The PFC reaches its greatest complexity in anthropoid primates (monkeys and apes), which appear to uniquely have many prefrontal areas that cover the entire anterior two-thirds of the frontal lobe and a well-defined granular layer 410,11. Previous analyses revealed that the transcriptomic differences between neocortical areas are most prominent in primates during the middle stages of fetal development, corresponding to post-conception weeks (PCW) 13 to 24 (hereafter referred to as ‘mid-fetal’)16,17—a crucial period for neuronal specification and the initial assembly of neocortical neural circuits18. Thus, we hypothesized that the molecular processes that govern the development and evolutionary diversification of the PFC could be revealed by differential regional gene expression analysis of the primate mid-fetal neocortex.

Fetal frontal upregulation of RA-related genes

Using human BrainSpan RNA-sequencing (RNA-seq) data17, we screened for genes that are differentially upregulated in the mid-fetal frontal lobe. The mid-fetal data consisted of tissue-level samples ranging in age from PCW 16 to 22, which included four prospective PFC areas (medial, mPFC or MFC; orbital, oPFC or OFC; dorsolateral, dlPFC or DFC; and ventrolateral, vlPFC or VFC) and the primary motor cortex (M1C). Gene expression in these frontal areas was compared to areas within the parietal (primary sensory cortex, S1C; and inferior parietal cortex, IPC), occipital (primary visual cortex, V1C), and temporal lobes (primary auditory cortex, A1C; superior temporal cortex, STC; and inferior temporal cortex, ITC) (Fig. 1a, Extended Data Fig. 1a). We identified 190 protein-coding genes, using stringent criteria, that were specifically upregulated in at least one area within a lobe in comparison with areas from other lobes, including 125 in the frontal lobe, which were able to differentiate the four brain lobes and most areas within them (as observed by principal component analysis (PCA)) (Fig. 1b, Extended Data Fig. 1a, b). Moreover, the first principal component (PC1), which accounts for the highest variability present in the data, corresponded to the anterior–posterior or frontal–temporal axis (Fig. 1b). Gene Ontology (GO) analysis of the frontally upregulated genes identified an enrichment of genes associated with categories such as ‘response to retinoic acid’, ‘synapse organization/assembly’ and ‘axon development/guidance’ (Fig. 1c, d, Supplementary Tables 1, 2), suggesting that these genes may have a role in frontal lobe patterning and circuit development. Many of the same frontal-lobe-enriched and retinoic acid (RA)-related genes were upregulated in the mid-fetal macaque frontal cortex using a PsychENCODE dataset19 (Extended Data Fig. 1c, d). We also observed prominent frontal enrichment of many of the same genes during the early fetal period (Extended Data Figs. 1, 2). Closer analysis of the spatiotemporal profile of the fetal frontally upregulated genes revealed the enrichment to be largely transient, with mainly conserved and some divergent expression patterns between human and macaque (Extended Data Fig. 4b).

Fig. 1: Predicted RA-signalling-associated genes are upregulated in human mid-fetal prospective frontal areas.
figure 1

a, Diagram of the analysed eleven areas of the human mid-fetal brain from the four lobes. b, PCA of genes specifically enriched in at least one area within a lobe. Ellipses are centred on the mean of the points of a given area and the size of the axes corresponds to their standard deviation on each component. c, GO terms associated with frontally upregulated genes and their unadjusted P value. Black text is used to highlight relevant GO terms. d, PC loadings with labelled genes being upregulated in the PFC (black dots) and M1C (grey dots). Genes highlighted in blue are associated with RA signalling and genes with red or green outlines are predicted ASD or SCZ risk genes, respectively. For reproducibility information, see Methods.

Analysing only the RNA-seq data that encompassed the five mid-fetal frontal areas17, predicted RA-signalling-associated genes—such as CBLN2, RXRG, CDH8, MEIS2 and RBP1—were among the genes upregulated in the PFC compared to the M1C (Fig. 1d, Supplementary Table 2; see our accompanying study20), whereas the RA-degrading enzyme CYP26B1 was upregulated in the M1C (Fig. 1d), consistent with previous microarray-based findings21. We also identified multiple predicted autism spectrum disorder (ASD) (group 1–3; and schizophrenia (SCZ) (score of 3 or higher from risk genes that were upregulated in the mid-fetal frontal cortex (Fig. 1d, Supplementary Table 2).

Mid-fetal PFC-enriched gradient of RA

RA is a diffusible biologically active derivative of vitamin A that is involved in neural tube patterning, neurogenesis, cell differentiation and synaptic function12,13,22,23,24,25,26,27. Moreover, alterations in RA signalling have been implicated in the pathophysiology of ASD28,29,30 and SCZ14,31,32. Given the enrichment of RA-related genes among the upregulated fetal frontal genes, we assessed the concentration of RA in different areas of the human and macaque mid-fetal neocortex, and in the neonatal mouse neocortex (approximately equivalent developmental ages), using an enzyme-linked immunosorbent assay (ELISA). We found that there was a PFC-enriched anterior–posterior gradient of RA concentration in the human and macaque mid-fetal neocortex, with homologous mPFC exhibiting the highest concentration within each species (Fig. 2). Overall, RA concentrations were significantly higher in prospective primate PFC areas compared to more posterior areas. Comparison across the three species identified higher concentrations of RA both in the mPFC and in all four PFC areas overall in humans as compared to the other two species (Fig. 2), as well as in macaques compared to mice (Fig. 2). Notably, the ITC—an association area within the temporal lobe that is thought to exhibit unique features and connectivity in humans33—had a higher concentration of RA among non-frontal areas in humans (Fig. 2), with a two-fold increase in RA concentration in humans compared to macaques.

Fig. 2: PFC-enriched anterior–posterior gradient of RA.
figure 2

RA concentrations in human (post-conception weeks (PCW) 16, 18, 19 and 21), macaque (four post-conception day (PCD) 110 brains) and mouse (four PD 1 brains) cortical areas (n = 3–4 for each sample area). One-way ANOVA with post-hoc Dunnett’s adjustment or two-tailed unpaired t-test: ****P < 1 × 10−4, ***P = 0.0005 (macaque PFC versus mouse PFC), **P = 0.005 (human non-PFC minus ITC versus human ITC), *P = 0.01 (human PFC versus macaque PFC), *P = 0.04 (human ITC versus macaque ITC). Centre value, average; error bars, s.e.m. The dashed red line in the macaque graph represents the human RA concentrations. A single assay was done for each brain sample. dmPFC, dorsomedial PFC; ENT, entorhinal cortex; PIR, piriform cortex; RSP, retrosplenial cortex; SSs, secondary somatosensory area; VISp, primary visual cortext; vmPFC, ventromedial PFC.

Expanded RA synthesis in the primate cortex

The observed enrichment and primate-specific lateral expansion of RA levels in the fetal PFC led us to systematically examine the spatiotemporal expression of RA-synthesizing enzymes (ALDH1A1–ALDH1A3) and RA-degrading enzymes (CYP26A1, CYP26B1 and CYP26C127,28) in human, macaque and mouse (for full descriptions, see Supplementary Results). In brief, we observed several potential sources of RA in the fetal cortex, including conserved expression of ALDH1A1 in the meninges and midbrain axons, and ALDH1A3 in the mPFC26. We also observed primate-specific expression of ALDH1A1 in astrocytes and frontal subplate neurons, and lateral frontal expression of ALDH1A1 and ALDH1A3 (Extended Data Fig. 6). These primate-specific expression patterns were complementary to the increase in and lateral expansion of RA in the mid- and late-fetal primate PFC and the lateral extension of PFC in anthropoid primates10,11.

Frontal RA signalling is mediated by RARB and RXRG

Given the fetal frontal cortical upregulation of both RA and RA synthesizing enzymes, we further assessed the expression of RA-dependent receptors and RA-responsive downstream genes in the developing human, macaque and mouse cortex. In addition to RXRG, which is upregulated in the human mid-fetal PFC21 (Extended Data Fig 1c), several genes that encode RA receptors were also detected in the developing human, macaque and mouse cortex or cultured human primary mid-fetal cortical neurons (Extended Data Figs. 5, 8, 9, 15b). Of these, only RARB and RXRG exhibited a higher anterior to lower posterior gradient of expression (Extended Data Figs. 8, 9a).

The RARB–RXRG heterodimer has previously been shown to mediate RA signalling in the adult mouse cortex and striatum, and is required for learning, locomotion and dopamine signalling34,35. To assess whether RARB and RXRG are required for RA signalling activity in the developing mouse frontal cortex, we generated constitutive Rarb and Rxrg double-knockout (dKO) mice (Extended Data Fig. 9d), which—consistent with previous findings34,35—are viable. Using a RARE-lacZ reporter line36 in which lacZ is under the transcriptional control of an RA response element, we identified a significant reduction of RA signalling in the mPFC of dKO mice at post-natal day (PD) 0 compared to control mice (Fig. 3a, Extended Data Fig. 10a). In addition, we found a less extensive reduction of RA signalling in the anterior cingulate area (ACA) and retrosplenial area (Extended Data Fig. 10a). Furthermore, expression of frontally enriched RA-regulated genes, Cbln220 and Meis2, which are induced by RA in human and chimpanzee cerebral organoids and repressed by RA receptor antagonists in human cortical neurons (Fig. 1d, Extended Data Figs. 11b, 15d), was reduced in the dKO mice (Extended Data Fig. 10b). Together, these findings indicate that RARB and RXRG mediate RA signalling in the developing mouse mPFC.

Fig. 3: Reduced RA signalling in the PFC of mice lacking Rarb and Rxrg leads to the downregulation of genes involved in synapse and axon development.
figure 3

a, β-Galactosidase histochemical staining in Rarb+/+Rxrg+/+ (wild type; WT); RARE-lacZ (blue) and Rarb−/−Rxrg−/− (dKO); RARE-lacZ (orange) brains at PD 0. Two-tailed Student’s t-test (WT versus dKO): ***P = 1 × 10−6 (left), 3 × 10−4 (right). Centre value, average; error bars, s.e.m. (n = 3 per genotype). Scale bars, 200 μm; 50 μm (inset). b, First two principal components calculated from the expression of differentially expressed protein-coding genes between WT and dKO littermates in at least one of the three frontal cortex areas (mPFC, MOs and OFC). c, GO terms associated with differentially expressed genes showing their z-score and unadjusted P values. The z-score represents the proportion of upregulated versus downregulated genes in the dKO compared to the WT that are associated with each GO term. Dark blue, z-score < −5; light blue, z-score −5 to 0; orange, z-score > 0. The size of the bubbles is proportional to the total number of differentially expressed genes associated with the given GO term. Black text is used to highlight relevant GO terms. d, Quantification of excitatory synapses marked by DLG4 (PSD95) in the mPFC, MOs, OFC, MOp and SSp regions of PD 0 WT and dKO mice brains. Two-tailed Student’s t-test: **P = 6 × 10−4 (mPFC); 2 × 10−3 (MOs). Centre value, average; error bars, s.e.m. (n = 3 per genotype). For reproducibility information, see Methods.

RARB and RXRG regulate frontal connectivity

To understand the functional importance of RA signalling through the RARB–RXRG heterodimer in the developing cortex, we performed RNA-seq analysis of different regions of the PD 0 mouse frontal cortex (mPFC; secondary motor cortex (MOs) and the adjacent parts of the primary motor cortex (MOp); and OFC) microdissected from dKO and wild-type littermates. We identified 4,768 differentially expressed protein-coding genes between the two genotypes in at least one of the areas, with the highest number of differentially expressed genes in the mPFC (Extended Data Fig. 12a, Supplementary Table 3). PCA based on the expression of these differentially expressed genes separated the wild-type and dKO mice along PC1, with the mPFC showing the greatest distance between wild type and dKO, further supporting the notion that the mPFC is most affected by the reduction in RA signalling (Fig. 3b).

The GO enrichment analysis revealed that terms associated with genes that are overexpressed in the wild-type compared to the dKO frontal cortex were highly related to the process of synaptogenesis and cellular components related to synapses and axons, whereas the genes that are overexpressed in the dKO were related to the regulation of the cell cycle (Fig. 3c, Extended Data Fig. 12c, d). In addition, when analysing differentially expressed genes in individual regions of the frontal cortex, only genes overexpressed exclusively in the wild-type mPFC were associated with the cellular components, axons and synapses (Extended Data Fig. 12c, d). Several of the genes overexpressed in wild-type mice that are related to axon guidance and synapse development exhibited an anterior enrichment in wild-type neonatal mouse cortex (Extended Data Fig. 13a, Supplementary Table 4). We also observed a significant enrichment of homologous genes specifically upregulated in the human mid-fetal frontal lobe among the genes that were downregulated exclusively in the frontal cortex and mPFC of the dKO mice (Extended Data Fig. 12b, e). Overall, these results suggest a possible role for RA signalling in the regulation of synaptogenesis and axon development, specifically in the mPFC.

We analysed the role of RA in synaptogenesis by quantifying synaptic puncta in multiple regions of the PD 0 cortex in dKO mice compared to wild-type littermate controls. We identified a significant reduction in DLG4 (also known as PSD95)-positive excitatory post-synaptic densities in the dKO mPFC and MOs compared to wild-type, but not in the OFC, primary motor area (MOp) and primary somatosensory area (SSp) (Fig. 3d, Extended Data Fig. 13b). Similarly, we identified a significant decrease in synaptophysin (SYP) and DLG4 co-immunolabelled puncta in deep layers of the mPFC (Extended Data Fig 13c) and a decrease in number of all and mushroom dendritic spines in mPFC pyramidal neurons labelled by the retrograde viral tracer AAVrg-Cag-Gfp (Extended Data Fig. 14e). Of note, there was no difference observed in the dendritic complexity of mPFC upper-layer neurons (Extended Data Fig. 14e). In addition, treating human primary cortical neurons with RA induced the expression of DLG4 (PSD95) mRNA, and DLG4 (PSD95) and SYP co-immunolabelled synaptic puncta, whereas applying the RA receptor inhibitor AGN193109 had the opposite effect (Extended Data Fig. 15e–j). Similarly, treatment with RA-soaked beads induced the expression of DLG4 (PSD95) in both human and chimpanzee cerebral organoids (Extended Data Fig. 11b, c). We assessed the role of RA signalling in regulating the expression of the gene encoding synaptic organizer CBLN2 and local connectivity in our accompanying study20.

We investigated the role of RA signalling in long-range connections from the mPFC using diffusion tensor imaging (DTI). We identified a reduction in long-range connections between the mPFC and thalamus in dKO compared to wild-type mice at PD 5 (Fig. 4a, b). There was no observable difference in connections between the left and right mPFC at PD 5 (Fig. 4a, c). Owing to the limitations of the technique and the developmental stage, we were unable to study the connections between the mPFC and striatum or basolateral amygdala (BLA) using DTI. Anterograde lipophilic axon-tracing experiments starting from the mPFC or the mediodorsal thalamus (MD) at PD 21 similarly identified a reciprocal reduction in PFC–MD connectivity, as well as a reduction in the number of fibres in the internal capsule (Fig. 4d, e, Extended Data Fig. 13d). To more fully examine mPFC connectivity in dKO mice, we used a retrograde viral tracer to study inputs into the PD 30 mPFC. We identified reduced inputs from the MD and anterior insula in dKO mice, whereas no obvious changes were identified in inputs from the contralateral mPFC, MOs or MOp, internal capsule, claustrum, piriform cortex, amygdala and ventral hippocampus (Extended Data Fig. 14a–d). Finally, we assessed for changes in dopaminergic innervation of the mPFC and found no significant difference between wild type and dKO (Extended Data Fig. 16f).

Fig. 4: Altered mPFC–MD connectivity in mice lacking Rarb and Rxrg.
figure 4

a, Representation of the number of streamlines generated as a connectivity measurement between the cortical areas, thalamus and internal capsule (IC) at PD 5 using DTI. b, Visualization and quantification of streamlines between the mPFC and the thalamus in WT and dKO brains. Two-tailed unpaired t-test: **P = 1 × 10−3. Centre value, average; error bars, s.e.m. (n = 5 per genotype). c, Quantification of select corticothalamic, corticocortical and corticospinal streamlines. Two-tailed unpaired t-test: P = 0.2 (SSp–Th), 0.3 (mPFC–mPFC), 0.5 (SSp–SSp) and 0.5 (corticospinal). Error bars, s.e.m. (n = 5). d, e, DiI placement in the mPFC (d) and medial thalamus (e) in WT and dKO brains at PD 21 with labelled processes in the MD (d; inset), and mPFC (e; inset). Two-tailed Student’s t-test: *P = 9 × 10−3, **P = 3 × 10−3. Centre value, average; error bars, s.e.m. (n = 3 per genotype). AU, arbitrary units. Scale bars, 1 mm; 100 μm (inset).

Although there is reduced RA activity in the outer shell of the striatum in the dKO brain (Extended Data Fig. 10a), we found that fibres traverse through the medial aspect of the striatum, suggesting that alterations in RA signalling in the lateral striatum do not affect the guidance of reciprocal mPFC–MD connectivity. RA signalling has previously been implicated in thalamic development37, so we assessed whether other thalamocortical connections were altered in the dKO mice at PD 5 using DTI. There was no observable difference in thalamocortical connectivity with the MOp, primary auditory area (AUDp) or SSp (Fig. 4a, c, Extended Data Fig. 16b) or in the formation of barrel fields in the SSp (Extended Data Fig. 16c). Given that Aldh1a3 is expressed in the ACA26, we also examined whether thalamocortical innervation was altered in the ACA and found no difference in the number of streamlines (Fig. 4a, Extended Data Fig. 16b). The corticospinal tract (CST) and connections across the corpus callosum between the left and right mPFC, SSp and MOp in the dKO showed no difference in the number of streamlines in dKO compared to wild type (Fig. 4a, c, Extended Data Fig. 16b). Furthermore, scalar indexes, which describe the microstructural integrity of white matter, were similar between wild-type and dKO mice in the corpus callosum, anterior commissure and internal capsule (Extended Data Fig. 16a). In addition, the width of the CST at PD 30 was slightly increased in the dKO mice (Extended Data Fig. 16d). The reduction of connections between the mPFC and thalamus was not due to cell death in the mPFC (Extended Data Fig. 16e). In summary, deletion of Rarb and Rxrg leads to a reduction of RA signalling specifically in the mPFC, as well as a selective reduction of reciprocal mPFC–MD connectivity.

RA signalling has previously been shown to be involved in the regulation of proliferation, cell cycle timing and cortical laminar development22,23,24, which could underlie our findings. Thus, we assessed whether the constitutive deletion of Rarb and Rxrg leads to changes in cortical size and laminar organization. The volume of the brain, cortex and frontal cortex at PD 5 were not significantly different when measured by magnetic resonance imaging (MRI) (Extended Data Fig. 17a). We also quantified the number of neurons using the upper-layer markers CUX1 and POU3F2 (BRN2), layer 4 marker RORB, and deep-layer markers BCL11B (CTIP2) and TBR1, and found no difference between wild type and dKO in the mPFC, MOs or visual cortex (Extended Data Fig. 17b). We did find a modest reduction in RORB-labelled neurons in the motor cortex (Extended Data Fig. 17b). Of note, the mouse mPFC lacks a granular layer 4 and had minimal RORB-labelled cells.

Ectopic RA expands MD connectivity

MD innervation of the primate frontal cortex is expanded laterally compared to rodents, with the primate PFC areas having a more prominent layer 410,11. Thus, we investigated whether the expansion of RA signalling is sufficient to increase and laterally extend MD innervation and alter layer 4 development in the mouse neonatal frontal cortex by genetically deleting the RA-degrading enzyme CYP26B1 in mice (Extended Data Fig. 18a). Consistent with previous findings in humans21 and mice24, whole-mount and serial tissue section in situ hybridization revealed that Cyp26b1 is upregulated in the prospective motor cortex (MOs and MOp) and anterior insula compared to the mPFC and OFC, defined by high expression of Cbln2, and SSp, defined by Rorb (Fig. 5a, b, Extended Data Fig. 18b). Although mice with constitutive deletion of Cyp26b1 died perinatally, it has previously been shown that genetic deletion of Cyp26b1 in mice resulted in the expansion of RA signalling in multiple organs prenatally38 and post-natally in the mouse mPFC24. To investigate the possible role of CYP26B1 in restricting RA signalling to the mouse medial frontal cortex, we generated Cyp26b1 KO mice that also contained the RARE-lacZ transgene. The lack of Cyp26b1 resulted in the spreading of RA signalling dorsolaterally toward the MOs and MOp regions of the RARE-lacZ reporter mouse line at post-conceptional day (PCD) 18 (Fig. 5c). This increase was not robustly observed in more posterior regions (Extended Data Fig. 18c).

Fig. 5: Increased RA signalling promotes mPFC–medial thalamic connectivity.
figure 5

a, Lateral (top) and coronal (bottom) views of whole-mount in situ hybridization (ISH) of the PD 0 mouse cortex with pseudo-colour merge of Cbln2, Cyp26b1 and Rorb. OB, olfactory bulb. b, Expression of Cyp26b1 in the PD 0 mouse brain using ISH. Cyp26b1 expression shows a gradient from the insula (AI) to the MOs (insets). Arrows indicate Cyp26b1 expression at the insula and MOs. Scale bar, 200 μm. AMY, amygdala; GP, globus pallidus; HIP, hippocampus; IG, indusium griseum; IN, insula. c, β-Galactosidase histochemical staining of the mPFC of Cyp26b1+/+ (WT); RARE-lacZ and Cyp26b1−/− (KO); RARE-lacZ mouse brains at PCD 18. Relative signal intensity was quantified in the boxed area compared with anterior mPFC expression (mPFC, MOs and MOp). Two-tailed Student’s t-test (WT versus Cyp26b1 KO): ***P = 1 × 10−4, 6 × 10−5 (mPFC, MOp), **P = 4 × 10−4; 1 × 10−3 (MOs, MOp), *P = 4 × 10−3, 2 × 10−2 (mPFC, MOp). Error bars, s.e.m. (n = 3 per genotype). d, e, Representative images and quantification of DiI labelling in PCD 18 frontal MOp/MOs (d) and medial thalamus (e) after DiI was placed in the MD thalamus or mPFC of WT and Cyp26b1 KO brains. Two-tailed Student’s t-test: ***P = 4 × 10−4, *P = 0.03. Error bars, s.e.m. (n = 3 per genotype and condition). Scale bars: 200 μm. f, Representative image and quantification of expression of Rorb in WT and KO mice in PCD 18 primordial motor cortex divided into 5 bins (1–5). Two-tailed Student’s t-test (WT versus Cyp26b1 KO): ***P = 2 × 10−4, *P = 0.01. Centre value, average; error bars, s.e.m. (n = 3 per genotype).

Next, we used the Cyp26b1 KO mice to study the effects of the expansion of RA signalling into the dorsolateral frontal cortex by inserting a crystal of 1,1′-dioctadecyl-3,3,3′,3′-tetramethylindocarbocyanine perchlorate (DiI) into the medial thalamus of fixed post-mortem wild-type and KO brains, which were collected at PCD 18 owing to the perinatal lethality of the KO mice. Histological analysis revealed that wild-type littermates had occasional thalamocortical axons within the medial and dorsolateral frontal white matter and cortex at this age. By contrast, KO mice showed precocious and robust innervation of both the medial and the dorsolateral frontal cortex by the axons originating from the medial thalamus (Fig. 5d, Extended Data Fig. 19b). When DiI was placed in the motor cortex, KO mice showed an increased signal in the developing thalamus (Fig. 5e, Extended Data Fig. 19c). We also observed a moderately enlarged frontal cortex with grossly normal cytoarchitecture of the cortical wall and cortical plate in the analysed areas of the Cyp26b1 KO cortex (Extended Data Fig. 20a, b).

Furthermore, we observed an upregulation and expansion in the laminar expression of the layer 4 marker Rorb in the motor cortex of neonatal Cyp26b1 KO mice (Fig. 5f, Extended Data Fig. 19a) and after misexpression of Aldh1a3 in the dorsolateral fronto-parietal cortex using in utero electroporation (Extended Data Fig. 20c). In summary, we identified that ectopic RA signalling in the perinatal mouse frontal cortex leads to the expansion of medial thalamocortical innervation as well as the increased laminar expression of Rorb, which are both characteristics of the lateral granular PFC in anthropoid primates10,11.


We have shown here that RA signalling is required for proper prefrontal gene expression, spinogenesis and long-range connectivity. We propose that the early- and mid-fetal cortical expansion of RA signalling underlies the lateral expansion of PFC areas and MD innervation in primates. As the expansion of the prefrontal and temporal association areas has been proposed to be one of the evolutionary underpinnings of complex cognition9,10,11, it will be important to explore whether RA signalling has a broader role in developmental specification and the expansion of association areas, as well as in disorders affecting cognition (see Supplementary Discussion for a detailed discussion).


Data reporting

No statistical methods were used to predetermine sample size.  Data collection was performed by independent investigators. Prior to data analysis, all experiments were randomized and analysed by independent blinded observers.

Analysis of human and macaque transcriptomic data

Developing human and macaque brain RNA-seq data (counts file) with the metadata information were obtained from BrainSpan ( and PsychENCODE (; projects17,19. The timelines of human and macaque development and associated periods were reported in a previous study39.

For human mid-fetal periods 4–6, a total of 73 mRNA samples corresponding to 11 prospective neocortical areas, comprising the pial surface, marginal zone, cortical plate (layers 2–6) and adjacent subplate zone, from windows 3 and 4 (PCW 16–22) were considered for analyses (Extended Data Fig. 1). The human neocortical areas under study are the orbital (oPFC or OFC), dorsolateral (dlPFC or DFC), ventrolateral (vlPFC or VFC), medial (mPFC or MFC) prefrontal cortex and primary motor cortex (M1C) from the frontal lobe; primary somatosensory cortex (S1C) and posterior inferior parietal cortex (IPC) from the parietal lobe; primary auditory cortex (A1C), posterior superior temporal cortex (STC) and inferior temporal cortex (ITC) from the temporal lobe; and primary visual cortex (V1C) from the occipital lobe. A TMM normalization procedure was applied (function normalizeCounts from tweeDEseq package in R) to the expression of 15,724 protein-coding genes that show sufficiently large counts (determined with function filterByExpr from edgeR package in R). To identify genes that are upregulated in a given brain lobe, we first applied RNentropy40, available as a package in R, to determine which genes are differentially expressed among the 11 neocortical areas. Then, we considered a gene to be specifically overexpressed in a given lobe if (1) there is at least one area in this lobe where the gene is significantly upregulated; (2) the gene is not upregulated in any area of the other lobes; and (3) the gene is under-expressed in at least 30% of the areas from the remaining lobes. Similarly, we identified genes that are specifically upregulated in the PFC compared to the M1C, or vice versa, by first running RNentropy pairwise comparisons between M1C and each of the prefrontal areas independently. Then, a gene was considered to be upregulated in PFC if (1) it is upregulated in a prefrontal area in at least one of the comparisons; (2) it is not upregulated in M1C in any of the comparisons; and (3) it is under-expressed in M1C in at least three of the comparisons. A gene was considered to be upregulated in M1C if (1) it is overexpressed in MOp in at least three of the comparisons; (2) it is not upregulated in any PFC area; and (3) it is under-expressed in a prefrontal area in at least one of the comparisons. Principal component analyses were performed using the prcomp function in R by centring the log2-transformed expression data of the selected genes. Significant GO terms were obtained with the goana function from the limma package in R, reporting the ones with at least 10 genes in the background and at least 5 in the dataset. Frontal lobe, PFC and M1C upregulated genes were characterized for association with RA signalling, autism spectrum disorder (ASD) and schizophrenia (SCZ) using the following criteria. Association with RA signalling was defined both by dysregulation in the PD 0 Rarb and Rxrg double-knockout frontal cortex RNA-sequencing dataset and by literature review identifying association with RA signalling. Association with ASD was based on Association with SCZ was based on a total score of 3 or higher on Information is provided in Supplementary Table 3.

Predicted ages for macaque samples were calculated via the TranscriptomeAge algorithm described previously19. To perform statistical comparisons, samples from various developmental periods were grouped (periods 4–6, 7–10 and 11–14) and a two-tailed Student’s t-test was used to compare gene expression levels between brain regions and species.


All studies using mice (Mus musculus) and rhesus macaques (Macaca mulatta) were performed in accordance with protocols approved by Yale University’s Institutional Animal Care and Use Committee and National Institutes of Health (NIH) guidelines. The animals were housed, and timed-pregnant prenatal and post-natal mouse and monkey brains were obtained in-house at the Yale Animal Resource Center.

Mice were reared in group housing less than five mice per cage at 25 °C and 56% humidity in a 12-h light:12-h dark cycle and provided food and water ad libitum with veterinary care provided by Yale Animal Resource Center. Both sexes were used and randomly assigned for all experiments. Animals were maintained on the C57BL/6J background. The day on which a vaginal plug was observed was designated as PCD 0.5 in mice. For timed pregnancies in monkeys, females were housed with males for three days and the middle day was designated after observation of a vaginal plug and subsequent pregnancy as PCD 1. By this method, the estimated age of a monkey fetus has a maximal variation of ±1 day in its 165-day gestation. All monkeys tested negative for herpes B virus and tuberculosis. RARE-lacZ (Tg(RARE-Hspa1b/lacZ)12Jrt) mice and timed-pregnant CD-1 mice for in utero electroporation were purchased from Jackson Laboratory and Charles River Laboratories, respectively. Sex of mouse samples in this study was not characterized. 

Post-mortem human and macaque tissue

This study was conducted using post-mortem human brain specimens or RNA-seq data generated previously17 from tissue collections at the Department of Neuroscience at Yale School of Medicine, the Brain and Tissue Bank for Developmental Disorders at the University of Maryland, the Clinical Brain Disorders Branch of the National Institute of Mental Health, the Human Fetal Tissue Repository at the Albert Einstein College of Medicine, the Birth Defects Research Laboratory at the University of Washington (R24HD000836), Advanced Bioscience Resources and the Joint MRC–Wellcome Trust Human Developmental Biology Resource (; MR/R006237/1). Tissue was collected after obtaining parental or next of kin consent and with approval by the institutional review boards at each institution from which tissue specimens were obtained, the Yale University and the NIH. Donated deidentified tissue was handled in accordance with ethical guidelines and regulations for the research use of human brain tissue set forth by the NIH ( and the WMA Declaration of Helsinki ( All available non-identifying information was recorded for each specimen. No obvious signs of neuropathological alterations were observed in any of the human or macaque specimens analysed in this study. The post-mortem interval was defined as hours between time of death and time when tissue samples were fresh-frozen or started to undergo fixation process. We strived to keep warm ischaemia to a minimum and most of the post-mortem interval was composed of cold ischaemic time.

Generation of Rarb, Rxrg and Cyp26b1 knockout mice using the CRISPR–Cas9 gene-editing technique

The overall strategy for the generation of Rarb and Rxrg KO mice follows a previously described protocol using the CRISPR–Cas9 genome editing technique41. For the construction of the templates of guidance RNAs, two sets of top and bottom strand oligomers (see Supplementary Table 4) directing the double-strand break at targeting sites were annealed and ligated into the BbsI site of a pX330-U6-Chimeric_BB-CBh-hSpCas9 vector42, which was purchased from Addgene (plasmid 42230). After amplification of the insert with T7-tagged primers (Supplementary Table 5), guidance RNAs were synthesized by T7 RNA polymerase. The coding region of Cas9 was PCR-amplified using pX330-U6-Chimeric_BB-CBh-hSpCas9 as a template and inserted into the pSP64 Poly(A) vector (Promega, P1241). Vectors were digested and linearized with EcoRI. Cas9 mRNA was synthesized by SP6 RNA polymerase. Guidance RNAs and Cas9 mRNA was purified by the MEGAclear Transcription Clean-Up Kit (Ambion, AM1908). Cas9 mRNA and two guidance RNAs were mixed at a concentration of 10 ng μl−1 or 100 ng μl−1 in the microinjection buffer (5 mM Tris-HCl pH 7.5; 0.1 M EDTA) and injected into the pronuclei of fertilized eggs from the B6SJLF1/J mouse strain purchased from The Jackson Laboratory. The fertilized eggs were then transferred to the uterus of females of the CD-1 mouse strain, purchased from Charles River Laboratories. The first generation (F0) mice with recombined alleles were identified by PCR with two primer sets designed outside and inside of the targeted area (Supplementary Table 5, Extended Data Fig. 9d), confirmed by genome DNA sequencing. The germline transmission in the F1 generation was confirmed by the same sets of PCR primers. For generation of Rxrg KO mice, a pair of guidance RNAs flanking whole exon 3 and 4 were designed to delete a large part of the DNA-binding domain43 (Extended Data Fig. 8d). For generation of Rarb KO mice, a pair of guidance RNAs was designed to delete the whole of exon 9 and a part of exon 10 (Extended Data Fig. 8d). As a result, α-helical sheets of H4 to H8 in the ligand-binding domain were deleted, and a frameshift occurred in the rest of the C-terminal region, which results in total abolition of receptor activity43. For generation of Cyp26b1 KO mice, a pair of guidance RNAs was designed to delete the whole of exon 3 and 6, as described previously, to abolish enzymatic activity44 (Extended Data Fig. 18a). All primer sequences are listed in Supplementary Table 5.

Human primary neocortical cultures and differentiation

Fresh tissues from prenatal human brain specimens (PCW 8-HSB#765, sex not determined; PCW 20-HSB#781, female; PCW 23-HSB#784, male) were maintained in ice-cold Hibernate-E (Thermo Fisher Scientific, A1247601) and processed within 12 to 18 h post-mortem interval. Primary cortical neural stem cells from PCW 8 cortical tissue were isolated from the dissected neocortical proliferative zones (that is, the ventricular zone and subventricular zone). Primary cortical neural progenitors and neurons from PCW 20 or PCW 23 cortical tissue were isolated from dissected neocortical walls, including cortical plate zones. In brief, the tissue was mechanically separated into small pieces, incubated with 2 mg ml−1 papain (Brainbits, PAP) for 20 min, and gently triturated to a single-cell suspension with 0.1 mg ml−1 DNase I (STEMCELL Technologies, 07900) in HBSS (Thermo Fisher Scientific, 88284). Cortical neural stem cell expansion and differentiation were performed as previously described45. Cells were plated onto poly-l-ornithine/laminin-coated wells at a density of 2 × 105 cells in a 24-well plate (IBIDI) and cultured with DMEM/F12 supplemented with 1× N2 (Thermo Fisher Scientific, 17502048), 1× B27 (Thermo Fisher Scientific, 17504044), 10 ng FGF2 (R&D Systems, 3718-FB) and 1% penicillin–streptomycin (Thermo Fisher Scientific, 15140122). The expanded cortical stem cells were replated onto poly-l-ornithine/laminin-coated wells at a density of 1 × 105 cells in a 24-well plate, and differentiation to neurons was induced two days after plating by FGF2 withdrawal. The isolated cortical neural progenitors and neurons were plated onto poly-l-ornithine/laminin-coated wells at a density of 2 × 105 cells in a 24-well plate (Ibidi) and cultured without FGF2. On day 4 after the FGF2 withdrawal of cortical neural stem cell culture and day 2 of cortical progenitor and neuron culture, the medium was replaced with neurobasal medium (Thermo Fisher Scientific, 21103049) containing 1× N2 (Thermo Fisher Scientific, 17502048), 1× B27 (Thermo Fisher Scientific, 17504044), 10 ng ml−1 BDNF (Abcam, 9794), 10 ng ml−1 GDNF (R&D Systems, 212-GD) and 1% penicillin–streptomycin (Thermo Fisher Scientific, 15140148). After neural differentiation and further maturation (Extended Data Fig. 15a), cortical neurons were exposed to varying doses of all-trans retinoic acid (Sigma-Aldrich, R2625-50MG) or the pan-retinoic acid receptor antagonist AGN193109 (Tocris, 5758) for another 14 days. See Extended Data Fig. 15a for a summary.

Human and chimpanzee cerebral organoid culture

All human and chimpanzee cell lines were authenticated by morphology or genotyping, and tested negative for mycoplasma contamination, checked monthly using the MycoAlert Mycoplasma Detection Kit (Lonza). For maintenance of human (cell line: HSB311 #36 (refs. 46,47)) and chimpanzee (cell line: 3651D (ref. 48)) cell lines, induced pluripotent stem (iPS) cells were dissociated to single cells with Accutase (Thermo Fisher Scientific, 00-4555-56) and plated at a density of 1 × 105 cells per cm2 in Matrigel (BD)-coated 6-well plates (Falcon) with mTeSR1 (STEMCELL Technologies, 85850) containing 5 μM Y27632, ROCK inhibitor (Sigma-Aldrich, Cat. SCM075). ROCK inhibitor was removed at 24 h after plating, and cells were cultured for another four days before the next passage. Cerebral organoids were generated by directed differentiation protocol as previously described45,46,47. Human and chimpanzee iPS cells were dissociated into single cells using Accutase (Thermo Fisher Scientific, 00-4555-56). Neural induction was directed by dual SMAD and WNT inhibition using neural induction medium supplemented with 100 nM LDN193189 (STEMCELL Technologies, 72147), 10 μM SB431542 (Selleck Chemicals, S1067) and 2 μM XAV939 (Sigma-Aldrich, X3004-5MG). The dissociated cells were reconstituted with the neural induction medium and plated at 10,000 cells per well in a 96-well v-bottom ultra-low-attachment plate (Sumitomo Bakelite). To increase the cell survival and aggregate formation, 10 μM Y-27632 (Sigma-Aldrich, SCM075) was added for the first day. After 10 days of stationary culture, organoids were transferred to a 6-well ultra-low-attachment plate (Millipore Sigma) and kept on an orbital shaker rotating at a speed of 90 rpm to enhance the nutrient and gas exchanges. From day 18, organoids were cultured with neural differentiation medium supplemented with 1× CD lipid concentrate (Thermo Fisher Scientific, 11905031), 5 μg ml−1 heparin (STEMCELL Technologies, 07980), 20 ng ml−1 BDNF (Abcam, 9794), 20 ng ml−1 GDNF (R&D Systems, 212-GD), 200 μM cAMP (Sigma-Aldrich, 20–198) and 200 μM ascorbic acid (Sigma-Aldrich, A92902) for further neuronal maturation (Extended Data Fig. 10a). On day 133, an RA-soaked bead was attached to each cerebral organoid by embedding with and solidifying the growth-factor-reduced Matrigel (Corning, 354230) and further cultured for 48 h before collection.

Preparation of RA-soaked beads

AG1-X2 Resin (150 or 200 μm in diameter, Bio Rad, 140–1231) were added with 1 M formic acid for 1 h and washed with distilled H2O (5 × 5 min). Beads were completely dried in a 37 °C incubator overnight. Dried beads were soaked in either 1 mg ml−1 or 5 mg ml−1 of all-trans retinoic acid (Sigma-Aldrich, R2625-50MG) dissolved in DMSO (Sigma-Aldrich, D8418) for 1 h and then washed twice in DMEM (Gibco) before being placed onto human or chimpanzee organoids.

In situ hybridization

Whole-mount and section in situ hybridization was performed as described previously49. Antisense digoxigenin (DIG)-labelled RNA probes were synthesized using DIG RNA Labeling Mix (Roche, 11277073910). Human and mouse ALDH1A1 (Clone ID 2988388 and 6477503, respectively), ALDH1A3 (Clone ID 6208628 and 6515355, respectively), RXRG (Clone ID 4635470 and 5707723, respectively), RARB (Clone ID 30341884 and 30608242, respectively) and mouse Rorb (Clone ID 5358124), Cbln2 (Clone ID 6412317), Cyp26b1 (Clone ID 6400154) cDNAs were purchased from GE Healthcare for template preparation. Mouse Meis2 DNA was a gift from J. Rubenstein. For macaque in situ hybridization, human probes were used because of high similarity between human and macaque transcripts (97.7% identity in ALDH1A1; 95.4% identity in ALDH1A3; 97.7% in RXRG; and 98.8% in RARB). Sections were obtained from PCW 21, 22 human brains and PCD 110, 140 macaque brains. In situ hybridization were repeated using these two sets. Images were taken using an Aperio CS2 HR Scanner (Leica) and processed by Aperio ImageScope (v., Leica). In Fig. 5a, Cbln2, Cyp26b1 and Rorb expression data were merged by converting colour whole-mount in situ hybridization into black-and-white images, then merged as separate RGB channels. Images were aligned manually by N.S. In Fig. 5f, Extended Data Fig. 19a, the cortical plate was divided into five equal bins and Rorb intensity in each bin was quantified using ImageJ (v.2.0.0-rc-69/1.52p).


Eleven neocortical areas were dissected from four fresh frozen post-mortem human mid-fetal brains (PCW 16, 18, 18 and 19) and four fresh frozen macaque brains (all four PCD 110) as described previously19. Twelve neocortical areas were dissected from three fresh mouse brains at PD 1, based on Paxinos50 and the Allen Mouse Brain Atlas51 ( Each human and macaque brain sample was further microdissected into three pieces and weighed. Each piece was independently homogenized using a Dounce homogenizer in three to four volumes of homogenizing buffer (isopropanol:ethanol 2:1; 1 mg ml−1 butylated hydroxytoluene), followed by centrifugation at 10,000 rpm for 10 min at 4 °C. Supernatant was used for the determination of both protein concentration by Pierce BCA Protein Assay Kit (Thermo Fisher Scientific, 23227), and RA concentrations using an ELISA colorimetric detection kit according to the manufacturer’s instructions (MyBioSource, MBS705877). This kit could not distinguish all-trans retinol, all-trans retinal or all-trans retinoic acid). Thus, the concentration of RA was the concentration of all forms.

Quantitative PCR with reverse transcription

Total RNA was isolated from freshly microdissected cortices after removal of the olfactory bulb and striatum (Extended Data Fig. 9a, b) using Trizol (Thermo Fisher Scientific, 15596026). cDNAs were prepared using SuperScript II (Thermo Fisher Scientific, 18064022) from three independent wild-type cerebral hemispheres. Quantitative PCR with reverse transcription was performed as described previously52 using the 7000HT Sequence Detection System (Applied Biosystems). At least three replicates per transcript were used for every reaction. The copy number of transcripts was normalized against the housekeeping TATA-binding protein (TBP) transcript level. For Rxra, Rxrb, Rxrg, Rara, Rarb, Rarc and Tbp primer sets, the correlation (R2) was higher than 0.98, and the slope was −3.1 to −3.6 in each standard curve. Primers to detect the expression of the genes above were designed in a single exon. Primer sequences are listed in Supplementary Table 5.

β-Galactosidase histochemical staining

Brains were dissected from PD 0 RARE-lacZ mouse pups and drop-fixed in 4% paraformaldehyde for 2 h at 4 °C, followed by embedding in OCT compound (Thermo Fisher Scientific, 23-730-572). Brains were sectioned at 20 µm by cryostat (Leica CM3050S) after they were frozen. β-Galactosidase staining followed the protocol described previously53. We used Red-gal (Sigma-Aldrich, RES1364C-A102X) for the chromogenic reaction. The intensity of β-galactosidase staining was quantified using ImageJ (v.2.0.0-rc-69/1.52p).

Nissl staining

Post-mortem brains were immersion fixed in 4% paraformaldehyde overnight at 4 °C, cryoprotected in 30% sucrose, embedded in OCT compound (Thermo Fisher Scientific, 23-730-572) and snap-frozen in dry ice chilled isopentane (2-methylbutane). Brains were sectioned at 15–20 µm by cryostat (Leica, CM3050S) after they were frozen. After PBS wash, sections were dehydrated using an increasing concentration of ethanol, followed by Cresyl violet, wash, and a second ethanol dehydration. Sections were cover-slipped with Permount (Thermo Fisher Scientific, 15820100).


Human (PCW 12, PCW 16, PCW 19, PCW 20, PCW 21, PCW 22, PCW 24, newborn, and adult (42, 47 and 79 years old; post-mortem interval (PMI) 5–15 h)), macaque (PCD 40, PCD 76, PCD 114, PCD 140, adult (4.5, 8 and 11 years old; PMI under 1 h)) and mouse (PD 0 and adult (4 months)) brain tissue samples, the cultured human cortical neurons and the cerebral organoids were fixed in 4% paraformaldehyde at 4 °C and cryoprotected in graded concentrations of sucrose (15%, 20%, 30%). Brain tissues and cerebral organoids were embedded in OCT compound (Thermo Fisher Scientific, 23-730-572) and snap-frozen in dry ice chilled isopentane (2-methylbutane). Embedded samples were sectioned at 15–20 µm for mouse brain and cerebral organoids, and 50 µm for macaque and human brain by cryostat (Leica, CM3050S). Tissue sections were mounted and dried overnight.

For the immunohistochemistry of mouse brain in Extended Data Fig. 16d–f, sections were treated with or without R-Buffer AG pH 6.0 (Electron Microscopy Sciences, 62707-10) for antigen retrieval, followed by washing in PBS (3 × 5 min) and incubation in blocking buffer (5% (v/v) normal donkey serum (Jackson ImmunoResearch Laboratories, 017-000-121), 1% (w/v) bovine serum albumin, (Sigma-Aldrich, A9647-100G), and 0.3% (v/v) Triton X-100 in PBS) for 1 h at room temperature. Sections were incubated for 24–48 h at 4 °C with diluted primary antibodies in blocking buffer. Sections were then washed in PBS (3 × 5 min) incubated with fluorescent secondary antibodies for 2 h at room temperature. After washing in PBS (3 × 5 min), Tissue was cover-slipped with Vectashield (Vector Laboratories, H-1000). For the immunohistochemistry of human, macaque and mouse brain sections in Extended Data Figs. 4d, e, 7b, antigen retrieval was performed using R-Buffer A pH 6.0 (Electron Microscopy Sciences, 62706-10) for post-natal tissue or R-Buffer AG pH 6.0 (Electron Microscopy Sciences, 62707-10) for prenatal tissue. Sections were incubated in 1% hydrogen peroxide/PBS to quench endogenous peroxidase activity. Sections were washed in PBS (3 × 10 min) and incubated in blocking solution containing 5% (v/v) normal donkey serum (Jackson ImmunoResearch Laboratories, 017-000-121), 1% (w/v) bovine serum albumin (Sigma-Aldrich, A9647-100G), and 0.3% (v/v) Triton X-100 in PBS for 1 h at room temperature. Primary antibodies were diluted in blocking solution and incubated with tissues sections overnight at 4 °C. Sections were washed with PBS (3 × 10 min) before being incubated with the appropriate biotinylated or fluorescent secondary antibodies (Jackson ImmunoResearch Labs) for 1.5 h at room temperature. All secondary antibodies were raised in donkey and diluted at 1:250 in blocking solution. Sections were subsequently washed in PBS and incubated with avidin–biotin–peroxidase complex (Vectastain ABC Elite kit; Vector Laboratories) for 1 h at room temperature. Finally, sections were washed in PBS (3 × 15 min) and signals were developed using a DAB Peroxidase (HRP) Substrate Kit (Vector Laboratories, NC9276270) according to the manufacturer’s protocol. Following washes in PBS, sections were dried, dehydrated and cover-slipped with Permount (Thermo Fisher Scientific, 15820100). For immunofluorescence, sections were washed in PBS with 0.3% (v/v) Triton X-100 and treated with the Autofluorescence Eliminator Reagent (Millipore, 2160) according to the manufacturer’s instructions, and cover-slipped with Vectashield (Vector Laboratories, H-1000). For the immunocytochemistry shown in Extended Data Fig. 15, the fixed primary cortical neurons were washed in PBS (3 × 15 min) and incubated with 0.3% (v/v) Triton X-100 and 10% (v/v) donkey serum (Jackson ImmunoResearch Laboratories, 017-000-121) in PBS for 40 min. For the immunohistochemistry shown in Extended Data Fig. 11a, b, the cerebral organoid sections were washed in PBS (3 × 15 min) and incubated with 0.5% (v/v) Triton X-100 and 10% (v/v) donkey serum (Jackson ImmunoResearch Laboratories, 017-000-121) in PBS for 2 h. Primary cortical neurons and the organoid sections were incubated with primary antibodies in 10% (v/v) donkey serum at 4 °C overnight. Samples were then washed in PBS (3 × 5 min), and incubated with fluorescent secondary antibodies for 2 h at room temperature in 10% (v/v) donkey serum. All sections and tissues processed for immunofluorescence were cover-slipped with Vectashield (Vector Laboratories, H-1000), after washing in PBS (3 × 5 min).

The sources of primary antibodies were anti-ALDH1A1 (1:500, Abcam, 52492), anti-BCL11B/CTIP2 (1:500, Abcam, ab18465), anti-cleaved caspase3 (1:500; Cell Signaling, 9611), anti-CUX1 (1:500, Santa Cruz Biotechnology, sc-13024), anti-GFAP (1:2,000, Sigma-Aldrich, G3893), anti-NR4A2 (1:100, R&D systems, AF2156), anti-GAD1 (1:50, R&D systems, AF2086), anti-NRGN (1:50, R&D systems, AF7947), anti-PAX6 (1:200; R&D Systems, AF8150), anti-POU3F2/BRN2 (1:500, Sigma-Aldrich, SAB2702086), anti-DLG4/PSD95 (1:500; Invitrogen, 51-6900), anti-SYP (1:2,000, Sigma-Aldrich, SAB4200544), anti-L1CAM (1:500; Millipore-Sigma, ABT143), anti-RORB (1:500, Novus Biologicals, NBP1-82532), anti-SATB2 (1:200, Abcam, ab92446), anti-SOX2 (1:200, R&D Systems MAB2018), anti-TBR1 (1:200, Abcam, ab31940), anti-TH (1:1,000, Immunostar, 22941), and anti-TUBB3 (1:400; TuJ1 clone, Abcam, 18207). Secondary antibodies: Alexa Fluor 488-, 594-, or 647-conjugated AffiniPure Donkey anti-IgG (1: 200; Jackson ImmunoResearch). For all microscopic analysis, images were acquired using an Aperio CS2 HR Scanner (Leica), ApoTome.2 microscope (Zeiss), LSM510 META (Zeiss), LSM800 or LSB880 confocal microscope (Zeiss) and assembled in Aperio ImageScope (Leica), Zeiss ZEN2009, ImageJ (v.2.0.0-rc-69/1.52p), Adobe Photoshop (v.12.0 x64), and Adobe Illustrator (v.23.1.1).

Western blotting

Five prospective neocortical areas (mPFC, dlPFC, vlPFC, oPFC and M1C) were dissected from fresh frozen brains of human and macaque, and three neocortical areas (mPFC, MOs/p, and OFC) from fresh mouse brains. Dissected tissues were lysed by sonication in RIPA buffer with protease inhibitors (Sigma-Aldrich, 11836153001). Following centrifugation at 10,000g, supernatant protein concentrations were quantified using the Pierce BCA Protein Assay Kit (Thermo Fisher Scientific, 23227). Supernatants were mixed with 4× Laemmli Sample Buffer (Bio Rad, 1610747), boiled, and electrophoresed on NuPage Bis-Tris gels (Thermo Fisher Scientific), followed by protein transfer to PVDF membranes (Bio Rad, 1620174). Blotted membranes were incubated in blocking buffer (TBS/5% milk) for 1 h and then transferred to blocking buffer with primary antibodies diluted at 1:1,000 at 4 °C overnight. Membranes were then washed in TBS/0.1% Tween20 (3 × 5 min), and incubated with secondary antibodies conjugated with HRP. Secondary antibodies were diluted at 1:1,000. After washing with TBS/0.1% Tween20 (3 × 5 min), protein bands were visualized using SuperSignal West Pico Chemiluminescent Substrate (Thermo Fisher Scientific, 34580). The sources of primary antibodies were anti-ALDH1A3 (1:2,000, Novus Biologicals, NBP2-15339) and anti-GAPDH (1:5,000, Abcam, 9485). Of note, two bands were identified using anti-ALDH1A3 antibody, a 56-kDa band likely to be ALDH1A3 and a 54 kDa band likely to be ALDH1A1.

Droplet digital PCR for quantification of gene expression

An aliquot of the total RNA that was extracted from meninges (human: PCW 19, 19 and 20; macaque: PCD 80, 80 and 110) or primary neural cultures were used for secondary validation through droplet digital PCR (ddPCR) analysis. Two hundred nanograms of total RNA was used for cDNA synthesis using SuperScript III First-strand synthesis Supermix (Thermo Fisher Scientific, 18018400) and subsequently diluted with nuclease-free water. ddPCR was carried out using the Bio-Rad QX100 system. After each PCR reaction mixture consisting of ddPCR master mix and custom primers/probe set was partitioned into 15,000–20,000 droplets, parallel PCR amplification was carried out. End-point PCR signals were quantified and Poisson statistics were applied to yield the target copy number quantification of the sample. A two-colour PCR reaction was used for the normalization of gene expression by the housekeeping gene TBP. All pre-designed assays used in validation can be found in Supplementary Table 5. For each region, the gene expression was compared between species by one-way ANOVA followed by Dunnett’s multi comparison test.

Quantification of post-synaptic and presynaptic puncta marked by immunostaining

For each region of both wild-type and dKO mice, using the 488-nm or 594-nm channels to detect synaptophysin (SYP) immunolabelling, DLG4 (PSD95) immunolabelling and DAPI nuclear counterstain, seven serial optical sections at 0.8-μm intervals over a total depth of 5 μm were imaged and the 2nd, 4th and 6th images were eliminated from further analysis to avoid overlap in counting54. The area of each image is 0.079 mm2. The number of SYP- and PSD95-immunolabelled puncta on each image was counted using ImageJ (v.2.0.0-rc-69/1.52p) using the automated Analyze Particles function using a threshold of 985 to 4,095, determined based on multiple wild-type and dKO images. At least two sections from each mouse were selected for counting, and at least three mice for each genotype were used. The total number and volume of SYP- and DLG4 (PSD95)-immunolabelled puncta and TuJ1 (anti-TUBB3 antibody) or MAP2 in cultured neurons were analysed with 20 z-stack (0.49 μm intervals) images per each 8 fields for each condition and the total number of DLG4 (PSD95)-immunolabelled puncta and TuJ1 immunolabelled neurons in cerebral organoids were analysed with 5 z-stack images using Volocity (v.6.3.1) and Spotfire (v.11.2.0) software.

Retrograde neuronal tracing with adeno-associated viruses

In brief, wild-type (n = 4) and Rarb;Rxrg dKO (n = 4) mice were anaesthetized by injecting ketamine/xylazine solution and head-fixed in the stereotactic frame. Mice were injected with burenorphine 30 min before the surgery. After lubricating the eyes and shaving the fur, an incision of less than 1 mm was made. A craniotomy was made with the round 0.5-mm drill bit at the desired co-ordinates (mPFC: ML ± 0.35, AP 1.5, DV 2.5 from bregma). Using a 0.5-µl Hamilton neuros syringe, we injected 50 nl of AAVrg-Cag-Gfp (Addgene, 37825-AAVrg) into the mPFC. To prevent the virus from spreading along the injecting tract, the needle was held in place for at least 10 min. After injections, the skin was sutured and mice were returned to the cage. After surgery, mice were injected with meloxicam for 48 h. Three weeks later, the mice were euthanized and brains were collected. The brains were coronally sectioned on a vibratome to obtain 70-µm thick sections. After staining the sections with anti-GFP antibody (1:500, Abcam, ab13970) and DAPI the sections were imaged with an LSM 800 microscope (ZEISS). The intensity and density of labelling was quantified on a scale of 0 to 3 with 0 being no labelling, 1 being weak or sparse cellular labelling (less than 10%), 2 being strong and less than 50% of cell labelled, and 3 being dense labelling. This discrete approach of quantification was used owing to variability in injection site.

Quantification of dendritic spines and arborization from viral tracing sections

The contralateral mPFC to the injection site was used for this analysis; wild type (n = 4) and Rarb;Rxrg dKO (n = 4). For Sholl analysis, the images of entire neurons were acquired at 20× magnification. For spine counts, z-stack images across entire dendritic thickness with 29–33 images per stack were obtained. The z-stack images were opened in Reconstruct (v., which is publicly available at, and a new series was recreated that enabled us to move across different stacks across z-planes in the same image. Whole dendrite was subdivided into segments of 10 µm and number of spines across whole thickness were traced for length and breadth of each spine. After tracing, the length and breadth of the spines their ratio was used to determine the spine subtype as described before56. After the analysis for each class of spine, standard deviation and P values were calculated using two-way ANOVA with Sidak’s multiple comparison method. For Sholl analysis, z-stack images were opened in ImageJ (v.2.0.0-rc-69/1.52p) and dendritic arbors were manually traced using the NeuronJ plugin57. Dendritic complexity was then quantified and plotted using the Sholl Analysis option.

Plasmid construction

For construction of expression vectors, full-length cDNAs (mouse Aldh1a3, Clone ID 6515355, purchased from GE Healthcare) were inserted into pCAGIG vector (pCAGIG was obtained from Addgene (plasmid 11159).

In utero electroporation

In utero electroporation was performed as previously described52. Plasmid DNA (4 μg μl−1) was injected into the lateral ventricle of embryonic mice at E13.5–E14.5 and transferred into the cells of the ventricular zone by electroporation (five 50-ms pulses of 40 V at 950-ms intervals). Brains were dissected at PD 0. Brains and tissue sections of electroporated mice were analysed for GFP expression after fixation with 4% paraformaldehyde for 12 h at 4 °C.

Mouse RNA-seq data generation and analysis

Post-mortem mouse brains were dissected at PD 0 in ice-cold sterile PBS, fresh frozen, and stored at −80 °C. Brains were incubated in RNAlater-Ice at −20 °C for 12–16 h prior to further dissection. mPFC, MOs, and OFC were microdissected based on Paxinos50 and the Allen Mouse Brain Atlas51 ( and RNA was isolated using RNeasy Plus Micro kit with additional on-column DNAase step (Qiagen, 74134). RNA quality and amount were quantified using the High Sensitivity RNA Screen Tape Assay (Agilent, 5067-5579), and the concentration was standardized to 10 ng μl−1. The SMART-seq v4 Ultra Low Input Kit (Takara Bio, 634890) was used to create cDNA, and the concentration was quantified using Quant-iT PicoGreen dsDNA Assay Kit (Thermo Fisher Scientific, P11496). Nextera XT DNA library Preparation Kit (Illumina, FC-131-1024) was used to create cDNA libraries for sequencing. Libraries were normalized and sequenced at the Yale Center for Genomic Analysis (YGCA) using the NovaSeq with 100-bp paired end reads. Reads from each library were mapped against the mouse assembly GRCm38 using STAR v.2.6.0a (gtf and fasta files downloaded from Ensembl version 94; parameters: --readFilesIn $j1 $j2 --outSAMattributes All --outFilterMultimapNmax 1 --outSAMstrandField intronMotif --outFilterIntronMotifs RemoveNoncanonical --quantMode TranscriptomeSAM --outFilterMismatchNoverLmax 0.1 --alignSJoverhangMin 8 --alignSJDBoverhangMin 1 --outSAMunmapped Within --outFilterType BySJout). Counts were obtained using featureCounts v1.6.2 with -p parameter.

To compare the gene expression patterns of three wild-type versus three Rarb and Rxrg dKO mice, a TMM procedure was applied (function normalizeCounts from tweeDEseq package in R) to the expression of 15,085 protein-coding genes that show sufficiently large counts (determined with function filterByExpr from edgeR package in R). We assessed differentially expressed genes in each brain region (mPFC, OFC and MOs) running RNentropy independently among wild-type and dKO mice per region. Genes overexpressed in a given condition are those that are both significantly upregulated in that condition and significantly downregulated in the opposite condition according to RNentropy. The same criterion was applied for the identification of downregulated genes. Genes with an inconsistent pattern of expression between regions were excluded. Principal component analyses were performed using the prcomp function in R by centring the log2-transformed expression data of the selected genes. Significant GO terms were obtained with the goana function from the limma package in R and plotted using the function GOBubble from the GOplot package in R. Fisher test enrichments calculated for RA-related genes (RA synthesis: Rdh10, Rdh5, Aldh1a1/Raldh1, Aldh1a2/Raldh2, Aldh1a3/Raldh3, Adh1, Adh5, Adh7; RA degradation: Cyp26a1, Cyp26b1, Cyp26c1; RA receptors: Rara, Rarb, Rarg, Rxra, Rxrb, Rxrg; RA binding: Ttr, Rlbp1, Rbp1, Rbp2, Rbp3, Rbp4, Fabp5), genes overexpressed in individual lobes of the mid-fetal human cortex based on Fig. 1, and neuropsychiatric-disease related genes (downloaded from ref. 30.) in up- and downregulated genes. Genes associated with the GO terms: ‘axon guidance’, ‘axon guidance receptor’, ‘axon development’ and ‘ephrin’ were manually screened for anterior-to-posterior gradient using developingmouse.brain-map.org58, gensat.org59 and a previous report60. Mouse RNA-seq data were deposited into the NCBI Gene Expression Omnibus (GEO) database (accession number GSE142851).

Diffusion-weighted MRI and tractography

Five PD 5 post-mortem dKO homozygotes and five wild-type C57BL/6 mouse brains were drop-fixed in 4% paraformaldehyde solution in 0.1 M PBS for 48 h. They were subsequently transferred to 0.1 M PBS and just before imaging to Fomblin (Sigma-Aldrich, 317926). The diffusion-weighted images were acquired on a Bruker BioSpin 9.4 T MRI (Bruker) using a standard 3D Stejskal-Tanner spin-echo sequence with 30 different angles of diffusion sensitization at a b value of 1,000 s mm−2 and the following parameters: repetition time = 2,000 ms; echo time = 25.616 ms; diffusion encoding duration = 4 ms. The in-plane resolution was 0.11 mm and slice thickness was 0.22 mm. Overall scanning time was around 24 h.

Image processing and tractography

Cerebral cortical regions of interest (ROI) and thalamus were manually defined according to Paxinos50 and the Allen Mouse Brain Atlas51 ( by D.A. and K.P. without prior knowledge of the experimental groups. Image preprocessing was done with Advanced Normalization Tools (ANTs, v. The reconstruction of axonal pathways was performed with MRtrix361 software (v.3.0.0-65-g91788533) using constrained spherical deconvolution62 and probabilistic tracking (iFOD2) with a FOD amplitude cut-off of 0.1. The thalamus was used as a seeding point and each cortical ROI was used as a termination mask. To evaluate the integrity of the major white matter tracts between the groups, both internal capsules, anterior commissure and corpus callosum were manually delineated according to Paxinos50 and the Allen Mouse Brain Atlas51 ( by D.A. and K.P. without prior knowledge of the experimental groups. Values of the fractional anisotropy (FA), apparent diffusion coefficient (ADC), radial (RD) and axial (AD) diffusivity were calculated using underlying scalar maps derived by MRtrix3.

Anterograde tracing of axons

For anterograde tracing of axons between mPFC and thalamus, brains were collected at either PCD 18 or P21, and fixed overnight in 4% paraformaldehyde at 4 °C. Brains were then hemidissected. A crystal of 1,1′-dioctadecyl-3,3,3′,3′-tetramethylindocarbocyanine perchlorate (DiI, Sigma-Aldrich, 24364) was inserted into the mPFC, MD nucleus of the thalamus, or medial thalamus under the stereomicroscope. The size of the crystal is around 200 μm. Brains were then placed in 1% paraformaldehyde in PBS and left for 14 days at 37 °C. Following DiI diffusion, the brains were sectioned coronally on a vibrating microtome (Leica) at 80-μm thickness and stained with DAPI. Sections were mounted onto glass and immediately sealed in VECTASHIELD Hardset Antifade Mounting Medium (Vector Laboratories, H-1400). Slides were analysed under a ApoTome.2 microscope (Zeiss) and intensity was quantified using ImageJ (v.2.0.0-rc-69/1.52p).

Processing, analysis and image visualization

To allow robust visualization and analysis, images depicting DiI tracing or immunohistochemistry using antibody against PSD95 (DLG4) have been inverted and/or pseudo-coloured, as in Figs. 4, 5. In addition, background was removed for in situ hybridization experiments and images were pseudo-coloured and superimposed in Fig. 5, Extended Data Figs. 4, 7, 9, 10 using Adobe Photoshop.

Reporting summary

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