Variable expression of MECP2, CDKL5, and FMR1 in the human brain: Implications for gene restorative therapies

Significance

MECP2, CDKL5, and FMR1 are X-linked neurodevelopmental genes responsible for Rett, CDKL5-, and fragile-X syndromes. Methods to treat the disorders have been centered on restoring protein production. For gene therapy, gene editing, and selective Xi-reactivation methodologies to work, it is critical to understand how much restoration is required and what regions in the brain must be targeted. Here, we profile the neurotypical human brain at single-cell resolution across multiple developmental stages, brain regions, and multiple donors. A surprising degree of variability is seen across cell types and donors—potentially defining a therapeutic window, with the low end delineating a minimum for neurotypical function and the high end informing toxicology margins. Potential biomarkers are also proposed.

Abstract

MECP2, CDKL5, and FMR1 are three X-linked neurodevelopmental genes associated with Rett, CDKL5-, and fragile-X syndrome, respectively. These syndromes are characterized by distinct constellations of severe cognitive and neurobehavioral anomalies, reflecting the broad but unique expression patterns of each of the genes in the brain. As these disorders are not thought to be neurodegenerative and may be reversible, a major goal has been to restore expression of the functional proteins in the patient’s brain. Strategies have included gene therapy, gene editing, and selective Xi-reactivation methodologies. However, tissue penetration and overall delivery to various regions of the brain remain challenging for each strategy. Thus, gaining insights into how much restoration would be required and what regions/cell types in the brain must be targeted for meaningful physiological improvement would be valuable. As a step toward addressing these questions, here we perform a meta-analysis of single-cell transcriptomics data from the human brain across multiple developmental stages, in various brain regions, and in multiple donors. We observe a substantial degree of expression variability for MECP2, CDKL5, and FMR1 not only across cell types but also between donors. The wide range of expression may help define a therapeutic window, with the low end delineating a minimum level required to restore physiological function and the high end informing toxicology margin. Finally, the inter-cellular and inter-individual variability enable identification of co-varying genes and will facilitate future identification of biomarkers.

The mammalian X chromosome has a unique pattern of inheritance and is subject to epigenetic regulation, including imprinting and dosage compensation (1). This sex chromosome is also known to be enriched for neurodevelopmental and neurobehavioral genes implicated in autism spectrum disorders (ASD) and intellectual disabilities (2). Because males generally have only one X-chromosome, males are more commonly and severely affected by X-linked disorders than females (who generally have two X-chromosomes). For example, the X-linked neurodevelopmental disorder, fragile X syndrome (FXS), can affect both boys and girls, but boys are generally more frequently and more severely affected (3). FXS has a worldwide incidence of ≈1:3,000 newborn males and ≈1:6,000 newborn females and is characterized by ASD, manifestations of an anxiety disorder, and a range of cognitive, motor, and developmental delays (4). FXS is caused by absent expression of FMR1, an RNA-binding protein that concentrates in the brain, especially at neuronal synapses where the protein is integral to translational repression of mRNAs and mRNA transport (5, 6). Interestingly, the FMR1 coding sequence is almost always normal in FXS and failed expression is caused by aberrant expansion of a CGG trinucleotide repeat in the 5′-UTR.

In a subset of X-linked disorders, however, affected individuals are predominantly female. A primary example includes Rett syndrome (RTT), a pervasive neurodevelopmental disorder characterized by seizures and developmental regression. RTT has a diagnostic prevalence of ≈1:10,000 girls and is caused by mutations in MECP2, a methyl-CpG-binding protein with key neurodevelopmental roles (7). Males with severe MECP2 mutations often die in utero, whereas heterozygous females survive to birth due to the presence of a second X-chromosome expressing a wild-type MECP2 allele. Similarly, CDKL5 deficiency disorder (CDD) mostly affects females (90% are females). CDD is characterized by complex patterns of neurological and intellectual deficits, ranging from epileptic encephalopathy to motor disabilities, and cortical visual impairment (8). The disorder is caused by mutations in CDKL5, cyclin-dependent kinase-like 5 protein that functions as a serine–threonine kinase (9). Altogether, RTT, CDD, and FXS are believed to affect ≈200,000 individuals in the United States alone. As the disease-causing mutations may have similar frequencies throughout the world, as many as 4 million people may be affected across the globe.

RTT, CDD, and FXS are neurodevelopmental rather than neurodegenerative. In all three cases, there is growing evidence that the neuronal defects can be reversed through restoration of the missing gene product (10–17). Thus, in recent years, there has been considerable pharmaceutical interest in developing disease-modulating treatments using various restorative approaches (18). One approach is gene therapy, which—in the case of RTT—introduces a normal copy of MECP2 into patient brain cells using the adeno-associated virus 9 vector. Although gene therapy has the potential to cure the disorders by restoring expression of the absent genes, a major challenge associated with AAV-mediated gene therapy is the limited penetration of AAV into the cortex, cerebellum, and brainstem. There is also the potential to over-deliver on a per-cell basis. This is particularly concerning for a disorder such as RTT, as overexpression of MECP2 causes another seizure disorder known as MECP2 duplication syndrome in humans (19). Human duplications of FMR1 and CDKL5 are also known to result in developmental delay, intellectual disability, and autism (20, 21). For therapeutic consideration, data from a recent study suggest that over-delivery and overexpression of MECP2 to even a small subset of cells in the mouse brain after AAV gene therapy may elicit neurotoxicity (22). Over-delivery of FMR1 and CDKL5 would be expected to have adverse outcomes as well. Alternative approaches include genome and RNA editing, which similarly use AAV to deliver their respective payloads, in these cases to edit the causal mutation in the endogenous MECP2 allele or to edit the aberrant RNA produced from the allele, respectively. Because AAV delivery is likewise required, a potential limitation is again a restricted brain penetration. Off-target effects have also been a concern.

Another approach under development for RTT, CDD, and FXS is X-reactivation. This method would be applicable only in females, as the approach aims to reactivate the normal but silent gene copy that is locked up on the inactive X chromosome (Xi) within the patient’s own cells. In females with two X-chromosomes, one is inactivated in early development to equalize gene dosages between XY males and XX females. The choice of which X-chromosome is inactivated occurs randomly, such that heterozygous females with a mutated X-linked gene would manifest defective gene expression in approximately half of their cells. Thus, ≈50% of the brain in a heterozygous female would lack the critical gene product (MECP2, CDKL5, or FMR1) despite carrying a normal copy of the gene on the Xi. In the reactivation approach, partial restoration would be achieved by treating brain cells with an antisense oligonucleotide (ASO) to deplete XIST RNA, the factor that is necessary to maintain the Xi in female cells. The ASO approach does not require a delivery vehicle such as AAV, but the extent to which an ASO would penetrate the human brain is not fully known.

For all of the above approaches, tissue penetration and delivery to brain regions remain a primary challenge. Because the brain is a rich composite of distinct cell types with high functional specialization, a major gap in our ability to effectively treat the disorders is knowledge of i) where and in what cell types MECP2, CDKL5, and FMR1 are expressed in the brain, ii) the extent to which expression varies among cell types, and iii) the range of expression between individuals. This information could ultimately inform strategies for drug delivery and dosing. Most importantly, they could help determine a minimum gene expression level for physiological function—which presumptively would occur at the lowest ranges of MECP2, CDKL5, and FMR1 values on a per-cell basis. Furthermore, they would also instruct us on how much expression might be toxic—which would presumptively occur beyond the highest ranges of MECP2, CDKL5, and FMR1 values on a per-cell basis.

Previous transcriptomic analyses of these genes have largely been limited to bulk brain, precluding the extraction of data from specific cell types, brain regions, and differences between individuals and single cells. On the other hand, single-cell technologies enable the mapping of cell identities and are enabling design and delivery of precision medicine. At present, knowledge of MECP2, CDKL5, and FMR1 expression at the single-cell level in the human brain remains limited. For RTT, several clinical and genetic databases exist, including the IRSA North American Database (23), the RettBASE (24), and the Italian Rett biobank (25). An MECP2 transcriptomic database, MECP2pedia, has also been created, focused mostly on bulk murine sequencing data (26). Similarly, previous studies on FXS have revealed subtle expression changes (27–29), highlighting the limitation of the conventional pooled-based assay to provide finer signals in complex tissues like the brain. The databases for FXS and CDD are more limited. Toward closing the knowledge gap, we collect and integrate existing single-cell/nucleus RNA-seq data from the human brain, collectively comprising over half a million of cells, to investigate MECP2, CDKL5, and FMR1 transcriptome in a cell-type-specific manner. A substantial range of expression is observed for all three genes, not only between cell types but also within and between female individuals. Using gene co-expression analyses, we also identify potential biomarkers for gene restoration therapies, based on autosomal and X-linked factors significantly co-expressed with MECP2, CDKL5, or FMR1 in a cell type–specific manner. The implications of our findings for therapeutics are discussed.

Results

Data Processing and Integration.

Rationale.

Our goals are twofold. First, we aim to assess the transcriptional landscape of three genes of interest (GOI) —MECP2, CDKL5, and FMR1—in different brain regions, cell types, developmental stages. Second, we aim to study transcriptional variation among female for multiple reasons. First, RTT and CDD affect more females. Second, X-reactivation methods rely on the presence of an Xi, which occurs only in females. Third, analytical challenges required us to limit examination to one sex. Because the two sexes significantly differ in a wide range of quantitative traits, sex is a confounding variable that would have to be accounted for in transcriptional comparisons. Another factor is the need to infer the sex when the information is unavailable. Though we relied on X-linked and Y-linked expression, and the sex-specific genes XIST and SRY (Materials and Methods), sex miscalls due to unclear measures, particularly for embryonic/fetal cells, may still occur. While including only female cells is not error-free, inferring both sexes would, to an uncertain degree, amplify the risk of miscalls. Given our focus on X-linked genes, misclassification of both sexes could have non-negligible effects on transcriptomic analyses. Thus, we largely focus on female brain expression patterns for MECP2 and CDKL5 but include analysis of the male brain for FMR1, given the greater male prevalence for FXS.

Transcriptomic data used for analyses.

Acquiring in-depth knowledge of the transcriptional patterns of these genes across multiple datasets requires harmonization. A major challenge is the multi-factorial variation arising from different labs, protocols, donors, and other heterogeneities (30, 31). It is necessary to correct technical differences to align cells of the same biological type across datasets. To accomplish this and address our questions, we utilized existing algorithms put together into our analytical pipeline for processing and downstream analyses (32–35). Here, we collected and integrated 10 publicly available single nucleus/cell RNAseq datasets originated from human brain specimens (SI Appendix, Fig. S1). Integrating diverse datasets that provide multiple variables of interest improves power and resolution of context-specific analyses (36). We preprocessed each dataset independently and labeled each cell by the pre-annotated label (SI Appendix, Fig. S2 A–J). Datasets comprising adult (37–40) and developing brains (41–45) were integrated separately via canonical correlation analysis in Seurat (32, 34) (Materials and Methods).

We undertook two strategies to assess batch of effects: 1): We assessed the experimental variables study batch id and technology in the integrated clustering map; 2) we assessed integration efficiency using the local inverse Simpson’s index (iLISI) to quantify the degree to which the distinct study variables are represented in a cell neighborhood (33). We conducted both (1) and (2) analyses on the adult batch-uncorrected data, the Seurat integrated assay, and the Harmony integrated assay. The Harmony integrated dataset was generated to test another established method for data harmonization (33). The unintegrated dataset is generated by merging the datasets on common genes, followed by log-normalization, variable features selection, and scaling. The UMAPs (Uniform Manifold Approximation and Projection) show that in the low-dimensional space, batch effects are widespread in the absence of a formal integration (SI Appendix, Fig. S3A). On the contrary, batch effects appear strongly reduced in both Harmony (SI Appendix, Fig. S3B) and Seurat-integrated (SI Appendix, Fig. S3C) datasets. A similar pattern can be seen for technologies, with substantial differences in the uncorrected dimensions and normalization upon integration (SI Appendix, Fig. S4 A–C). While batch effects are reduced in the integrated data, biological variation at the cell type level appears well preserved in both the developing (SI Appendix, Fig. S5 A and B) and the adult brain (SI Appendix, Fig. S6 A and B).

Comparison of the iLISI measures also shows differences between the unintegrated and integrated data (Dataset S1). Specifically, integration results in higher iLISI scores, reflecting improved harmonization (Dataset S1). Seurat showed an overall better performance than Harmony. Nevertheless, residual batch structures may still occur but these involve relatively small clusters that remained separated to some degree. Residual batch effects may somewhat persist when integrating distinct datasets (31). Thus, compared to the uncorrected data, batch effects are largely minimized in the harmonized data while biological variation is conserved at the cell type level.

Identification of cell types.

We performed clustering analysis (Materials and Methods) on each harmonized dataset to identify the distinct cell communities in the developing (SI Appendix, Fig. S5A) and adult brain (SI Appendix, Fig. S6A). Differential gene expression (DGE) between cell clusters was conducted to aid cell type identification. The data indicate substantial heterogeneity and might also suggest potential cell type–specific marker genes as a resource for future studies (Datasets S2 and S3). We also verified the expression of established marker genes (SI Appendix, Fig. S6C). We validated our inferred cell type labels using two approaches: 1) We measured concordance between the original labels and our cell type designations; 2) in the adult brain, we re-assigned cell types using the tool BRETIGEA, which relies on marker genes to determine cell types on external datasets (46). We found that the original labels were largely concordant with our inferred cell type labels (average concordance ≥70% for the developing brain, ≥90% for the adult brain; SI Appendix, Fig. S7 A and B). Moreover, BRETIGEA-based classification was >80% concordant with our determined cell types. As cell identities are defined at the cluster level upon integration, these good concordance rates support our data harmonization. Nevertheless, identifying and aligning cells of the same biological type across studies is challenging for embryonic/fetal cells due to transcriptionally unresolved cell type ambiguities. Thus, for consistency with original studies, we reassigned the original cell type labels in the integrated developing dataset (Fig. 1A).

Fig. 1.

MECP2, CDKL5, and FMR1 expression levels in the developing brain. (A) UMAP of cells in the integrated developing brain dataset across all studies. (B) Bar plots of cell type fraction within the distinct brain regions from the integrated developing brain datasets, per (A). (C) Bar plots of cell type fraction within gestational weeks, GW6-13 and GW16-27, pooled from multiple donors. (D) Bar plot of average expression (log-scale) of each GOI within each available brain region.

The integrated embryonic/fetal brain dataset included 60,320 female cells, sampled from an initial population of ≈280,000 cells from multiple donors. The integrated adult brain dataset included 88,470 female cells, sampled from an initial population of ≈300,000 cells from multiple donors. The datasets comprised the major cell types neuronal and glial, and progenitor cells, spanning 10 distinct brain regions. Collectively, the dataset includes neurodevelopmental stages spanning gestational week (GW) 6 to 27 and ages 4 to 60, enabling a comprehensive single-cell expression survey of MECP2, CDKL5, and FMR1.

Embryonic and Fetal Brain: Cell Type– and Stage-Specific Profiles.

We examined profiles from GW6-10 to GW26 (Dataset S4). GW6-10 includes the first trimester of neurodevelopment, a crucial early period when basic neuroanatomical structures are laid down. GW17-18 is a neurodevelopmentally sensitive window linked to neuropsychiatric conditions (43). The dataset comprises germinal zones and cortical plate, two fundamental areas involved in early cell migration and molecular gradients during neurogenesis (47). A rich diversity of cell types was observed, including neurons, microglia, neuroepithelial cells, and radial glial and intermediate progenitor cells (Fig. 1A). Representation of various cell types differed across brain regions (Fig. 1B). Consistent with substantial neurodevelopmental changes during gestation, the two neurodevelopmental time windows, GW6-13 and GW16-27, demonstrated an increase in the relative fraction of neuronal and glial populations, along with a parallel decrease in stem cell lineages (Fig. 1C).

MECP2, CDKL5, and FMR1 expression levels in regions of the embryonic/fetal brain.

We then examined expression differences for the GOI in various brain regions (Dataset S5). Significantly, differences were detected between regions, particularly in the germinal zones and cortical plate (Fig. 1D). We performed down-sampling to 1,000 single cells per region, but still found a similar pattern of GOI expression (SI Appendix, Fig. S8). For CDKL5, by far the greatest expression came from the cortical plate. For FMR1, the strongest expression came from the cortical plate and germinal zones. By contrast, MECP2 expression was relatively high across multiple regions, though the greatest degree of expression was observed in the central and occipital cortices. The GOI also exhibited substantial variation in the number of cells/regions with expression (Fig. 2A). For instance, while MECP2 was most highly expressed in the occipital cortex and central cortex, its expression was evident in no more than ≈43% of cells in those regions. For CDKL5, the highly expressing cortical plate showed expression in only ≈13% of cells. For FMR1, the highly expressing germinal zones and cortical plates exhibited expression in only ≈14% of cells. We conducted DGE upon downsampling. Indeed, MECP2 was one of the ≈500 genes showing higher expression in the occipital cortex compared to all other regions combined (Log₂FC = 0.15, P_Adj = 8.6e-4). Within the occipital cortex (N = 1,702 cells), excitatory neurons (N = 745) and radial glial cells (N = 855) comprised the vast bulk of cells in the region, and neurons showed higher MECP2 expression than radial glial cells (Log2FC = 0.14, P ≤ 0.05, P_Adj ns; Fig. 2B).

Fig. 2.

Regional and developmental timepoint analysis. (A) Dot plots of scaled average expression of each GOI within each brain region. Telencephalon was excluded as no specific anatomic subregion was predefined. The size of each dot corresponds to the fraction of single cells expressing the gene. The degree of expression is indicated by the heatmap. (B) Violin plot of expression levels of MECP2 in excitatory neurons and radial glial cells populating the occipital cortex. The occipital cortex is chosen because our analysis shows that it exhibits upregulation of MECP2 expression during neurodevelopment, in comparison to other regions. (C) Average MECP2, CDKL5, and FMR1 expression in NPCs/progenitor cells, neurons, and glial cells at each time point during neurodevelopment (GW = gestational week). Only cells with a precisely assigned single time point (as per original metadata) were used. For more robust assessment, time points with ≤30 single-cells were removed.

MECP2, CDKL5, and FMR1 expression levels at distinct neurodevelopmental time points.

We assessed GOI expression across time points from GW6 to GW26. Substantial changes in expression were evident in major cell types across neurodevelopment (Dataset S6 and Fig. 2C). Of interest were transcriptional differences among three main cell types: i) progenitors, comprising stem and radial glial cells, intermediate and cycling progenitor cells; ii) neurons; and iii) glial cells, comprising OPCs, oligodendrocytes, microglia, and astrocytes. In neurons, MECP2 exhibited higher expression at GW10 than other available time points (|Log₂FC| = 0.22; P ≤ 0.05). We captured similar MECP2 profiles in neural stem cells, with a detectable peak at GW10 (P ≤ 0.05; P_Adj ns) relative to all other time points combined. In contrast, FMR1 exhibited a peak at GW17 in NPCs (|Log₂FC| = 0.17; P_Adj ≤ 0.05) and neurons (|Log₂FC| = 0.11; P_Adj ≤ 4.3e-11). CDKL5 showed the least variation across neurodevelopment in NPCs and glia (Fig. 2C), with no significant differences detected at any time point relative to all others. Finally, in glial cells, there were no significant GOI expression changes between time points. To gain more power, we re-grouped the time points into two wider windows: GW6-13 and GW16-27 (Dataset S7). In this case, the only GOI exhibiting a significant difference was MECP2, with greatest upregulation in neurons within GW6-13 of neurodevelopment compared to GW16-27 (Log2FC = 0.16; P_Adj ≤ 2e-40) and relative to NPCs and glial cells.

Adult Brain.

General transcriptional landscape.

The transcription landscape in the post-natal and adult brain is of particular interest, as therapeutic restoration would generally occur only following diagnosis in girls and women, ranging from toddlers to adults who have lived with RTT, CDD, and FXS for decades. Transcriptome analyses of MECP2, CDKL5, and FMR1 have been performed in the human brain and neural models but mostly in bulk samples (17, 48–56). Sub-anatomic and cell type–specific landscapes in the human brain are still largely unknown for these genes. Just as for the embryonic brain, our analysis here revealed differences in the GOI expression between distinct adult brain regions (Fig. 3A). MECP2 showed highest expression in the cerebellum, prefrontal/frontal cortex, anterior cingulate cortex, substantia nigra, and primary visual cortex. By contrast, CDKL5 showed least expression in the cerebellum and relatively high expression almost everywhere else. FMR1 had similar expression all throughout the brain regions.

Fig. 3.

MECP2, CDKL5, and FMR1 expression levels in the adult brain and in the GTEx datasets. (A) Average expression (log-scale) of MECP2, CDKL5, and FMR1 per adult brain region. Analysis is conducted per Fig. 1D. (B) Bulk tissue gene expression data (log10(TPM+1)) of MECP2, CDKL5, and FMR1 in each available tissue (sampled from both male and female donors) from the GTEx resource (https://gtexportal.org/home/).

Analysis of bulk expression datasets from ≈1,000 donors, 50 tissues, and multiple brain regions in GTEx (57) revealed similar findings (Fig. 3B). We observed highest MECP2 expression in the cerebellum and relatively lower CDKL5 expression in the cerebellum and substantia nigra, but the bulk dataset did not provide sufficient granularity to make conclusions regarding other brain regions and subregions or cell types. This analysis also indicates that MECP2 (as well as FMR1) shows substantial expression in a number of non-CNS tissues, including the lung, gut, and female reproductive organs (Fig. 3B), reaffirming the idea that, while RTT has major CNS manifestations, it is also a systemic disorder.

Neuronal and glial expression of MECP2, CDKL5, and FMR1 in each brain region.

We returned to the single-cell data and investigated differences between cell types of each region. We observed variation in both proportion of expressing cells and degree of expression within each region, with potential significant heterogeneity between neuronal and glial cell types (SI Appendix, Fig. S9). The cell fraction expressing the GOI widely ranged from ≈5.5 to ≈94% throughout the brain regions. For instance, the median neuronal and glial cellular detection rate of MECP2 across brain regions was ≈44%. In contrast, CDKL5 had a median detection rate of ≈69% across the brain regions, while for FMR1, it approached ≈40% (SI Appendix, Fig. S9). To gain more detailed knowledge, we performed DGE between neuronal (excitatory and inhibitory neurons) and glial cells (astrocytes, oligodendrocytes, OPCs, microglia) within each region (Dataset S8). This approach allowed us to identify genes with marked differences between neuronal vs. glial cells and, in parallel, display the regional identity of these patterns. We found that MECP2 expression in neurons was somewhat lower than glial cells in the primary motor cortex (PMC: |Log₂FC| = 0.13; P_adj = 8e-143) and primary visual cortex (PVC: |Log₂FC| = 0.13; P_adj = 2e-6) (Fig. 4A and Dataset S8). Interestingly, within these two regions, CDKL5 exhibited opposite patterns, with overall higher expression in neurons than glial cells in both PMC (|Log₂FC| = 0.77; P_adj = 8.8e-23) and PVC (|Log₂FC| = 0.7; P_adj = 7.3e-5) (Fig. 4A and Dataset S8).

Fig. 4.

Expression variability between cell types and donors in the adult brain. (A) Dot plots of (unscaled) average expression of MECP2 and CDKL5 in glia and neurons in the primary motor cortex and primary visual cortex, computed on randomly downsampling to the less represented cell type within each region. Log-normalized expression values are used. The size of each dot corresponds to the fraction of single cells expressing the gene. The average expression is indicated by the heatmap. (B, Left) bar plots of % of grouped neuronal and grouped glial cells expressing MECP2 in single-nucleus RNAseq data of human, chimpanzee, marmoset, and rhesus dorsolateral prefrontal cortices (dlPFC) (58). (Right) bar plots of % of excitatory neurons, inhibitory neurons, astrocytes, microglia, OPCs, and oligodendrocytes expressing MECP2 in the dlPFC in each of the four species. (C) Violin plots of normalized MECP2 expression levels in grouped neuronal and grouped glial cells and in the distinct cell subtypes in the human and rhesus. (D) Violin plots of normalized expression for CDKL5 (Top) and MECP2 (Bottom) in neurons vs. glia in 3 donors assessed for CDKL5 and 2 donors assessed for MECP2. (E) Variance partitioning analysis to estimate the contribution of cell type to variation in GOI gene expression within each donor. Bar plot of fraction of variation in MECP2, CDKL5, and FMR1 expression within each donor explained by variation across cell types. Residuals indicate unexplained expression variation. (F) MECP2 and CDKL5 average expression levels in inhibitory neurons within each of 9 donors (donors represented on the x axis). Average expression (log-scale) of MECP2 (red line) and CDKL5 (blue line) is shown. (G) MECP2 average expression levels in excitatory neurons within each of 9 donors (donors represented on the x axis). Average expression (log-scale) of MECP2 is shown. CDKL5 is not shown, as there was no evidence for statistically significant differences across donors. (H) MECP2 and FMR1 average expression levels in excitatory neurons upon downsampling within each of 9 donors (donors represented on the x-axis). Average expression (log-scale) of MECP2 (red line) and FMR1 (black line) is shown.

A high neuronal expression of CDKL5 in the PVC may support its role in excitatory visual synapses, explaining why individuals carrying CDKL5 mutations may manifest deficits in visual attention and acuity (59). A biologically important observation may be the reciprocal patterns of MECP2 and CDKL5 in PMC and PVC in glia and especially in neurons (Fig. 4A). This may suggest potential co-regulation between these two genes in these disease-associated regions, a hypothesis supported by previous evidence (60). Higher expression of CDKL5 in neurons than glia cells was also detected in the prefrontal/frontal (|Log₂FC| = 0.72; P_Adj = 9e-14), somatosensory (|Log₂FC| = 0.96; P_Adj = 0.04), auditory (|Log₂FC| = 0.92; P_Adj = 1.5e-8), middle temporal gyrus (|Log₂FC| = 0.85; P_Adj = 3.8e-15) and anterior cingulate cortex (|Log₂FC| = 0.76; P_Adj = 6.7e-5) (Dataset S8). In the case of FMR1, expression was greater in neurons in PVC (|Log₂FC| = 0.23; P_Adj = 2.5e-5) (Dataset S8). Other GOI fold-change signals in other regions were weaker or insignificant after adjustment of covariates and multiple testing correction. Altogether, these data indicate that the GOI may exhibit significant and distinct patterns of variation between neuron and glial cell types across regions of the human brain. This may also suggest that, in pathological states characterized by altered cell type composition, abnormal MECP2, CDKL5, and FMR1 dosages could occur and have a differential impact on distinct brain areas.

MECP2 expression in neurons vs. glia in the human and non-human primate (NHP) brains.

We detected higher MECP2 expression in neurons than other cell types in the occipital cortex of the fetal/embryonic brain. However, we captured marginally higher MECP2 levels in grouped glial cells than neurons in PMC and PVC of the adult brain (Fig. 4A and Dataset S8). Higher MECP2 expression in glial cells was surprising, as prior studies reported little or no expression in glial cells, and lower glial than neuronal expression—and thus, it has been assumed that MECP2 could play a larger role in neurons than in glia in disease pathogenesis (61–76).

Given this discrepancy, we investigated further. We reasoned that grouped glial cells are a more complex composite of distinct cell types than grouped neurons. In both PMC and PVC, we computed: i) average expression in neurons and within each glial cell type; ii) fraction of cells expressing MECP2; and iii) DGE between neurons vs. each glial cell type. We found that, in both regions, MECP2 was expressed in more neurons than glial cells, as expected. However, to our surprise, glia appeared to exhibit overall higher MECP2 expression levels on average (Dataset S9). We downsampled the data to the under-represented cell type (glial) and replicated a marginally higher MECP2 signal in glia than neurons in PMC (|Log2FC| = 0.14; P_Adj = 1e-9), but not in PVC (|Log2FC| = 0.11; P = 4.3e-5; P_Adj NS). To confirm, we repeated these analyses in the PRJNA4340042 dataset (40), comprising ≈11K nuclei from prefrontal (PFC) and anterior cingulate cortices (ACC) from four female donors. Again, we observed more neurons with MECP2 expression than glial, but glia could exhibit stronger MECP2 expression when averaged (Dataset S10). We detected higher MECP2 levels in PFC microglia compared to all other cell types combined (|Log2FC| = 0.22; P_Adj = 1e-9).

As the high glial cell expression of MECP2 was unexpected, we turned to the NHP brain to ask whether a similar trend could be observed. We analyzed single-nucleus RNAseq data (≈611 K nuclei) from adult chimpanzee, marmoset, and rhesus dorsolateral prefrontal cortices (dlPFC) (58) and compared them to the adult human cortex. In human dlPFC, MECP2 was expressed by ≈56% of neurons and ≈20% of glial cells. Similarly, more neurons than glial cells expressed MECP2 in the chimpanzee (≈62.5 vs. ≈18.4%), marmoset (≈48.3 vs. ≈22%), and rhesus (≈64% vs. ≈24%) (Fig. 4B). These patterns remained discernible when looking at each cell subclass separately (Fig. 4B). In each species, we sampled 1,000 neurons and 1,000 glial cells and conducted DGE testing. We found a marginally higher MECP2 expression in glia than neurons in humans (|Log2FC| = 0.1; P_Adj = 3.7e-5) and a stronger difference in rhesus (|Log2FC| = 0.23; P_Adj = 5.5e-13), but not in the chimpanzee and marmoset. To gain deeper insights, we identified all MECP2-expressing cells and sampled 1,000 MECP2-expressing neurons and 1,000 MECP2-expressing glial cells in each species. In all four species, MECP2 expression was higher among MECP2-expressing glial cells than among MECP2-expressing neurons (|Log2FC| range: [1, 1.32]). The strongest difference was observed in rhesus (Fig. 4C). Therefore, both neurons and glia cell types may contribute to the MECP2 dosages in a bulk cell population within the primate brain. Altogether, our analyses show that MECP2 i) is expressed in both neurons and glial cells; ii) is predominantly expressed in neurons; but iii) MECP2-expressing glial cells may exhibit higher MECP2 levels than neurons; and iv) the high MECP2 expression in glia occurs in relatively small subsets of cells.

Within an Adult Brain: High Variability for MECP2 in Inhibitory Neurons and for FMR1 in Multiple Cell Types.

Intra-cell type MECP2, CDKL5, and FMR1 variability.

Single-cell analysis also permits the study of expression variability within one cell type and between distinct cell types. Variability across cells of the same biological type was assessed by modeling the relationship between the coefficient of variation and average expression using an established method for scRNA-seq (77) (Materials and Methods). Among the three GOI, FMR1 exhibited variability in both neuronal (excitatory neurons: minLogBioVar = 0.27, maxFDR < 2e-4; inhibitory neurons: LogBioVar = 0.35, FDR = 9.8e-8) and glial cell types (microglia: minLogBioVar = 0.22, maxFDR = 0.014; astrocytes: minLogBioVar = 0.24, maxFDR = 5e-3) within multiple unrelated adult donors. MECP2 expression demonstrated high variability in inhibitory neurons within two donors (minLogBioVar = 0.26, maxFDR = 0.02), while no instances of variability were detected for CDKL5 (SI Appendix, Fig. S10). A source of transcriptional heterogeneity across cells of the same type is the distinct subtypes that may co-exist within the cell population. To account for this phenomenon, we sub-clustered cells of the same biological type within each donor at resolution 0.5. We selected the largest cluster as a transcriptionally more homogeneous group of cells and performed hyper-variable gene analyses as per above. We replicated the significant variability of FMR1 in astrocytes (LogBioVar = 0.24, P = 5e-4, FDR = 0.01) and microglia (LogBioVar = 0.53, P = 2e-11, FDR = 1e-9), and MECP2 in inhibitory neurons (LogBioVar = 0.46, P = 2.3e-3, FDR = 0.07). To test further, we used a distinct feature selection method which characterizes variable expression using the relationship between dropout rate and expression (78). This method also identified FMR1 as highly variable. Therefore, within any individual brain, we observed events of high MECP2 variability in inhibitory neurons, and high FMR1 variability in both neurons and glial cells. The observation that FMR1 is more frequently variable than CDKL5 and MECP2 might suggest more complex transcriptional regulation. This phenomenon warrants further investigations considering the non-negligible variability in neurological trait expressivity across females with a confirmed diagnosis of FXS. Another facet of this phenomenon may link to the X-inactivation status of FMR1: Such variability, for example, could result from variable escape from X chromosome inactivation (XCI), especially as escape from XCI in humans may be more plastic than originally thought (79, 80). The degree to which variable escape occurs in tissues and individuals is unclear, and the use of single-cell screenings and unconventional biological models may reveal novel escapees.

Inter-cell type MECP2, CDKL5, and FMR1 variability.

To assess GOI transcriptional differences between cell types within an individual, we conducted DGE between neurons (excitatory and inhibitory neurons) vs. glial (astrocytes, microglia, oligodendrocytes, OPCs) cells within each female donor (Fig. 4D). We tested only genes detected in at least 30% of cells in either of the two cell classes within each donor. We downsampled cells to match the under-represented cell type and found CDKL5 to be significantly more expressed in neurons than glial cells (min|Log₂FC| = 0.8; P_adj ≤ 0.05) within 3 out of 6 (50%) donors in which CDKL5 was tested (Fig. 4D). On the other hand, MECP2 manifested marginally higher overall expression in glia than neurons (min|Log₂FC| = 0.11; P_adj = 4.9e-3) in 2 out of 4 (50%) tested donors (Fig. 4D). These data line up well with our results on region-specific neuronal vs. glial comparison, where we found instances of higher CDKL5 expression in neurons than glia, and instances of marginally higher MECP2 expression in glial than neuronal cells (SI Appendix, Fig. S9 and Dataset S8). It should be noted that, in any given snapshot across any region of the adult brain, more neurons demonstrate MECP2 expression than glial cells—but glial cells may show higher MECP2 expression among cells that express the gene (Fig. 4C, SI Appendix, Fig. S9). Microglia, OPCs, oligodendrocytes, and astrocytes may therefore demonstrate higher MECP2 expression than neurons, whereas this tends not to be the case for CDKL5 and FMR1 (which showed stronger neuronal expression in general). High MECP2 expression in some glial cells suggests that glial anomalies may contribute significantly to the etiologies of all three disorders.

Variance partitioning analysis of MECP2 and CDKL5 expression.

To complement these results with data on expression variance, we conducted a variance partitioning analysis to assess the potential impact of cell type on the MECP2 and CDKL5 expression variability in each donor. Specifically, variation across cell types was modeled via linear mixed-effect models (Materials and Methods). We found that on average, cell type explains ≈9% and 3% of the variation in CDKL5 and MECP2 expression, respectively, in a donor, after accounting for all other variables (Fig. 4E). In summary, cell type may account for a non-negligible fraction of variation in CDKL5 and MECP2 expression within an adult female brain.

Analysis of FMR1 Expression in Males.

As FXS affects both males and females, investigating FMR1 expression profiles in both sexes, as well as sex differences in FMR1 transcription is of interest. To address this, we retrieved PFC and ACC data from the 12 male donors in PRJNA4340042 (40). We processed the data consistently as performed for each female dataset (Materials and Methods) and calculated: i) average FMR1 expression in each cell type; ii) fraction of FMR1-expressing cells; iii) DGE between neuronal vs. non-neuronal cells; iv) intra-individual variability per cell type; v) DGE between males vs. females in neurons and glia. We found that in both sexes, FMR1 is predominantly detected in neurons than glial cells (Datasets S11 and S12). In males, neurons (excitatory + inhibitory) and glial cells (microglia + astrocytes + OPCs + oligodendrocytes) have highly similar overall FMR1 expression levels (|Log2FC| = 0.03; P_Adj = 0). Intriguingly, we detected slightly different patterns in females, with higher FMR1 expression in grouped neurons than grouped glial cells (|Log2FC| = 0.12; P_Adj = 2e-29). These female-specific patterns align with our observations of higher FMR1 expression in neurons than glia in the female integrated data assay.

To gain insights into the FMR1 expression variability in males, we conducted similar analyses as above: In each male donor, we assessed i) variability across cells of the same biological type; ii) DGE between neurons vs. glial cells. We found FMR1 variability in both excitatory (LogBioVar = 0.54; FDR = 0) and inhibitory neurons (LogBioVar = 0.5; FDR = 4.3e-5) within 2 donors (ID: 5242, 6032). No instances of hyper-variable FMR1 were detected in glia. Contrarily to our female integrated dataset, we could not replicate these FMR1 patterns in males with an alternative method (78). This suggests that FMR1 expression variability in males may be lower than in females. We then conducted DGE between neurons vs. glial cells, considering only genes with minimum detection rate of 10% in either of the two cell classes within each donor and upon down-sampling. We found slightly higher FMR1 levels in neurons than glial cells (|Log2FC| = 0.14; P_Adj = 5.4e-13) in only 1 out of the 10 (10%) male donors in which FMR1 was tested. In another donor, FMR1 exhibited the opposite behavior, with higher glial expression (|Log2FC| = 0.14; P_Adj = 0.02). We found no significant FMR1 expression changes between neurons and glia in other male donors. These results align with our previous results: in most males, the FMR1 expression levels are highly similar between neurons and glial cells. Finally, we compared FMR1 expression between males and females and found no significant sex differences after correcting for confounding variables. We took a step forward and assessed female and male FMR1 expression in the GTEx bioresource. In agreement with our data above, we mostly observed similar FMR1 levels in the two sexes (SI Appendix, Fig. S11).

Analysis of Variability of MECP2, CDKL5, and FMR1 Expression between Individuals.

Also of significant biomedical interest is the degree of variability between individuals. Previous studies have indeed characterized inter-individual transcriptional variation across cell types and in the brain, e.g., (81). At present, however, the information for MECP2, CDKL5, and FMR1 expression is not well documented. Variability may correlate with disease risk or severity, and the magnitude of the differences may inform various therapeutic strategies regarding the degree to which the gene products need to be restored.

Inter-individual variation in MECP2, CDKL5, and FMR1 expression.

To investigate inter-individual variation for our GOI, we searched for significant down- or up-regulation by DGE between each individual and the remaining set of individuals per cell type, while correcting for confounding variables. Significant signals were seen for MECP2 (av. expression range: MECP2_ExNeu[0.35,0.9]; MECP2_InNeu[0.41,1.1]) and CDKL5 (CDKL5_ExNeu[0.87,2.75]; CDKL5_InNeu[1,1.56]) in neurons. Within the same donor (‘H200.1023’), MECP2 (|Log₂FC| = 0.13; P_adj = 2.7e-4) and CDKL5 (|Log₂FC| = 0.18; P_adj = 4e-3) both exhibited significantly lower expression in inhibitory neurons than the average expression of all remaining donors (Fig. 4F). A similar pattern was seen in the donor ‘H18.30.001’, where MECP2 had lower than average expression (|Log₂FC| = 0.17; P_adj = 7.4e-4). Other signals involved MECP2 in excitatory neurons, with events of significantly higher than average expression (|Log₂FC| = 0.3; P_adj = 1.2e-16) (Fig. 4G), and lower CDKL5 expression in oligodendrocytes (|Log₂FC| = 0.14; P_adj = 4e-2; CDKL5_Oligo[0.83,1.14]). We also detected a MECP2 and FMR1 signal in excitatory neurons (P ≤ 0.01) (Fig. 4H) in downsampled data. Other individual- and cell type–specific signals had smaller effect sizes or were statistically insignificant.

Estimating GOI transcript dosages per cell.

For therapeutic purposes, knowing the number of GOI transcripts within a healthy cell could inform on the tolerated dosages and neurotoxicity margins. However, the detection of a transcript in a cell has a stochastic component and only a fraction of transcripts may be captured (e.g., ≈8 to ≈32% of mRNAs for 10× Genomics). Thus, scRNAseq may not detect the real transcriptomic dosages present within a cell. For example, using the SCT sequencing-depth adjusted counts and cells sequenced with 10X Genomics, we sampled 200 cells among GOI-expressing cells within each donor. We estimated the average number of MECP2 transcripts per cell to range from ≈1.5 to over 3, and the average number of CDKL5 transcript molecules per cell to range from ≈2 to over 3 between donors. If these captured transcripts represent a low fraction of 8% of MECP2 and CDKL5’s transcriptomes, we could estimate that within sampled cell populations of similar size, the average MECP2 and CDKL5 transcript load could reach ≈6,000 and ≈7,000 molecules, respectively. If this model holds, we would expect ≈30 to 35 MECP2 and CDKL5 molecules per cell, on average, within a donor. Clearly, owing to the stochastic detection in each cell and the use here of in silico approaches and assumptions, these estimates need caution and would need experimental validation. The dosage of GOI molecules present in a cell at a given time may be significantly more variable, including be several fold higher, and depending on multiple factors. On the other hand, comparing average expression between donors upon downsampling, reveals differences of up to ≈2.5 fold for MECP2 [av. expr. range 0.43-1.1] and CDKL5 [0.96-2.3], and up to ≈1.8 fold for FMR1 [0.38-0.7]. Altogether, these data suggest that individuals in the population may exhibit different MECP2 and CDKL5 expression in the brain.

Co-Varying Genes and Potential Biomarkers.

MECP2, CDKL5, and FMR1 are involved in disparate biological processes (82–87). The variability in expression of MECP2, CDKL5, and FMR1 provided an opportunity to identify co-varying downstream genes that could be potential biomarkers for RTT, CDD, and FXS, and be especially valuable for candidate disease-modifying drugs undergoing clinical trials.

To identify putative biomarkers, we analyzed co-expression between each GOI and other genes per cell type in the adult brain. We employed scLink, a recent algorithm tailored for scRNA-seq data (88). Arbitrary stratification of correlation strengths could drive inconsistencies and low reproducibility (89). Moreover, large gene sets might incorporate sparse patterns, lowly expressed genes or false zero detection, which hinder accurate measurement of co-expression and increase computational burden (88, 90). To select a robust set of genes co-expressed with a GOI in each cell type, we implemented a multiphase strategy. First, we run an initial round of measurement of gene correlation matrix; to remove potentially weak relationships, we identified all genes correlated with at least a GOI with an absolute coefficient higher than 0.4 and used these to infer co-expression networks and re-measure correlative strengths. Positive correlations were more frequent throughout the genome (SI Appendix, Fig. S12). Second, in each cell type, we identified all genes with an edge (i.e., an interaction) with at least a GOI at all regularization λ values tested (88, 91). Third, for each gene pair (consisting of a GOI and another gene) in the network, we estimated the statistical significance of the correlation using bootstrapping followed by multiple testing correction (Materials and Methods).

We ranked the top correlated genes with each GOI and visualized their co-expression and anti-correlation networks within each cell type (Dataset S13; top 50 genes in neurons in Figs. 5–7). A total of 364 genes were co-expressed with MECP2 in neurons. 35 were specific to excitatory neurons, while 291 were restricted to inhibitory neurons. 38 (10.4%) were shared. Intriguingly, UBE3A, an autosomal imprinted gene, was the only gene significantly anti-correlated with MECP2 in neurons (ρ = −0.35; P = 5e-3) (top 50 Fig. 5; Dataset S13). In contrast, PTPRD was anti-correlated with MECP2 in microglia (SI Appendix, Fig. S13 and Dataset S13). Notably, PTPRD is a receptor protein tyrosine phosphatase associated with neurodevelopmental disorders and playing important neurobiological roles (92).

Fig. 5.

Potential biomarkers for Rett syndrome. Analysis of genes correlated or anti-correlated with MECP2 within excitatory and inhibitory neurons. Barplots of scLink’s correlation coefficient between each GOI and the top 50 positively and negatively correlated genes (if any). Barplots for glial cell types are shown in SI Appendix, Fig. S13. Correlation networks using igraph on the top 25 (or all if <25 are available) positively and negatively correlated genes: The GOI is placed as the central node; network connections were assigned distinct colors: i) gray for ρ < 0, light green for 0 ≤ ρ ≤ 0.2, violet for 0.2 < ρ ≤ 0.5, orange for ρ >0.5; ii) edge widths (corresponding to absolute scLink’s correlation coefficient) were uniformly augmented of a factor of 0.3; iii) vertex label sizes (i.e., gene names) assigned to scLink’s correlation coefficients; iv) vertex sizes (i.e., nodes) were uniformly multiplied by a constant factor of 9. Complete correlation tables are in Dataset S13.

Fig. 6.

Potential biomarkers for CDKL5 syndrome. Analysis per Fig. 5 legend. Glial cell types are shown in SI Appendix, Fig.S14. Complete correlation tables are in Dataset S13.

Fig. 7.

Potential biomarkers for FXS. Analysis per Fig. 5 legend. Glial cell types are shown in SI Appendix, Fig. S15. Complete correlation tables are in Dataset S13.

We detected 10 and 19 genes co-expressed and anti-correlated, respectively, with CDKL5 in neurons. BCYRN1 is a gene anti-correlated with CDKL5 in neurons. A single gene, the guanine nucleotide exchange factor RASGEF1B, showed anti-correlation in both excitatory and inhibitory neurons (Fig. 6 and Dataset S13). UBE3A was co-expressed with CDKL5 in neurons, thus exhibiting a behavior antithetical to MECP2. This pattern could reflect the potential MECP2 activity as a repressor of CDKL5(60).

By contrast, FMR1 co-expressed with 223 genes in neurons, with 36 and 170 restricted to excitatory and inhibitory neurons, respectively, and 17 (≈8%) shared (top 50 Fig. 7; Dataset S13). Our analyses revealed that BCYRN1 is also anti-correlated with FMR1 in neurons. It was reported that FMRP may bind to BCYRN1(93, 94). We identified a subset of 3 genes, MDGA2, MEG3, and RIMS2, significantly anti-correlated with FMR1 in astrocytes (SI Appendix, Fig. S15 and Dataset S13). Intriguingly, MEG3 is an imprinted lncRNA gene. MEG3 has been described in diverse contexts (95), including neurodevelopment, however its functions in the human brain remain unknown. On the other hand, the gene RIMS2 encodes a presynaptic protein which might be involved in autism (96). Compared to neurons, more genes co-expressed with our GOI in glial cells (top 50 SI Appendix, Figs. S13–S15 and Dataset S13).

Our approach enabled building networks of top correlated genes with each GOI in each cell type (Figs. 5–7). The strengths of these relationships varied between GOI and cell types, involving distinct gene sets. Notably, ≈3.3% (N = 316) of genes that co-expressed with at least a GOI in a cell type are X-linked, such as XIST, and biomedically relevant genes such as RPS4X, TSPAN6/7, ATRX, USP9X, DDX3X, KDM6A, TMSB4X, NLGN4X, and OGT. Co-fluctuation of neighboring genes may result from cis-acting genetic effects and concurrent regulation. Combining in silico analyses with functional genomic screens will be essential to assess the power of these genes as predictors of disease and their druggability for therapeutics.

GO analysis of genes co-expressed with MECP2, CDKL5, and FMR1.

Gene co-expression may indicate shared involvement in biological function and might have key roles in disease (97). To get insights into the potential biological roles of these relationships, we performed functional enrichment analyses for each GOI (SI Appendix, Fig. S16). We found enrichment for terms involving a wide range of bioprocesses and molecular functions, such as at the RNA and protein level, metabolic processes, and synaptic components (Datasets S14–S16). Overrepresented terms also included nervous system development and phenotypes associated with neurodevelopmental delay, intellectual disabilities, autistic behavior, seizures, and altered neuronal morphologies (e.g., microcephaly). GOI-coexpressed genes also enriched for KEGG (Kyoto Encyclopedia of Genes and Genomes) pathways of neurodegeneration, such as Parkinson and Alzheimer diseases (SI Appendix, Fig. S16). The diffuse enrichment of these terms highlights the critical neurological roles of these genes. Co-expression is not an unbiased indication of direct regulatory interactions, but it may highlight genes involved in coordinating mechanisms in disease. These data may provide a foundation for further studies of downstream therapeutic targets.

Replication of MECP2 and FMR1 co-varying genes.

We sought to replicate the inferred MECP2 and FMR1-covarying genes in other RNA-seq data from human brain samples from the respective disease model. There are currently no accessible data from CDD donors. For RTT, we used data from temporal and cingulate cortex samples from females diagnosed with RTT (98). Using the NCBI raw data, we kept genes with at least 10 read counts, performed CPM + TMM normalization and log2(x+1) transformation. We then calculated Pearson’s correlation between MECP2 and any other gene. We identified 5,512 genes with minimum correlation of 0.2 and overlapping with our inferred MECP2-covarying genes. We found that ≈60% of these genes show correlative patterns with MECP2 concordant with our inferred patterns.

For FXS, we used RNAseq data from organoids derived from FXS male patients, and data from the fetal cortex from FXS fetuses from the same study (56). Using genes with at least 10 read counts in each dataset, we merged the two datasets on common genes. We performed CPM + TMM normalization, log2(x+1) transformation, and batch correction (99). We then computed Pearson’s correlation between FMR1 and any other available gene. We identified 1,896 genes with minimum correlation coefficient of 0.2 and overlapping with our FMR1-covarying genes. We found that ≈58% of these genes show correlative patterns with FMR1 concordant with our inferred patterns. We also replicated BCYRN1 and MEG3 anti-correlation with FMR1.

Discussion

Here, we have used existing single nucleus/cell RNA-seq data to profile the human brain across multiple developmental stages, in various brain regions, and in multiple donors. We focused on three X-linked genes—MECP2, CDKL5, and FMR1—whose mutations result in devastating neurodevelopmental disorders associated with autism. For all three genes, expression is observed in both neuronal and glial cells, but there is considerable variability within a cell type, between cell types, and between brain regions. We also observe considerable inter-individual variability, with differences of twofold to threefold between healthy individuals. The wide range of expression may help define a therapeutic window in drug development efforts, with the low end informing a minimum for normal physiological function and the high end a potential toxicity margin. Our analysis also identified co-varying genes for MECP2, CDKL5, and FMR1 within each cell type. Because co-varying genes can either share regulation and/or be part of the same functional networks, they may serve as a foundation for identifying biomarkers for RTT, CDD, and FXS in the future. As the whole network homeostasis may depend on the normal functioning of the single members, aberrant expression of a single gene within a network of co-varying genes might systemically dysregulate the entire network with profound cellular effects. The goal of restorative therapies is to normalize expression not only of the mutated gene, but also of the entire downstream network.

Documenting the behavior of MECP2, CDKL5, and FMR1 across functionally distinct cell types and regions of the brain can also be informative for understanding RTT, CDD, and FXS. For instance, we captured higher CDKL5 expression in neurons than glia in the PVC, potentially helping to explain visual attention deficits in CDD. When designing gene restoration therapies for CDD, strategies that achieve higher CDKL5 restoration in neurons could be prioritized. Single-cell assays could be employed in the clinics to monitor the GOI in accessible tissues via minimally invasive procedures. For example, GOI molecular profiles could be measured in cells and tissues outside the brain and then correlated via cross-cell type and cross-tissue comparisons with publicly available human brain data to identify cell types and tissues exhibiting highly concordant GOI levels with those in the brain. Likewise, genes co-varying with the GOI could be used as proxy of the GOI levels to monitor gene restoration therapies.

In the case of RTT, we found that MECP2 is expressed in more neurons than glial cells, in line with prior understanding. However, we also observed higher MECP2 expression in glial cells on average, due to very high levels expressed in a small subset of glia. We observed similar findings in NHPs. The high-level MECP2 expression in some glial cells was unexpected, as earlier investigations in rodents reported the presence of MECP2 protein in neurons but absence in glia (61–64). Robust neuronal MECP2 expression was also detected in the monkey prefrontal cortex (65). These pioneering studies were mostly limited to immunohistochemical assays and mostly examined protein levels. Later studies documented the presence of MECP2 mRNA and protein in both neurons and multiple glia cell types (100–106). Some studies found that levels in astrocytes were lower than in other glial cell types (100), while others show lower levels in microglia than astrocytes and neurons (102). Still others found that astrocytes expressed MECP2 at lower levels than neurons (100, 101, 106). On the other hand, other studies found that MECP2 levels in glia and cerebellar granule neurons were similar, while other neuronal classes showed higher expression (100). Regardless, the presence of substantial MECP2 levels in glia has been repeatedly documented, and the roles of glia as an integral component of RTT pathophysiology have been demonstrated (100–103, 106–113). Astrocytes and microglia carrying defective MECP2 have consequences for neurological function. Notably, restoration of MECP2 in astrocytes has been shown to rescue neurological and neurobehavioral traits and prolong lifespan in mice (103). Taken together with prior work, our analysis suggests that MECP2 deficiency in some glial cells may contribute significantly to RTT. We note, however, that our analysis focused only on MECP2 transcript levels. Relevant to this, some studies have documented a low correlation between MECP2 transcript and protein levels (61, 106), implicating post-transcriptional regulation as a significant factor.

Limitations of the Study

We have used single-cell analytical methods to gain insight into the expression pattern of three ASD GOI and to inform future approaches to RTT, CDD, and FXS neurotherapies. Although single-cell transcriptomics is revolutioning precision medicine, we caution against over-interpreting the data (30). A large fraction of the transcriptome may be unprofiled due to technical limitations. Stochastic detection due to sampling variation, sequencing depth, and baseline expression could also affect detection power and sparsity. Other issues concern the cell type inference. While unsupervised clustering paralleled to DGE may aid classification of cell types based on established marker genes, uncertainty for rare or under-represented cell types may still be a challenge. Rare cell types and subtypes, or cell states altered by disease or experimental conditions could escape profiling with standard protocols. Lastly, because single-cell RNA-seq assays generally lack spatial data, we could not study the spatial context of GOI expression within subregions of the brain (114).

Materials and Methods

Here below is a brief description of Materials and Methods. The detailed Materials and Methods are provided in SI Appendix.

Data Preprocessing and Integration.

Each dataset was processed independently. Genes undetected in all cells, and "LOC" genes were removed. Cells with hemoglobin gene expression detection rate ≥5% were removed. Cells with <500 unique features or a detection rate of mitochondrial expression >5% were discarded. Fetal/embryonic and adult datasets were normalized and integrated independently in Seurat. Cell clusters were identified and cell types inferred. Analyses of DGE between groups of interest were performed using MAST. DGE P-values were adjusted using Bonferroni.

Gene Co-Expression and Functional Enrichment analyses.

For each cell type, gene co-expression relationships were inferred with scLink. A first round of gene–gene correlations was implemented. Genes correlated with at least a GOI with absolute coefficient >0.4 were identified and used to infer co-expression networks. Genes with an edge with a GOI were identified. Statistical significance of correlation was assessed using bootstrapping and P-values were adjusted using FDR. Significant genes were displayed using igraph functionalities. Genes co-expressed with a GOI in at least one cell type were identified; the resulting 3 lists (one per GOI) were used for over-representation analysis of gene ontology terms in g:Profiler.

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (XLSX)

Dataset S02 (XLSX)

Dataset S03 (XLSX)

Dataset S04 (XLSX)

Dataset S05 (XLSX)

Dataset S06 (XLSX)

Dataset S07 (XLSX)

Dataset S08 (XLSX)

Dataset S09 (XLSX)

Dataset S10 (XLSX)

Dataset S11 (XLSX)

Dataset S12 (XLSX)

Dataset S13 (XLSX)

Dataset S14 (CSV)

Dataset S15 (CSV)

Dataset S16 (CSV)

Data, Materials, and Software Availability

The datasets used for this study include Allen Brain Atlas datasets (37), PRJNA434002 (40), GSE97930 (38), GSE126836 (39), RRID_SCR_002001 (41), GSE132672 (42), phs001836 (43), GSE104276 (44), GSE131258 (45) are all publicly available on the Gene Expression Omnibus (GEO) repository (115) and/or dedicated websites. The processed single-nucleus RNAseq data of adult human, chimpanzee, macaque, and marmoset dlPFC (58) were downloaded from http://resources.sestanlab.org/PFC/.