Mouse experimental model and details of other animal experiments
All animal procedures in this study followed experimental protocols approved by the respective institutions.
Mouse procedures in this study followed experimental protocols approved by the Institutional Animal Care and Use Committee (IACUC) at the University of California, San Francisco (UCSF). The Nkx2-1-Cre;Ai14 mouse line is the result of crossing the C57BL/6J-Tg(Nkx2-1-cre)2Sand/J (Nkx2-1-Cre; Jackson Laboratory stock no. 008661) and B6.Cg-Gt(ROSA)26Sortm14(CAG-tdTomato)Hze/J (Ai14; Jackson Laboratory stock no. 007914) strains. Mouse housing and husbandry were performed in accordance with the standards of the Laboratory Animal Resource Center (LARC) at UCSF. Mice were group housed in a 12-h light–dark cycle, with access to food and water ad libitum. Nkx2-1Cre;Ai14 mice were crossed, and the date of a positive vaginal plug was considered as E0. Pregnant dams were euthanized, and the brains of embryos from their litters were extracted and assessed for tdTomato fluorescence at the following developmental stages: E15 (n = 16 embryos from 2 litters), E17 (n = 7 embryos from one litter) and E18 (n = 5 embryos from 2 litters).
Rat procedures in this study followed experimental protocols approved by the IACUC at UCSF. Rat housing and husbandry were performed in accordance with the standards of the LARC at UCSF. Rats were group housed in a 12-h light–dark cycle, with access to food and water ad libitum. Long Evans outbred rats were purchased from Charles River (strain code 006). After rat crossing, the day of positive vaginal plug was considered E0. Pregnant dams were euthanized, the brains of embryos from their litters were extracted at E18 and brains from two litters were pooled for sequencing.
Pig samples (E62, E73 and one year and two months of age) were collected at the Swine Teaching and Research Center at the University of California, Davis. Pig procedures were performed in accordance with the standards of the Animal Welfare Act and under conditions approved by the Association of Assessment and Accreditation of Laboratory Animal Care (AAALAC), and followed experimental protocols approved by the IACUC at the University of California, Davis.
Opossum samples (n = 2, P20) come from a pedigreed, breeding colony of grey short-tailed opossums (Monodelphis domestica) that was established by founder individuals purchased from the Southwest Foundation for Biomedical Research, and is maintained by the laboratory of K.E.S. at the University of California, Los Angeles (UCLA). Opossums were euthanized by CO2 inhalation, followed by decapitation. These procedures are in accordance with the AVMA Guidelines for the Euthanasia of Animals 2013, and all animal procedures were approved by the IACUC at UCLA.
Sugar glider (n = 2, P40) (Petaurus breviceps) experiments were performed with the approval of the IACUC committee at Princeton University. Captive-born, adult sugar gliders were purchased through the US pet trade and subsequently housed in a breeding colony at Princeton University. Sugar gliders were fed a diet of dried food, fruits and protein daily, and housed in breeding pairs or trios. The colony was kept under a 12-h light–dark cycle (temperature, 20–27 °C; humidity, 30–70%). Adult female sugar gliders were checked for pouch young by manual palpation of the maternal pouch and visual inspection. Pouch young identified during inspection were collected by briefly anaesthetizing the mother with isoflurane and gently everting the pouch to expose the neonate. Joeys were gently detached from the nipple, euthanized and processed in the laboratory. More details about the sugar glider colony and husbandry can be found a previous study35.
Ferret samples (P1, P14, P22 and 33 months of age) were a gift from the University of Iowa National Ferret Resource and Research Center and were collected under experimental protocols approved by the IACUC at the University of Iowa. New Zealand white rabbit embryos (n = 3, PCD23) were obtained from BioIVT (USDA NY-TEC-0004) and procedures were performed in accordance with the standards of the Animal Welfare Act and under the AAALAC.
Harbour porpoise (Phocoena phocoena) brain samples were collected from stranded individuals by the Marine Mammal Center and obtained under the National Marine Fisheries Service (NMFS) permit no. 23922.
Dissections and dissociation
Dissections of embryonic and postnatal brain samples were performed in Hibernate-E (Gibco A1247601) under a stereo dissection microscope (Olympus SZ61 or Leica MZ10). The entire dissection process was done on ice, using Hibernate-E culture medium. The LGE, the striatum and the mantle zone of the MGE were manually dissected and collected as a striatal sample. The motor and somatosensory cortex was collected independently from ventral regions and collected as a cortical sample. The sex of the embryos was not determined, and thus the results reported are assumed to include animals of both sexes.
For the E73 pig sample, the brain was embedded in low-melting-point agarose and vibratome-sectioned at 300 μm in artificial cerebrospinal fluid (125 mM NaCl, 2.5 mM KCl, 1 mM MgCl2, 1 mM CaCl2, and 1.25 mM NaH2PO4) before dissection under stereo dissection microscope.
For single-cell dissociation of samples, dissected regions were cut into small pieces and incubated with a prewarmed solution of papain (Worthington Biochemical) prepared according to the manufacturer’s instructions. After 30–60 min of incubation, samples were gently triturated with wide-orifice pipette tips. Once the samples were dissociated to a single-cell suspension, DMEM + 0.1% bovine serum albumin (BSA) was added to quench papain and cells were pelleted at 300g for 5 min at 4 °C. Finally, cells were resuspended in phosphate-buffered saline (PBS) supplemented with 0.04%–0.1% BSA. Mouse samples were then sorted for tdTomato expression on a BD FACSAria Fusion.
Single-cell sequencing of developing species
With the exception of sugar gliders, scRNA-seq of all developing species was completed using the 10x Genomics Chromium X controller and v.3.1 high-throughput RNA capture kits. Sugar glider sequencing was completed using the 10x Genomics Chromium Next GEM Single Cell 3’ Reagent Kit v.3.1 (10x Genomics, CG000315). Samples were loaded at approximately 30,000–100,000 cells per well (10,000 cells per well for sugar gliders) and library preparation was completed following the manufacturer’s instructions. Libraries were sequenced on Illumina HiSeq and NovaSeq X platforms. Sequencing was performed at the UCSF Center for Advanced Technology, supported by UCSF PBBR, RRP IMIA and NIH 1S10OD028511-01 grants.
Alignments and gene models
Illumina BCL files were converted to Fastq files using bcl2fastq2. Genes were quantified with the CellRanger v.7.0.1 count function, using the mRatBN7.2 genome for rats, mOrycun1.1 for rabbits, mMonDom1.pri genome for opossums, ASM1176430v1.1 genome for ferrets, SusScrofa11.1 for pigs and mm10-2020A for mice. For sugar gliders, we used an early draft assembly of the ‘SugarGlider_lib1_testN_pseudohap’ genome36 (an updated but unannotated version of this is now available: https://www.dnazoo.org/assemblies/petaurus_breviceps). Annotations were lifted over from Trichosurus vulpecula (GCF_011100635.1) using Liftoff (https://github.com/agshumate/Liftoff; -a 05 -s 05 -d 5). To improve the recovery of missing transcripts, we used the ReferenceEnhancer37 package to update the SusScrofa11.1 and mMonDom1.pri gene annotations by manually examining read alignment to inhibitory interneuron marker-gene boundaries to improve 3′ end annotation and incorporating intergenic read pile-ups. For sugar gliders, we observed an excessive number of genes with read pile-ups just downstream of the 3′ UTR. Consequently, we applied a uniform 3-kb extension downstream of all genes, but extensions were truncated wherever they would overlap other genes.
Quality control
CellRanger output was input into CellBender v0.2.138 to remove ambient RNA and the CellBender filtered output matrix was used for downstream processing. Additional filtering was performed on individual sequencing lanes according to the distribution of read counts. Droplets with low gene counts and high gene counts (greater than 6,000 genes) were removed from the datasets. Droplets with high mitochondrial and ribosomal reads were removed on the basis of species-specific distributions. Doublets were removed using Scrublet (v.0.2.3)39.
Clustering and assignment of cell types
Analysis of developmental data was based on the Scanpy package and tutorial40 and to be consistent with a previous report6. Counts were normalized and log-transformed. The data were then scaled for each gene and PCA was performed. Batch correction was applied to species with more than one age collected using Harmony integration41. Leiden clustering was applied with resolution set to 1. Inhibitory neurons were then isolated as previously described6. The data were reset to raw, and normalization, log transformation, scaling, PCA and Leiden clustering at a resolution of 1 or 0.5 were repeated on the isolated inhibitory neurons. Remaining clusters containing excitatory neurons and/or doublets were removed when applicable and clustering was recalculated. Individual Leiden clusters were subdivided by increasing resolution for selected clusters or merged to initial classes. Initial classes and nomenclature were assigned on the basis of markers in previous work6.
Cross-species integration of developmental data
Individual species were randomly downsampled to include a maximum of 1,000 cells per initial class from original clustering to account for rare populations and unequal sampling across species. Mouse and rat gene names were converted to human orthologues. Subsetted data were reset to raw and datasets were combined. Genes expressed in all species and CRABP1 (a total of 10,709 genes) were retained for downstream analyses. Counts were normalized, log-transformed and scaled. PCA was performed and Harmony41 was used to integrate the data across species on the basis of batch and species. Leiden clustering was applied with resolution set to 1. Individual Leiden clusters were subdivided by increasing resolution for individual clusters or merged to initial classes. Initial classes and nomenclature were assigned on the basis of markers in previous work6. The Pearson correlation coefficient of integrated clusters and original clusters across species was calculated based on the basis of the average PCA of the initial class in each species. The Jaccard index was calculated to determine the similarity between the original and the integrated class labels for each cell in the integrated dataset.
Differential expression of cortical and striatal TAC3 populations
Differential expression analysis was performed according to the decoupleR tutorial42. Pseudobulking was first performed using decoupleR based on batch and class. Differential expression of MGE_CRABP1 initial classes was performed using the pyDEseq2 package43. Contrast was set between cortical and striatal regions.
Pseudotime calculation
For pseudotime and trajectory inference, we converted anndata objects to R object with Zellkonverter and used the R package Slingshot44. Scanpy-based UMAP dimensional reduction and interneuron classes were used as input for all datasets. Lineage reconstruction was performed with the slingshot() function, with the ‘Progenitor’ class being designated as the beginning of pseudotime with start.clus = “Progenitor”. All initial classes were set as end points with the end.clus parameter. Averaged pseudotime values per cell were calculated with slingAvgPseudotime(). Differential expression analysis across pseudotime between species was calculated using TradeSeq45.
Analysis of published dataset of developing macaques
Macaque data were taken from a previous study46. We first isolated the LHX6_CRABP1 population, normalized, log-transformed, scaled, batch-corrected and performed Leiden clustering. Clusters were labelled on the basis of gene expression of known markers. Owing to the presence of MGE_CRABP1 populations in the A1C and insula dissections, we next isolated cells from A1C and insula dissections. We again normalized, log-transformed, scaled, batch-corrected and performed Leiden clustering. We examined marker-gene expression and renamed dissections A1C and insula to include striatum owing to the presence of medium spiny neurons in these dissections.
Analysis of published human dataset
Human data were taken from a previous study47. We normalized, log-transformed and scaled raw counts to examine gene expression in pre-labelled clusters.
Individual analyses of published datasets of adult marmosets, humans and mice
Adult mouse striatal interneurons were curated from the previously published ABC atlas27 class ‘08 CNU-MGE GABA’ (n = 18,849 cells), which we subsetted further to include only striatal cells (anatomical_division_label = “STR”), comprising 4,785 cells derived from a single ABC dataset (WMB-10Xv3-STR). This dataset initially contained cells from ABC subclasses ‘054 STR Prox1 Lhx6 Gaba’, ‘055 STR Lhx8 Gaba’, ‘056 Sst Chodl Gaba’, ‘057 NDB-SI-MA-STRv Lhx8 Gaba’ and ‘058 PAL-STR Gaba-Chol’, but after further examination, we found that most of the cells from ABC subclass 057 originated from outside the striatum. Excluding subclass 057 resulted in a final dataset of 3,586 cells.
Adult human striatal interneurons were curated from the neuron dataset of a previously published28 atlas by first selecting for the ROI terms ‘Basal Nuclei (BN) – Body of the Caudate – CaB’, ‘Basal Nuclei (BN) – Putamen – Pu’ and ‘Basal Nuclei (BN) – Nucleus Accumbens – NAC’, totalling 88,044 cells. These cells were further subsetted to include only cells from the supercluster terms ‘CGE interneuron’ (n = 711), ‘MGE interneuron’ (n = 238) or ‘Splatter’ (n = 6,716). The remaining 7,665 cells were re-embedded (UMAP from 30 PCs) using 2,000 variable features (Seurat ‘vst’) and reclustered using Louvain (at arbitrary resolution 0.8) into 22 clusters. These clusters were used to further exclude putative medium spiny neurons on the basis of MEIS2 expression, producing a final dataset of 6,617 striatal interneurons.
Adult marmoset striatal interneurons were curated from the previously published24 Marmoset Census dataset of 6,249 striatal GABAergic neurons (‘striatum.GAD’). Again, we excluded MEIS2-expressing cells (in this case using marmoset census cluster 02). Using the same method as for the adult human data, the remaining 3,930 cells were re-embedded and clustered into 22 clusters, from which we identified 6 small clusters (totalling 551 cells) that appeared to be contamination from excitatory (SLC17A6– and SLC17A-expressing) neurons (possibly from doublets or rare co-transmitting neurons); we excluded these, leaving 3,379 cells in the final dataset.
After cell selection in each species, the data were reprocessed using a typical Seurat (v.5.0.3) pipeline: NormalizeData (natural log of [1 + reads per 10,000 cells]), FindVariableFeatures (2,000 features using Seurat ‘vst’), ScaleData/PCA, FindNeighbors (using 30 PCs), FindClusters (Louvain resolution 0.8) and RunUMAP (using 30 PCs). In all cases, we found our re-embeddings were consistent or coherent with the previously published cell groupings and labels.
Integration analysis of published adult datasets
For cross-species integrations, we used scVI (ref. 48; scvi-tools v1.2.0). We used the same set of striatal inhibitory interneurons from ref. 24 (marmoset) and ref. 28 (human) as selected above. However, after a preliminary version of the analyses described below, we dropped a mixed cluster of 98 marmoset cells and a cluster of 575 putatively hypothalamic or mixed human cells, because these never showed cross-species homologies and are unlikely to be striatal, leaving a final set of 3,281 marmoset and 6,042 human cells. We used a similar set of cells from ref. 27 (mouse) to that detailed above, but for the sake of comparative analysis, we also expanded the set of mouse cells to include any cell found in the ABC class ‘08 CNU-MGE GABA’ subclass ‘055 STR Lhx8 Gaba’, which included cells from hypothalamus, pallidum and striatum-like amygdala nuclei (sAMY) dissections. We also included in the starting set any GABAergic or cholinergic cells that were derived from striatal dissections, which included cells that the ABC atlas had annotated as being non-striatal, although we excluded cells from class ‘09 CNU-LGE GABA’ and ‘05 OB-IMN GABA’. These additional inclusions produced a set of 3,770 mouse cells (or 184 more cells than the original mouse set). However, to allow proper modelling of donor effects in our scVI model, we also removed any cells from donors that contributed fewer than 30 cells, which left 3,553 cells (and removed nearly all cells from sAMY dissections).
During revision of this manuscript, a new large and high-quality dataset of marmoset cells became publicly available as an unannotated pre-release of the HMBA. This early pre-release contained 282,806 high-quality curated cells, which we included in our initial set as well. To find homologous cells in this dataset specifically, we performed an initial scVI integration with all of the starting sets. For all scVI integrations, we used scanpy (v.1.10.3) to find highly variable genes (flavor = “seurat_v3”, n_top_genes = 4,000, batch_key = “experiment”, subset = True). We used an scVI model to integrate across ‘experiment’ (so ignoring species commonality across the two marmoset datasets) with a covariate for the individual donor marmosets to model donor effects, and we trained with n_hidden = 256, n_latent = 64, n_layers = 3 and dispersion = “gene_batch” for 250 epochs. To annotate HMBA cells that did not integrate with the homologous types, we found k = 10 nearest neighbours on the scVI latent space, and clustered using scanpy leiden (flavor = “igraph”, resolution = 1, n_iterations = 2), and we calculated the ‘experiment’ entropy for each cluster (scipy.stats.entropy; scipy v.1.11.2). After manually examining the consequences of using different cut-offs, we removed any HMBA cells that were present in clusters with entropy < 0.1 (which were clusters that generally failed to incorporate non-HMBA cells). This pre-filtering removed 267,945 cells (94.7% of the total), leaving 27,737 cells.
After pre-filtering HMBA cells (‘first pruning’), we found new variable genes, redid scVI integration and redid clustering with the same parameters. We used these clusters to identify, select and remove MEIS2-expressing cells from the expanded mouse data and HMBA data, removing a further 3,296 (mostly HMBA) cells from the data. After this ‘second pruning’, we repeated variable gene selection, integration and Leiden clustering with the same parameters. These Leiden clusters for this final integration were used to annotate the HMBA data on the basis of the co-clustering cells from the other datasets, and the UMAP from this integration is used in the integration figures.
During editing of this manuscript, an expanded and newly annotated version of the marmoset HMBA data became publicly available, with annotations from a draft cross-species Basal Ganglia taxonomy (https://alleninstitute.github.io/HMBA_BasalGanglia_Consensus_Taxonomy). We verified that the striatal cells were included in the initial release that we used in our analyses, and we found a reassuring correspondence between our clustering/annotations and the taxonomy at the ‘Group’ level (the finest annotation present). We added those labels in Extended Data Fig. 10b. We applied the taxonomy Group labels to our initial dataset to check the extent to which our integrated cluster entropy-based pre-filtering had selected cells that the taxonomy annotated as striatal inhibitory interneurons, finding that non-targeted groups were efficiently removed and that the fraction kept was between 98.5% and 100% for all desired groups except for STR SST-ADARB2 GABA and STR SST-RSPO2 GABA. We also attempted an integration that included those excluded groups, but those groups still failed to integrate with cells from the other datasets.
Pearson correlation of gene expression in published adult datasets
For correlation analysis, we used the same curated cell sets that were used for the individual dataset analysis and annotation, but we dropped several clusters for clarity. As in the integration analysis, marmoset ‘Mixed’ and human ‘Hypothalamus’ and ‘Mixed’ clusters were dropped, and mouse Th-type cells (almost entirely ABC subclass 055) were split into their ABC cluster identities. The small number of cells from subclass ‘054 STR Prox1 Lhx6 Gaba’ that we annotated as Th were dropped for clarity. Pearson correlations were calculated on the metacells (cell-averaged normalized expression by species and cluster) using all 1:1 homologues, and so correlations are among log-transformed data.
SAMap analysis of published adult datasets
For SAMap, we followed the authors’ tutorial with only minor variations. We used protein-coding transcript (CDS) sequences matching the original reference assemblies (GENCODE 35 for human, GENCODE M21 for mouse), although we used only the more complete marmoset reference that was used for the HMBA data (RefSeq GCF_011100555.1, mCalJa1.2.pat.X) for either marmoset dataset. For all six pairwise directional comparisons between the three species, we ran tblastx (NCBI BLAST v.2.9.0). For each gene pair in the tblastx results, we chose the transcript pair with the highest mapping score, and we used that for the gene pair mapping score (the ‘maps’ for SAMap). We ran SAMap (v.1.0.15; sam-algorithm v.1.0.2), specifying our own clusters and using pairwise = True, and we calculated scores using samap.analysis.get_mapping_scores with n_top = 0 (to use all cells in each cluster). We used the same datasets and clusterings as were used for the Pearson correlation (with the same additional clusters being dropped). For the version with marmoset HMBA data, we used the data subset and annotations we generated after scVI integration. For aesthetic reasons, we set the self-self SAMap scores to 1 (which are 0 by default).
Gene-module discovery and projection using Hotspot
Hotspot (hotspotsc v1.1.1; https://github.com/YosefLab/Hotspot) was used for unsupervised partitioning of genes into modules following the authors’ recommendations29. For each dataset, only genes present with non-zero counts were used, and Hotspot was run using the model ‘danb’ on the PCA embedding. Genes were selected from Hotspot autocorrelations with a false discovery rate less than 0.05 and partitioned using agglomerative clustering, as previously shown by the authors using the default ‘min_gene_threshold = 30’.
To address computational constraints on hotspot module discovery in the larger developing macaque (109,111 cells) and developing mouse (71,023 cells) datasets, we subsetted 50% of the cells. To make the subsets representative, we sampled within high-resolution clusters: for macaque, we used the 124 ‘hires_leiden’ clusters already present in the data; for mouse, we reclustered (Leiden at resolution = 10) to obtain 157 distinct clusters used only for this purpose. All other analyses (including module scores/projections) use the full datasets.
Gene-module scores were calculated using PC1 following PCA using only the genes within the module, min–max scaled from 0 to 1. For cross-species module projections, we found the unsupervised gene module most correlated (Spearman) with a given marker gene of interest. We then found the gene orthologues in other species (subsetting genes if necessary) and calculated module scores.
Immunohistochemistry tissue processing
Samples were fixed in 4% paraformaldehyde (PFA) in PBS overnight at 4 °C with constant agitation. The PFA was then replaced with fresh PBS (pH 7.4) and samples were cryopreserved by incubation for 24 h in 10% sucrose diluted in PBS (pH 7.4), followed by a 24-h incubation in 20% sucrose, and finally a 24-h incubation in 30% sucrose before being embedded in OCT (Tissue-Tek, VWR). Tissue was then frozen at –80 °C, cryosectioned at 20 μm and stored at −80 °C until use. Slides from P22 ferret were a gift from A. Kriegstein. Sections from P2 naked mole rat were a gift from S. Villeda.
Adult Nkx2-1-Cre;Ai14 mice were transcardially perfused with PBS followed by 4% PFA in PBS; their brains were dissected out and post-fixed overnight at 4 °C in 4% PFA in PBS with constant agitation. Fifty-micrometre sections were obtained on a vibratome, and preserved in freezing buffer (30% ethylene glycol, 28% sucrose in 0.1 M sodium phosphate buffer) at −20 °C until use.
Rhesus macaque sections from previous studies were provided by the Primate Center at the University of California, Davis and prepared for histology as stated above. All animal procedures conformed to the requirements of the Animal Welfare Act, and protocols were approved before implementation by the IACUC at the University of California, Davis.
RNAscope was performed following the manufacturer’s instructions for the Advanced Cell Diagnostics RNAscope Multiplex Fluorescent Reagent Kit V2 Assay (ACD, 323120). For immunostaining of the Nkx2-1-Cre;Ai14 line, rabbit anti-RFP (Rockland 600-401-379) was diluted 1:1,000 in blocking buffer (PBS + 5% BSA + 0.3% Triton X-100) and incubated for one hour at 37 °C. Alexa-dye-conjugated goat secondary 594 antibody was diluted 1:500 in blocking buffer and incubated for 30 min at 37 °C.
After RNA in situ hybridization, TrueBlack lipofuscin autofluorescence quencher (Biotium, 23007) was applied to adult ferret, adult pig and seven-month-old macaque samples according to the manufacturer’s directions. Slides were mounted using Prolong Gold Antifade plus DAPI (Invitrogen, P36931).
Imaging
Entire tissue sections were imaged single plane in widefield mode on a Leica DMI8 inverted microscope (Leica Microsystems) connected to a Flash4 V3 camera (Hamamatsu), with a 5× (0.12 NA), 10× (0.4 NA) or 20× (0.8) Plan Apochromat objective. Stitching was performed automatically in LAS X. Regions of interest were imaged at higher resolution using a laser scanning confocal microscope (Stellaris, Leica) using a 100× oil (1.4 NA) Plan Apochromat objective with a pixel size of 116.25 µm × 116.25 µm. Z-stacks of the entire volume of cells were acquired using optimal sections. Images from all samples were acquired under the same imaging settings with minimal adjustment of illumination intensity.
Image processing
The 100× images are maximum intensity projections produced from volumes acquired on the confocal microscope. Brightness and contrast were manually adjusted in Fiji for publication. Image processing of scans was performed using Imaris v.10.1.0 (Bitplane). Files were converted using the Imaris v.10 Converter and imported into Surpass mode for interaction. Images were saved as 3,500-dpi PNG files for publication.
Statistics and reproducibility
All RNAscope and immunofluorescence experiments were independently reproduced at least twice on non-consecutive sections. The images are representative of the results.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
