Cortical arealization of interneurons defines shared and distinct molecular programs in developing human and macaque brains

Abstract

Cortical interneurons generated from ganglionic eminence via a long-distance journey of tangential migration display evident cellular and molecular differences across brain regions, which seeds the heterogeneous cortical circuitry in primates. However, whether such regional specifications in interneurons are intrinsically encoded or gained through interactions with the local milieu remains elusive. Here, we recruit 685,692 interneurons from cerebral cortex and subcortex including ganglionic eminence within the developing human and macaque species. Our integrative and comparative analyses reveal that less transcriptomic alteration is accompanied by interneuron migration within the ganglionic eminence subdivisions, in contrast to the dramatic changes observed in cortical tangential migration, which mostly characterize the transcriptomic specification for different destinations and for species divergence. Moreover, the in-depth survey of temporal regulation illustrates species differences in the developmental dynamics of cell types, e.g., the employment of CRH in primate interneurons during late-fetal stage distinguishes from their postnatal emergence in mice, and our entropy quantifications manifest the interneuron diversities gradually increase along the developmental ages in human and macaque cerebral cortices. Overall, our analyses depict the spatiotemporal features appended to cortical interneurons, providing a new proxy for understanding the relationship between cellular diversity and functional progression.

How cortical interneurons gain regional specification remains elusive. Here authors recruit 685,692 interneurons from the developing telencephalons of humans and macaques to characterize the shared and distinct molecular programs regulating the cortical arealization of primate interneurons.

Introduction

Cortical interneurons (InNs) are primarily derived from the ganglionic eminences (GEs) via the tangential migration process and play a crucial role in the development of the central nervous system^1–6, particularly in the regulation of corticogenesis, the establishment of cortical circuitry, and the formation of higher-order cognition. Accumulating evidence suggests the dysfunction of cortical interneurons is likely to contribute to neurodevelopmental and neuropsychiatric disorders, such as autism spectrum disorder, epilepsy, schizophrenia, and major depressive disorder^7–10. A comprehensive understanding of cortical interneurons involves the exploration of their morphology, electrophysiology, and transcriptome, leading to the definition of multimodal cell types^11–14. Notably, the advent of single-cell and single-nucleus mRNA sequencing (scRNA-seq and snRNA-seq) technologies has significantly advanced the investigation of interneuron transcriptomic types (t-types). These types can mostly be mapped to interneurons with established morphological and electrophysiological properties, underlining the utility of t-type in defining diverse interneuron cell types^15–30.

The current understanding of cortical interneuron generation is that the medial, caudal, and lateral ganglionic eminences (MGE, CGE, and LGE), respectively, give rise to interneurons that exhibit morphological, molecular, and functional diversity^31–34. For instance, the MGE-derived interneurons (mgeInN) are originally characterized by high expression of NKX2-1. However, in the adult cerebral cortex (CTX), they predominantly express either calcium-binding protein parvalbumin (PV, encoded by PVALB) or neuropeptide somatostatin (SST or SOM), which are considered as two major terminal classes of mgeInNs^35–39. The CGE mainly generates the vasoactive intestinal peptide (VIP) and LAMP5 PAX6 [synonymous with 5-hydroxytryptamine receptor 3A (Htr3a)-expressing, but lacking VIP] cortical interneurons (cgeInN), with specific expression of NR2F1/2, PROX1, and SP8^40–42. Also, the recent reports claimed the dorsal LGE-derived interneurons (lgeInN) emerged in the cortical frontal lobe and persisted in the deep white matter, expressing MEIS2, FOXP2 and PAX6^18,43. Beyond these intrinsically encoded diversities, the interaction of cells traveling from their origins with the environment of distant destinations triggers transcriptomic changes accordingly, leading to an additional layer of heterogeneity¹⁷. However, there is limited understanding regarding the extent to which the transcriptomic alterations accumulate during long-distance migration and after settling in functionally distinct cortical destinations. Besides, the connection between transcriptional alterations of cortical interneurons and their implications for cell type refinement and brain functions, along with the subsequent exploration of cellular malfunctions linked to human brain diseases, has lagged behind. Addressing these knowledge gaps is critical for advancing our understanding of the intricate processes governing cortical interneuron development and their relevance to neurological disorders.

In this study, we curated publicly available scRNA-seq and snRNA-seq datasets from developing human and macaque telencephalon to characterize the interneuron maturation and regional specification along migration. Our inclusion of multiple anatomically and functionally distinct brain regions in both species helped identify human-specific characteristics during cell migration and arealization. The collected brains spanned from embryonic weeks to late-fetal stages, incorporating 21 time points to provide high-resolution depictions of the temporal regulation process. The integrated datasets are interactively accessible at http://www.braininfo.ac.cn/mammal_inn_migration/. After in-depth integrative and comparative analyses, we refined interneuron t-types and strikingly observed a small set of interneuron progenitors emerging in CTX in both human and macaque species. To delve deeper, we conducted multilevel analyses for three interneuron lineages separately, including the quantification of cell diversity, the identification of spatiotemporally regulated genes, and the exploration of shared and distinct features between human and macaque species. Taken together, our findings depict an enhanced picture for illustrating the molecular dynamics underlying cortical arealization of the interneurons in developing human and macaque brains, shedding light on species-specific characteristics and potentially uncovering key elements relevant to the formation of higher-order cognition in human species.

Results

Interneuron atlas of developing human and macaque brains

To dissect the cortical arealization of interneurons, we systematically collected public scRNA-seq and snRNA-seq datasets from human and macaque telencephalon (Supplementary Data 1)^{17,18,20,44–47}, including 13 cortical areas and 6 subcortical regions [4 areas from frontal lobe (FC), 4 areas from motor-somatosensory lobe (MSC), 3 areas from temporal lobe (TC), 1 area from occipital lobe (OC), insula, MGE, CGE, LGE, hippocampus (HIP), amygdala (AMY), and striatum (STR)] (Fig. 1a). Also, to gain insights into temporal regulation, we thoughtfully selected donors or specimens to ensure an even distribution across the prenatal developmental stages, resulting in the recruitment of 21 ages ranging from 7 gestational weeks (GWs) to 33 GWs, of which the comparative ages were calculated for macaque species using TranslatingTime based on the heterochrony of cortical neurogenesis (Fig. 1b)⁴⁸. In total, 76 subjects and 307 samples generated from 6 independent resources were categorized into our interneuron atlas, in which each sample was subjected to rigorous quality controls and cell type manual elaboration procedures to ensure only high-quality interneurons were included (Supplementary Fig. 1), leading to 282,339 and 403,353 interneurons recruited from human and macaque species, respectively.

Fig. 1. Overview of interneuron taxonomy in human and macaque embryonic and fetal brain development.

a Diagram of human and macaque brain samples from cerebral cortex and subcortex regions. Cortex was collected with more refined sub-areas including 4 FC regions (DFC, MFC, OFC and VFC), 4 MSC regions (S1C, M1C, PCC and IPC), 3 TC regions (STC, ITC and A1C) and 1 OC region (V1C). b Schematic illustration of cell numbers in 4 human transcriptome datasets and 2 macaque transcriptome datasets across broad developmental phases. c The inner- to outermost circles respectively depict the UMAP plot labeled by 5 prominent interneuron classes; subtype proportions; canonical lineage marker expressions; species, region and stage makeup in each subtype; the nomenclature of 5 classes and 50 interneuron subtypes. Clockwise from the top left, the surrounding four corner insets show the distribution of dissected regions across three age trimesters and marker gene expression using UMAP embedding. Each dot represents a cell. Five major cell classes include RGCs (radial glial cells, 1–9), inIPCs (interneuron intermediate progenitor cells, 10–15), mgeInNs (MGE-derived interneurons, 16–27), cgeInNs (CGE-derived interneurons, 28–42) and lgeInNs (LGE-derived interneurons, 43–50). The organization of 3 primary interneuron lineages depends on the presumed birthplace and the expression of lineage marker genes. DFC dorsolateral frontal cortex, OFC orbital frontal cortex, MFC medial frontal cortex, VFC ventrolateral frontal cortex, S1C primary somatosensory cortex, M1C primary motor cortex, PCC posterior cingulate cortex, IPC inferior parietal cortex, STC superior temporal cortex, ITC inferior temporal cortex, A1C primary auditory cortex, V1C primary visual cortex, GE ganglionic eminence, MGE medial ganglionic eminence, LGE lateral ganglionic eminence, CGE caudal ganglionic eminence, FC frontal cortex, MSC motor-somatosensory cortex, TC temporal cortex, OC occipital cortex, NCX neocortex, HIP hippocampus, AMY amygdala, STR striatum, E embryonic, GW gestational week.

We introduced a top-down hierarchical strategy to first group the cells into radial glial cell (RGC) expressing VIM, HOPX and HMGA2, interneuron intermediate progenitor cell (inIPC) expressing MKI67 and ASCL1, mgeInN (LHX6 and NKX2-1), cgeInN (NR2F2, PROX1 and ADARB2) and lgeInN (MEIS2) classes based on the outcome of unsupervised clustering and the subsequent enrichment test for cell type canonical gene markers (Fig. 1c, Supplementary Fig. 1g). Uniform manifold approximation and projection (UMAP) dimensionality projection manifested a clear segmentation of the five main classes, of which the RGC and inIPC groups denoting the early stages assembled densely on the left side (Fig. 1c), in contrast to the remaining three classes that stood on the right were believed to be the relatively late stages of interneuron fate, resembling the developmental dynamics of interneuron-genesis. This accordance was also observed in the distributions of natural ages and harvested regions, with the second trimester positioned between the first and third trimesters, and the cortical interneurons were congregated on the right side of UMAP, in contrast to the GE cells on the left (Fig. 1c and Supplementary Fig. 1d). No obvious separation was found between human and macaque species on UMAP (Supplementary Fig. 1c–e), suggesting the evolutionarily shared molecular programs in the primate genus.

Beyond the top interneuron classes, we eventually defined 41 subtypes, with distinguishable transcriptomic identities confirmed by a list of literature reported gene markers (Fig. 1c; Supplementary Data 2). To further ensure the definition of interneuron subtypes more rational in biology rather than mathematical, we conducted parameterized simulations by clustree and pairwise correlation analysis to assure the commitment to a novel subtype was computationally and statistically reproducible (Supplementary Fig. 2)⁴⁹. In agreement with prior studies^17,18, 83.0% (34 of 41) subtypes were reported before, whereas the remaining 7 subtypes remarkably expanded previous definitions, owing to the improved statistical power by means of our most comprehensive collection. Furthermore, the comparative analysis of each subtype between human and macaque species reasserted distinguishable correspondences and substantial similarities, suggesting the evolutionary conservation and shared molecular program governing interneuron fate (Supplementary Fig. 3). Lastly, the comparison to human and macaque adults showed concordance in the prototypical transcriptomic signature and maturation progression for mgeInNs and cgeInNs, as evidenced by the loss of expression of DLX family genes and the increased expression of GAD1/2, SST, PVALB, VIP, and ADARB2 in the adult (Supplementary Data 2).

Six previously published mouse datasets^23,50–54, covering critical areas such as the CTX and GEs, spanning crucial developmental stages from early embryonic periods (E13.5 to E18.5) through early postnatal phases (P0, P1, P4), were supplemented to enrich the interneuron atlas of developing human and macaque brains (Supplementary Data 1, Supplementary Fig. 4c). A similar hierarchical strategy was employed for the annotation of interneurons in mice, yielding 5 subtypes of RGCs, 5 subtypes of inIPCs, 8 subtypes of mgeInNs, 3 subtypes of cgeInNs, and 8 subtypes of lgeInNs. UMAP plots showed that interneurons exhibited a consistent intermingling within species and across species, with a clear division into five distinct classes, aligning with the cell type annotations (Fig. 1c; Supplementary Fig. 4a, b, d; Supplementary Fig. 5a). Hierarchical clustering of these subtypes revealed that most subtypes were composed of cells both from GE subdivisions and CTX (Supplementary Fig. 4e). Cross-validation using cellHint⁵⁵ and MetaNeighbor⁵⁶ confirmed the robustness and repeatability of cross-dataset cell types (Supplementary Fig. 4f–i). The subtype classification from the original datasets and our annotated subtype data essentially displayed a one-to-one or one-to-many relationship. This demonstrated that the cell classification in our study was highly reproducible across different datasets, both at a broad level and with finer resolution.

Nearest-neighbor analysis revealed that cells of the same identity grouped together across all three species, however, with a greater number of consensus subtypes observed in humans and macaques (Supplementary Fig. 5b). Interestingly, a relatively higher level of discrepancy was observed in cgeInNs, particularly in mice, though visible correspondence between different subtypes was still evident. The cross-species comparison highlighted a substantial population of primate LHX6 NKX2-1 CRABP1 TAC3 cells mostly abundant in striatum that had previously been reported^18,57. Mice had a single ancestral type of LHX6 NKX2-1 CRABP1 TAC3- that exhibited strong homology to the LHX6 NKX2-1 CRABP1 TAC3/TAC3- in primates. The expression pattern was compared across the three principal interneuron classes, revealing a set of class-specific gene markers shared across species, whereas many gene markers were found to exhibit species-specific enrichment (Supplementary Fig. 5c, d). Besides, the spatial expression of the conserved class markers such as Lhx6, Nkx2-1, Sst, Meis2 and Ebf1 readily segregated the MGE and the LGE regions of E14.5 mouse (Supplementary Fig. 5f). The mgeInNs were highly conserved across species (Supplementary Fig. 5e). Several human-enriched and primate-enriched genes were identified (Supplementary Fig. 5g, h). Specifically, SCGN (secretagogin) particularly stood out by a pattern of high expression levels in the cgeInNs of humans, coinciding with very few Scgn+ cells in the mouse CGE²⁰.

A small set of interneuron progenitors emerged in cerebral cortex

Similar as neural cells in the cerebral cortex, the RGCs in the subpallium undergo a process of proliferation, transitioning to inIPCs before committing to the interneuron fate. We sought to conduct an in-depth survey of inIPCs to attain a comprehensive understanding of their intrinsic features and their role in regulating cortical development. Of the brain-wide distribution, small proportions of inIPCs were strikingly found in CTX (3855, 4.0%), STR (5752, 5.9%) and AMY (1747, 1.8%), even though the majority were identified in the GE (85,282, 88.1%) (Fig. 2a, Supplementary Fig. 6h). The presence of cortical inIPCs was not occasional, because there were promising numbers of inIPCs (203 ± 47, i.e., Mean ± SEM) consistently found in different developmental ages for multiple cortical areas of human and macaque species (Fig. 2b, Supplementary Fig. 6g, h). Besides, we also detected triple-positive of ASCL1, DLX2 and TOP2A cells at the second trimester, third trimester and infancy from the publicly available human MERFISH data for prefrontal cortex (PFC) and primary visual cortex (V1C)⁵⁸. The majority were primarily present in the second trimester (92.9%) of the InNs (54.4%, Supplementary Fig. 6a, b). The presence of cortical inIPCs was also substantiated with our immunostaining (Supplementary Fig. 6d). Further exploring the inIPCs from GE and CTX, we found they assembled closely on UMAP, whereas the more meticulous comparison convinced that the ganglionic inIPCs stood in-between RGCs and cortical inIPCs, implying a plausible origin of these cortical inIPCs in the GEs, followed by subsequent migration to the CTX (Fig. 2a). In recent studies^59–63, the cortical ventricular zone was thought to be an alternative source to produce cortical interneurons, some of which were transcriptionally and morphologically similar to cortical interneurons derived from GE. Some cortical-derived inIPCs were collected from previous scRNA-seq studies, including the STICR-labeled (the STICR lentiviral vector allowed for clonal lineage tracing) cortical germinal zone cells in the primary human cell vitro cultures or following transplantation into the cortex of early postnatal immunodeficient mice⁶⁰ and cells from human pluripotent stem cell (PSC)-derived regionalized dorsal forebrain organoids⁶¹, and then integrated with inIPCs in this study using Harmony⁶⁴. The integrative analyses of our annotated inIPCs in CTX with the in vitro and in vivo lineage-tracing datasets and dorsal cortical organoids all illustrated that the cortical inIPCs born from CTX were largely distinct from those cortical inIPCs, which might be due to cortical inIPCs being predominantly post-mitotic and quiescent cells, especially in macaques, while cortical-derived inIPCs were mostly proliferative, marked by MKI67 (Fig. 2c, d). Human and macaque ganglionic inIPCs assembled closely with cortical-derived inIPCs on UMAP, since dividing cells with different origins might cluster together by cell cycle phases rather than their identities. Yet, GE RGCs exhibited distinct genetic program with dorsal RGCs (Supplementary Fig. 6e). Regional identities of different inIPC cell types in GEs were not evident until they exited cell cycle and committed towards interneuron lineages (Supplementary Fig. 6f). A relatively large number of post-mitotic inIPCs in CTX hinted a ventral origin via a long-distance journey. However, we could not rule out the possibility that cortical RGs have the capacity to generate InNs during embryonic development.

Fig. 2. Transcriptomic variation of interneuron progenitors in GE and CTX.

a Visualization of human and macaque inIPCs in GE (left) and CTX (right) using UMAP. b Histogram showing the number of ganglionic and cortical inIPCs between humans and macaques across the entire age scheme. c Mapped human and macaque inIPCs from CTX integrated with human cortical-derived DLX2+ inIPCs from in vitro lineage-tracing dataset (left), xenografted human cortical-derived DLX2+ cells (middle), and inIPC-like cells collected at weeks 8, 10 and 15 from the control group of dorsal cortical organoids (right). The lower row shows the expression of MKI67 in the corresponding dataset of the upper row. d Bar plots indicating the cell quantity (top) and proportion (bottom) of MKI67+ and MKI67- cells in four datasets, including human and macaque inIPCs from CTX (this study), human cortical-derived DLX2+ inIPCs from in vitro lineage-tracing dataset (I), xenografted human cortical-derived DLX2+ cells (II), and inIPC-like cells collected at weeks 8, 10 and 15 from the control group of dorsal cortical organoids (III). e Volcano plot showing differential expressions of genes for inIPCs from GE and CTX in humans and macaques. P-values were determined by two-sided Wilcoxon Rank Sum test. f Enrichment of GO terms for the GE-enriched and CTX-enriched genes for inIPCs. Significance was determined by hypergeometric test with False Discovery Rate correction (adjusted P value < 10⁻⁵). g Heatmap illustrating differentially expressed genes with high specificity for cortical lobe, species and subtype. Selected genes with synapse function are labeled blue and TFs are labeled red. GE ganglionic eminence, CTX cerebral cortex, GW gestational week, FC frontal cortex, MSC motor-somatosensory cortex, TC temporal cortex, OC occipital cortex, inIPC interneuron intermediate progenitor cell.

Next, we performed differential expression analysis in-between cortical and ganglionic inIPCs without separating different ages in order to portray the global picture of their transcriptomic alterations, but confining to ASCL1 DLX1 subtype. Considering the imbalanced numbers of cells in GE and CTX, an equal number of cells were sampled and 100 times of such samplings were carried out to detect reliable and reproducible genes, resulting in over a thousand genes (CTX, 2260 in humans and 180 in macaques; GE, 178 in humans and 1388 in macaques) exhibiting statistically significant enrichment in either CTX or GE (Supplementary Data 3). For instance, ATRNL1, NRXN3, SYT1 and NGRN were significantly enriched in CTX, whereas cell proliferation-related genes, e.g., TOP2A, MKI67, CENPF and UBE2C, were significantly enriched in GE of human and macaque species (Fig. 2e). The subsequent systematic GO (gene ontology) analysis unveiled the CTX-enriched genes were strikingly correlated with telencephalon development and synapse formation, whereas GE-enriched genes were remarkably associated with cell mitosis and proliferation (Fig. 2f, Supplementary Data 4)⁶⁵. Together, our results might indicate that the cortical inIPCs have adjusted their molecular program during and after the cortical migration.

To elucidate the regional diversification of cortical inIPCs, we performed comprehensive comparisons across the cortical lobes for each subtype in both human and macaque species (Supplementary Data 5). In particular, we focused on genes showing distinguishable enrichment in either lobe or species because they were always top-ranked and more representative (Fig. 2g). These included 18 transcription factors (TFs) and overlapped prominently with synapse genes^66–68. Hence, we hypothesized that the TFs play a crucial role in regulating regional diversification of cortical inIPCs, and likely the subsequent formation of synapses and circuits during interneuron generation and maturation.

Diversification of MGE-derived interneurons

To depict the brain-wide map of mgeInNs, we analyzed the cell relative proportions in all our collected brain regions, including MGE, CGE, LGE, multiple neocortical regions, HIP, AMY, and STR. Surprisingly, the mgeInNs labeled by canonical markers LHX6, SST, MAF were also detected in CGE and LGE, accounting for ~31.3% (30,120 cells) in humans and ~19.5 % (24,462 cells) in macaques (Fig. 3a, Supplementary Fig. 7a), as also indicated by the normally distributed Shannon entropies with a mean of about 1, where positive and greater values implied higher level of the richness metric for subtype distributional profiles (Supplementary Fig. 7b). Further supporting this, the re-analysis of external in-situ sequencing data of GW12 human brain revealed the consistent expression profiles (LHX6, SST, and MAF) in CGE and LGE²¹, as also evidenced by the in situ hybridization (LHX6) of embryonic day (E) 18.5 mice from Allen brain atlas (Supplementary Fig. 7e, g). Beyond that, we observed the early interneuron precursor marker NKX2-1 was expressing restrictedly in MGE, unlike the expressions of LHX6, SST, MAF and RELN, which were not confined to a single GE subdivision (Supplementary Fig. 7e, f). Thus, we believe the mgeInNs were generated in MGE, but their appearance in the CGE and LGE might be the result of short-distance movement within GEs, in agreement with the viewpoint of previous studies^21,40,69. Furthermore, previous studies have reported the lineage markers were not exclusively confined to their place of origin, although they appeared in lower abundance compared to their birthplace in the human fetal subpallium, consistent with our findings (Supplementary Fig. 7). This possibly indicates the highly migratory characteristics of interneurons within the GEs^21,40.

Fig. 3. Molecular mechanisms for the diversification of mgeInNs across cortical regions and species.

a The percentages of mgeInNs originating from different areas, including MGE, CGE, LGE, FC, MSC, TC, OC, insula, HIP, AMY and STR of the developing humans (left) and macaques (right). b Correlation of area expression enrichment of mgeInNs. c, d Venn diagrams showing shared or regional-specific genes across the three regions (CTX, CGE and LGE) compared with MGE (left). Regional-specific genes for mgeInNs were identified by differential expression analysis for humans (c) and macaques (d). The GSEA of GO terms ranked genes from high to low based on the values of log2 fold changes. The positive enrichment scores represent CTX enrichment and negative values denote GE enrichment. e Volcano plots showing differential expressions of species-specific genes identified in the MSC lobe for MAF lineage subtypes (top) and NKX2-1 lineage subtypes (bottom). NS, not significant; orange dots, adjusted P-value < 0.01 and log2 fold change > 1 for humans; blue dots, adjusted P-value < 0.01 and log2 fold change > 1 for macaques. P-values were determined by two-sided Wilcoxon Rank Sum test. f The size of the bar indicates the proportion of cell subtypes derived from MGE in CTX by developmental ages in humans (top) and macaques (middle). Unavailable data are shown in dark gray. Temporal changes in the abundance of cell types among cortical mgeInNs in humans and macaques are shown (bottom). g Radar plots visualizing the counts of specifically expressed genes in FC, MSC, TC and OC at human GW18 and GW20 (top) and macaque GW16, GW20&21, GW26 and GW33 (bottom). MGE medial ganglionic eminence, LGE lateral ganglionic eminence, CGE caudal ganglionic eminence, FC frontal cortex, MSC motor-somatosensory cortex, TC temporal cortex, OC occipital cortex, HIP hippocampus, AMY amygdala, STR striatum, E embryonic, GW gestational week.

Besides, we employed two computational approaches to quantify the cellular differences of mgeInNs harvested from different telencephalic regions. Firstly, we carried out pairwise transcriptomic correlation analyses among GEs, cortical lobes, insula, HIP, AMY and STR in human and macaque species (Fig. 3b). The intra-correlations among GEs and among cortical lobes were positively high, in contrast to their negative inter-correlations, as well as those negative correlations with HIP, AMY and STR. Of note, the obvious higher intra-correlations among macaque cortical lobes might suggest the increased divergence in cortical lobes gained in human species. As for the high correlations within GE subdivisions, it might be due to the short-distance movement from their birthplace and the similar ganglionic environment. On the other hand, we compared mgeInNs in MGE to those in cortical lobes, and to those in CGE and LGE. Particular attention was given to avoid overestimating the difference, as the comparison could be distorted by the different waves of interneuron generation. Only mgeInN subtypes with comparable numbers among the four regions (MGE, CTX, CGE and LGE) in the second trimester were retained and then equal number of cells from the four regions were sampled for 100 times for following differential expression analysis. Interestingly, we found more genes were uniquely up-regulated in CTX, accounting for 161 and 98 genes in humans and macaques respectively, denoting more divergences gained for cortical mgeInNs (Supplementary Data 6). Moreover, gene set enrichment analysis (GSEA) found that the CTX-enriched genes were more correlated with biological processes accompanied with synapse formation⁷⁰, but the GE-enriched genes were more associated with the elementary functions, i.e., cell differentiation (Fig. 3c, d). To extend this conclusion at spatial resolution, we have identified some genes that displayed robust frontal or caudal enrichment in mgeInNs from multiple cortical layers of PFC and V1C, based on the prior MERFISH data during the second trimester (Supplementary Fig. 8d)⁵⁸. Moreover, the spatial transcriptomic data from the coronal section of E14.5 mouse brain were generated using the 10X Genomics Visium V2 platform to address the absence of GE regions in the aforementioned dataset (Supplementary Fig. 8a). Robust cell type decomposition (RCTD)⁷¹ was applied to estimate the fractions of cell types contributing to each spot in the manually assigned region of the CTX, MGE, LGE and STR and unraveled that the MGE region was predominantly populated by mgeInNs, the LGE region by lgeInNs, and excitatory neurons were the predominant cell type in the cortex (Supplementary Fig. 8b, c). Besides, RGCs and inIPCs segregated between the ventricular zone and subventricular zone of GEs (Supplementary Fig. 8c)²⁰. Interestingly, area-specific genes that may be up-regulated in CTX including Nxph2 and Pdzrn4 were sparsely expressed in MGE. Additionally, area-specific genes like Gad2 showed equal and high-intensity levels across spots in MGE (Supplementary Fig. 8f–i). These findings could further elucidate that the cortical arealization genes of mgeInNs have transformed their expression patterns since leaving GEs and bore distinct molecular programs with exposure to the varying environmental circumstances. Taken together, these analyses indicate that different molecular programs have been activated for mgeInNs to adapt themselves to cortical or ganglionic migration.

Next, to further explore the evolutionary divergence, we chose mgeInN subtypes to carry out differential gene expression analysis. To further minimize the imbalance of cell collection, NKX2-1 and MAF were represented by combing their respective subclasses. 566 and 717 genes respectively passed our stringent statistic tests and displayed up-regulations in either human or macaque motor-somatosensory lobes (Fig. 3e, Supplementary Fig. 9a, b). Importantly, genes reported at the bulk level were consistently detected as differential expression between human and macaque species⁴⁷, e.g., COX7C and PTMA (Supplementary Fig. 9c), and owing to the enhanced capability of single-cell resolution we uncovered many additional differentially expressed genes (Supplementary Data 7).

To investigate the machinery of temporal regulation, we focused on cortical mgeInNs to portray their dynamic features along developmental times (Fig. 3f). We calculated the relative ratios for each cortical mgeInN subtype (i.e., a total of 12 mgeInN subtypes defined in Fig. 1c) in each developmental age in human and macaque species, and interestingly found the developmental dynamics of cell subtypes. For example, LHX6 NKX2-1 and LHX6 NKX2-1 LHX8 ZIC1 cell subtypes were not detectable after GW10 in either human or macaque species and E14.5 in mouse (Supplementary Fig. 4e), whereas many other mgeInN subtypes, e.g., SST+ long-range projection neuron (LHX6 SST NPY)⁷², became more pervasive in the mid-fetal stage, even though some were detected unstably due to inadequate depth of single-cell sequencing. Moreover, Shannon entropy was employed to quantitatively measure the diversity of mgeInN subtypes (Fig. 3f, Supplementary Data 8)⁷³. We found the entropies gradually increased, denoting the increase of mgeInN diversity along developmental times. Taken together, we could conclude that the temporal dynamics of mgeInN subtypes seemed to be temporally programmed, and the increased complexity of cortical mgeInN types might provide a proxy for neural circuit formation. These results align with prior studies^23,31, illustrating a fundamental principle that applies to both human and non-human primates.

Furthermore, to simultaneously dissect spatial and temporal molecular dynamics of mgeInNs, we chose the subtypes in MAF class with robust detection to conduct transcriptomic comparisons across different regions, ages and species. In human brains, we identified more regional transcriptomic variations in GW18 compared with those in GW20, particularly with an extensive reduction in the MSC (Fig. 3g, Supplementary Data 9). Yet, the regional variations were remarkably decreased in macaque brains, specifically in FC, when comparing GW16 (E64/65) to other later developmental times. Taken together, these developmental reductions are consistent with the hour-glass model we previously proposed to summarize the transition mode for inter- and intra-regional variations that shifted from a high level, transient repression to subsequent restoration of high level in the developmental human and non-human primate species during three stages: embryonic and early to mid-fetal, late-fetal, and after late childhood stages^19,47.

Temporal regulation of CGE-derived cells

Next, we carried out the aforementioned analysis workflow to depict the brain-wide map of cgeInNs. Overall, significant amounts of cgeInNs were found outside GEs, confirming their migrations into cortical lobes, insula, HIP, AMY and STR in human and macaque species (Fig. 4a). Within GEs, we found higher proportion of cgeInNs identified in humans compared with those in macaques (Fig. 4a), which might be due to the elder ages of the recruited macaques (Fig. 1b). In addition, we found the cgeInNs tended to be concentrated in their birthplace, accounting for ~58.6% (12,817 cells) in humans and ~45.8% (2903 cells) in macaques, which contrasted with the high proportions of ganglionic migration in mgeInNs (Supplementary Fig. 9d). We assumed the expressed canonical gene markers could represent the existence of cgeInNs, and consequently found that NR2F2, NR2F1, PROX1 expressed in MGE and LGE in addition to CGE (Supplementary Fig. 7c). Much more than that, the re-analysis of external in-situ sequencing data of GW12 human brain revealed the sparse distribution of NR2F2 in MGE and LGE (Supplementary Fig. 7e)²¹. Lastly, pairwise transcriptomic correlation analysis replicated the regional transcriptomic similarity patterns observed in mgeInNs (Fig. 3b), suggesting the same principle could be applied to the molecular programs governing the cortical and ganglionic migration of cgeInNs (Fig. 4b).

Fig. 4. Transcriptomic heterogeneity of cgeInNs.

a The percentages of cgeInNs originating from different areas, including MGE, CGE, LGE, FC, MSC, TC, OC, insula, HIP, AMY and STR of the developing humans (left) and macaques (right). b Correlation of area expression enrichment of cgeInNs. c Temporal changes in the abundance of cell types among cortical cgeInNs in humans and macaques are shown. d Comparison of CGE-derived subtypes distribution between the GE and CTX. The average percentage of a cell type occupying the FC, MSC, TC, OC and insula lobes was defined as the proportion in the CTX. Species are distinguished by the same colors in Fig. 4c. e CRH expression level in humans (left) and macaques (right) is visualized via UMAP. f CRH expression of cgeInNs in humans, chimpanzees, macaques, marmosets and mice at different developmental timepoints spanning from early embryo to adulthood. g Immunostaining for CRH in the frontal lobe slide of the mouse (E18.5) and human (GW23) brain. Immunofluorescence experiments in (g) were repeated three independent times, with similar results. Scale bar, 25 μm (middle); 50 μm (left and right). h Heatmap illustrating the gene expressions of CRH signaling pathway from prenatal period 2 to period 13 in bulk RNA-seq data. Genes in CRH signaling pathway were divided into G_pre group (n = 41) and G_post group (n = 50). i Individual cell score overlay for cortical expression activities of selected genes in G_post group across period 2 to period 13. We combined cgeInNs from the CTX of our integrated human datasets and a snRNA-seq dataset of human prefrontal cortex spanning from late gestation (GW22 and GW38) to adult stages. Twenty-seven overlapping G_post genes identified in bulk RNA-seq data (h) and scRNA-seq/snRNA-seq data (Supplementary Fig. 10f) were used. j CRH pathway activities scored per cell between mgeInNs, cgeInNs and lgeInNs. Enrichment scores were computed for mgeInNs (MGE-derived interneurons), cgeInNs (CGE-derived interneurons) and lgeInNs (LGE-derived interneurons) in human cortical areas. k Individual cell score overlay for activities of selected genes in G_pre group between CTX and GE region. Only cgeInNs for human species were included. Nineteen overlapping G_pre genes identified in bulk RNA-seq data (h) and scRNA-seq/snRNA-seq data (Supplementary Fig. 10f) were used. A two-sided Wilcoxon test was used in (j, k) to test the statistical significance (^***P < 2.22e − 16). The center line for each box plot in (j) represents the median, the box edges indicate the upper and lower quartiles, and the whisker represents 1.5-fold of the interquartile range. Data in (i, k) are presented as Means ± SDs. MGE medial ganglionic eminence, LGE lateral ganglionic eminence, CGE caudal ganglionic eminence, FC frontal cortex MSC motor-somatosensory cortex, TC temporal cortex, OC occipital cortex, HIP hippocampus, AMY amygdala, STR striatum, E embryonic, GW gestational week.

We then used Shannon entropy to quantify the diversity of cgeInNs of CTX over developmental times (Fig.4c, Supplementary Fig. 10a). We found their entropies steadily increased. In contrast to mgeInNs (Fig. 3f), the cgeInNs displayed lower levels of initial entropies but a higher magnitude of alteration, in addition to lower diversity of cgeInNs in CTX. These observations suggest a distinct propensity for environmental adaptation in cgeInNs and mgeInNs during and after long-distance journeys. The cgeInNs could be divided into two groups based on the expression levels of NR2F2 and PROX1 in both humans and macaques. In macaques, the NR2F2 group co-expressed DCN, while the PROX1 group co-expressed CRH. However, in human cgeInNs, we only defined a single PROX1 CRH VIP subtype, with the rest classified under the NR2F2 category, which showed high expression of SCGN and CXCR4 (Supplementary Fig. 10b, e). Then, we compared the percentages of each cgeInN subtype between GE and CTX, and found the PROX1 CRH, PROX1 CRH VIP and PROX1 CRH SP8 subtypes were more abundant in CTX (Fig. 4d), implying that these corticotropin-releasing hormone/factor (CRH or CRF) identifiers seemed to be specified after cell migration into CTX. Next, we labelled CRH expressed cells in UMAP (Fig. 4e), and noticed these cells assembled in a very limited domain, a distal margin of cgeInNs dominated area. In general, CRH is thought to mediate stress-induced behaviors, and mounting prior studies unveiled that CRH gene was highly expressed in parvocellular neurons of the paraventricular nucleus of the hypothalamus (PVN), but low levels were detected in other brain regions, including AMY, HIP, cerebral cortex, and olfactory bulb^74–76. Besides, it was reported that CRH neurons were undetectable in mouse cerebral cortex before birth⁷⁷, and CRH derived from placenta was secreted into maternal circulation during the third trimester of human pregnancy⁷⁸. Hence, we sought to extend these findings by investigating CRH expressions in cgeInNs of the developing and adult cerebral cortex, yielding developmental datasets from humans^{20,44–46,79}, macaques^17,18 and mice²⁸, and adult datasets from humans⁷⁹, chimpanzees, macaques, marmosets¹⁶ and mice^29,52,53 (Fig. 4f). In line with previous studies, cortical CRH+ cells were consistently detected in juvenile and adult humans, adult non-human primates, as well as juvenile and adult mice. Moreover, the cortical CRH+ cells were reproducibly detected, albeit with relatively lower abundance compared to postnatal stages, at the second and third trimesters in human and macaque species, e.g., human GW25/GW34 (1,042 cells, 13.3%) and macaque GW26/GW29/GW33 (natural age E93/E100/E110, 16,053 cells, 75.6%). To further validate the presence of CRH+ cells before birth, immunofluorescence was conducted at human GW23 dorsolateral frontal cortex (DFC, Fig. 4g). In mice, we discovered the presence of cortical CRH+ cells before birth, notably at E18.5 (equivalent to translated age GW18) with around 2.2% (17 cells), as confirmed by immunofluorescence in the mouse E18.5 frontal lobe (Fig. 4f, g). Besides, the PROX1 CRH VIP subtype was primarily observed in the CTX of postnatal mice (Supplementary Fig. 4e), and the spatially colocalized expression of CRH and PROX1 in prenatal E14.5 and E16.5 mice were universally absent, while a small number of CRH + PROX1+ spots occurred from birth onwards (E18.5, P0 and P4 coronal sections, Supplementary Fig. 10c, d). Although the counts were notably lower than the CRH+ cells detected in late-fetal human and macaque species, comparatively higher counts were observed at the first postnatal week (P1, P4, and P7), aligning with translated ages corresponding to human late-fetal stages. Thus, we believe that cortical CRH+ cells are more distributed than previously recognized, although their employment in human and non-human primate during late-fetal stage distinguishes from their postnatal emergences in mice. It is worth to note that the lack of support from the developing chimpanzee and marmoset, due to ethical restrictions and limited resources, could potentially be addressed in the future by exploring alternatively in vitro systems.

Furthermore, we recruited 91 genes involved in CRH signaling pathway (accession number: WP2355 in WikiPathways⁸⁰) and quantified their expression using BrainSpan dataset¹⁹, which was composed of multiple human cortical regions spanning from early embryo to adulthood. Hierarchical clustering analysis revealed their gene expression profiles could be categorized into two groups, specifically with 41 genes assigned to G_pre group showing high levels of expression during the prenatal periods (Period 2-7) while the remaining 50 genes constituting G_post group exhibiting high expression during the postnatal periods (Period 8-13) (Fig. 4h). Analyzing gene contents in G_pre and G_post further, we realized the CRH receptors (CRHR1 and CRHR2), and CRH binding protein (CRHBP) were concurrently included in G_post, while some inhibitors annotated in WP2355, such as IL2, KRT14 and RELA, were down-regulated in G_post (Fig. 4h, Supplementary Fig. 10e, f). We further explored the expression profiles of CRH pathway genes in cortical areas using our recruited scRNA-seq data and observed their higher expression activities in cgeInNs than those in mgeInNs and lgeInNs (Fig. 4j). Therefore, only cgeInNs were included in the subsequent analysis. Nineteen of the 33 genes in the G_pre group and 27 of the 44 genes in the G_post group recapitulated the temporal expression transition in the prenatal and postnatal periods (Supplementary Fig. 10f). To characterize the expression mode of G_pre and G_post genes in greater detail, we separately compared G_pre and G_post activities. Notably, G_pre scores were significantly higher in developing CTX compared with GE (Fig. 4k). Moreover, cortical expression activities of G_post genes increased with time and the transition points occurred in the 7^th period (Fig. 4i), which corresponded to human late-fetal stage from 24 to 38 PCW (post-conceptional weeks), agreeing with the identification of cortical CRH+ cells in human late-fetal stages (Fig. 4f). Together, our findings lead to the hypothesis that the initiation of CRH activation might start before birth in human species.

Cortical migration of LGE-derived cells

The lateral ganglionic eminence in rodents and primates is known to predominantly generate both striatal medium spiny neurons and olfactory bulb interneurons^20,34,81. However, the debate about whether LGE contributes to the neocortical population of interneurons continues. Notably, LGE has been shown to give rise to 30% of PV and SST cortical interneurons in the human fetal brain²¹. In developing macaques, a recent study reported interneurons probably derived from dorsal LGE showed remarkable enrichment in dorsomedial cortical frontal lobe, and their further investigation found these cells were more concentrated in deep white matter and might be contributed by the Arc and Arc-ACC migratory streams¹⁸. We sought to expand these findings through surveying our more comprehensive interneuron atlas-that was more developmental ages and anatomical regions in human and macaque species, so as to explore the temporal and evolutionary regulation principle. Consistent with prior study, our brain-wide analysis found most lgeInN subtypes distributed solely in GE and STR, except that MEIS2 PAX6 SCGN and MEIS2 FOXP2 TSHZ1 subtypes were distinctly detected in FC, MSC, STR and GE (Fig. 5a), in contrast to relatively few numbers of cells identified in OC, TC, insula, AMY and HIP. Mouse MEIS2 SP8 PAX6, the homologous subtype of the primate MEIS2 PAX6 SCGN, barely expressed SCGN and also accounted for the largest proportion of lgeInNs in the cortex (Supplementary Fig. 12b). We have performed co-immunostaining to assess the expression of GAD1, MEIS2, and PAX6 in sections of the human GW23 DFC and V1C, aiming to characterize the distribution of the MEIS2 PAX6 SCGN subtype. Notably, PAX6 was expressed in the majority of GAD1 + MEIS2+ cells, thereby substantiating the MEIS2 PAX6 SCGN subtype as the predominant lgeInN subtype within the cortex. Additionally, there was a significant increase in the number of GAD1 + MEIS2 + PAX6+ cells in DFC compared with V1C (Supplementary Fig. 11f, g).

Fig. 5. Transcription regulation of the development of lgeInNs in different cortical lobes.

a Alluvial plot illustrating the distribution pattern of 8 major types of lgeInNs from different areas. b Species composition of MEIS2 PAX6 SCGN and MEIS2 FOXP2 THSZ1 (representing the most abundant LGE-derived subtypes in the cortical region) in FC, MSC and GE across the developmental ages. c Developmental trajectory of MEIS2 PAX6 SCGN subtype cells in FC, MSC and GE of E93 macaque. Pseudotime for each cell (upper middle) and the composition of regions along the lineage progression (lower middle) are shown. Fitted curves along the pseudotime illustrate the divergent expression patterns of CCND2, MEIS2, and NR2F2 at the branch point. d Scatterplot of representative GO terms enriched in MSC fate. Bubble color indicates the q-value and the circle size indicates the frequency of the GO term in the underlying Gene Ontology Annotation database. e The ratio of cells that expressed genes of MEIS2 + /ETV1+ (G1) and NR2F1 + /NR2F2 + /PROX1+ (G2), in GE fate for GE region, FC fate and MSC fate for E93 FC and E93 MSC region. Data are presented as Means ± SDs; n = 5. Monocle analysis was repeated 5 times using downsampled MEIS2 PAX6 SCGN cells from GE, FC and MSC lobes, with individual replicates shown as dots ^***P < 0.001 for comparison of cells of GE fate, FC fate and MSC fate in G1 or G2 expression pattern. ^##P < 0.01 for comparison of G2 cells of MSC fate in FC region and MSC region. A two-tailed t test (two groups) and ANOVA (three groups) were used to test the statistical significance. f The number of MEIS2 PAX6 SCGN subtype in 4 regions of the frontal lobe across distinct samples. g The proportions of lgeInNs in DFC, MFC, OFC and VFC of humans (top) and macaques (bottom) were deconvolved from humans (top) and macaques (bottom) transcriptomic profile of bulk RNA-seq, respectively using prenatal macaque scRNA-Seq dataset (E110). Each value represents a bulk-tissue sample. n = 16 in DFC, n = 17 in MFC, n = 14 in OFC and n = 8 in VFC for human bulk samples. n = 3 in DFC, n = 12 in MFC, n = 4 in OFC and n = 3 in VFC for macaque bulk samples. The value in each bulk sample was normalized to a proportion relative to the maximum proportion of lgeInNs. The center line for each box plot in (g) represents the median, the box edges indicate the upper and lower quartiles, and the whisker represents 1.5-fold of the interquartile range. GE ganglionic eminence, FC frontal cortex, MSC motor-somatosensory cortex, TC temporal cortex, OC occipital cortex, HIP hippocampus, AMY amygdala, STR striatum, E embryonic, GW gestational week, DFC dorsolateral frontal cortex, OFC orbital frontal cortex, MFC medial frontal cortex, VFC ventrolateral frontal cortex.

To investigate cell emergence over developmental times (Fig. 5b), we noticed MEIS2 PAX6 SCGN cells and MEIS2 FOXP2 TSHZ1 cells were reliably detected after GW18 in FC and MSC, whereas they could be detected as early as GW13 in GE, implying that the necessary accumulation was required before cells could initiate cortical migration. Moreover, the matched result was observed through re-analyzing the recently published snRNA-seq dataset of human neocortices during early fetal (11/12 PCW), early mid-fetal (14/15 PCW) and late mid-fetal (17/18 PCW)⁸², confirming the presence of MEIS2 PAX6 SCGN cells in CTX during late mid-fetal stage (Supplementary Fig. 11a).

Next, we utilized Monocle3⁸³ to reconstruct the developmental trajectory of lgeInNs. The MEIS2 PAX6 SCGN cells in LGE were identified as olfactory bulb precursors, while the other lgeInNs in LGE belonged to striatal precursors (Supplementary Fig. 11b–e). To investigate MEIS2 PAX6 SCGN differences across cortical regions, Monocle⁸⁴ was used for trajectory inference. The MEIS2 PAX6 SCGN in GE was rationally included as a referenced and internal root in the analysis of cell trajectory. In E93 and E110 macaques, we consistently observed the “Y”-shaped cell trajectories, of which the stem was predominantly represented by cells from GE, and the forked branches were enriched in either FC or MSC. This confirmed regional distinction not only demonstrated the precision of trajectory analysis but also suggested that the cells at the stem developed into two different fates-that was Fate1 enriched with FC and Fate2 enriched with MSC (Fig. 5c, Supplementary Fig. 12c). To further elucidate the molecular regulation of distinct cell fates, we conducted a differential expression analysis and identified a list of genes specific to the stem or the forked branches (Supplementary Fig. 12c–e). For instance, cells in GE, Fate1, and Fate2 branches highly expressed CCND2, MEIS2/ETV1, and NR2F2/NR2F1/PROX1 respectively, which were markers of and were required for cell cycling, products of olfactory bulb or neocortical neuron (Fig. 5c, Supplementary Fig. 12c–e). The GO analysis further emphasized that genes enriched in Fate2 were significantly associated with chemical synaptic transmission, postsynaptic events, and cognition (Fig. 5d). Moreover, we replicated the above analysis by downsampling the total cell count, and quantified the percentages of MEIS2 + /ETV1 + (G1) and NR2F1 + /NR2F2 + /PROX1 + (G2) cells of distinct fates in FC and MSC region for E93 macaques to test the robustness of the “Y”-shaped cell trajectory. As expected, there was a clear distribution discrepancy between Fate1 and Fate2, indicating that FC and MSC employed different preferences in cell fates (Fig. 5e). Overall, despite cortical MEIS2 PAX6 SCGN cells being derived from GE and consistently detected in FC and MSC, their transcriptomic differences were not ignorable and might be due to the so-called adaptive alterations.

It is worth noting that the frontal lobes of three macaques (E93, RMB683 and RMB691) were more accurately dissected and four subdivisions were consequently recruited including DFC, medial (MFC), orbital (OFC) and ventrolateral (VFC) frontal regions. Surprisingly, few MEIS2 PAX6 SCGN cells were found in any of the three macaque VFC samples, in contrast to immense cells found in DFC, MFC and OFC (Fig. 5f, Supplementary Fig. 12a). To confirm such observation, we employed CIBERSORT to carry out deconvolution analysis⁸⁵, in which the averaged transcriptomic profile in bulk RNA-seq was theoretically deconvoluted using single cell signature so as to predict cell type proportions. For both human and macaque species, our result illustrated the relatively lower percentages of MEIS2 PAX6 SCGN cells in VFC compared with DFC, MFC and OFC (Fig. 5g, Supplementary Fig. 12f). We assumed this observation might suggest there was few ventrolateral migration stream to frontal lobe, consistent with the prior model of Arc and Arc-ACC migratory streams¹⁸.

Discussion

By collecting and analyzing extensive single-interneuron transcriptomic datasets from developing human and macaque species, we have uncovered shared and distinct molecular programs that regulate interneuron fate and adaptation within cortical destinations. While previous research has extensively examined these versatile interneurons in humans, macaques, and other mammals, our integration of human and macaque data across multiple cortical regions throughout various developmental stages before birth offers a unique resource. This allows us to understand how and when interneurons gradually adapt to their final positions and integrate into cortical circuits. Additionally, it facilitates the further exploration of evolutionary convergence and divergence between human and non-human primates.

Despite the value of our findings, certain limitations must be acknowledged. These include constrained datasets, uneven sample distribution across developmental time points and brain regions, and the indistinctive use of scRNA-seq and snRNA-seq. To address these challenges, it is crucial to prioritize the generation of more standardized resources, particularly for underrepresented specimens. This approach will not only allow for fair comparisons but also contribute to a more comprehensive understanding of the spatial, temporal, and evolutionary principles governing interneurons in the future.

Although previous studies have proposed multiple migration routes to describe the journey of interneurons towards various brain regions, including cerebral cortex, olfactory bulb, amygdala, striatum and others, limited attention has been given to explaining the movement of interneurons within ganglionic eminence^35,86. In our study, we observed that interneurons exhibited a gradient distribution within ganglionic eminence rather than concentrating in their birthplace ready for migration. This pattern was consistently pervasive across different interneuron lineages. Interestingly, we noted fewer transcriptomic alterations were accompanied by interneuron migration within the ganglionic eminence subdivisions compared to cortical migration. This finding suggested that the microenvironment, including interactions with repressors or activators, might play a crucial role in triggering molecular program transitions. Generally, interneurons initiate cortical migration shortly after becoming postmitotic³¹. However, we found intermediate progenitors of interneurons in the cerebral cortex, which differed from cortically born interneurons and exhibited extensive transcriptional alterations compared to their ganglionic parents. A longstanding debate and uncertainty persist regarding the lineage potential of cortical RGCs, particularly their capacity to generate cortical interneurons during the late mid-fetal stages. However, the findings have been inconsistent. In mice, individual cortical progenitors underwent a GABAergic ‘switch’ to generate interneurons that were transcriptionally similar to the olfactory bulb interneurons⁶³. Moreover, human cortical RGCs in the second trimester could produce cells closely resembling MEIS2 + PAX6 + interneurons born from LGE and SCGN + SP8+ interneurons born from CGE through tripotential IPCs^58,60,61. Conversely, labeling proliferating newborn neurons in cortical slice cultures did not substantiate the claim that the vast majority of human cortical interneurons were produced in the cortical wall⁶⁹. In our study, the UMAP visualization indicated that most cortical-derived inIPCs assembled closely with ganglionic inIPCs while segmented from cortical inIPCs, which might be due to the cortical-derived inIPCs and ganglionic inIPCs being mostly proliferative, while cortical inIPCs were predominantly post-mitotic (Fig. 2c, d, Supplementary Fig. 6e). The inIPC genesis within the cortical domain was not clear and reserved for elucidation in subsequent research.

Our study endeavors to integrate developmental timelines, cell fates, migration destinations, and transcriptomic alterations in a cohesive manner. A novel aspect of our approach involves the application of Shannon entropy to measure cellular diversities, revealing increasing profiles highly correlated with developmental times. Notably, we observed distinct increase rates of entropy between mgeInNs and cgeInNs. We postulated that these differences in entropy dynamics could be attributed to varying adaptation abilities. Besides, we found minimal differences between humans and macaques, compared with mice, suggesting conserved and intrinsic features accompanying brain development. On the other hand, the cortical arealization of interneurons led to measurable transcriptional changes, albeit a larger variation reported for excitatory neurons¹⁷. Along the developmental axis, we noted areal heterogeneity for MAF interneurons, particularly evident in late mid-fetal stages, aligning with the previously proposed developmental hourglass model^19,87. Furthermore, despite the consistent presence of lgeInNs in the anterior and dorsal parts of the frontal lobe, we observed their rarity in VFC, agreeing with the Arc and Arc-ACC models proposed previously¹⁸.

Our analysis unveiled the stable presence of CRH cortical interneurons after GW18 in humans and macaques, contrasting with the sparse findings in mice at E18.5. Also, we identified 91 genes associated with the CRH signaling pathway exhibiting an on-off pattern with a transition time in the late-fetal stage of developing human brains. Currently, research on CRH neurons has predominantly focused on the hypothalamus^88,89, leaving uncertainties about the origin and function of CRH cortical interneurons. Despite commonly being classified as a subpopulation of CGE-derived cells⁹⁰, there is a pressing need for a more comprehensive knowledge of the developmental processes of CRH cortical interneurons. Although there is a well-established connection between the release of CRH and stress-induced behavior, the role that CRH-expressing cortical interneurons play in the context remains unclear. To address this issue, future studies could involve additional cell collections from the developmental hypothalamus in both human and non-human primates, providing valuable insights into the genesis and function of CRH cortical interneurons.

Methods

Data collection

We developed a database containing 10X transcriptomic data from previously published literature on prenatal GE and CTX development of macaques and humans. Two public macaque datasets using scRNA-seq were employed, and they were comprised of multiple sub-area tissues (frontal, primary motor, parietal, temporal, primary visual cortex and GEs) at E37, E40, E42, E43, E50, E54, E62, E64, E65 (n = 3), E77, E78, E80 (n = 2), E90, E93, E100, E110 (n = 2)^17,18,47. Three publicly available human scRNA-seq datasets utilized here included data from samples of GEs (GW9-GW18), separate medial, lateral and caudal GEs (GW18)²⁰, and interneurons isolated from neocortical areas, STR, HIP and the whole GE across the first (GW7-GW13)⁴⁶ and second developmental trimester (GW14-GW27)⁴⁴. To enrich the analysis of the cortex during the third developmental window of human fetus, a recent snRNA-seq dataset was complemented⁴⁵. Finally, the integrated datasets we utilized provided comparable developmental ages for both human and macaque species, allowing for meaningful inter-species comparisons. Overall, 14,198 single-cell transcriptomes of 403,353 cells from developing macaque and 282,339 cells (or nuclei) from developing humans were applied for downstream analysis. This study encompassed the most complete developmental timeline ranging from the start of neurogenesis (GW7) to mid-gliogenesis (GW33).

We have also curated and integrated six previously published mouse datasets, covering critical areas such as the CTX and GEs, spanning from early embryonic periods (E13.5 to E18.5) to early postnatal phases (P0, P1, P4)^23,50–54. These data were obtained using single-cell sequencing technologies, specifically 10X Genomics or Drop-seq. After rigorous filtering and processing, we assembled a total of 70,072 interneurons for further analysis. Notably, the developmental ages of mice were comparable to that in the integrated human and macaque datasets, providing a solid foundation for cross-species comparisons and developmental trajectory studies.

Fetal brain collection

De-identified tissue samples were collected with prior patient consent in strict compliance with legal and institutional ethical regulations. The female human DFC and V1C samples at GW23 were obtained from the Obstetrics and Gynecology Hospital of Fudan University with approval by the Ethics Committee of Obstetrics and Gynecology Hospital of Fudan University (Ethics Review number, 2021-161-C1).

C57BL/6 J (C57) wild-type mice at different ages (P0 (n = 5), P4 (n = 5), and 6–8 weeks (n = 21, male: female=1:2)) were purchased from the Laboratory Animal Research Center, Tongji Medical College, Huazhong University of Science and Technology (Wuhan, China). E14.5/E18.5 mice were bred from the purchased 6-8 weeks mice. All mice were housed under standardized conditions, with ambient temperatures maintained at 20-23 °C, humidity levels ranging from 30-60%, and a 12-hour light/dark cycle. All animal procedures were conducted in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals and approved by the Animal Research Ethics Committee of Tongji Medical College of Huazhong University of Science and Technology ([2024] IACUC no. 4160; Wuhan, China).

Age translation

The equivalent ages across humans, macaques and mice were translated by https://www.translatingtime.org/, which modeled the timing of neural development of 19 mammalian species⁴⁸. Briefly, the post-conception day (PCD) of macaque and mouse was converted into the PCD of humans, then into the PCW or GW. We divided the prenatal development roadmap of humans into three stages: the first trimester (GW1-GW13), the second trimester (GW14-GW28), and the third trimester (GW29-end of pregnancy).

Quality control and removal of cell cycle effect

To gain better clusters and a more detailed understanding of a larger number of cell types, we applied stringent quality control metrics to remove low-quality cells and outliers. We selected cells based on the following criteria: (1) the number of detected genes is between 100 to 5000, (2) the number of unique molecular identifiers (UMIs) ranges from 100 to 15,000, (3) the fraction of mitochondrial gene counts per barcode is less than 10%, (4) the fraction of erythrocyte gene counts per barcode is less than 3%. The cell cycle stage was defined by a set of G2/M and S phase-dependent genes using the function CellCycleScoring. The cell cycle effects were regressed out by function ScaleData using vars.to.regress.

Classification of subtypes

Cell clustering and subtype definitions were performed respectively for each sample from different sources of data. The filtered digital gene expression matrix was firstly subset based on the experimental information and translated age labels, and then loaded into R for downstream analysis using the R package Seurat (version 4.3.0)⁹¹. Raw counts in cells were log-normalized by the function NormalizeData (primarily read depth, scaling factor = 10000), and then the top 2000 variable genes were identified by FindVariableFeatures for principal-component analysis (PCA). The normalized data were further scaled for each gene between cells by ScaleData. PCA (RunPCA) was performed to compute the 50 most variance-explaining principal components (PCs) and the PCA coordinates were used for data dimensionality reduction and unsupervised clustering. The interrelationship of cell populations under multiple resolutions was visualized by R package clustree⁴⁹, and resolution 3 was selected for the most distinguishable and explicit clusters. UMAP in two-dimensional coordinates by RunUMAP was applied for visualization.

The cell identity of each cluster was assigned by manual annotation using prior known markers for interneurons^17,18,92 and differentially expressed genes for each cluster using the FindAllMarkers function. The average expression and proportion of all curated marker genes for each cluster were examined and visualized using Dotplot. We defined and annotated a diverse set of neuronal populations (RGCs, excitatory neurons, InNs, excitatory IPC and inIPC) and non-neuronal clusters of cell types (glial IPC, oligodendrocytes, astrocyte, microglia, vasculature and endothelial cells) in humans and macaques at different developmental ages from different resources. We hierarchically organized cell types based on the transcriptomic correlation between cell clusters. RGCs in non-GE regions were discarded. After the first-round production of major cell types, interneuron clusters identified by the cardinal markers (for example, DLX1, DLX2, DLX5, GAD1, GAD2) were separated to retain interneuron-only matrices. Further precise annotation identified several lineage-specific clusters of MGE, CGE and LGE. Clusters with potential heterogeneous subtype profiles were sub-clustered to clarify their genuine identities and remove multiplets.

10X Visium spatial transcriptomics

Tissue preparation and sectioning

Freshly collected P0/P4 mouse brains and E14.5/E18.5 mouse heads were sectioned into appropriate sizes, embedded in OCT, and rapidly frozen in isopentane. The heads/brains were then coronally sectioned (10 µm thickness) using a Leica CM1950 cryostat (Leica Microsystems). After sectioning to the target locations, including MGE, LGE, and cortical regions, a portion of the tissue was sampled to measure RNA RIN values (all four tissue samples in this study had RIN values of 10). The remaining sections were stored frozen at −80 °C. Depending on the sample sizes, sequencing was performed on pairwise combinations of the four samples, with one group comprising P4 and E14.5, and another group comprising P0 and E18.5. Subsequently, tissue sections were subjected to methanol fixation, HE staining, imaging, and destaining according to the recommended experimental procedure by 10X Genomics (CG000614). Following the 10X Genomics experimental flow (CG000495), probe hybridization and probe release were carried out, and the samples were transferred to the 10X Visium CytAssist slide v2. Library construction was performed using the Visium CytAssist Spatial Gene Expression v2 kit (PN-1000521 for Mouse, 6.5 mm). The DNA libraries were then subjected to high-throughput sequencing using the PE-150 mode.

Spatial data processing

The 10X Genomics Space Ranger 3.1.1 pipeline was run with the FASTQ files, the CytAssist image and the brightfield H&E image. The read alignment was performed using the mouse reference genome mm10-2020-A provided by 10X Genomics. Raw Space Ranger outputs were screened to retain genes expressed in at least 5 spots and spots with more than 600 genes. The spot-gene count matrix was normalized using the SCTransform with default parameters. Dimensionality reduction was performed by using PCA and the top 20 PCs were used for following neighbor identification, Louvain clustering, and UMAP visualization. Clusters of interest based on manual matching anatomical region and marker genes were subset from the capture area of each slide separately. White matter, non-cortical and non-GE areas were filtered out. The remaining subsets were merged with the merge function to enable an integrated matrix and analysed in the same manner.

Spatial deconvolution

RCTD was applied for deconvolution of spatial distribution of major cell classes by using annotated scRNA-seq data as input⁷¹. Three scRNA-seq datasets with the anatomical areas of somatosensory cortex (E12.5, E13.5, E14.5, E15.5, E18.5, P1 and P4) and GEs (E12.5, E13.5, E14.5 for MGE and E12.5, E14.5 for LGE) were included^23,51,52. As the CGE region was not included in the coronal section, single-cell data from the respective area has been omitted. The major cell classes (apical progenitors, intermediate progenitors, excitatory neurons, glia cells, GE RGCs, inIPCs, mgeInNs, cgeInNs, lgeInNs) were annotated. We subsampled the single-cell dataset for each cell class according to following criteria: (i) If a cell class had ≤ 1000 cells, select all cells; (ii) if a cell class had >1000 cells, randomly select 1000 cells. RCTD was run with full mode, allowing each barcode to potentially contain any number of cell classes. The results of RCTD full mode were standardized using the normalize_weights function, ensuring that the proportions of various cell classes in each spot sum up to 1. By applying this algorithm, we mapped the cell class information from the scRNA-seq data onto the Visium data, enabling the visualization of the normalized probabilities of different scRNA-seq cell classes within each Visium spot.

Evaluation of subtype replicability using MetaNeighbor and cellHint

CellHint was applied to recapitulate the cell-type relationships across datasets and calculate the cross-dataset distance matrix⁵⁵. Cell types annotated in current study were merged with those of the respective original studies. Then, the merged object was employed in a standardized single-cell pipeline encompassing normalization, the detection of highly variable genes, data scaling, PCA, and construction of the neighborhood graph. During the harmonization process, the function cellhint.harmonize was used with a canonical PCA-based matrix as the search space. Each cell type was assigned to the most analogous cell type between the two datasets being compared. Finally, the function cellhint.treeplot was used to visualize as a tree plot or a cell type hierarchy based on the assignment matrix among cell populations.

We conducted MetaNeighbor analysis⁵⁶ to identify possible links across the cell type annotation in this study and that in the human dataset⁴⁴. Seurat objects were transformed into SingleCellExperiment objects. Shared genes, assays, and metadata columns between the two object types were reserved. Subsequently, the variableGenes function was employed to identify genes exhibiting high variance, and the differential genes of different classes of interneurons were supplemented for optimization of feature selection. The unsupervised MetaNeighborUS function with the fast_version set to TRUE was used to assess which cell types replicated across datasets in a statistical framework. In brief, cells from the reference dataset (cell type assignment in this study) voted for the closest neighboring cell type in the target dataset (cluster labels provided by the original authors). Then, in the target dataset, the ranking of each cell was aggregated into a ranking of cell types and expressed as the Area Under the Receiver Operating Characteristic Curve (AUROC). This metric reflected the degree of similarity between a target cell type and the reference cell type.

Integration of datasets within and across species

All transcriptome data were integrated with scVI⁹³ and scANVI⁹⁴. The top 2000 highly variable genes were selected using the sc.pp.highly_variable_genes function in Scanpy with batch_key set as the laboratory of origin. We trained the scVI model using the raw count and individual sample was used for correcting batch effects. The annotated five major classes of interneurons were incorporated as ‘labels_key’ for scANVI. The scANVI latent space was used to generate a neighbor graph for dimensionality reduction and then the function sc.pl.umap was used to generate UMAP graph.

Identification of homologous cellular counterparts

After integration, a 10-dimensional representation of each cell was used for nearest neighbor analysis. To achieve a more balanced representation, downsampling (2000 cells) was performed to equalize the cell counts across various cell types among different species. The mutual nearest neighbors (MNNs) were used for the Sankey plot for the comparison of cell type similarities in developing humans, macaques and mice. The links in Sankey diagram connecting the left and middle columns and the middle and right columns delineate the cell types that share the highest number of MNNs between human and macaque and between macaque and mouse, respectively. In order to depict the neighbor relationships encompassing all cell types of the three species, certain cell types, exemplified by human LHX6 SST NPY and human LHX6 MAF DCN were presented with their two closest neighbors.

Simulation to detect stable differentially expressed genes

We intended to identify genes differentially expressed in pairwise comparisons either within GEs or between GE and the whole cortex, for mgeInNs across different species. Differential expression analysis was performed by randomly downsampling 30% of the cells per ident at a time, which was repeated 100 times. Then, marker genes of the selected cells were calculated using Seurat FindAllMarkers by means of the Wilcoxon test, restricting genes that were detected in at least 10% of cells of the ident. The marker genes with adjusted P value (p_val_adj) < 0.05 and log2 fold change > 1.25 of each ident were defined as differentially expressed. Besides, genes stably present in more than 50 repeats of the differential gene result lists were retained for further analysis. Ribosomal genes were removed.

GO, GSEA and pathway enrichment analyses

Following the process above, 146 genes were found to be differentially expressed across distinct species-subtype idents in four major cortical lobes (FC, MSC, OC, TC). Among them, 18 genes were TFs and 25 genes were apparently enriched in synapse cellular components (http://geneontology.org/)^66–68. The functions enrichGO and GSEA from the R package clusterProfiler were employed to conduct enrichment analysis⁹⁵. The significant GO terms with adjusted P value < 10^-5 were retained. We performed GSEA on the gene sets sorted by log2 fold change between compared groups. A representative subset of the GO terms sorted by semantic similarity was visualized by REVIGO Web server (available at http://revigo.irb.hr/) ⁹⁶. To score CRH pathway activities in individual cells, we used the AddModuleScore function in the R package Seurat.

Cell type deconvolution

Utilizing scRNA-seq data as prior information, along with the deconvolution of bulk RNA-seq data, facilitates the inference of the cellular composition of tissue samples. To unravel the composition of bulk samples, the initial step is to obtain cell type signature. Major cell classes in DFC, VFC, OFC, and MFC from the prenatal macaque scRNA-Seq dataset (E110) were included, and a random subset containing 500 cells for each cell type was taken from each region to equalize the cell number across different samples. A specificity metric, denoted as “Spec”, was employed to quantify the transcriptome profile of discrete samples with respect to specified cell types⁷³. We measured the highly specific gene expressions for six cell types, including excitatory neurons, mgeInNs, cgeInNs, lgeInNs, astrocytes and oligodendrocytes. The calculation of Spec values was used to initially optimize the bin size for a given gene and subsequently arrange genes in descending order of Spec scores. The cell type signature matrix was constructed by the 50 top-ranked genes for each cell type and was input to perform cell type deconvolution using CIBERSORT R script⁸⁵. The other input file consisted of the expression matrices from human or macaque bulk-tissue samples across four frontal regions spanning from the early embryonic period to adulthood¹⁹. Bulk samples of a particular frontal region were back-convoluted with the corresponding single-cell data. The proportion of lgeInNs in four frontal cortical lobes was converted to a proportion relative to the maximum proportion value of lgeInNs (Fig. 5g). Additionally, dynamic changes in the cell composition of lgeInNs relative to all aforementioned six cell types throughout developmental periods were presented (Supplementary Fig. 12f).

Trajectory analysis

Monocle (v2), a method based on a machine learning minimum spanning tree algorithm, was employed to depict the pseudo‐temporal evolution of MEIS2 PAX6 SCGN cell type in FC and MSC cortex area⁸⁴. Seurat function FindVariableFeatures identified the genes with significant differences in expression between cells for temporal resolution of transcriptome dynamics. The root state of the reduced-dimension DDTree diagram was determined by known marker genes, prior knowledge and actual gestational time of each cell. BEAM analysis on branch point was conducted to identify the genes that were regulated in a branching-dependent manner along the pseudo‐timeline. These genes were then categorized into clusters with different dynamic trends.

The Seurat object containing some inIPCs and lgeInN subtypes in LGE was integrated and clustered following the Seurat pipeline and then converted into a cell_data_set object to infer the developmental trajectory with the Monocle (v3) R package⁸³. The function get_earliest_principal_node was used to automatically select the “root_state” of the tree. The order_cells function was utilized to arrange the cells and allocate pseudotime values. This indicated an early bifurcation of MEIS2 PAX6 SCGN cells and other lgeInNs with striatal potentials which diverged in three distinct fates including TSHZ1 D1-expressing medium spiny neurons (D1-MSNs), D1-MSNs expressing ISL1 and D2-MSNs expressing PENK (Supplementary Fig. 11b−e). Variable genes were identified with the graph_test function using the Moran’s I statistic and genes with significant q-value from the autocorrelation analysis were grouped into modules. Five modules of genes that were changing across the developmental trajectory were identified and module 5 was enriched in MEIS2 PAX6 SCGN cells and module 1 and 3 in striatal lineage.

Lobe expression specificity

Human GW18 and GW20 and macaque GW16, GW20&21, GW26 and GW33 had comparable numbers of mgeInNs for MAF class in 4 cortical lobes (FC, MSC, TC and OC), so they were included in the following analysis. The snRNA-seq data in Velmeshev et al. 2023 were ruled out⁴⁵. The Python package CELLEX was used to compute the expression specificity of 4 cortical lobes independently for both humans and macaques⁹⁷. Gene expression matrix and corresponding cell annotations were set as input. The output result comprised expression specificity weights (float) for each gene in different lobes. This analysis was repeated 100 times, with a downsampling of 3000 cells at each gestational age in each iteration. The set of genes with enrichment scores > 0.2 were regarded as the most specifically expressed genes for each lobe. The count of significant genes associated with a particular lobe was recorded for each computation, and the average across the 100 iterations was calculated.

Immunofluorescence

Frozen human and mouse brain tissues were embedded in OCT (Sakura Finetek) and then snap-frozen in dry ice. Cryosections were cut at −20 °C to a thickness of 30 μm using a cryostat (RWD) and mounted on adhesive slides. The sections were then incubated at 37 °C for 10 minutes, fixed with 4% paraformaldehyde (Biosharp) for 30 minutes, washed in PBS (Biosharp), and permeabilized with PBS containing 0.5% Triton X-100 (Biosharp) for 20 min. Subsequently, the sections were blocked with 5% BSA (Abbkine) for 30 minutes at room temperature and incubated overnight at 4 °C with the primary antibody. After washing off the primary antibody with PBS, the sections were incubated with Alexa Fluor secondary antibody (1:500) at room temperature for 1.5 hours. Nuclei were visualized using DAPI, and the sections were covered with glass coverslips. Final images were acquired using confocal microscopy (Orthogonal single-photon confocal microscope LSM800, Zeiss).

Antibody

The following antibodies were used for immunofluorescence: rabbit anti-CRH (1:400, Proteintech, cat.#26848-1-AP); chicken anti-KI67 (1:1000, EnCor, cat.#CPCA-Ki67); anti-ASCL1 (1:100, santa cruzs,cat.#sc-390794); chicken anti-GAD1 (1:400, Abcam, cat.#ab26116); rabbit anti-MEIS2 (1:300, Atlas Antibodies, cat.#HPA003256); mouse anti-PAX6 (1:100, Abcam, cat.#ab78545); guinea pig anti-DLX2 (1:200, Oasis Biofarm, cat.#OB-PGP017); mouse anti-PROX1 (1:300, Novus Biologicals, cat.#NBP1-30045SS). Secondary antibodies included Alexa Fluor 488 goat anti-rabbit IgG (Abbkine), Alexa Fluor 594 goat anti-mouse IgG (Abbkine), Alexa Fluor 488 goat anti-guinea pig IgG (Yeasen), Alexa Fluor 405 goat anti-chicken IgG (Abcam), and Alexa Fluor 405 goat anti-rabbit IgG (Abbkine).

Reporting summary

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

Author contributions

M.Li conceived the project and designed the experiment. M.Li and Q.L. coordinated the research. X.F. and Y.G. analyzed the data and performed data visualization. Y.Z. and Y.W. helped for the tissue procurement. F.C. performed the immunofluorescence staining. Y.S. painted the brain diagrams of human and macaque species. M.Li and Q.L. interpreted the results. M.Li and X.F. wrote the manuscript with the proofreading assistance from Y.S. and M.Liu. All authors edited the manuscript.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

Datasets analyzed in the course of this study are at your disposal. Human datasets contained four resources. Shi dataset²⁰ was downloaded from Gene Expression Omnibus (GEO) under the accession code GSE135827. Bhaduri dataset⁴⁴ was deposited in Neuroscience Multi-omic (NeMo) Archive (RRID:SCR_002001) and available at https://data.nemoarchive.org/biccn/grant/u01_devhu/kriegstein/transcriptome/scell/10x_v2/human/processed/counts/. Velmeshev dataset⁴⁵ was downloaded from UCSC Cell Browser, collection “Human Cortical Development” (https://pre-postnatal-cortex.cells.ucsc.edu). Braun dataset⁴⁶ was available at European Genome Phenome Archive database under the accession code EGAS00001004107. Macaque datasets were obtained from GSE169122¹⁸, GSE226451¹⁷ and NCBI BioProjects (accession code PRJNA448973)⁴⁷. Mouse data such as Mayer dataset²³, Loo dataset⁵⁰, Lee dataset⁵¹, Di Bella dataset⁵², Yuan dataset⁵³, and Abad dataset⁵⁴ were downloaded from GEO under the accession code GSE104158, GSE123335, GSE167013, GSE153164, GSE204759 and GSE244477. The integrated datasets are interactively accessible and deposited at http://www.braininfo.ac.cn/mammal_inn_migration/. The spatial transcriptome sequencing data generated in this study have been deposited in the National Genomics Data Center (NGDC) database under accession code PRJCA030245. All data substantiating the findings of this study are meticulously presented within the main body of the paper and its Supplementary Information. Source data are provided with this paper.

Code availability

All codes generated during this study are available at https://github.com/lmfeng/mammal_inn_migration.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-025-56058-8.