Decoding transcriptional identity during neuron-astroglia cell fate driven by RAR-specific agonists

Abstract

How cells respond to different signals leading to defined lineages is an open question to understand physiological differentiation leading to the formation of organs and tissues. Among the various morphogens, retinoic acid signaling, via the RXR/RAR nuclear receptors activation, is a key morphogen of nervous system development and brain homeostasis. Here, we analyze gene expression in ∼80 000 cells covering 16 days of monolayer mouse stem cell differentiation driven by the pan-RAR agonist all-trans retinoic acid, the RARα agonist BMS753 or the activation of both RARβ and RARγ receptors (BMS641 + BMS961). Furthermore, we have elucidated the role of these retinoids for driving nervous tissue formation within 90 days of brain organoid cultures, by analyzing >8000 distinct spatial regions over 28 brain organoids. Despite a delayed progression in BMS641 + BMS961, RAR-specific agonists led to a variety of neuronal subtypes, astrocytes, and oligodendrocyte precursors. Spatially resolved transcriptomics performed in organoids revealed spatially distinct RAR isotype expression leading to specialization signatures associated with matured tissues, including a variety of neuronal subtypes, retina-like tissue structure signatures and even the presence of microglia.

Graphical Abstract

Graphical Abstract.

Introduction

How cells respond to different signals to develop along defined cell lineages is an open question to understand physiological cell differentiation leading to the formation of organs and tissues but also to unveil the mechanisms that govern aberrant cell transformation leading to disease. Among the various morphogens, all-trans retinoic acid (ATRA), a vitamin A metabolite, play a key role in vertebrate embryogenesis, and notably in the development of the nervous system [1]. ATRA has been shown to act via its interaction with its cognate retinoic acid receptor (RAR) targets, RARα, RARβ, and RARγ, encoded by three different genes and expressed in addition through various isoforms [2]. During mouse development, all three RAR receptors present spatially distinct expression signatures, indicative for specialized functions (reviewed in [1]), and such distinct spatial signatures are also retrieved in the adult brain [2, 3].

Previous studies, including ours, focused on the power of retinoids action for driving cell differentiation in vitro, demonstrated that the use of specific RAR agonists could lead to a variety of cell types [4–6]. P19 embryonal carcinoma cells treated with the RARα specific agonist BMS753 were shown to give rise to neuronal precursors in a period of 48 h, but they not progress in differentiation when treated with the RARβ-specific agonist BMS641 or the RARγ agonist BMS961 [6]. Furthermore, we have previously reported that in 10 days of culture conditions both the RARα agonist BMS753 and the combined use of the RARβ and the RARγ agonists (BMS641 + BMS961) lead to a complex differentiation process generating a variety of neuronal subtypes but also oligodendrocyte precursors and GFAP (+) astrocyte cells [7]. The analysis of the RAR-driven gene programs revealed that such synergistic action is the consequence of the hijacking of RARα-controlled programs, in agreement with previous studies describing a redundancy for the activation of certain genes by distinct RARs, and notably on Rar-null mutant lines [5, 8, 9].

Concerned by the fact that RARs are known to present anatomically distinct expression signatures both in the embryo but also in the adult brain, we aimed in this study to address the capacity of both the RARα specific agonist BMS753 and the synergistic action of the RARβ and RARγ agonists (BMS461 + BMS961) to generate distinct cell types during neuroectodermal differentiation but also to evaluate their potential spatially distinct expression in in vitro three-dimensional brain organoid cultures. We have first shown a large variety of cell types generated during 16 days of mouse stem cell differentiation in presence of either of the aforementioned ligand conditions, thanks to the use of single-cell transcriptomics strategy. Then we have setup mouse brain organoid cultures (3 months) in presence of either vitamin A, ATRA, BMS753, or the combination of BMS461 + BMS961, to further enhance our understanding of the role of distinct RARs for generating distinct cell-types structured within distinct spatial regions.

Materials and methods

Cell culture

Embryonic stem cells (ESCs) were grown in Dulbecco's Modified Eagle Medium (DMEM) supplemented with 4.5 g/l glucose and GlutaMAX-I (11594446, Thermo Fisher Scientific), 15% fetal bovine serum tested from embryonic stem cells (FBS-ES), 5 ng/ml Leukemia inhibitory factor (LIF) recombinant mouse protein (15870082, Thermo Fisher Scientific), 1% penicillin–streptomycin, 1% of Minimum Essential Medium complemented with non-essential amino acids (MEM-NEAA), and 0.02% β-mercaptoethanol.

For cell differentiation assays, ESCs were seeded onto poly-D-lysine–coated culture plates (0.1%) at a concentration of 100 000 cells per well (for cells collected at 4 days of differentiation), 40 000 cells per well (for cells collected at 8 days treated with ATRA), 20 000 cells per well (for cells collected at 8 days treated with ethanol, BMS753, or BMS961 + 641; for cells collected at 16 days treated with ATRA), 10 000 cells per well (for cells collected at 16 days treated with ethanol, BMS753, or BMS961 + 641). The cells are seeded in DMEM (4,5 g/l glucose) w/GLUTAMAX-I, 10% FBS-ES, 1% penicillin–streptomycin, 1% MEM-NEAA, and 0.02% β-mercaptoethanol.

Next day, same medium is refreshed and retinoid ligands are added. ATRA was added for a final concentration of 1 μM. For treatment with RAR subtype–specific agonists, cells were incubated with BMS753 (RARα-specific, 1 μM) or BMS641 + BMS961 (RARβ + RARγ-specific, 0.1 μM each). After 8 days of treatment with either of the aforementioned retinoids, the medium was replaced by Neurobasal medium (11570556, Thermo Fisher Scientific) and B27 without vitamin A (11500446, Thermo Fisher Scientific) and cultured for eight more days. Half medium was refreshed every other day.

Immunocytochemistry

At the end of the cell differentiation assay, medium is aspirated completely and cells are fixed in 4% paraformaldehyde followed by 3 × 5 min washes in Phosphate-buffered saline (PBS). Cells were permeabilized (Triton 0.1% in PBS; 15 min at room temperature), washed 3 × 5 min in PBS and blocked (0.1% Triton and 1% BSA in PBS for 1 h at room temperature). Cells were incubated with the primary antibodies overnight at + 4°C [anti-β III tubulin/anti-TUBB3 (ab14545, Abcam), and anti-MAP2 (ab32454)]. Cells are washed 3 × 5 min in PBS followed by incubation with a secondary antibody for 1 h at room temperature [(donkey anti-mouse IgG [H + L] antibody Alexa 555 (Invitrogen A-31570), and donkey anti-rabbit IgG [H + L] antibody Alexa 488 (Invitrogen A-21206)], respectively. Cells are washed 3 × 5 min in PBS and finally mounted on microscope slides with mounting medium containing DAPI (4′,6-diamidino-2-phenylindole; P36962, Thermo Fisher Scientific).

RNA extraction and cDNA generation

Total RNA was extracted using RNeasy Mini Kit (74104, Qiagen). RNA (0.5–1 μg) was used for reverse transcription (4368814, Applied Biosystems). Transcribed complementary DNA (cDNA) was diluted five-fold and used for real-time quantitative PCR with QuantiTect SYBR Green PCR Kit (204145, Qiagen). Comparative Ct (ΔΔCt) method was used for performing relative quantitation of gene expression. The list of used oligonucleotides is enumerated below.

Single-cell RNA sequencing and processing

The Chromium instrument and single-cell 3′ Reagent Kit v3 (PN-1000128; 10× Genomics) was used to prepare individually barcoded single-cell RNA-Seq whole transcriptome libraries following the manufacturer’s protocol. The target number of cells was 8000 per condition. Libraries were sequenced on Illumina Novaseq 6000 using paired-end 150 cycles v3 kits (Read 1: 28 cycles; Index Read 1 (i7): 10 cycles; Read 2: 90 cycles).

Single-cell transcriptome

Data from nine conditions have been processed using Cell Ranger 7.0.1 for primary analyses, aligned to the mouse reference genome mm10 2020A. Cells were filtered out on the grounds of doublets (DoubletFinder V.2)(McGinnis et al., 2019), >5% of mitochondrial counts content, novelty score >0.8, and cells with <200 genes. These cells from different conditions were merged and integrated by a selected group of genes (200 features). Read counts were normalized by SCTransform method and processed downstream for Uniform Manifold Approximation and Projection (UMAP). Clusters were identified (KNN method) based on gene expression similarities. Differential expression analysis was performed and filtered out genes with a log2 fold change < 0.25 and a P-value > 0.05 using the Wilcoxon test. All these steps were performed in R (R core team 2024) using the package “Seurat” (Version 5) (Hao et al., 2024) and customized scripts.

For Fig. 1E, the fraction of cells per sample assigned to each cluster was calculated and further normalized to the total levels per time-point. These frequencies were clustered by Euclidean distance and displayed in a heatmap (MeV: Multiple Experiment Viewer). Supplementary Table S1A and B show the values. For Supplementary Table S1A, the percentage of cells for each cluster was calculated dividing the number of cells per retinoid treatment & per cluster by the total number of cells per cluster. The addition of all percentage in the cluster (addition by rows) will bring 100%. For Supplementary Table S1B, first, the number of cells per retinoid treatment and per cluster is divided by the number of cells per treatment and multiplied by 100. The addition of all percentage in the treatment-time point (addition by columns) will bring 100%. Second, this obtained percentage was divided by the sum of percentages for each retinoid treatment in different time points (e.g. sum of frequencies for cluster 1 in BMS753 treatment at day 4, 8, and 16) and multiply by 100. In this manner, the information is considered at each time point in order to evaluate changes across time.

Figure 1.

Transcriptome at single-cell level reveals retinoid-driven stem cells specialization differences in a temporal dependency. (A) Scheme illustrating the mouse stem cell differentiation strategy, where cells were treated with different retinoid agonists (all-trans retinoic acid: ATRA, RAR-α agonist BMS753, or combination of the RAR-β agonist BMS641 and the RAR-γ agonist BMS961) during 8 days, then exposed to maturation media (Neurobasal medium: NB, supplemented with B27 devoid of vitamin A). Cells were collected at 4, 8, or 16 days for single-cell transcriptome assays. (B) Representative micrographs assessed in either bright-field (top panels) or immunofluorescence (bottom panels) from cells treated with ethanol (EtOH), ATRA, or the combination of BMS641 and BMS961 analyzed after 16 days of differentiation. Cells were stained for the neuronal precursor marker TUBB3 (red), and the marker for mature neurons MAP2 (green). Nuclei were stained with DAPI (blue). (C) UMAP for all three conditions (ATRA, BMS753, and BMS641 + BMS961) assessed at three time-points (4, 8, and 16 days). Color-code correspond to 17 nonsupervised clusters (D) UMAP illustrated in (C) stratified per treatment and time-point. (E) Heatmap representation of the relative number of cells per cluster and normalized across time-points per treatment. UMAP clusters and conditions are clustered by “Euclidean” distance. (F) Cell-type annotations per corresponding UMAP cluster assessed by hypergeometric probability assay. Colored boxes in (E) and (F) make reference to the group of clusters retrieved associated with either early, intermediate or late stages during differentiation.

In the Fig. 1F, a P-value in a negative logarithmic scale in base 10, scaled ×10 was calculated by a hypergeometric test from the differential expression analysis and the target gene sets, using a customized “R” script. These “gene set” collections were manually curated from different publications or databases (Supplementary Table S2).

Pseudo time analysis was performed using the python package “STREAM” [10] in all the nine samples together (nine treatments and nine time points). To perform a dimension reduction, the next parameters were set up: 2000 variable genes were selected, lineal method “mlle,” 10 principal components and 50 neighbors. The trajectory was inferred by the UMAP. The next parameters were used to tunning the trajectory: epg_alpha = 0.02, epg_mu = 0.1, epg_lambda = 0.02, and epg_ext_par = 0.8 for the branch extension. A flat tree was obtained with a dist_scale = 0.5. Ribbon rooted trees were graphed with the parameter distance scale = 2 showing the 3 treatments together (Fig. 2A), or the different timepoints (Fig. 2B). Percentage of cells contained in each branch by different treatment conditions, or compared to by different clusters from in the UMAP projection in Fig. 1 C were calculated, depicted and clustered (method “complete”) in a heatmap (Fig. 2C and D).

Figure 2.

Pseudo-temporal cells stratification reveals a delayed cell specialization progression in RARβ+RARγ-driven trajectories relative to the RARα activation. (A) Pseudo-time stratification of single-cell transcriptomics data assessed in all three treatments. (B) All 14 branches assessed in panel (A) are displayed per treatment and colored on the grounds of their associated time-points. (C) Fraction of cells per pseudo-time branches (in percent) displayed in the context of their corresponding treatment and the collected time-point. (D) Fraction of cells in pseudo-time branches per UMAP clusters displayed in Fig. 1. In panels (C and D), red square boxes correspond to fraction of cells associated with the BMS641 + BMS961 treatment. (E) Heatmap of cell fate markers covering transitions from intermediate progenitors, toward either astroglia cells, or mature neurons. Gene expression for each gene was normalized along the branches represented in panel (A). (F) Gene expression of four gene markers displayed in the context of the pseudo-time branches organization, as provided by STREAM. Expression levels correspond to normalized counts in logarithmic scale base 2. tB = terminal branch, iB = intermediate branch, R = root.

A hypergeometric test to obtain the probability of finding the highly expressed genes from each branch (showed in the Fig. 2A) in different gene sets was performed. The fold change was calculated dividing the mean value for determinate gene and branch, over the total mean counts for the determinate gene. Genes with fold change >0.5 in logarithmic values on base 2 were selected. The P-values obtained were transformed to a negative logarithmic scale on base 10, scaled ×10, clustered and depicted in Fig. 2E using a customized R script. Details from these gene sets can be found in Supplementary Table S2.

Cut&tag epigenomics assays

Cut&Tag experiments were carried out as previously described (Kaya-Okur et al. [11] and Li et al. [12]), with some modifications. Briefly, 250 000 cells were captured with BioMagPlus Concanavalin A (Polysciences, 86057-3) and incubated with the appropriate primary antibody overnight at 4°C [anti-Histone H3 (trimethyl K4) antibody (Abcam, ab8580; 1:50), anti-Histone H3 (acetyl K27) antibody (Abcam, ab4729; 1:50); anti-Histone H3 (trimethyl K27) antibody (Abcam, ab192985; 1:50); Anti-RNA Pol II monoclonal antibody: diagenode (C15100055) (1:20)]. For the experiment in which the antibody RNA Pol II was used, the mixture cells-beads were fixed with 1% Paraformaldehyde (PFA) in 1× PBS for 15 min before the addition of the primary antibody. A secondary antibody was incubated at room temperature for 1 h (Donkey anti-rabbit IgG antibody: Sigma-Aldrich, SAB3700932; 1:100). Fusion protein ProteinA–Tn5 was prepared in-house (plasmid addgene #124601) by following a previously described protocol (Li et al. [12]). Purified PA-Tn5 was loaded with the following adapters compatible with Illumina sequencing (hybridized with the complementary Mosaic sequence [PHO]CTGTCTCTTATACACATCT):

• MOS-universal: TACACTCTTTCCCTACACGACGCTCTTCCGATCTAGATGTGTATAAGAGACAG

• MOS-index: GTTCAGACGTGT.GCTCTTCCGATCTAGATGTGTATAAGAGACAG

Loaded PA-Tn5 complex (0.16 μM) was incubated for 1 h at RT, followed by tagmentation step for 2 h at 37°C. DNA was extracted by phenol/chloroform (Sigma–Aldrich 77617) followed by ethanol precipitation. Finally, the tagmented DNA was amplified using Illumina primers (NEBNext^® Multiplex Oligos for Illumina, NEB #7335S) with the following cycling conditions: 98°C for 30 s, 20 cycles of 98°C for 10 s, and 65°C for 75 s; then final extension at 65°C for 5 min and hold at 12°C. DNA libraries were cleaned using 1× volume of SPRI-select beads (Beckman Coulter, B23318) and Illumina sequenced (150-nt pair-end sequencing; Illumina NovaSeq 6000).

Cut&tag histone modification sequencing data were aligned to the mouse genome (mm9) using Bowtie2 (v2.4.1) under the default parameters. Chromatin state analysis has been performed with ChromHMM (v1.14). Bedgraph files and peak annotations were obtained with MACS2 and visualized within the IGV genome browser (v.2.4.15).

Mouse brain organoid

Mouse brain organoid protocol has been derived from Lancaster et al., 2013. ESCs should be present in culture as EBs (embryoid bodies). On day 0, the EBs are dissociated with acutase (A6964-100ML; Sigma) for 10 min. Cells are seeded in P96 (ultra-low attachment) plate at 5000 cells/well in 150 μl medium/well [DMEM (4.5 g/l glucose) w/w GLUTAMAX-I, 10% FBS-ES, 1% penicillin–streptomycin, 1% MEM-NEAA, and 0.02% β-mercaptoethanol, SB 431542 (1/1000)] and incubated for 2 days at 37°C. On day 2, same medium is refreshed by taking out 75 μl and adding 150 μl fresh medium per well. On day 4, medium is refreshed by taking out 150 μl medium and adding 170 μl Neural induction medium/well (DMEM/F12, 1% N₂ supplement, 1% glutamax supplement, 1% MEM-NEAA, 1% penicillin–streptomycin, and 0.1% heparin solution). On day 6, the EBs are embedded in Matrigel GF reduced (11523550; Fisher scientific) for 45 min before transferring to P6 (ultra-low attachment) plates in maturation medium [DMEM/F12 medium, Neurobasal medium, 2% B27 without vitamin A supplement, 1% Glutamax supplement, 1% penicillin–streptomycin, 0.5% N₂ supplement, 0.5% MEM-NEAA, 0.025% insulin (I9278-5ML, Sigma)]. On day 9, medium is changed completely to maturation medium and adding 2% B27 with vitamin A supplement (11530536, Thermo Fisher Scientific) or retinoid ligands for a final concentration of [(ATRA (1 μM), BMS753 (1 μM); BMS641, BMS961, BMS641 + 961 (0.1 μM)] and placed on an orbital shaker at 75–80 rpm. Two-thirds of medium is changed twice a week.

Brain organoids cryo-sectioning and spatial transcriptomics library preparation

Organoids are fixed with 4% paraformaldehyde, 6% sucrose for 20 min, washed 3 × 5 min with PBS and suspended in 30% sucrose at 4°C overnight. Several organoids are placed in plastic molds and submerged in OCT (Optimal Cutting Temperature) embedding medium and frozen before being cryo-sectioned using a cryo-star. Frozen brain organoid molds are cryosectioned and deposited on top of in-house manufactured DNA arrays as described previously (Lozachmeur et al. [13]). Briefly, in-house manufactured DNA arrays are composed of a double-barcoded part encoding their unique position within the array, and a poly-T extremity allowing to capture messenger RNA.

During organoids cryosectioning, slides were kept inside the cryostat chamber in order to conserve the integrity of the RNA and at −80°C for long storage. Deposited tissue sections were processed for library preparation as described in our in-house protocol (Lozachmeur et al. [13]). At the end of the process, libraries were used for Illumina sequencing (150 nts paired-ends sequencing; NovaSeq; >200 million reads per library).

Spatial transcriptomics processing

Primary analysis has been performed with our in-house developped tool SysISTD (SysFate Illumina Spatial transcriptomics Demultiplexer: https://github.com/SysFate/SysISTD). SysISTD generates as outcome a matrix presenting read counts per gene IDs in rows and spatial coordinates in columns. To focus the downstream analysis to the physical positions corresponding to the analyzed tissue, we used an in-house R script taking as entry an image of the DNA array scanned with the TRICT filter, releaving the presence of the fiducial borders introduced during the manufacturing of the DNA arrays.

The focused matrix was quantile normalized with our previously described tool MULTILAYER [14, 15]. Then, the normalized matrix was further processed for revealing cell/tissue density contrast information by integrating normalized counts with gexel (gene expression element, in analogy to pixel) value levels retrieved in the TRITC tissue image. The obtained normalized and pixels intensity adjusted matrix was used for all downstream analysis.

Gexels classification were performed by processing the normalized read count matrices to obtain the pseudo-time information using the python package “STREAM” (Chen et al., 2019 [10]), in the same way as the single-cell data (described above) except for the next parameters: For dimension reduction a loess_frac = 0.1, with a previous normalization by “lib_size,” two principal components and 15 neighbors, to tunning the trajectory: epg_alpha = 0.01, epg_lambda = 0.08, epg_trimmingradius = 0.1, and the branching was optimized with incr_n_nodes = 30. Before this step, gexels with total number of counts lower than the first quartile (0.25) of the total read counts in the matrix were filtered out.

Gexels presenting upregulated levels for each retinoid receptor (Rara, Rarb, or Rarg) across treatment samples were selected using the second quartile of the normalized counts from all pixels for each respective gene. The pixels with lower number of counts were filtered out like indicated before.

Gene regulatory network reconstruction

Highly expressed genes for each condition (except for the intersection of downregulated) were analyzed with our previous described tool TETRAMER (Cholley et al. [16]) within the “Cytoscape” platform. Briefly, a gene regulatory network (GRN) was generated by using a collection of transcription factor–target gene relationships (CellNet database; available within TETRAMER), which has been integrated with the differentially expressed genes retrieved within the gexels presenting high levels of Rar receptors.

TETRAMER inferred a master regulator index for each TF retrieved in the network, which were used for reconstructing a gene co-regulatory view among the top ranked factors. The betweenness centrality was calculated for this gene co-regulatory network considering the direction of the edges, this topological parameter provides the centralized structure to the network. Cell/tissue annotations for genes composing this gene co-regulatory network has been performed with GeneMania [17]. Obtained gene associations to the cell/tissue annotations were merged with cell type annotations obtained from Supplementary Table S2, and displayed in a binary heatmap format with MEV.

Statistical analysis

Statistical analysis was performed as described in the figure legends. Hypergeometric test was performed in a customized R script as described by Kachitvichyanukul et al. [18]. Using the equation:

Other statistical tests inherent to the respective tool are described in the user manual (SEURAT, STREAM, TETRAMER, GeneMania, and Multilayer).

Results

Synergistic activation of RARβ and RARγ receptors give rise to a variety of neuronal subtypes in mouse embryonic stem cells

With the aim to address potential cell differentiation differences driven by the action of either the pan-RAR agonist ATRA or the use of RAR specific agonists, we have cultured mouse ESCs in monolayer conditions and treated them with either the pan agonist ATRA, the RARα specific agonist BMS753, or the combination of the RARβ and RARγ agonists; BMS641 + BMS961. After 8 days of continuous treatment, cells were exposed for other 8 days to neurobasal medium presenting the B27 supplement but devoid of vitamin A, to promote cell specialization (Fig. 1A). In aggreement with our previous observations [7], the use of either of these conditions allowed to obtain mature neuronal cell differentiation, as confirmed by immunofluorescence using the mature neuronal marker MAP2 (Fig. 1B).

While in our previous study we have demonstrated that the synergistic activation of the RARβ+RARγ receptors recapitulate neuronal cell differentiation like in the case of the use of the RARα agonist, in this study we focus our attention to the potential differences in neuronal subtypes that can be generated by using different RAR receptors. To address this question we have collected samples during retinoids treatment (4 days), before adding the neurobasal medium (8 days), and at the end of the 16 days of cell differentiation. Single-cell transcriptomic assays performed in 79 596 cells across all samples and treatment conditions were analyzed together by using a Uniform Manifold Approximation and Projection for dimension reduction (UMAP) revealing up to 17 distinct cell clusters (Fig. 1C). Stratification of this global UMAP visualization by treatments and time-points revealed differences between treatment conditions. Notably, cluster 8 has been retrieved in early stages of ATRA treatment (d4: 4%, d8: 19% of cells) and in late stages of BMS641 + BMS961(d8: 2%, d16: 7% of cells), but it is preferentially enriched in BMS753 treatment (d8: 2%, d16: 61% of cells) (Fig. 1D and Supplementary Table S1). Furthermore, cluster 12 is mainly seen in BMS641 + BMS961 treatment (d4: 69% of cells); cluster 4 appears preferentially in ATRA at day 4 (50% of cells) and in BMS641 + BMS961 at day 8 (34% of cells); and cluster 3 is retrieved in BMS641 + BMS961 at day 16 (32% cells) and in ATRA treatment at day 4 (20% of cells) (Fig. 1D and Supplementary Table S1).

A classification strategy applied to the relative enrichment levels of cells per cluster and adjusted across timepoints allowed to highlight a correlation with the timeframe. Indeed, clusters 3, 7, 1, 5, and 12 were preferentially associated with early events; clusters 4, 6, 16, 11, and 14 to intermediate transitions; and cluster 15, 17, 9, 13, 10, 2, and 8 to late events; and this in despite of the treatment in use (Fig. 1E). Importantly, this analysis also revealed discrepancies between ATRA and the other ligand treatments, including a specific ATRA enrichment in cluster 15, and temporally desynchronized behavior relative to ATRA in clusters 10, 2, 8, 16, and 11; indicative of a delayed maturation process driven by the specific ligands relative to the pan-RAR agonist. Furthermore, clear differences between the BMS641 + BMS961 and the BMS753 treatment were observed. Clusters 4 and 6 presented a preferential enrichment in day 8 for BMS641 + BMS961 treatment (cl4-d8: ∼81% cells per condition adjusted to timepoint in BMS641 + BMS961; cl6-d8: ∼82% in BMS641 + BMS961); and cluster 3 and 11 in day 16 (cl3-d16: ∼62%; cl11-d16: ∼46% of cells in BMS641 + BMS961). In contrary, cluster 11 displayed a preferential enrichment in day 8 for BMS753 treatment (d8: 60% BMS753); as well as cluster 17 in day 16 (d16: ∼95% BMS753) (Fig. 1E and Supplementary Table S1).

A cell-type annotation strategy, based on public gene expression signatures, allowed to clearly state that clusters associated with the earlier timepoint corresponded to proliferative precursors (clusters 12, 1, 5, and 3) (Fig. 1F and Supplementary Table S2). The intermediate cluster 6 presented enriched signatures associated with neurons, including Dopaminergic and GABAergic subtypes, as well as oligodendrocyte precursors and astrocytes. Finally, clusters associated with day 16 presented enhanced signatures for the aformentioned neuronal subtypes, as well as new others including chrolinergic (clusters 8, 13, and 15) and serotonergic neurons (clusters 8 and 15). Importantly, among the previously preferentially associated clusters to the BMS641 + BMS961 treatment, cluster 6 appeared to be enriched in Dopaminergic, and GABAergic neuronal subtypes, as well as signatures associated with Oligodendrocyte precursors and astroglia (Fig. 1F). In contrary, cluster 11—preferentially enriched to the BMS753 treatment—presented weak signatures associated with Oligodendrocyte precursors and astroglia, while cluster 17 presented signatures associated with Dopaminergic neuron subtype (Fig. 1F).

Overall, mouse ESCs treated with the combination of the RARβ and RARγ agonists recapitulated the differentiation potential driven by either the pan-agonist ATRA or the RARα agonist BMS753; with some differences on the grounds of the obtained neuronal subtypes. Despite these differences, several of the identified clusters during this analysis appeared associated with multiple cell-types, suggesting that a further celltype stratification strategy might be required for better resolve such potential differences.

Disentangled RARβ+RARγ driven trajectories during differentiation reveals a delayed progression relative to RARα activation

To further stratify the cell clusters obtained by UMAP classification, we have taken advantage of STREAM, a pseudo-time stratification methodology [10]. At first, we have stratified at once all cells collected in all three timepoints and issued from all treatment conditions. This analysis gave rise to 14 pseudo-time trajectories, from which, pseudo-time S2 corresponded to the main root (R) composed of cells harboring a strong transcriptomics signature similarity (Fig. 2A). This main root gave rise to four intermediate branches (iB: S4, S12, S6, and S5), leading in a second time to other nine terminal branches (tB: S1, S8, S7, S3, S14, S13, S11, S9, and S10) (Fig. 2A). Cells issued from the BMS753 treatment were preferentially enriched in the terminal branches S13 (d16: 79%), S7 (d16: 86%), and in the intermediate branch S4 (d16: 53%) (Fig. 2B and C, and Supplementary Table S3A). Terminal branches S9 (d4: 96%), S3 (d4: 59%; d8: 24%), S14 (d8: 57%), S8 (d16:85%), and the intermediate branch S5 (d4: 81%) were enriched for cells issued from ATRA treatment. Finally, the BMS641 + BMS961 treatment appeared to contribute preferentially to only two intermediate branches (S1_d8: 45% and S6_d16: 68%), or retrieved together with BMS753 in multiple other branches [S2 (d4:27%;d8: 18%), S4 (d16:33%), S12 (d16:18%), S11 (d16:17%), and S10 (d16:20%) (Fig. 2B and C, and Supplementary Table S3A)] suggesting that this treatment is redundant with the outcome of the BMS753, as demonstrated in our previous study [7].

A direct comparison between clusters obtained by the UMAP classification and the pseudo-time progression provided by STREAM allowed to further stratify their composition. For instance, cluster 6, initially described as BMS641 + BMS961 specific, presented a strong overlap with pseudo-time branches S1 (42%), S4 (11%), S8 (25%), and S3 (13%) (Fig. 2D and Supplementary Table S3B). Among all these four branches, only S1 and S4 has been associated with BMS641 + BMS961, while S8 and S3 to BMS753 and/or ATRA treatment (Fig. 2C). Similarly, cluster 11 (associated with BMS753) has been enriched in pseudo-time root S2 (36%), the intermediate branch S6 (37%) and the terminal branch S14 (11%). Or, as illustrated in Fig. 2C, the branch S6 has been enriched in cells treated with BMS641 + BMS961 and branch S14 with ATRA.

In the other hand, events taking place in the late phase of differentiation appeared more specific between clusters and pseudo-time branches. For instance, the terminal branch S11 and cluster 17 appeared specifically enriched for cells issued from the BMS753 treatment; branch S13 corresponded to clusters 8 and 15; S14 is mainly enriched in cluster 8, S8 in cluster 9, S7 in cluster 13, and pseudo-time branches S9 and S10 in cluster 10 (Fig. 2D).

Cell-type annotation assay revealed that the terminal branches S13 and S14 (mainly associated with ATRA and BMS753 treatment) are strongly enriched for gene expression signatures related to forebrain development, amacrine cells, retina, retinal ganglion cells, callosal neurons, corticofugal neurons, as well as to cholinergic, glutamatergic, or dopaminergic neuronal markers (Supplementary Fig. S1). This neuronal specialization characteristics are further confirmed by the presence of gene expression signatures associated with the neuronal markers Calb1, Rbfox3, or Ncam1; further supported by the presence of the neuronal differentiation markers Sox4 and Sox11 (Fig. 2E and F, and Supplementary Fig. S2).

Terminal branch S7 (mainly associated with ATRA and BMS753 treatment) presented signatures related to astroglial cells including radial glial cell, astrocyte, and oligodendrocyte precursors in addition to Muller glial, pigmented epithelia, or neuroepithelial cell markers (Supplementary Fig. S1) and confirmed by the gene over-expression signatures associated with the astrocyte differentiation markers S100b, Gfap, Vim, Notch1, Fabp7, or Pax6 (Fig. 2E and F, and Supplementary Fig. S2). Furthermore, pseudo-time branches associated with BMS641 + BMS961 treatment were enriched in gene expression signatures associated with radial glial cells (S6; presenting the over-expression of markers like Notch1, Pax6, Nr2e1, Nes, Mcm2, or Eomes; Fig. 2E and F, and Supplementary Fig. S2), Intermediate progenitors, neuroeptihelial cells, oligodendrocyte/proliferative precurors (S4; over-expressing genes like Cdx2 or Eomes), arguing for a less differentiated status than those observed in branches S13 and S14, associated with ATRA or BMS753 treatment (Supplementary Fig. S1).

Repressive/active epigenetic signature transitions at the promoter level controls cell differentiation progression in RARα or RARβ+RARγ activation

Considering the differences observed in single-cell transcriptomics assays performed between BMS753 and BMS641 + BMS961, we aimed at evaluating for potential epigenetic signatures retrieved at promoter regions (5 kb around transcript start site) able to confirm such observations. To do so, we have performed Cleavage Under Targets and Tagmentation (Cut&tag [11]) assays after 16 days of BMS753 or BMS641 + BMS961 driven cell differentiation, targeting the repressive histone modification mark H3K27me3, the active promoter markers H3K27ac and H3K4me3, as well as the RNA polymerase II (RNAP2). The obtained chromatin enrichment signatures were integrated for revealing promoter co-occurring combinatorial events. As illustrated in Fig. 3A, a total of 10 chromatin co-occuring events were retrieved, which were associated with functional promoter status as following: Promoters being in a transcribing active status (presenting both the H3K27ac and the H3K4me3 markers and if possible also traces of RNAP2 or RNAP2 and H3K27ac); promoters presenting an active enhancer status (presenting only H3K27ac); promoters being repressed (enriched in H3K27me3), or simply being devoid of any of these markers and thus considered as inactive (Fig. 3A).

Figure 3.

Epigenetic signature transitions at the promoter level controls cell differentiation progression in RARα or RARβ+RARγ activation. (A) Chromatin state analysis performed on chromatin enrichment signatures (Cut&tag) assessed for the repressive mark H3K27me3, RNAP2, and the active promoter markers H3K27ac and H3K4me3; all collected in mouse stem cells differentiated during 16 days in presence of either the RARα ligand BMS753, or the combination of the RARβ and RARγ agonists (BMS641 + BMS961). Ten different combinatorial states were retrieved (ChromHMM), which were classified based on their promoter functional roles. (B) Comparison between the promoter chromatin status assessed in samples issued from the BMS753 and those issued from BMS641 + BMS961. The illustrated heatmap displays the number of promoters sharing the indicated chromatin state for either conditions. Diamond symbols indicate promoters presenting an inactive or repressed state in BMS753 and being active in BMS641 + BMS961 (orange diamond), or the opposite (green diamond). (C) Heatmap comparing the promoter chromatin state transitions between BMS753 and BMS641 + BMS961 and the corresponding gene expression associated with the pseudo-time branches displayed in Fig. 2. (D) Cell type annotation (hypergeometric probability gene set enrichment assay) of the genes associated with the most relevant promoter chromatin state transitions observed in panel (C). (E) Example of promoter chromatin state transitions between BMS753 and BMS641 + BMS961 conditions. Promoter enrichment levels were assessed in a window of 5 kb around the transcription start site. E: chromatin state.

Once promoter status are characterized, we have classified signatures found in either of the retinoids treatment to retrieve for potential changes. Indeed, while the majority of genes are presenting a similar histone modification signatures within their promoters, we have identified several genes passing from an inactive/repressed status in BMS753 into a trancscriptionally active status in presence of BMS641 + BMS961 (E3/E1:312 genes; E3/E2:192 genes; E3/E10:108 genes; E6/E1:35 genes; E7/E1:82 genes; E7/E2:49 genes; E7/E10:18 genes; E9/E1:29 genes), as well as others doing the oposite (E1/E8:517 genes; E1/E9:72 genes; E4/E8:23 genes; E4/E9:41 genes; E5/E3:127 genes; E5/E8:1117 genes; E5/E9:1216 genes) (Fig. 3B).

The comparison of this information with the pseudo-time branches described in Fig. 2 led to the identification of a strong enrichment of promoters passing from an active status in BMS753 into a repressive/inactive status in the presence of BMS641 + BMS961 in branches S13 and S7 (E1/E8 transitions). Furthermore, branches S11, S8, and S1 were preferentially enriched for a promoter activation transition in the combination of the RARb + RARg agonists (E3/E1); confirming the results obtained from the single-cell transcriptomics (Fig. 3C).

Finally, a cell type annotation analysis confirmed that active promoters retrieved in the BMS753 condition but repressed in the BMS641 + BMS961 treatment (E1/E8) were preferentially associated with genes related to a variety of neuronal subtypes; while active promoters in BMS641 + BMS961 treatment but repressed in BMS753 failed to present a distinct signature, with the exception that were enriched for markers associated with neuroepithelial cells (E3/E1), indicative for a less differentiated state than that observed in BMS753 (Fig. 3D).

Examples of such promoter status transitions are illustrated in Fig. 3E, where genes like Gad1 (Glutamate Decarboxylase 1; known as marker for GABAergic neurons), or Chat (Choline O-Acetyltransferase; known to catalyze the biosynthesis of the neurotransmitter acethylcholine) were found to present an active promoter status in BMS753, but rather repressed in BMS641 + BMS961 treatment (E1/E8 transition). In contrary, A repressive to active transcriptional behavior (E7/E1) has been observed in the promoter associated with genes like Pgm2l1 (Phosphoglucomutase 2 Like 1; coding for an enzyme strongly expressed in the brain, accounting for the elevated levels of glucose-1,6-biphosphate), or the gene coding for the Proenkephalin protein (Penk), a neuropeptide that plays a role in the perception of pain and response to stress, and which its expression under the control of RARγ in DARPP32-positive striatal-like medium spiny neurons [19] (Fig. 3E). Similarly, genes like Cntn4 (Contactin4, an axon associated cell adhesion component, playing a role in neuronal network formation and plasticity), Dkc1 (Dyskerin Pseudouridine Synthase 1, a protein involved in the telomerase activity), Pde1a (Phosphodiesterase 1A, a protein known to be expressed in neurons where it hydrolyzes cGMP into cAMP), or Gata3 (transcription factor known to play a major role in the development of serotonergic neurons [20]), were shown to present an inactive promoter status in presence of the BMS753 treatment but present a rather active state in presence of the BMS641 + BMS961 (Fig. 3E).

Overall, these disentangle analyses performed either in the single-cell transcriptomics data with the help of pseudo-time inference strategies, or the assessment of promoter status by the capturing of relevant histone modification signatures, confirm the capacity of the BMS753 and the combination of the BMS641 + BMS961 RAR ligands to drive the emergence of a variety of cell types during cell differentiation, and notably a delayed progression in differentiation in the case of the synergistic activation of the RARβ and RARγ receptors.

Long-term brain organoid cultures in presence of RAR-specific agonists provide distinct nervous tissue differentiation signatures

The aforementioned differentiation assays performed in two-dimensional monolayer cultures demonstrated (i) a delayed progression toward maturation in presence of the combination of the RARβ and RARγ agonists BMS641 + BMS961 relative to the use of the RARα agonist BMS753; and (ii) subtle differences in neuronal subtypes generated in these two conditions. Considering the major relevance of the retinoids action during brain development [1, 21], and notably the use of vitamin A (the natural precursor of the pan-RAR agonist, the ATRA) for maturation of cerebral organoids in in vitro three-dimensional cultures [22], we have envisioned that the use of RAR-specific agonists could be considered as an option for generating specialized cerebral organoids. To address this hypothesis, we have followed the previous protocol described by M. Lancaster for generating mouse brain organoids [21], with the slight difference that in day 9 we have transferred the matrigel embeeded organoids into flasks instead of bioreactors (Fig. 4A). Furthermore, to address the potential performance of retinoids action for organoids maturation we have complemeted the maturation media (day 9) with either vitamin A, ATRA, the RARα agonist BMS753, or the combination of the RARβ and RARγ agonists (BMS641 + BMS961) (Fig. 4A and Supplementary Fig. S3A).

Figure 4.

Long-term brain organoid cultures in presence of RAR-specific agonists provide distinct nervous tissue differentiation signatures. (A) Scheme illustrating the strategy in use for generating mouse brain organoids. After 90 days of culture, multiple brain organoids are collected for spatial transcriptomics assay. (B) Left panels: Microscopy image from a double-barcoded DNA array slide presenting a cryosection of multiple brain organoids per treatment, in which captured cDNA is visualized by dCTP-Cy3 incorporation. Right panels: Corresponding spatial transcriptomics landscapes assessed with the help of the in-house double-barcoded DNA arrays (Lozachmeur et al. [13]). Heatmap landscapes correspond to the normalized read counts per gexel (gene expression element; in analogy to pixel). BG: Brain organoid. (C) Left panels: Gexels stratification per treatment condition with the help of a pseudo-time strategy (STREAM [10]). Right panels: the number of gexels per pseudo-time branches. (D) Spatial localization of the identified branches in panel (C). Gray gexels present low normalized read-counts in panel (B); thus, they are not included in the pseudo-time analysis. (E) Heatmap displaying gene set enrichment levels for cell/tissue type annotations performed over all pseudo-time branches across all treatments. (F) Spatially resolved differential gene expression (relative to the mean expression across all gexels) of gene markers associated with glutamatergic (Slc17a7; coding for VGLUT1), dopaminergic (Atp13a2 also known as Park9), GABAergic (Gad1), or cholinergic (ache coding for acetylcholinesterase) neuronal subtypes, as well as the Astrocyte marker S100b. BG: brain organoid.

Brain organoids were traced over time by reverse transcription assessed by quantitative PCR (RT-qPCR) targeting the depletion of pluripotency markers, the transient emergence of neuronal precursor signatures, as well as the establishment of neuronal subtype markers (Table 1). Among the evaluated pluripotency markers, Oct4 appeared strongly downregulated in both vitamin A and ATRA treatment after 90 days of differentiation. In contrary, the use of either BMS753 or BMS641 + BMS961 gave rise to mild detected downregulation, indicative for the presence of pluripotent cells even after 90 days of culture in these two conditions (Supplementary Fig. S3B). In the other hand, strong levels of the Klf4 transcript were detected in all conditions and across all time-points, indicative for the presence of neuronal progenitors [23, 24]. Neuronal precursors were confirmed also by the detection of transcripts associated with Ascl1, Nestin, and Tuj1, which presented high levels during the first 30 days, but were gradually reduced under all treatment conditions. Furthermore, transcripts associated with the gene Gad67 (GABAergic neurons), Glut1 (glutamatergic neurons), Th (tyrosine hydroxulase, marker for dopaminergic neurons), and Tph2 (tryptophan hydroxylase, marker for Serotonergic neurons) were detected preferentially under vitamin A treatment (Supplementary Fig. S3B). Indeed, ATRA and BMS753 treatment conditions presented barely two-fold levels of the Tph2 transcript relative to the day 0 control, and completely absent in BMS641 + BMS961. Finally, transcripts targeting the astrocyte marker Gfap has been detected in all four treatment conditons treatments and across all timepoints, while the transcript Olig2, specific for oligodendrocytes remained undetectable in all treatments (Supplementary Fig. S3B).

Table 1.

List of oligonucleotides used for RT-qPCR assays

No.	Oligonucleotide name	Sequence
1	36b4-F	AATCTCCAGAGGCACCATTG
	36b4-R	CCGATCTGCAGACACACACT
2	Sox2-F	TGCCAATATTTTTCGAGGAAAG
	Sox2-R	CGGTCCTAAATTTAACTGCAGAAT
3	Nanog-F	GCCTCCAGCAGATGCAAG
	Nanog-R	GGTTTTGAAACCAGGTCTTAACC
4	Oct4-F	GCATTAGCATCACCATGCAG
	Oct4-R	CCGGTTACAGAACCATCCTC
5	Klf4-F	CGGGAAGGGAGAAGACACT
	Klf4-R	GAGTTCCTCACGCCAACG
6	Ascl1-F	GCTCTCCTGGGAATGGACT
	Ascl1-R	CGTTGGCGAGAAACACTAAAG
7	Nestin-F	CTGCAGGCCACTGAAAAGTT
	Nestin-R	TCTGACTCTGTAGACCCTGCTTC
8	Tubb3-F	GGCAACTATGTAGGGGACTCAG
	Tubb3-R	CCTGGGCACATACTTGTGAG
9	Gfap-F	CAGCTTACGGCCAACAGTG
	Gfap-R	GCCTCAGGTTGGTTTCATCT
10	Olig2-F	AGACCGAGCCAACACCAG
	Olig2-R	AAGCTCTCGAATGATCCTTCTTT
11	Gad67-F	TGGAGATGCGAACCATGAG
	Gad67-R	GAAGGGTTCCTGGTTTAGCC
12	Glut1-F	CTGGGGTCCTTGTGCAGTAT
	Glut1-R	CCGAAGCTGCCATAGACATAG
13	Th-F	AAGATCAAACCTACCAGCCG
	Th-R	TACGGGTCAAACTTCACAGAG
14	Tph2-F	GAGCTTGATGCCGACCAT
	Tph2-R	TGGCCACATCCACAAAATAC

With the aim to perform a more in-depth analysis, after 90 days of culture in presence of retinoids, brain organoids were cryosectioned and tissue sections were deposited in in-house manufactured DNA arrays presenting 2048 unique molecular positional barcodes and poly-T ends for capturing messenger RNA [13]. This in-house spatial transcriptomics strategy allowed to capture local gene expression signatures (spots of ∼100 microns of diameter) across multiple brain organoids (BGs) cryosectioned together (Fig. 4B). Gene expression readouts assessed per physical position, herein called as “gexels” (gene expression elements, in analogy to pixels) were stratified in multiple pseudo-time branches with the help of STREAM [10]. As illustrated in Fig. 4C, pseudo-time stratification of gexels based on their gene expression levels issued from vitamin A treatment assessed in six different brain organoids allowed to identify up to eight branches. Among them, the branch S6 appeared as the most frequent (∼150 gexels), followed by branches S8 and S7 (∼70 gexels each); all three retrieved in five over six analyzed organoids (Fig. 4D). Similar pseudo-time stratification performed over all three other treatment conditions revealed up to six branches for ATRA treatment, seven for BMS753, but only four in the case of BMS641 + BMS961, and this despite a similar number of analyzed brain organoids (Fig. 4B and C). Indeed, gexels associated with all four branches were retrieved in all seven analyzed brain organoids issued from the BMS641 + BMS961 treatment, indicative for an homogeneous differentiation behavior; an aspect that is also retrieved in all other treatments (Fig. 4D and Supplementary Fig. S4).

Cell-type annotation performed on the identified pseudo-time branches revealed signatures ranging from proliferative precursor states, passing through intermediate progenitors leading to glial precursor cells giving rise to astrocytes or driven toward neuronal cell specialization (Fig. 4E). Indeed, traces of proliferative precursor signatures were found enriched in the root S4, and branches S0, S7, and S8 from vitamin A. Furthermore, the most abudant—in the number of gexels—branch 6 appeared enriched in mature neuronal signatures, specifically corresponding to glutamatergic and dopaminergic subtypes (Fig. 4E); further confirmed by the large number of over-expressed gexels associated with the markers Slc17a7 (coding for the vesicular glutamate transporter 1, VGLUT1) or Atp13a2 (also known as the Parkinson disease 9 gene, Park9), respectively (Fig. 4F). Similarly, the pseudo-time branch S4 from ATRA is specifically enriched in Radial glial cell signatures, S2 in newly formed oligodendrocytes, S5 and S6 in glial cell fate signatures, while branch S1 present signatures associated with proliferative precursors, but also to GABAergic specialized neurons (Fig. 4E), further suppored by the presence gexels over-enriched in the specialized marker Gad1 (Fig. 4F).

Pseudo-time branches from organoids treated with the BMS753 ligand appeared enriched in specialized neurons (Dopamineric: S1, S2, and S6; GABAergic: S1 and S2; Glutamatergic: S1), as well as glial cells (S4), similarly to what has been observed in either the vitamin A or the ATRA treatment conditions. Finally the use of the BMS961 + BMS641 treatment conditions gave rise to organoids presenting pseudo-time branches having similar enrichment signatures like the aforementioned conditions but with a weaker confidence. Notably, branch S2 appeared enriched in signatures associated with Astrocyte cells, but also in specialized cholinergic neurons, further supported by over-expression signatures associated with the gene markers S100b or Ache (Acetylcholinesterase) respectively (Fig. 4E and F); demonstrating that brain organoid cultures in presence of RAR-specific agonists can lead to well differentiated tissues, presenting a variety of cell-types.

Finally, further functional nervous tissue annotation signatures were enriched in various of the analyzed branches, among them the presence of callosal projection neurons (S5 and S7 ATRA; S2 and S6 BMS753; S4 and S9 vitamin A); bipolar cells (S5 and S7 ATRA; S1 and S2 BMS753); pigmented epithelia (S7 ATRA; S1 and S2 BMS753; S0, S4, and S9 vitamin A; S5 BMS641 + BMS961); retinal ganglion cells (S4, S2, S7 ATRA; S4 BMS641 + BMS961; S1, S2, S6 BMS753; S4, S9 vitamin A); photoreceptors (S0, S2 BMS753; S2 BMS641 + BMS961; S4 vitamin A; S7 ATRA); retina (S2 BMS641 + BMS961; S7 ATRA; S2 BMS753; S4, S7 vitamin A); amacrine cells (S4 BMS753; S1 BMS641 + BMS961; S7 ATRA; S9 vitamin A); or Corticofugal neurons (S7 ATRA; S4 vitamin A) (Fig. 4E).

It is also worth to mention that strong microglia signatures were also detected in vitamin A-treated organoids (S0, S4, S9, and S8); in ATRA treatment (S7, S5, S6, and S2); in BMS753 (S7, S6, S2, and S1); and in a less pronounced manner in BMS641 + BMS961 treatment (S5); in aggreement with brain organoid cell differentiation cultures where no means to inhibit the endodermal/mesodermal differentiation pathways are used [25].

Overall, this detailed analysis of the various tissular regions disected by spatial transcriptomics performed on brain organoid cultures demonstrated that the RAR-specific agonist BMS753 (RARα specific), as well as the synergistic activation of the RARβ and RARγ receptors (BMS641 + BMS961) could perform as well as the pan RAR-agonist ATRA, or the lagely used vitamin A supplement.

Reconstruction of the RAR-driven gene regulatory programs involved in cell specialization during brain organoids progression

Stratification of the various gexels composing the spatial transcriptomic maps assessed under the various retinoids treatment revealed a variety of gene expression signatures associated with distinct differentiated cell types. Considering that these processes are expected to be driven by the action of the added retinoids, we aimed at evaluating the expression of each of the RAR receptors. As illustrated in Fig. 5, brain organoids collected at 90 days of BMS753 treatment presented expression signatures for all three Rara, Rarb, and Rarg receptors (Fig. 5A). Such expression signatures were observed in nearly all nine analyzed brain organoids, covering up ∼60 gexels for Rarg, ∼75 gexels for Rara, and ∼60 gexels for Rarb (Fig. 5B). Similar observation was made for these receptors in brain organoids treated with BMS961 + BMS641, where ∼80 gexels were associated with upregulated transcripts for Rara, ∼55 gexels for Rarb, and ∼70 gexels for Rarg (Fig. 5A and B). RAR expression across the analyzed brain organoids represented ∼60% of the total gexels; with <5% of them corresponding to presence of all three RARs in the same spatial regions (Fig. 5C). Barely ∼25% of gexels over-expressed two over three RAR receptors, and ∼30% corresponded to gexels presenting only one of the RARs being over-expressed (Fig. 5C).

Figure 5.

Reconstruction of the RAR-driven gene regulatory programs involved in cell specialization during brain organoids progression. (A) Gexels presenting high expression levels of the genes Rara, Rarb, or Rarg in brain organoids under the retinoid agonist BMS753 (top panels) or BMS961 + 641 (low panels). (B) Gexel quantification for each Rar gene and the correspondences to pseudo-time branches displayed in Fig. 4. (C) Fraction of gexels expressing all three RARs, two over three or only one of them, in both treatment conditions. (D) Venn diagrams comparing highly expressed genes (log2 fold change > 1 respect to the logarithmic media) between the treatments MBS753 and BMS961 + BMS641. (E) Left panel: Ranking of transcription factors in BMS753 treatment, based on their capacity to control downstream factors in a GRN reconstructed from the differentially expressed genes displayed in (D). The master regulatory index (MRI) corresponds to the fraction of genes that can be controlled by a given TF within the GRN (TETRAMER [20]). MRI inferred to the top ranked TFs are displayed in blue solid line. In gray, a randomized GRN of the same size has been used for evaluating the MRI confidence. Right panel: Confidence associated with the TF computed from the randomized GRN. (F) Same as (E) for the differentially expressed genes assessed from brain organoids under the BMS961 + BMS641 treatment. (G) Gene co-regulatory network of the top ranked master TFs obtained in (E) and (F). TFs specific to the MBS753 treatment are displayed in blue, those specific to BMS961 + 641 in green and those in common for both treatments in orange. The network was organized from outside to inside by the topological parameter “betweenness centrality” (most incoming connections at the network center). (H) Cell/tissue type annotations associated with the most relevant TFs retrieved in the gene co-regulatory network. BG: brain organoid.

Gexels presenting over-expressed transcripts for at least one of the RAR receptors in either of treatments were compared on the grounds of their differentially expressed genes. This comparison revealed 949 common upregulated genes, 1164 upregulated genes specific to the BMS753 treatment, and 855 specific to BMS961 + BMS641 (Fig. 5D). Surprisingly, >7000 genes were found commonly downregulated, where only 1573 genes were specific to BMS753 and 1959 to BMS961 + BMS641; indicative for major differences between these two conditions in the context of active players (Fig. 5D). For this reason, we focused in revealing active master transcription factors by using a GRN reconstruction strategy, followed by a modeling of transcription activation cascade, with the help of our previously described tool TETRAMER [16]. This modeling strategy applied to brain organoids treated with BMS753 allowed to infer up to 83 confident master transcription factors (P-value < 0.25) presenting a transcription regulation capacity of ∼40%, meaning that each of these TFs are able to control 40% of the total nodes composing the GRN (Fig. 5E). Similarly, 63 confident master transcription factors were inferred from the analysis of the GRN associated with the BMS961 + BMS641 treatment (Fig. 5F). The comparison of the identified master transcription factors for both treatment conditions revealed a set of 45 common players, displayed as part of a gene co-regulatory network, allowing to reveal the transcription regulation relationship between all these factors. (Fig. 5G).

These master transcription factors were associated with functions like central nervous system development, ventral spinal cord, neural tube development, brain morphogenesis, or forebrain generation (Fig. 5H). Furthermore, several of these players were also associated with specialized neuronal subtypes, retina-related differentiation tissues, as well as mesodermal differentiation giving rise to microglia (Fig. 5H). Among the key transcription factors specifically over-expressed under BMS641 + BMS961 conditions, we have retrieved the Retinal homeobox factor (RAX), known to be expressed in radial glial cells and described as being essential for eye and forebrain development [26]. Similarly, the expression of the sine oculis homeobox homolog 2 (SIX2) has been detected, a transcription factor that has been seen to be expressed in the adult differentianting cells of the mouse retina, and from a broader view, in the lens formation in vertebrates [27]. Furthermore, the transcription factor MAFG has been associated with cataract or lens defects [28]; the proinflammatory factor BATF2 has been previously shown to be expressed in retinal microglia [29]; and FOXD1 has been shown to be required for the specification fo the temporal retina [30].

Overall, the analysis of gexels presenting RAR receptors expression unveiled spatially distinct expression signatures for each of the RAR isotypes, with a rather low fraction of regions presenting their redundant expression. Furthermore, it confirmed the presence of gene expression signatures associated with well-differentiated nervous tissue generation. Notably the use of either the RARα agonist BMS753 or the synergistic activation of the RARβ and RARγ receptors (BMS641 + BMS961) gave rise to the activation of a common pool of transcription factors and a subset of others specific to the treatment, and several of the transcription factors specific to the RARβ and RARγ activation appeared to be associated with eye development.

Discussion

From a historical point of view, developmental processes were studied from a tissue level (morphogenesis), a cellular level (cell specialization) but also from the characterization of molecular components retrieved at precise concentrations in a spatio-temporal fashion (morphogen gradients). Among all these levels of granularity, morphogens were early on shown to be able to control both cell specialization and tissue organization, hence providing a mechanistic view of the events controlling morphogenesis.

Retinoic acid signaling, via the RXR/RAR nuclear receptors activation, is a key morphogen during nervous system development [1], but also for brain homeostasis. Indeed, RAR/RXR receptors were shown to be expressed in multiple brain areas in the adult (RARα: hippocampus, cerebellum, and cortex; RARβ: striatum and spinal cord; RARγ: hippocampus; RXRα: hippocampus, RXRβ: cerebellum; RXRγ: Striatum, limbic cortex, and spinal cord [31]), and their deletion or a deficiency in retinoic acid supply in mouse models leads to developmental malformations but also deficits in synaptic plasticity, memory, or accumulation of amyloid-β in the adult mice [32–35].

Despite these findings, brain morphogenesis remains still poorly understood notably from the angle of the gene regulatory programs implicated in its landscaping. Previous studies, including ours, focused in driving neuronal differentiation in vitro by the use of the pan-RAR agonist ATRA, but also in decorticating the role of each RAR receptors thanks to the use of specific agonists. Previously we have shown that, in monolayer differentiation assays, mouse stem cells could give rise to neuronal precursors when treated with the RARα specific agonist BMS753, but not in presence of the RARβ agonist BMS641 or the RARγ agonist BMS961 [6]. Furthermore, we have discovered that the combination of the BMS641 and BMS961 agonists can rescue the neuronal precursors differentiation by hijacking the RARα-driven gene programs [7].

Herein we aimed at addressing first, the influence of these synthetic RAR agonists in long-term monolayer cultures, as a way to interrogate cell heterogeneity during cell specialization; and second, explore their influence in nervous tissue formation.

Cell heterogeneity has been addressed with the help of single-cell transcriptomics assays performed in 16 days monolayer in vitro cultures in presence of either the synergistic action of the BMS641 + BMS961 ligands or the RARα specific agonist BMS753. This comparative study revealed a variety of specialized neuronal cell types retrieved in both treatment conditions, including the presence of GABAergic, dopaminergic, cholinergic, and even serotonergic subtype signatures, as well as a delayed progression in maturation of cells treated by the combination of BMS641 + BMS961, confirmed by the enrichment in signatures associated with intermediate progenitor states. By decoding the chromatin epigenetic status after 16 days of differentiation revealed some differences in promoter activity between both treatment conditions. Notably, the presence of repressive histone modification signatures in promoter regions associated with the gene Gad1 or Chat in samples treated with the BMS961 + BMS641 supports the hypothesis of a delayed accessibility of these promoters, potentially mediated via the presence of RARα/RXR repressed complexes. Indeed, our previous study in P19 cells demonstrated a higher yield of differentiation progression driven by BMS961 + BMS641 ligands in a Rara−/− null background [7].

Three-dimensional brain organoid cultures during 90 days where the use of vitamin A has been replaced by either the pan-RAR agonist ATRA or the use of the combination of BMS641 + BMS961 ligands or the RARα specific agonist BMS753, revealed spatially distinct RAR isotype expression, in agreement with the observations in the mouse brain [31]. Despite a delayed differentiation behavior in presence of the BMS641 + BMS961 ligands, the obtained organoids presented gene expression signatures associated with well matured tissues, including a variety of neuronal subtypes but also tissue structure signatures like retina and even the presence of microglia cells.

One aspect that remained open in this study is the reason for having a rather exclusive expression of RAR receptors per gexels across tissues. While this observation is in agreement with previous studies performed in mouse models demonstrating the highly regionalized and/or cell type-specific expression of RAR receptors, explanations other than the retinoid gradient concentration across the organoids, are required. Counting with spatially resolved in vitro systems like the use of brain organoids might allow to trace the molecular mechanism behind this localized RAR receptors selection. This being said, novel spatially resolved technologies, like spatial Cut&tag [36] are essential to interrogate for local epigenetic barriers they might account for such differences.

The presented results are in the front-end of research aiming to generate complex tissues derived in vitro. Indeed, while all brain organoid assays rely systematically in the use of vitamin A as a key component within the maturation medium, the use of RAR-specific ligands remain unexplored. Our results clearly suggest that the use of synthetic agonists can mimic the performance of the classical vitamin A compound, but it also paves the way for future studies in which further other RAR/RXR-specific ligands can be used for driving tissue specialization.

Supplementary data

Supplementary data is available at NAR Molecular Medicine online.

Conflict of interest

None declared.

Funding

This work was supported by the institutional bodies CEA, CNRS, Université d’Evry-Val d’Essonne. Aysis Koshy has been supported by the “Fondation pour la recherche medicale” (FRM; funding ALZ°201912009904); Ariel Galindo-Albarra is supported by EU grant HORIZON N° 101070740; Maria Grazia Mendoza-Ferri has been funded by the Institut National du Cancer (INCa: Funding N° 2020-124; N° 2022-078).

Data availability

All datasets generated on this study have been submitted to the NCBI Gene Expression Omnibus database (GEO; http://www.ncbi.nlm.nih.gov/geo/). Bulk Cut&Tag datasets are available under the accession number GSE286245; Single-cell transcriptomics data under the accession number GSE286249 and double-barcoded spatially resolved transcriptomics data under the accession number GSE286250. Furthermore, processed spatial matrices as well as the corresponding scanned images are available in Mendeley Database: https://data.mendeley.com/datasets/ywpxbdbpb4/2.