Background

A number of neurodevelopmental syndromes are caused by mutations in genes encoding proteins involved in epigenetic regulation [1, 2]. Loss of function of proteins encoded by such genes is expected to result in alterations of epigenetic marks at specific genomic loci. To test this hypothesis, we elected to study the X-linked gene KDM5C, encoding histone H3 lysine 4 (H3K4) demethylase. Mutations in the KDM5C gene (MIM No: 314690) were first described as causing X-linked intellectual disability (XLID) in 2005 [3]. To date, 21 different KDM5C mutations have been identified in XLID patients. The prevalence of KDM5C mutations in patients with XLID is estimated to be ~3% [3–10]. The clinical features most consistently reported in males with mutations include mild to severe intellectual disability (ID), epilepsy, short stature, hyperreflexia, aggressive behavior and microcephaly. In addition, a mutation in KDM5C was identified in one male case of autism [11]. Female mutation carriers are usually unaffected but sometimes demonstrate mild ID or learning difficulties [7].

KDM5C is a member of the evolutionarily conserved KDM5 family of four proteins, KDM5A/B/C and D. KDM5A/C/D demethylate tri- and di-methylated forms of H3K4, whereas KDM5B is capable of demethylating all three forms (tri-, di-, and mono) of H3K4 methylation [12, 13]. The KDM5C protein contains several conserved functional domains, including the Bright/ARID DNA binding domain; the catalytic JmjC domain; the JmjN domain responsible for protein stability; the zinc finger-C5HC2 domain; and two PHD domains, responsible for histone binding [14] (Figure 1). Mutations leading to XLID have been found in most of the functional domains of this protein [15]. KDM5C is ubiquitously expressed in almost all human tissues including white blood cells, with the highest levels of expression found in the brain and in skeletal muscle [3, 15].

Figure 1
figure 1

Schematic diagram of the human KDM5C protein. The diagram shows the functional domains and the positions of 5 mutations as well as the p.R1546Q variant of unknown clinical significance.

Full size image

A significant effort has been invested in elucidating the role of KDM5C mutations in the ID phenotype. In zebrafish, downregulation of KDM5C leads to an increase in neuronal cell death and a decrease in total dendritic length [13]. Chromatin immune precipitation (ChIP) of HeLa cells revealed that KDM5C co-localizes with a transcriptional repressor REST, in the promoters of a subset of REST target genes, suggesting that loss of KDM5C activity impairs REST-mediated neuronal gene regulation [16]. ChIP-sequencing of a panel of chromatin remodeling proteins in the leukemia cell line K562 had shown that KDM5C along with other transcriptional repressors binds to a wide range of promoters, including those that are active, competent, and repressed [17]. The specific molecular mechanism by which loss of function of KDM5C causes impairment in neuronal development is not understood, but epigenetic deregulation is presumed to play an important role.

The KDM5C protein is likely to play a role not only in ID but also in sex-specific differences in brain function. The X-linked human KDM5C and its mouse ortholog Kdm5c escape X–inactivation [18, 19]; and, not surprisingly, Kdm5c has higher expression levels in XX females compared to XY males in mouse adult brain [20]. This difference has been shown to be associated with sex chromosome complement (XX vs. XY), rather than gonadal sex of the animals [21]. Interestingly, there is a Y-linked functional homologue of KDM5C, namely KDM5D, in both human and mouse. The homologue in murine neurons, Kdm5d has been shown to be expressed at lower levels than Kdm5c and is not able to compensate for Kdm5c differences between females and males [21].

Recent studies suggest there is interplay between histone modifications and DNA methylation [22, 23]. This relationship is bidirectional; histone modifications are more labile while DNA methylation is more stable [22]. In embryonic development, the formation of histone marks precedes and guides de novo DNA methylation, either by recruiting de novo DNA methyltransferase enzymes (H3K9 methylation) [24], or by protecting DNA from de novo methylation (H3K4 methylation) [25, 26].

We hypothesized that in patients with KDM5C mutations an aberrant increase of H3K4 tri- and di-methylation leads to decreased DNA methylation at genomic sites critical for normal neurodevelopment. We also proposed that the sites exhibiting decreased DNA methylation due to KDM5C mutation would also exhibit sexually dimorphic patterns of DNA methylation correlating with KDM5C and KDM5D dosage in normal females and males. Since KDM5C escapes X-inactivation [18], we enriched our sample set by including 47,XXX, 47,XXY and 45,X (Turner syndrome) individuals. In agreement with our hypothesis, we identified a significant loss of DNA methylation at specific genomic loci in blood samples of male patients carrying KDM5C mutations, suggesting these genes are epigenetic targets of KDM5C. To our knowledge, this is the first report of significant DNA methylation alterations in association with a mutation in a human histone modifying enzyme. Furthermore, we have shown that some genes with downstream loss of DNA methylation in individuals with KDM5C mutation also demonstrate positive correlation for DNA methylation levels across individuals with varying KDM5C/KDM5D dosage due to sex chromosome number in both blood and brain.

Methods

Research subjects

This study was approved by research ethics boards of the Greenwood Genetic Center (Greenwood, SC, USA) and the Hospital for Sick Children (Toronto, ON, Canada). All research subjects and/or their caregivers provided informed consent. Blood samples were collected from 10 patients with XLID and confirmed KDM5C mutations [5], two patients with XLID and KDM5C sequence variant, 19 unaffected control males (three unaffected relatives/16 unrelated individuals) (Additional file 1: Table S1), 13 control females, 11 females with Turner syndrome (45,X karyotype), three males with 47,XXY karyotype and three females with 47,XXX karyotype. DNA was extracted from 5 ml of blood using phenol-chloroform and ethanol precipitation.

Methylation microarray

The HumanMethylation27 BeadChip (Illumina, San Diego, CA) containing 27,578 individual CpG sites covering >14,000 genes was used for genome-wide DNA methylation analysis. Genomic DNA was sodium bisulfite converted with the EpiTect Bisulfite Kit according to the manufacturer’s protocol (Qiagen, Germantown, MD). Labeling, hybridization and scanning were performed at the Centre for Applied Genomics (TCAG) at The Hospital for Sick Children, Toronto, Canada. The methylation status of the interrogated CpG sites was measured from the intensity values of the methylated (M) and unmethylated (U) probes, as the ratio of fluorescent signals β = Max(M,0)/Max(M,0) + Max(U,0) + 100]. DNA methylation β values are continuous variables between 0 (absent methylation) and 1 (completely methylated) representing the ratio of combined locus intensity. The β values were extracted using the Methylation Module in Illumina Bead Studio after background normalization. All samples included in the analysis passed quality control metrics including bisulfite conversion control intensity values in green channel >4000 and 99% of probes had p-values of detection of signal above background <0.01 [27, 28]. 6 publically available datasets (accession numbers: GSE19711, GSE36064, GSE27097, GSE25395, GSE20067, GSE20236) were used to assess DNA methylation at FBXL5 (cg02630888), SCMH1 (cg03387723), and CACYBP (cg16743289) CpG sites. Sex specific DNA methylation analysis in brain was performed using published dataset (GEO Accession No: GSE15745) [29]. If the QC information was available, only samples with appropriate QC metrics (BS control intensity values in green channel >4000 and ≥95% of probes with p-values of detection of signal above background <0.05) were included in the analysis.

Statistical analysis

Microarray probes with detection p-value ≥ 0.01 [28], 2,984 cross-reactive probes, and 907 probes overlapping SNPs in queried CpG [30] were excluded, leaving in total 23,837 sites for downstream statistical analysis. A non-parametric Mann–Whitney U test was used for group comparisons. To adjust for multiple comparisons, a permutation-based method controlling the false discovery proportion (FDP) γ in the data, for a pre-specified confidence levels α was applied [31]. 1000 random permutations of the mutation labels among the data cases, while maintaining the sample sizes (10 vs. 19) were generated. For each permutation Mann–Whitney U test-based p-values for all CpG sites were ranked from smallest to largest (from most significant to least significant). The distribution of the 1000 “best” permutation-based p-values was used to determine p-values cut offs for different confidence levels based on the percentiles of this distribution. The cut off p-values were used to determine the number of significant CpG sites at different γ and α level. The Principal Component Analysis was performed in R statistical package using median centering of the data as well as with missing-value imputation to the 10 nearest neighbors.

Bisulfite pyrosequencing

Targeted DNA methylation analysis was performed using pyrosequencing as described by Tost and Gut [32]. Pyrosequencing assays containing two PCR primers and one sequencing primer were designed to target CpG sites of interest using PyroMark Assay Design Software (Qiagen). One of the PCR primers had a universal tag which annealed to the universal biotinylated primer. Genomic DNA was sodium bisulfite converted the same way as for the Illumina microarray and amplified using Hot-Start Taq-polymerase (Qiagen). Amplicons were analyzed on a Q24 pyrosequencer (Qiagen) as specified by the manufacturer; % of methylation was quantified as the ratio of C to C + T using PyroMark Q24 Software (Qiagen). Pyrosequencing primers are shown in Additional file 1: Table S2.

Quantitative real-time RT-PCR

To analyze tissue specific expression of FBXL5, SCMH1 and CACYBP we performed real-time RT-PCR expression analysis. Total RNA was purchased from Clontech for 8 tissues and 4 brain regions. Lymphoblastoild cell line established from control sample C2 RNA was extracted using RNeasy kit (Qiagen). The cDNA for quantitative real-time PCR was synthesized using VersoTM cDNA kit (Thermo Fisher Scientific, Epsom, UK). Mx3005P QPCR System (Agilent, Santa Clara, CA) with SsoFast™ EvaGreen® Supermix (Bio-Rad, Hercules, CA). Expression of each gene was determined using the comparative Ct method and normalized to the expression of the GAPDH housekeeping gene. Primers sequences are shown in Additional file 1: Table S2.

Results

Methylation microarray profiling in patients with KDM5Cmutations

To determine whether altered genome-wide DNA methylation patterns are associated with KDM5C mutations, we performed genome-wide DNA methylation analyses using the Illumina Infinium HumanMethylation27 BeadChip array containing 27,578 CpG sites and covering >14,000 genes in DNA from blood samples of XLID patients and controls. Our study group was comprised of 10 XLID patients from 5 families with 5 different KDM5C mutations. Two of these mutations were frameshift mutations, resulting in premature stop codons and 3 were missense mutations, predicted to be damaging by both Polyphen and Sift algorithms [4, 5]. The number of affected individuals per family ranged from one to three. The control group was comprised of 16 age and ethnicity matched normal unrelated males and three unaffected male relatives (not carrying the mutation) from the family with the p.V504M mutation. We also tested two brothers with ID who carried a p.R1546Q sequence variant of unknown significance (VUS), predicted to be benign/tolerated by Polyphen and Sift respectively. During the course of this study, this KDM5C variant was reclassified as a benign variant, as it was found in a phenotypically normal maternal grandfather in another XLID family. Clinical and demographic information for the XLID patients and controls and locations of the KDM5C mutations are summarized in Additional file 1: Table S1 and Figure 1.

Methylation ratios (beta-values) for each of 27,578 CpG sites were calculated as ratios of methylated CpG to the sum of methylated and unmethylated CpG. The reliability of Illumina Infinium data was demonstrated by high correlation (R2 = 0.95-0.99) among samples as well as high correlation of microarray data with bisulfite pyrosequencing (R2 = 0.77-0.98) (Additional file 2: Figure S1).

Cross-reactive probes and probes overlapping SNPs in the queried CpGs [30] were excluded, leaving 23,837 sites for statistical analysis. As a first step in our exploratory analysis, we used principal component analysis (PCA) and unsupervised hierarchical clustering to create a high-level summary of the data for all 23,837 CpG sites. Examination of these results did not reveal a clear separation between the mutation cases, VUS cases, unaffected relatives, and controls (Additional file 2: Figures S2 - S3). This suggests that if there are DNA methylation differences associated with KDM5C mutations they occur at specific loci but not across all CpG sites analyzed.

To identify the differentially methylated CpG sites, we compared beta-values between 10 KDM5C mutation cases and 19 controls using the non-parametric Mann–Whitney U test. After generating an initial p-value for each of the 23,837 CpG sites, we observed that as many as 6,124 sites in a given sample were below the 0.05 significance level. To adjust for multiple testing we used a permutation-based method that estimated false discovery proportion (FDP) levels for pre-specified confidence levels [31]. We generated 1000 random permutations of the mutation labels among the data cases, while maintaining the sample sizes (10 vs. 19), and have identified significant CpG sites at four FDP levels for three confidence levels (Additional file 1: Table S3). The top candidates (53 CpG sites from 51 genes) with the lowest FDP = 0 and the highest confidence level of 99.5% (unadjusted p-values < 2.00e-07) are shown in Additional file 1: Table S4.

To address the potential effect of relatedness on differences in DNA methylation and controls we performed additional analysis including only unrelated individuals by randomly selecting 5 mutation cases from each of the five available families and 17 unrelated controls (one randomly chosen unaffected relative and 16 unrelated controls). To identify differentially methylated CpG sites, we performed the Mann–Whitney U test on all 72 possible combinations of 5 cases vs. 17 controls. We observed a large degree of consistency between the CpG sites identified in this analysis (5 unrelated cases vs. 17 unrelated controls) with the CpG sites identified in the general setting (10 cases vs. 19 controls). For example, the CpG sites found to be significantly different between cases and controls across all 72 combinations at 95% confidence level exactly coincided with the 53 CpG sites shown in Additional file 1: Table S4 (FDP =0 at 99.5% confidence in the complete dataset of 10 cases vs. 19 controls).

We further tested if DNA methylation differences at 53 CpG sites with p-values < 2.00e-07 could differentiate cases either from controls or from the variant of unknown significance (p.R1546Q). A variety of unsupervised methods including hierarchical clustering (Figure 2), principal component analysis (Additional file 2: Figure S4), and K-means and K-median clustering (not shown) were capable of unambiguously differentiating KDM5C mutation cases from controls. Two samples, carrying the p.R1546Q variant at the C-terminal end of the KDM5C protein, consistently showed the same DNA methylation levels as control samples confirming the benign nature of this variant.

Figure 2
figure 2

Hierarchical unsupervised clustering of 31 study samples. DNA methylation levels at the 53 most significant CpG sites were used for hierarchical unsupervised clustering. The results show a clear distinction between the 10 KDM5C mutation cases (top part), and the remaining 21 samples (bottom part), which is also indicated by the large distance between the two branches of the dendrogram on the left. DNA methylation levels corresponding to mutations are generally lower than for the controls. Interestingly, the two cases of variant of unknown clinical significance (p.R1546Q) have methylation profiles very similar to the 19 controls. Also, mutation cases from the same family do not always cluster together. The clustering was based on complete linkage and cosine distance metric.

Full size image

Multiple studies have shown that at some genomic loci DNA methylation levels could be affected by sequence polymorphisms in cis and subsequently be heritable [29, 33–36]. We determined whether DNA methylation of the top 53 CpG sites were dependent not only on KDM5C mutation but also on single nucleotide polymorphisms (SNPs). For these analyses, we took advantage of two published datasets that used the same Illumina HumanMethylation27 microarray platform and reported a number of methylation quantitative trait loci (mQTL) in lymphoblastoid cell lines (LCLs) [33, 34]. At these mQTLs DNA methylation depends on genotypes of single nucleotide polymorphisms (SNPs) located in cis or trans in relation to CpG. We did not find any mQTLs at the 53 CpG sites identified in our study, suggesting that the DNA methylation alterations we observed are more likely to be associated with KDM5C mutations rather than other genetic variation that exists between cases and controls.

In agreement with the prediction of cross-talk between H3K4 methylation, and DNA methylation, we observed an over-representation of CpG sites with loss of DNA methylation versus gain at all levels of FDP and at all confidence intervals tested (Additional file 1: Table S3, p < 2.2e-16, Fisher Exact test). All 53 top CpG candidates with the lowest FDP = 0 and the highest confidence level of 99.5% (unadjusted p-values < 2.00e-07) exhibited loss of DNA methylation. The 13 CpGs with the greatest loss of DNA methylation (delta beta ≤-0.2) are shown in Table 1. In addition, in the volcano plot where delta beta is plotted against p-values, we observed an enrichment of CpG sites with negative delta beta, reflecting the loss of DNA methylation. Interestingly, this asymmetry was limited to CpG islands (Figure 3A). Also, averaging all microarray probes and comparing cases and controls we observed a small (<1%) loss of DNA methylation in the KDM5C mutation group (p = 0.035) also for CpG islands, but not for non-CpG island probes (Figure 3B). This genome-wide loss of DNA methylation was observed only in unique sequences, but not at LINE-1, the most common non-LTR retrotransposon, comprising about 17% of the human genome [37] as determined by pyrosequencing (Additional file 2: Figure S5).

Table 1 Top 13 CpG sites with loss of DNA methylation in KDM5C mutation cases with the lowest false discovery proportion level =0, the highest confidence level of 99.5% and delta beta ≤ −0.2
Full size table
Figure 3
figure 3

Loss of DNA methylation at CpG islands associated with KDM5C mutations. A) Volcano plot of DNA methylation differences between controls and KDM5C mutation cases (delta beta, X-axis) versus statistical significance –log10(p) in CpG islands (left) and non-CpG islands (right). Negative delta beta represents loss and positive delta beta gain of DNA methylation in KDM5C mutation cases. Black lines represent p-values at three levels of confidence 95% (6.69E-06), 99% (1.07E-06) and 99.5% (2.00e-07). B) Boxplot of the mean methylation of all microarray probes within CpG islands (left) and non-CpG islands (right) in controls (C) and in KDM5C mutations cases (K).

Full size image

Illumina HumanMethylation27 coverage on average extends to two CpG sites per gene. Of the 53 most significant CpG sites based on permutation analysis, only two genes C2orf3 (GCFC2) and TSPYL5 had two CpG sites meeting the permutation p-value cut off. In order to assess the genomic extent of the DNA methylation changes for each significant gene we have evaluated DNA methylation levels, delta beta differences, and p-values in cases and controls for additional array probes for the 53 top significant genes. For a majority of the genes the second probe did not exhibit significant loss of DNA methylation. Only 8 genes had a second significant probe with p ≤ 0.05. The absolute delta beta differences for the second probes were relatively small ≤ 0.05, with the exception of the STMN1 gene having a delta beta −0.16 and −0.14 in two CpG sites (Additional file 1: Table S5). The CpG sites with significant loss of DNA methylation tended to be located just on the edge of CpG island several hundred base pairs upstream of TSS. In contrast, CpG sites without DNA methylation differences were predominantly unmethylated in both cases and controls and located a few 100 bp downstream of TSS (Additional file 1: Table S5, Figure 4 and Additional file 2: Figures S6 - S9).

Figure 4
figure 4

Regional DNA methylation in FBXL5 promoter. A) Screenshot from the UCSC genome browser showing location of FBXL5 promoter exon1, intron1, CpG island, Illumina microarray probes and pyrosequencing assays and tracks for H3K4me1 and H3K4me3 in lymphoblastoid cell line (GM12878) and ES cell line (H1 h-ES) (Broad Institute Histone data). B&C) Boxplots of Illumina DNA methylation data for two microrray probes within FBXL5 gene. The Y axis shows DNA methylation levels presented as C/C + T and ranging from 0 to 1. The bottom and the top of the box are 25th and 75th percentiles respectively; the whiskers are within the 1.5 interquartile range (IQR) of the data, and the circles, are outlier data points above or below 1.5 IQR. C are controls (N = 19), K are cases with KDM5C mutations (N = 10), V are cases with p.R1546Q sequence variant. D) DNA methylation upstream of FBXL5 transcription start site as determined by pyrosequencing assays 1&2 (pyro1&2). The order of CpG sites are shown from the furthest upstream to the transcription start site. Each column is an average for each of three groups of 1)10 cases with KDM5C mutations (mutation), 2)19 controls (control), and 3) two individuals with the p.R1546Q variant of unknown significance (VUS). The arrow shows the CpG sites overlapping the microarray probe. P-values were determined by Kruskal-Wallis test between mutation cases and controls. **** is p <0.0001. E) DNA methylation upstream of FBXL5 transcription start site and downstream of pyrosequencing assays 1&2 as determined by pyrosequencing assay 3 (pyro 3). The order of CpG sites are shown from the furthest upstream to the transcription start site. Each column is an average for each of two groups of 1) 5 cases with KDM5C mutations (mutation), 2) 5 controls (control). Error bars are standard deviations.

Full size image

Targeted validation by pyrosequencing

We have previously shown that validation of Illumina HumanMethylation27 CpG methylation by bisulfite pyrosequencing demonstrates high correlation between the two methods and methylation determined by Illumina microarray is a good predictor of regional CpG methylation in CpG islands [38]. To validate the DNA methylation differences found on the array we tested five loci: top three candidates FBXL5 (delta beta = −0.48), SCMH1 (delta beta = −0.45) and CACYBP (delta beta = −0.28) with the largest DNA methylation differences and two loci with smaller differences DYDC1 (delta beta = −0.1) ZMYND12 (delta beta = −0.08). Using pyrosequencing, we validated the direction of the differences for all 5 loci and observed high correlation DNA methylation levels between pyrosequencing and Illumina for overlapping CpG sites (R2 ranging from 0.77 to 0.98, Additional file 2: Figure S1). For FBXL5, SCMH1, CACYBP and ZMYND2 the assays contained >1 CpG site, and pyrosequencing data showed that DNA methylation changes affected not only the index CpG site from the array but also multiple adjacent CpGs (Figures 4 and Additional file 2: Figures S6-S9). The samples carrying the p.R1546Q variant consistently exhibited the same DNA methylation patterns as controls. For FBXL5 (8 CpGs) and ZMYND2 (5 CpGs) all sites tested exhibited DNA methylation differences between cases and controls (Figures 4, Additional file 2: Figure S8). For SCMH1 (5 CpGs), the CpG sites more distant from the TSS exhibited overall higher DNA methylation in controls and larger differences in DNA methylation between cases and controls (Additional file 2: Figure S6). For CACYBP, it was not possible to design a pyrosequencing assay overlapping the Illumina CpG site. Therefore, we designed an assay covering 2 CpGs ~100 bp upstream of the Illumina site. In this assay one CpG site exhibited significant DNA methylation differences between cases and controls consistent with the difference found on the array (Additional file 2: Figure S7).

For FBXL5 we designed an additional assay to test DNA methylation within a CpG island in closer proximity to the TSS. We found very low overall DNA methylation levels in both cases and controls (<10%) with no significant differences. Interestingly, based on ChIP sequencing data for histone marks from the Broad Institute [39], sites that were hypermethylated in controls and exhibited significant loss of DNA methylation in KDM5C mutations cases frequently coincided with enhancer mark H3K4me1 in embryonic stem cell (ES) and lymphoblastoid cell lines. In contrast, the mark of active promoters H3K4me3 was shifted more towards the TSS, where DNA was hypomethylated (Figure 4, Additional file 2: Figures S6-S9). Thus, it is possible that sequences exhibiting loss of DNA methylation in patients with KDM5C mutations are involved in the regulation of downstream genes through enhancer activity.

Ubiquitous expression of FBXL5, CACYBP and SCMH1in human tissues

We focused the next phases of our downstream analysis on the three top candidate genes FBXL5, SCMH1 and CACYBP with the highest delta Z scores (Table 1). Interestingly, the three top CpG sites identified by methylation array to be significantly hypomethylated in KDM5C mutations cases are within promoters of genes involved in ubiquitin-mediated protein degradation [40–43]. FBXL5 is an iron sensing E3 ubiquitin ligase that regulates iron homeostasis [44]. CACYBP is part of a ubiquitin ligase complex regulating beta-catenin, which is important for cell-cell adhesion and transcription regulation through Wnt-signaling [45]. SCMH1 is part of a polycomb group complex 1 (PcG1) involved in transcriptional silencing [46] and proteosomal degradation for the Geminin protein, important for regulation of replication and maintenance of undifferentiated states [42]. Little is known however about tissue specific expression of these genes in human. We tested the expression of FBXL5, SCMH1 and CACYBP in several somatic tissues including brain, kidney, heart, muscle and lymphoblastoid cells, as well as several brain regions. We observed ubiquitous expression across tissues and brain regions (0.6-12% of GAPDH expression level) (Additional file 2: Figure S10), consistent with the function of these genes in pathways important for multi-systemic physiological processes.

Loss of DNA methylation associated with KDM5Cmutations is not due to altered blood cell counts

There are no reported blood cell-related phenotypes associated with KDM5C mutations [7]. However, white blood cells consist of functionally distinct cell populations in varying proportions and it has been shown that at some loci in different white blood cells DNA methylation patterns can vary substantially [47]. As we did not have blood cell counts for either cases or controls in our study, the possibility that observed DNA methylation differences could be due to differences in the proportion in different cell types cannot be completely ruled out. To test this, we checked DNA methylation levels for the three top candidates FBXL5, SCMH1 and CACYBP using GEO dataset (GSE35069). This dataset analyzed genome-wide DNA methylation using Illumina Methylation450 array in DNA from whole blood, peripheral blood mononuclear cells, and granulocytes, as well as in 7 isolated cell populations (CD4+ T cells, CD8+ T cells, CD56+ NK cells, CD19+ B cells, CD14+ monocytes, neutrophils, and eosinophils) in six healthy males [47]. The beta values for probes overlapping Illumina27 with significant loss of DNA methylation in KDM5C mutations cases were extracted for analysis. We did not observe differences between cell types at three CpG sites analyzed at a magnitude that could explain loss of DNA methylation in KDM5C mutations cases (Additional file 2: Figure S11). For FBXL5, DNA methylation was very uniform across all cell types, with a maximum difference of 2%, for SCMH1 and CACYBP the biggest differences were 22% between CD19+ B cells and CD4+ T cells and 13% for CD19+ B cells and neutrophils, respectively. This comparison strongly suggests that the observed loss of DNA methylation found in individuals with KDM5C mutations are not due to differences in the proportion of blood cell types.

DNA methylation patterns at FBXL5, SCMH1 and CACYBPpromoters in general population

In our discovery dataset of 10 KDM5C mutations cases and 19 controls, we observed no overlap in DNA methylation levels between cases and controls for three genes FBXL5, SCMH1 and CACYBP. We wanted to test the frequency of loss of DNA methylation in these three genes in the general population. For these analyses we used 6 Illumina Infinium HumanMethylation datasets of white blood cells from GEO NCBI database, comprising a total 946 control samples passing QC (Additional file 1: Table S6). Three datasets (GSE36064, GSE27097, GSE20236) [48, 49] investigated DNA methylation association with age and included only control samples, while the other studies investigated DNA methylation association with disease including diabetes/nephropathy (GSE20067) [27], ovarian cancer (GSE19711) [27] and trisomy 21 (GSE25395) [50]. Due to common nature of diabetes we have included samples with this disease into our control dataset, but have excluded ovarian cancer samples because cancer as well as cancer therapies are known to alter epigenetic marks [51] and somatic KDM5C mutation was previously identified in cancer [52, 53]; trisomy 21 cases were also excluded due to overlapping ID phenotype with KDM5C mutations cases [50]. None of the 946 control samples exhibited loss of DNA methylation comparable with KDM5C mutations cases (Figure 5).

Figure 5
figure 5

DNA methylation at the three top significant CpGs in KDM5C mutations cases and population controls. DNA methylation microarray data at three CpG sites within CpG-rich promoters of three genes FBXL5 (A), SCMH1 (B) and CACYBP (C) as determined in 6 published studies using Illumina methylation27 array. AF (Aging in females, n = 93), AP1 (aging pediatric 1, n = 398), AP2 (aging pediatric 2, n = 79), CO (cancer ovarian, n = 257), DB (diabetes, n = 99), DS (Down syndrome, n = 21), K-C are controls from our study (N = 16), K-M are KDM5C mutations cases. For CO and DS only control samples were included.

Full size image

We have also analyzed the association of DNA methylation levels at these sites with age, ethnicity, and sex for the 6 datasets. These data were analyzed separately within each study to avoid possible batch effects. We did not observe any significant association of DNA methylation at these three loci with ethnicity or age. There was a small (median differences = 0.01-0.04) but highly significant increase of DNA methylation in females compared to males in FBXL5 and CACYBP (p < 1.00E-04) and a trend towards significance for SCMH1 (p = 0.07), in the diabetes study where samples of both sexes were included [27] (Figure 6). We also assessed the consistency of the observed sex-specific differences in 9 additional autosomal CpG sites (top candidates with the largest DNA methylation loss, delta beta ≤ −0.2 from Table 1). Similar to differences described for FBXL5, CACYBP and SCMH1, we observed a significant (q-value ≤ 0.05) increase of DNA methylation in females compared to males in 5 loci and a trend towards significance (q-value ≤ 0.1) in three out of 9 tested loci with the delta beta differences ranging from 0.01 to 0.06 (Additional file 1: Table S7).

Figure 6
figure 6

Sex-specific DNA methylation differences in FBXL5, CACYBP and SCMH1 promoters in blood and four brain regions. Boxplots show DNA methylation levels in three CpGs located in the promoters of FBXL5, CACYBP and SCMH1 in 99 (48 males/51 females) blood samples (GSE20067), 133 (90 males/43 females) frontal cortex samples (FCTX), 127 (85 males/42 females) temporal cortex samples (TCTX), 121 (86 males/35 females) cerebellum samples (CBL) and 125 (87 males/38 females) pons samples from neurologically normal individuals (GSE15745). The Y axis shows DNA methylation levels presented as C/C + T and ranging from 0 to 1. The bottom and the top of the box are 25th and 75th percentiles respectively, the whiskers are within the 1.5 interquartile range (IQR) of the data, and the circles, are outlier data points above or below 1.5 IQR. P-values were calculated using Kruskal-Wallis test. **** is p <0.0001, *** is p < 0.001, ** is p < 0.01, * is p < 0.05, NS is p > 0.1.

Full size image

DNA methylation comparison at FBXL5, SCMH1 and CACYBPbetween brain and blood

Since brain tissue from individuals with KDM5C mutations is not available for study, we took an alternative approach to assess whether the genomic targets we identified might be demonstrated to be functionally important in brain. In this regard, we investigated whether DNA methylation levels in brain are similar to those in blood and whether sex-specific DNA methylation differences we found in blood are also observed in the brain. For this analysis we used a published Illumina HumanMethylation27 dataset for four brain regions (temporal cortex, frontal cortex, cerebellum and pons) of neurologically normal individuals (GEO Accession No: GSE15745) [29]. DNA methylation at three tested CpG sites exhibited overall hypermethylation (methylation level > 50%) in brain and blood for FBXL5 and SCMH1 with the exception of cerebellum which had intermediate methylation levels in brain (30-40%) and high methylation levels in blood (70%). CACYBP had intermediate levels of DNA methylation in both brain and blood (30-50%) (Figure 6). Furthermore, we observed a small but statistically significant increase of DNA methylation in females compared to males in three brain regions at FBXL5 and in four brain regions at CACYBP, which was similar to sex-specific difference found in blood (Figure 6, Additional file 1: Table S8). In conclusion, these data suggest that DNA methylation at FBXL5 and CACYBP can be regulated by similar mechanisms in both blood and brain and the observed sex- specific differences could be the result of differences in KDM5C/KDM5D dosage between males and females.

DNA methylation levels at FBXL5, SCMH1 and CACYBP and KDM5C/KDM5Ddosage in blood

KDM5C is an X-linked gene that escapes X-inactivation in humans and mouse [18, 19, 54], and has a functional Y-linked homologue KDM5D[55]. Interestingly, in mouse the degree of Kdm5c’s escape from X –inactivation is highly variable across different tissues. The level of transcript from the inactive X allele is 20-70% of the active X allele [56]. It is not known if the same variability is present in humans. In mouse brain, Kdm5c/Kdm5d are expressed in a sex-specific fashion i.e., the expression of Kdm5c is significantly higher in female brains than in male brains, and the expression of Kdm5d in males is not sufficient to compensate for the female bias in Kdm5c expression [21]. Furthermore, in human tissues, KDM5D is reported to be expressed at lower levels than KDM5C. However, as commercially available RNA mixed from several individuals was used for this experiment the proportion of male cells present in these samples is not known [3]. Our observation of increased DNA methylation in females compared to males in the top three affected loci (Figure 6), led us to hypothesize that DNA methylation at these loci might depend on sex chromosome dosage and specifically on the dosage of the X and Y linked homologues, KDM5C and KDM5D, respectively. To further investigate this, we assessed FBXL5, SCMH1 and CACYBP DNA methylation levels using targeted pyrosequencing assays in blood samples with different sex chromosome constitutions and reflecting variation in KDM5C/KDM5D dosage, including 47,XXX (KDM5C/KDM5C/KDM5C, N = 3), 47,XXY (KDM5C/KDM5C/KDM5D, N = 3), 46,XX (KDM5C/KDM5C, N = 16), 46,XY (KDM5C/KDM5D, N = 19) and 45,X (KDM5C/0, N = 11) in comparison to males with KDM5C mutations (0/KDM5D, N = 10) and female carriers of KDM5C mutation (KDM5C/0, N = 4). We found that the DNA methylation levels at these three genes generally correlated with KDM5C/KDM5D dosage for all three genes analyzed (Figure 7). 47,XXX females exhibited the highest DNA methylation closely followed by 47, XXY males, 46,XX females and 46,XY males for the majority of analyzed CpG sites. There were less differences for SCMH1 between 46,XX females and 46,XY males suggesting that for this gene other factors might be involved in equalizing DNA methylation between the two sexes.

Figure 7
figure 7

DNA methylation at FBXL5, SCMH1 and CACYBP promoters correlates with KDM5C/KDM5D dosage. Boxplots show DNA methylation levels for CpG sites within the FBXL5, SCMH1 and CACYBP promoters in blood in 7 group of samples with different dosage of KDM5C/KDM5D. CpG numbering corresponds to Figures 4D, 6D and 7D. CACYBP CpG#2 boxplot is not shown, as no correlation with KDM5C/KDM5D dosage was found for this site. The Y axis is % of DNA methylation. The X axis shows groups of samples, numbered from 1 to 7. Information for each group regarding sex chromosome constitution, X-chromosome inactivation (XCI, Xa is active and Xi is inactive X-chromosome respectively), presence/absence of KDM5C mutation, functional KDM5C/KDM5D dosage, and number of samples is shown in the table beside the graph.

Full size image

DNA methylation levels were similar for 45,X females and female carriers of KDM5C mutations but significantly lower than in the four groups described above. This suggests that a single functional copy of KDM5C, without additional activity from KDM5D, is not sufficient to achieve levels of DNA methylation present in females and males with normal karyotypes. Males with KDM5C mutations exhibited the lowest DNA methylation of the 7 analyzed groups. They had significantly lower DNA methylation than females with only one functional copy of KDM5C (45,X and KDM5C mutation female carriers) (Figure 7), suggesting that KDM5D alone is not sufficient to compensate for absence of functional the KDM5C.

The differences among groups with normal or extra copies of KDM5C/KDM5D (the 47,XXX, 47, XXY, 46, XX and 46, XY) were substantially smaller than the differences for cases missing one copy of functional KDM5C (45, X, females and males with KDM5C mutations). This could be due to the fact that two copies of KDM5C or one copy each of KDM5C and KDM5D is close to saturation of H3K4 demethylase activity at target promoters. Another possible explanation of small differences between 46, XX females and 46, XY males observed in both brain and blood, is that KDM5D is expressed only at slightly lower levels than KDM5C from inactive X-chromosome in human, making 46,XX and 46,XY relatively close in their levels of H3K4 demethylase activity, in contrast to larger differences between males with KDM5C mutation vs. 45,X females, where the difference is between KDM5D and KDM5C expressed from an active X. The observation that KDM5C mutation female carriers exhibit DNA methylation levels similar to individuals with 45,X karyotype and the fact these females have highly skewed X-chromosome inactivation [5], suggests that in carriers the wild type KDM5C is expressed from the preferentially active X-chromosome.

Based on these data we suggest that DNA methylation levels at FBXL5, SCMH1 and CACYBP promoters correlate with H3K4 demethylase activity of the proteins KDM5C/KDM5D due to an inverse relationship between H3K4 methylation and DNA methylation [22].

Discussion

Advances in molecular technologies have helped to identify genetic causes in many cases of syndromic and non-syndromic forms of intellectual disability (ID) [57–59]. However, the molecular pathogenesis of ID still remains incompletely understood. It has been suggested based on known genetic etiologies that perturbed neuronal homeostasis altering synaptic outputs could be a key component of the cognitive impairment phenotype [60]. This notion is strongly supported by the types of functions attributable to genes mutated in ID which include basic cellular functions, such as transcription, translation, RNA biogenesis, protein turnover, and cytoskeletal dynamics [60]. One important emerging mechanism in ID is epigenetic dysregulation that ultimately affects transcription of multiple genes [1, 57]. KDM5C is one of more than 20 epigenetic regulators involved in ID [1]. The identification of specific downstream targets exhibiting aberrant epigenetic marks in response to mutation of an epigenetic regulator will have an important impact on our understanding of the molecular pathogenesis of ID.

Previously, profiling of mRNA in lymphoblastoid cell lines of 12 males with KDM5C mutations compared to 5 controls identified 11 upregulated genes. These transcriptional changes were not very consistent among KDM5C mutations samples, and a combination of at least 6 genes was required to distinguish cases from controls [15]. None of these 11 genes exhibited DNA methylation changes in our dataset. Thus, it is likely that these expression differences are more tissue and developmental stage specific, than the DNA methylation patterns, potentially reflective of disrupted binding of KDM5C specifically in lymphoblastoid cell lines. DNA methylation patterns can be maintained by DNMT1 through replication and multiple cell divisions [61], and thus if they occur early in development they could be represented in multiple lineages including peripheral blood, but not be reflective of gene expression patterns in all differentiated lineages. This has, in fact, been observed in neurodevelopmental syndromes such as Immunodeficiency–centromeric instability–facial anomalies (ICF), Fragile-X, Angelman and Prader-Willi syndromes [62]. Further, the regions where we found loss of DNA methylation associated with KDM5C mutations, coincided with an enhancer mark H3K4me1 in ES and lymphoblastoid cell lines. In contrast, DNA sequences more proximal to the TSS were hypomethylated in both controls and cases and coincided with the active promoter mark H3K4me3 (Figure 4, Additional file 2: Figures S6-S9) [39]. These data suggest that regions affected by loss of DNA methylation have enhancer-driven rather than basal promoter function. As enhancers are involved in the control of spatial and temporal gene expression [63], the relationship between loss of DNA methylation at identified sites and expression at downstream genes is likely to be more complex than a simple inverse correlation.

The mechanism of the observed loss of DNA methylation associated with loss of function mutations in KDM5C is not completely clear; however, it is unlikely to be the direct consequence of loss of KDM5C function, as KDM5C is not known to possess DNA methyltransferase activity. Based on the current literature, the most plausible mechanism is that a deficiency in H3K4 demethylase activity leads to increased H3K4 methylation, which protects DNA from de novo DNA methylation at KDM5C downstream target loci. In mouse ES cell, Dnmt3L recruits de novo methyltransferases to DNA associated with unmethylated forms of H3K4, and contact between Dnmt3L and the nucleosome is inhibited by all forms of H3K4 methylation [26]. Biochemical assays have shown that the human de novo methyltransferase DNMT3A interacts with histone H3 unmethylated at K4, whereas di- and tri-methylation inhibit this interaction [64]. There is also evidence from a yeast model system lacking endogenous DNA methyltransferases and ectopically expressing mouse Dnmt3a and Dnmt3L that depletion of H3K4 methylation results in increased DNA methylation [65]. In humans, a correlation of increased tri-methylation in H3K4 with reduced DNA methylation at promoters has been shown in fibroblast cells from normal individuals [66].

Our data demonstrating DNA methylation alterations in individuals with mutations in the KDM5C gene further support the link between H3K4 methylation and DNA methylation. These data show, for the first time, a functional consequence of loss of function of H3K4 demethylase resulting in significant alterations of DNA methylation at specific gene targets at a genome-wide level. Furthermore, in agreement with cross-talk of H3K4 methylation and DNA methylation, there was significantly more loss than gain of DNA methylation resulting from mutations in KDM5C. In mammals several H3K4 demethylases have been described. Apart from four enzymes of the KDM5 family, KDM1A and B specifically act to demethylate di- and mono-methylated forms of H3K4, and KDM2B similarly to the KDM5 family specifically demethylates tri- and di-methylated forms of H3K4 [67, 68]. At this point it is not clear if the loss of DNA methylation associated with KDM5C mutations is specific to KDM5C loss of function or would also be observed in the context of loss of function of other H3K4 demethylases. In support of the concept that loss of DNA methylation could be a common association of loss of H3K4 demethylase activity, mouse oocytes deficient in the H3K4 demethylase Kdm1b demonstrated a global increase in H3K4 di-methylation and failed to generate normal DNA methylation marks at several imprinted loci [68]. However, as we did not observe loss of DNA methylation at imprinted genes in patients with KDM5C mutations, the genomic sites demonstrating loss of DNA methylation for each H3K4 demethylase are likely to be specific, possibly reflecting the binding sites of these proteins.

Further support for the inter-dependence of histone methylation and DNA methylation in humans comes from cancer research. Mutations in IDH1 and IDH2, frequently found in gliomas and acute myeloid leukemias (AML), are characterized by enzymatic gain of function and subsequent production of hydroxyglutarate, which inhibits several histone demethylases, including H3K9, H3K27, H3K36 and H3K4 [69]. The somatic mutations in IDH1/IDH2 are associated with genome-wide hypermethylation in AML compared either to normal bone marrow or to AML caused by mutations in other genes. However at this point it is not clear which specific histone marks contribute directly to this DNA hypermethylation phenotype [70]. Interestingly, two genes, PABPN1 and ZNF532, demonstrating loss of DNA methylation in our study (Additional file 1: Table S4) were found to be hypermethylated in AML with IDH1 mutation [70]. These data suggest that there could be some common mechanism regulating DNA methylation of these two genes in opposite directions in the context of loss of function of KDM5C and gain of function of IDH1.

The genes FBXL5, SCMH1 and CACYBP on which we focused in our downstream analysis have exhibited a surprisingly large degree of DNA methylation differences between cases and controls reminiscent of the DNA methylation alterations at imprinted loci in disorders affecting neurodevelopment, such as Prader-Willi and Angelman syndromes [71, 72]. Further, loss of DNA methylation at these sites was not observed in 946 population control blood samples from publically available datasets, suggesting that altered DNA methylation at these three genes could be used for establishing pathological authenticity of new missense mutations, as in silico predictions of effects on protein function are often inconclusive, and functional experiments are expensive and labor intensive. Thus, as KDM5C mutation cases are frequently indistinguishable from other genetic causes of ID based on clinical phenotype alone [15], DNA methylation analysis could complement KDM5C sequencing to provide more accurate molecular diagnosis leading to improved patient management.

Interestingly, the three top candidate genes are part of ubiquitin-ligase protein degradation pathways. Synaptic network remodeling, a vital part of central nervous system function, depends on ubiquitin-mediated protein degradation at the postsynaptic membrane [73]. Genes involved in ubiquitination pathways have already been implicated in a number of other neurodevelopmental disorders, such as Angelman syndrome (loss of function of maternal copy of UBE3A) [71], and autism (copy number variants in PARK2, RFWD2, FBX040) [74]. Furthermore, 7% of XLID genes identified to date are components of the ubiquitin pathway [75].

We have shown that these three genes are ubiquitously expressed in human tissues at relatively low levels compared to the house-keeping gene GAPDH. Although it is currently not clear how loss of DNA methylation at these sites affect gene expression, we propose that abnormal expression of these genes at specific cell types/developmental stages causes disturbances in downstream pathways such as degradation of target proteins. FBXL5 has been recently discovered to be a component of an E3 ubiquitin ligase complex that targets IRP2, an iron regulatory protein 2 important in intracellular and plasma iron homeostasis [43, 76]. IRP2 regulates RNA stability and translation by binding to iron responsive - RNA stem loop structures, in a number of genes involved in iron uptake, storage and utilization [44]. While FBXL5 loss of function leads to embryonic lethality in Fbxl5−/− mice, associated with aberrant iron accumulation and increased oxidative stress [77], Irp2 −/− knockout mice exhibit a neurological phenotype associated with locomotor abnormalities accompanied by iron accumulation in white and grey matter [78]. Thus, abnormal expression of FBXL5 could result in abnormal iron accumulation, thereby contributing to seizures and/or ID phenotypes observed in males with KDM5C mutations. SCMH1 is a member of the Polycomb-group 1 complex, which not only is a transcriptional repressor, but also acts as an E3 ubiquitin ligase for the geminin protein involved in DNA replication and maintenance of undifferentiated cellular states. Specifically, SCMH1 has been shown to provide an interaction domain for geminin [42]. Recently a genome-wide association study implicated SCMH1 in the regulation of human height [79], thus it is possible that loss of DNA methylation at SCMH1 is important for short stature associated with KDM5C mutations. CACYBP (Sip) is part of the SCF-like complex, involved in ubiquitin-mediated degradation of the transcriptional activator β-catenin [40, 41]. β-catenin is a signaling molecule playing an important role in neurodevelopmental processes such as neural crest development, development of cortical and hippocampal neuroepithelium, and dendrite spine morphogenesis [80–82]. In addition, it has been implicated in seizure susceptibility [83]. Further, CACYBP was shown to dephosphorylate ERK1/2 [84], extracellular signal-regulated kinases, important in many aspects of early brain development and implicated in 16p11.2 and 22q11 deletion syndromes phenotypes [85].

As the described DNA methylation changes in our study were identified in blood samples, an important question that cannot be directly addressed by our data is the issue of whether parallel changes occur in brain. As de novo DNA methylation is an important process in epigenetic reprogramming occurring at early stages of embryonic development [22, 86], we suggest that loss of DNA methylation in the blood of patients with KDM5C mutations could at least in part result from abnormally high H3K4 di/trimethylation in the embryo, protecting DNA from de novo methylation. We expect that this state is maintained through differentiation into multiple lineages. In support of this, we found parallel sex-specific DNA methylation differences in both brain and blood at FBXL5 and CACYBP, whereas SCMH1 exhibited this difference only in blood, but not in brain. We suspect that these observed sex-specific differences are due to KDM5C/KDM5D dosage rather than the effects of sex hormones, as the DNA methylation at tested targets correlates better with sex chromosome constitution than with gonadal sex, e.g. the highest DNA methylation was observed in 47,XXX females, followed by 47,XXY males, 46,XX females, 46,XY males, 45,X females. The lowest DNA methylation is seen in males with KDM5C mutations (Figure 7). In addition we observed that female carriers of KDM5C mutations have DNA methylation levels similar to 45,X females, reflecting the fact that they have only one functional copy of KDM5C.

Based on DNA methylation comparison between brain and blood, we propose that the epigenetic status of FBXL5 and CACYBP is regulated by KDM5C in both brain and blood, and that their deregulation in brain can contribute to the intellectual disability and seizure phenotypes in individuals with KDM5C mutation. We propose as well that KDM5C contributes to sex-specific differences in brain function. In contrast, SCMH1 might be responsible for other aspects of the clinical phenotype associated with KDM5C mutation such as growth abnormalities. Furthermore, the KDM5C-mutation associated targets identified here could play a role in Turner syndrome. It has been previously suggested that X-linked genes escaping X-inactivation such as KDM5C are likely to be implicated in neurocognitive phenotypes of 45,X females with Turner syndrome, who in spite of normal cognitive abilities, frequently have problems in spatial reasoning and emotion recognition [87, 88]. Our observation of loss of DNA methylation at the FBXL5, SCMH1 and CACYBP promoters in 45, X females compared to XX females and XY males, but to a lesser degree than in males with KDM5C mutations, supports this hypothesis and suggests that deregulation of epigenetic targets of KDM5C could be relevant to the mild neurodevelopmental impairments found in females with Turner syndrome. Similarly, loss of DNA methylation at these three genes found in female carriers of KDM5C mutations could contribute to learning difficulties frequently observed in such individuals [5, 7].

In summary these data provide new opportunities to address the molecular basis, both genetic and epigenetic, of ID. An important area for future investigation would be to establish both spatial (tissue-specific) and temporal (developmental stage- specific) maps of KDM5C targets, and to annotate how loss of KDM5C function impacts expression of these targets through embryonic development and in diverse tissues. Validation of the affected molecular pathways, described here such as abnormal iron homeostasis or β-catenin dysregulation could also, in an animal model of KDM5C mutations, provide a framework for potential therapeutic developments for patients with KDM5C mutations.

Conclusions

We have, for the first time, identified significant multilocus loss of DNA methylation in individuals with loss of function mutations of a gene encoding histone modifying enzyme, specifically a histone H3K4 demethylase KDM5C. We have validated changes in three loci with the most prominent changes: FBXL5, SCMH1 and CACYBP. We have also demonstrated that loss of DNA methylation at these three genes is specifically associated with KDM5C mutations and is not observed in >900 control blood samples. In addition we have shown that DNA methylation at these three genes correlates with dosage of KDM5C and its Y-linked homologue KDM5D in blood of individuals with different sex chromosome complements. Finally we observed parallel sex-specific differences in several brain regions for FBXL5 and CACYBP, suggesting that these genes could play an important role in the ID phenotype of individuals with KDM5C mutations.