurons originating from the A-dLGE; these neurons contribute to the RMS and the Arc. LV, left ventricle; Str, striatum. g, Schematic of a macaque brain. Multiple streams extend from the anterior pole of the dLGE, or the SVZ at later stages, to the RMS leading to the OB, the Arc extending dorsomedially and the Arc–ACC subsidiary stream extending to the ACC.

We performed immunofluorescence microscopy to visualize the spatial distribution of the LGE_FOXP2/TSHZ1 and LGE_MEIS2/PAX6 classes, using combinations of MEIS2 together with FOXP2/FOXP4 and SCGN/SP8/PAX6, respectively. Both populations appeared to emanate from the dorsal LGE (dLGE) but showed complementary distributions. LGE_FOXP2/TSHZ1 cells immunoreactive for FOXP2, FOXP4 and SCGN were found mainly in the dorsolateral dLGE (DL-dLGE; Extended Data Figs. 10 and 11) but were not detected in the anterior dLGE (A-dLGE) or rostralmigratory stream (RMS). Instead, cells of this class migrated directly into the striatum, via the lateral migratory stream (LMS) to the outer OB and ventromedially to cortical superficial white matter (Extended Data Figs. 1012). Consistent with trajectory analysis, markers of dLGE origin (ETV1, SCGN and SP8) were downregulated, whereas FOXP2, CASZ1, OPRM1 and projection neuron markers were upregulated, as the cells differentiated and migrated into the striatum (Extended Data Figs. 9 and 10). The expression of TSHZ1, LYPD1, PCDH8 and CASZ1, the absence of expression for the canonical medium spiny projection neuron markers NPY and FOXP1, and the results of RNA velocity analysis all imply that the LGE_FOXP2/TSHZ1 initial class also explains the previously unknown developmental origin of recently described striatal projection neurons in adult mice, eccentric spiny projection neurons (eSPNs)21 and amygdala ITCs (Fig. 2 and Extended Data Fig. 9). This linkage is consistent with reports that cells in mouse dLGE initially express SP8 and maintain TSHZ1 expression as they migrate via the LMS to become amygdala ITCs19,25. This developmental perspective suggests that these cells are not eccentric deviations from canonical spiny projection neuron development; instead, the LGE_FOXP2/TSHZ1 class converges on a similar striatal and amygdaloid projection neuron transcription profile despite its distinct origin.

By contrast, we observed MEIS2+PAX6+SP8+SCGN+ cells representing the LGE_MEIS2/PAX6 class continuously from the anterior end of the dLGE along the RMS to the OB granule cell layer (Fig. 4c and Extended Data Figs. 11 and 12). Notably, we observed dense parenchymal chains26 of these cells radiating from the dLGE at PCD80 (n = 3 hemispheres; Fig. 4c and Extended Data Fig. 12h, i). At PCD120, we found large numbers of LGE_MEIS2/PAX6 precursors that express SCGN extending dorsomedially and caudally in the Arc migratory stream27 in addition to the RMS (Extended Data Fig. 13). These cells were densest in chains running along the entire striatum in the primary tier of the Arc with fewer cells radially27. Unexpectedly, we also observed a robust stream diverted from the Arc that stretched from the A-dLGE into the anterior cingulate cortex (ACC; Fig. 4d, e). This stream, referred to as the Arc–ACC, appeared to be bounded by TH+ fibres in superficial white matter (Fig. 4d). Cells from the Arc and Arc–ACC were common in dorsomedial cortex deep white matter but were rarely found lateral or ventral to thestriatum; however, many CGE-derived MEIS2SP8+NR2F2+ neurons were observed throughout the white matter (Extended Data Fig. 13), highlighting regional heterogeneity in the composition of white matter INs.

We confirmed that LGE_MEIS2/PAX6 neurons from the A-dLGE also contribute to the RMS, Arc and Arc–ACC in perinatal humans (Fig. 4f) and postnatal macaques (Extended Data Fig. 13). We further found that these neurons persist postnatally in the deep white matter of the cingulate cortices and the superior corona radiata (Extended Data Fig. 13l, m). By contrast, in postnatal day 2 (P2) mice, we identified only rare instances of LGE_MEIS2/PAX6 cells in deep white matter (Extended Data Fig. 14a–c), consistent with recent reports that sparse MEIS2+HTR3A+ neurons in the white matter integrate into cortical circuitry perinatally28. Instead, the vast majority of these cells appeared in the anterior SVZ and RMS in mice (Extended Data Fig. 14). Overall, we found that neurons derived from the dLGE are more widely distributed than previously recognized in primates, representing a major source of neurons in the primate Arc migratory streams and persisting in the deep white matter.

Our analysis identified a third presumed dLGE-derived class in and around the striatum, insula and claustrum, which we refer to as striatum laureatum neurons (SLNs or Str-SLN_TH/SCGN). Likely derived from the LGE_MEIS2/PAX6 initial class, SLNs are named for the wreath shape they form around the striatum. At both PCD120 and 7 months postnatally, SLNs were immunoreactive for PAX6, MEIS2, SP8, TH and SCGN but not for FOXP2, NKX2-1 or NR2F2, which is also characteristic of TH+ PCGs (OB-PGC_TH/SCGN) of the OB (Fig. 5a–d and Extended Data Figs. 13f and 15). This distribution matches observations of TH+ cells circumscribing the primate striatum and their reported absence in rodents and illuminates their molecular identity and origin29. Indeed, we did not identify MEIS2+PAX6+SCGN+TH+ cells along the mouse striatum border or the claustrum (Extended Data Fig. 14e–g). Instead, these cells in mice were restricted to the OB, olfactory tract or olfactory tubercle, matching the macaque olfactory peduncle domain (Fig. 5e, f). We found that SLNs form a reticule at the white matter boundaries of the caudate and putamen of macaques, exist in humans and persist throughout life (Fig. 5a–h).

Fig. 5: Primate TH+ SLNs and ancestral olfactory populations.
figure 5

All immunofluorescence images in this figure are of MEIS2, SCGN and TH. a, Seven-month-postnatal macaque coronal section including the remnant RMS. b, TH+ SLNs at the border of the striatum. MEIS2+SCGN+TH+ peristriatal SLNs are indicated by yellow arrows. c, SLNs in the claustrum with long, straight processes among dense TH+ midbrain–cortical fibre synapses including one TH+ process (orange arrowheads) and one TH process (white arrowheads). d, Anterior olfactory nucleus at the olfactory peduncle, with MEIS2+SCGN+TH cells (blue arrows) including SCGN+TH fibres entering the ventral cortex (inset) and triple-positive cells (yellow arrows). e, Coronal section of mouse P2 OB showing SLN-analogous TH+SCGN+ cells. f, Mouse P2 coronal section showing the olfactory tubercle (OT) and striatum (Str) outlined by dotted lines, with only MEIS2+TH+ or MEIS2+SCGN+ cells. g, Photograph of a brain coronal slab from an 88-year-old human. The box shows the approximate location of the inset section block. h, TH+SCGN+ human SLN, with double-positive processes highlighted with white arrowheads. Only two cells were observed across this section. i, Schematic summarizing the unequal scaling of cortical and olfactory structures. j, Schematic summarizing migration of dLGE neurons and unequal evolutionary scaling of their destinations. Relative values in the bar plots are arbitrary.


By identifying transcriptional regulatory programmes distinguishing the earliest specification of initial classes, our study provides a resource for identifying conserved molecular mechanisms that specify cell type diversity. This resource can support rational in vitro derivation of these populations from pluripotent stem cells and interpretation of the cellular substrates of genetic disorders of neural development. TAC3-expressing striatal INs represent an exceptional case in which an evolutionarily novel initial class of neurons emerges in differentiating progenitors. A limited number of gene networks distinguish the TAC3 initial class from the related MGE_CRABP1/MAF class, consistent with a recent model of cell type evolution. Under this model, an ancestral cell type is partitioned into distinct subtypes by changes in transcription factor expression that enable genomic individuation of sister cell classes that still share many regulatory complexes and developmental trajectories30. However, the conservation of nearly all other initial classes of INs between macaques and mice suggests that evolutionary diversification of primate INs arises mainly by radiation of conserved initial classes of new-born neurons and may be shaped by the expanded diversity of primate regional destinations7.

The neurons of the dLGE appeared to be particularly affected by primate brain reorganization. In both macaques and mice, the LGE_MEIS2/PAX6 class is among the latest-born INs and migrates to olfactory structures and deep white matter. However, the absolute migration distance of late-born A-dLGE neurons to the OB is more than two orders of magnitude longer in new-born macaques than in mice and increases further as the brain expands after birth27,31. Similarly, the volume of white matter is more than three and five orders of magnitude larger in macaques and humans than in mice, respectively32, whereas the relative size of the primate OB is markedly smaller (Fig. 5i, j)33. Thus, in mice, the birthplace is only several cell lengths from any point in the adjacent deep white matter. In macaques, however, these homologous cells traverse histologically distinct dorsal migratory streams and apparently reuse the chain migration strategy. OB granule cells derived from this class contribute to adult plasticity34, and myelination is delayed for up to two decades in human frontal lobe white matter35, potentially linking these cells to white matter plasticity. Notably, abnormal accumulations of frontal lobe white matter neurons have been reproducibly associated with schizophrenia and autism36. With their prolonged migration to far-flung and ever-changing destinations, the A-dLGE neurons we identified here may be particularly vulnerable to environmental influences, and the markers we identified will be useful for assessing the molecular heterogeneity of disease-associated populations.

Finally, we identified SLNs, another likely OB sister type, which are redistributed to peristriatal regions and show a molecular resemblance to dopaminergic OB TH+ PGCs. Future studies can examine whether this primate striatal population partly explains the human-specific increase in TH-expressing striatal neurons37,38 and whether these neurons produce dopamine themselves or have an auxiliary role to compensate for increased demands on midbrain dopaminergic neurons39,40. Crick and Koch41 speculated that, in the claustrum, hitherto undiscovered sparse INs resembling intraglomerular OB cells with dendrodendritic synapses could contribute to binding information. Molecular access to SLNs will enable future circuit-level studies of this rare claustrum population. Together, our results highlight contrasting models for diversification of primate INs by specification of an entirely novel initial class and by redistribution of conserved initial classes that supply the OB into primate white matter migratory streams and peristriatal locations.



The Primate Center at the University of California, Davis, provided nine specimens of cortical tissue from PCD40, PCD50, PCD65 (n = 3), PCD80 (n = 2), PCD90 and PCD100 macaques. All animal procedures conformed to the requirements of the Animal Welfare Act, and protocols were approved before implementation by the Institutional Animal Care and Use Committee at the University of California, Davis. PCD40 represents embryonic Carnegie stage 20 and marks the approximate beginning of neurogenesis of both excitatory neurons and INs, whereas PCD100 is the approximate end of excitatory neurogenesis in the cortex42. Macaque data from PCD110 (n = 2) were taken from ref. 9. In addition, we used public mouse datasets, including for embryonic day 13.5 (E13.5) and E14.5 ganglionic eminences, which were enriched for DLX6+ cells10, and three samples from the 10x Genomics E18 mouse cortex example dataset with 1.3 million cells (GSE93421; samples 1, 3 and 4). We also used mouse public datasets for E14, including neonatal cortex, subcortex11 and whole-brain developmental structures as well as adult structures43,44; P9 striatum45; and adult OB12. In total, we analysed single-cell transcriptomes from 109,112 cells from developing macaque, 76,828 cells from developing mouse and 141,065 total mouse cells. De-identified human tissue samples were collected with previous patient consent in strict observance of legal and institutional ethical regulations in accordance with the Declaration of Helsinki. Protocols were approved by the Human Gamete, Embryo, and Stem Cell Research Committee and the Committee on Human Research (institutional review board) at the University of California, San Francisco.

Single-cell RNA sequencing tissue processing

For the PCD40 to PCD100 macaques, dissections were performed in PBS under a stereo dissection microscope (Olympus SZ61). A number of regions were difficult to distinguish at earlier time points because key anatomical landmarks were still forming. Accordingly, presumptive regions were dissected such as motor versus somatosensory cortex before the appearance of the central sulcus or the anterior ends of the MGE and LGE. For single-cell dissociation, samples were cut into small pieces and incubated with a prewarmed solution of papain (Worthington Biochemical Corporation) prepared according to the manufacturer’s instructions for 10 min at 37 °C. After 30–60 min of incubation, samples were gently triturated with glass pipette tips, and the PCD100 macaque samples were further spun through an ovomucoid gradient to remove debris. Cells were then pelleted at 300g and resuspended in PBS supplemented with 0.1% BSA (Sigma). Samples for MULTIseq were prepared in strip tubes and were maintained at 4 °C for the labelling protocol, as described in McGinnis et al.46. Single-cell RNA sequencing was completed using the 10x Genomics Chromium controller and version 2 or 3 3-prime RNA capture kits. Most samples were loaded at approximately 10,000 cells per well; up to 25,000 cells were loaded per lane for multiplexed samples. Transcriptome library preparation was completed using the associated 10x Genomics RNA library preparation kit. Multiseq barcode library preparation was completed as described in McGinnis et al.46. Following library preparation, libraries were sequenced on Illumina HiSeq and NovaSeq platforms.

Alignments and gene models

Fastq files were generated from Illumina BCL files using bcl2fastq2. Genes were quantified using Kallisto release 0.46 (ref. 47) and the RheMac10 genome assembly, newly annotated using the comparative annotation toolkit48, as well as the transcript annotations of Mus musculus ENSEMBL release 100. A custom Kallisto reference for each species was created for the quantification of exons and introns together, in which introns were defined as the complement of exonic and intergenic space. The Kallisto index used k-mers of length 31. Public data were downloaded as raw fastq files or as BAM files that were converted back to fastq files. All data were processed from raw reads using the same Kallisto pipeline to minimize annotation and alignment artefacts.

Quality control

Kallisto–Bus output matrix files (including both introns and exons) were input to Cellbender (release 0.2.0;, which was used to remove probable ambient RNA only. Only droplets with a greater than 0.99 probability of being cells (not ambient RNA), as calculated by the Cellbender model, were included in further analysis. Droplets with fewer than 800 genes detected, or greater than 40% ribosomal or 15% mitochondrial reads, were filtered from the dataset. Doublets were then detected and removed from the dataset using Scrublet (release 0.2.2; using threshold parameter 0.5).

Clustering and determination of homologous cell types

Much of the analysis pipeline was based on scanpy infrastructure and AnnData data structures49. Counts in cells were normalized by read depth, log transformed and then scaled for each gene across all cells. Principal-component analysis was then performed using the top 12,000 most variable genes by applying the original Seurat variable gene selection method implemented in the scanpy package, with the 100 most variance-encompassing principal components used for the following steps. Batch correction was limited to the requirement that highly variable genes be variable in more than one sequencing sample and by application of batch-balanced k-nearest neighbours (BBKNN)50 using the Euclidean distance of principal components to find 3 neighbours per batch in the developmental data and 12 neighbours per dataset in the developmental and adult merged mouse data. Leiden clustering using BBKNN-derived k-nearest-neighbour graphs was then applied according to the KNN graph with the scanpy resolution parameter set to 10 (or 7 in the developmental mouse dataset). Glia, along with excitatory progenitor and neuron clusters, were removed from the dataset in non-ganglionic eminence batches if their expression value was below the mean for two or more of the following genes: GAD1, GAD2, DLX1, DLX2, DLX5 and DLX6. Cajal–Retzius cells (RMTW_ZIC1/RELN) met this threshold and served as a useful outgroup. These cells were considered to be derived from the rostromedial telencephalic wall (RMTW) on the basis of ZIC1 and RELN expression even though they are known to have multiple origins51. After non-INs were removed, scaling, principal-component analysis and the following steps were repeated with this final IN dataset.

High-resolution Leiden clusters partitioned continuous differentiation trajectories of postmitotic initial classes into subclusters based on maturation stage. These high-resolution clusters were then manually merged to initial classes by using hierarchical clustering of cluster gene expression averages and distinctness of individual Leiden cluster markers as a guide. The nomenclature for merged clusters incorporates the presumptive spatial origin of the initial classes and specific marker genes. Spatial origin for each class was inferred according to the expression of canonical marker genes for the RMTW, MGE, LGE, CGE and VMF, such as LHX5, NKX2.1, MEIS2, NR2F2 and ZIC1 and was supported by immunostaining and the enrichment of these genes in cells from region-specific dissections. For merged species analysis, genes were normalized and scaled within species and were then merged for downstream analysis using BBKNN with 25 neighbours across and within species, the mutual nearest neighbours of which were used for Sankey plot comparison of developing macaques and mice. After clustering, the mean expression in each class was calculated for each gene that was among the original 12,000 most variable one-to-one orthologues from each dataset showing variability in both species (6,227 genes). These classes were then compared across species by Pearson correlation of their gene expression vectors.

Trajectory analysis of activating and inactivating macaque genes

We applied scVelo’s dynamical model (release 0.2.3)52 to derive a shared latent time based on RNA velocity using spliced and unspliced counts from Kallisto. Next, we used the related CellRank (release 1.3.1) package15 to derive absorption probabilities for immature cells in the transition cluster to likely initial classes. This step was necessary because new-born neurons, like children, are more similar to each other than to their mature state. By using adjacency along the paths of differentiation, it is possible to infer which mature state is likely to absorb a given immature cell. We then classified the new-born neurons as cells below the 0.5 quantile of latent time for that class. Recent studies have indicated that these transcriptionally immature neurons correspond to new-born neurons as labelled by classical nucleoside-based methods53. To identify genes that were activated or inactivated along trajectories, we used linear regression implemented in SciPy based on latent time values (x) versus gene expression values (y). This yielded linear regression coefficients and two-tailed P values for each gene, which were corrected for multiple-hypothesis testing using the Holm–Sidak method implemented in the statsmodels (release 0.12.2) package to derive q values. The gene sets were compared by calculating the Jaccard indices of set intersections, defined as the number of intersecting elements between two sets divided by the number of elements in the union of the two sets.

Linking developmental and adult data

Similarly to the reassignment of macaque transition cells, we also used CellRank-derived absorption probabilities, with equally weighted KNN and RNA velocity kernels, to estimate the precursor states of adult cells. Because absorption probabilities are biased by cell numbers in terminal states, and the goal this time was not to assign each developmental cell to a terminal state, we subsampled a maximum of 1,000 cells per class, with the rarest class having 707 cells from MGE_CRABP1/MAF, and we report the class identity of the 100 developing cells with the highest probability of being absorbed into each terminal class. This enabled us to provide an estimate of which developing class was the likely origin of the terminal classes, which is reflected in the weights of the edges in the Sankey diagram in Fig. 2. We also calculated the mean absorption probability for cells in each initial class to each terminal state to alleviate compositional effects, which we present as a heatmap. Notably, RMTW_ZIC1/RELN and VMF_TMEM163/OTP were not included because they are excitatory cortical and hypothalamic classes, respectively.

Immunohistochemistry tissue processing and imaging

Mouse, macaque and human tissues for histology were fixed in 4% paraformaldehyde in PBS overnight at 4 °C with constant agitation. The paraformaldehyde was then replaced with fresh PBS (pH 7.4) and samples were cryopreserved by incubation for 24–48 h in 30% sucrose diluted in PBS (pH 7.4) before being embedded in a mixture of OCT (Tissue-Tek, VWR) and 30% sucrose. Tissue was then frozen at –80 °C and was cryosectioned at 16–20 μm. For RNAscope RNA in situ hybridization, fixed cryosections were stained according to the protocol for the Advanced Cell Diagnostics RNAscope Multiplex Fluorescent Reagent Kit V2 Assay (ACD, 323120). For immunostaining, antigen retrieval was performed by placing tissue slides in 95 °C citrate buffer and then allowing them to cool at room temperature. Antibodies were diluted in blocking buffer (0.1% Triton X-100, 5% donkey serum and 0.2% gelatine in PBS). Sections were incubated with primary antibodies overnight at room temperature under bright light to photobleach autofluorescence in a light box54. The primary antibodies and dilutions used are recorded in Supplementary Table 7.

Alexa dye-conjugated donkey secondary antibodies were incubated in the dark at room temperature for 1 h. All tiled scans were acquired using an Evos M7000 microscope. All images were stitched using a custom Python script and ImageJ’s max correlation grid/collection stitching (release 1.2). They were then processed using ImageJ (release 1.53c) Rolling Ball background subtraction and manual brightness/contrast adjustment within an ImageJ macro. Image quantification of CRABP1+ cells was conducted using a custom ImageJ macro with the CRABP1+ area automatically thresholded using maximum entropy. Positivity for other gene products was classified manually for every cell in at least five random areas in the striatum or MGE and was defined as >1 puncta within CRABP1+ areas not clearly belonging to another cell.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this paper.