Epigenomic and phenotypic characterization of DEGCAGS syndrome

Abstract

Developmental Delay with Gastrointestinal, Cardiovascular, Genitourinary, and Skeletal Abnormalities syndrome (DEGCAGS, MIM #619488) is caused by biallelic, loss-of-function (LoF) ZNF699 variants, and is characterized by variable neurodevelopmental disability, discordant organ anomalies among full siblings and infant mortality. ZNF699 encodes a KRAB zinc finger protein of unknown function. We aimed to investigate the genotype-phenotype spectrum of DEGCAGS and the possibility of a diagnostic DNA methylation episignature, to facilitate the diagnosis of a highly variable condition lacking pathognomonic clinical findings. We collected data on 30 affected individuals (12 new). GestaltMatcher analyzed fifty-three facial photographs from five individuals. In nine individuals, methylation profiling of blood-DNA was performed, and a classification model was constructed to differentiate DEGCAGS from controls. We expand the ZNF699-related molecular spectrum and show that biallelic, LoF, ZNF699 variants cause unique clinical findings with age-related presentation and a similar facial gestalt. We also identified a robust episignature for DEGCAGS syndrome. DEGCAGS syndrome is a clinically variable recessive syndrome even among siblings with a distinct methylation episignature which can be used as a screening, diagnostic and classification tool for ZNF699 variants. Analysis of differentially methylated regions suggested an effect on genes potentially implicated in the syndrome’s pathogenesis.

Introduction

Developmental Delay with Gastrointestinal, Cardiovascular, Genitourinary, and Skeletal Abnormalities syndrome (DEGCAGS, MIM #619488) was delineated in 2021-22, using genotype-first approaches, in 14 individuals with biallelic, ZNF699 (MIM #609571, chromosome 19p13.2), Loss-of-Function (LoF) variants. The clinical presentation is characterised by wide phenotypic variability that includes a range of neurodevelopmental disabilities, discordant organ anomalies even among full siblings and infant mortality caused by complications of rare, multiple, intestinal atresias [1, 2].

ZNF699, which encodes a KRAB-Zinc Finger Protein (KZFP), is part of a gene family that expanded by segmental and individual gene duplications to reach hundreds of members in the genomes of higher vertebrates. The human genome encodes around 370 KZFPs, 60% of which are in six gene clusters on chromosome 19. The Krüppel associated box (KRAB) domain is the group-defining N-terminal constituent of KZFPs. KZFPs act as major repressors of transposable elements (TEs) via KRAB interaction with the heterochromatin scaffold KRAB-associated protein 1 (KAP1). Each KZFP is also characterised by a specific number of Zinc Finger (ZF) domains in the C-terminal which recognise and bind specific genomic positions. KZFP genes emerged in tetrapods around 400 million years ago and their evolutionary trajectory is intimately associated with that of TEs, which constitute the bulk of their genomic targets. However, the genomic position of the ZNF699 binding sites is, so far, unknown.

While ZNF699 seems to be conserved in many mammals, it has been lost in several rodent species [3, 4].

It is known that many rare genetic disorders are associated with a distinct DNA methylation pattern, a heritable modification to the genome that does not involve DNA sequence changes, known as episignature [5]. In recent years, episignatures have been utilized as stable and reliable biomarkers for the diagnosis of genetic syndromes and for the reclassification of variants of uncertain significance (VUS) and are being implemented in clinical diagnostic laboratories across the world [6–8].

Our objectives were to investigate the genotype-phenotype spectrum of DEGCAGS and, considering that KZFPs act as major epigenetic silencers of TEs [4], to assess peripheral blood DNA for the existence of a diagnostic episignature, to facilitate the diagnosis of such a highly variable condition that lacks a pathognomonic clinical finding.

Material and methods

Cohort collection

We used a genotype‐first approach through the GeneMatcher (GM) [9] data sharing platform and data mining of aggregated DNA sequences from individuals with ultra‐rare disorders across multiple research and diagnostic laboratories worldwide as well as a collaborative call through the European Reference Network on Rare Congenital Malformations and Rare Intellectual Disability (ERN-ITHACA). Individuals with biallelic ZNF699 variants were included. All 12, newly identified, affected individuals had been assessed by consultants in clinical genetics as part of a standard, diagnostic setting. PubMed was queried with the search term ‘ZNF699’. Further inclusion criteria included (i) English language, (ii) availability of the full text, and (iii) description of the patient’s phenotype or pathogenic variant. Articles without phenotypic data were excluded from the phenotypic analysis. In total, four publications describing 18 affected individuals [1, 10, 11] were included (Supplementary Table S2). Clinical and imaging data of individuals with a biallelic ZNF699 variant (new affected individuals and literature review) were collected using an extensive spreadsheet containing more than 100 items (Supplementary Table S1).

GestaltMatcher analysis (comparison to random control)

We used GestaltMatcher (GM) to investigate whether the face phenotype descriptors of affected individuals cluster together more significantly than by chance [12, 13]. We used four published photographs of three individuals and 49 unpublished ones from two newly identified affected individuals (Patient 1 and Patient 8), at different ages (Supplementary Table S3). We compared the five ZNF699 affected individuals to the control distribution in the GestaltMatcher Database (GMDB). We first conducted a statistical analysis utilizing the mean pairwise cosine distance. We randomly sampled 1,555 facial images from individuals representing 328 different syndromes from the GMDB. We then built two control distributions: one for distances between subjects diagnosed with the same syndrome and another for distances between randomly selected affected individuals. In the end, we compared the mean pairwise distance between affected individuals within the given ZNF699 cohort (C) to these two distributions. All the pairwise distributions were sampled randomly 100 times. To avoid bias caused by comparing images from the same patient, we ensured that images from the same patient were not sampled within the same iteration.

We first determined the threshold between two distributions, same syndrome (“controls”) and from random affected individuals (“cases”), by Receiver Operating Characteristic (ROC) analysis. The 5-fold cross-validation resulted in the threshold c = 0.915, corresponding to a sensitivity of 0.851, a specificity of 0.862, and an Area Under the Curve (AUC) of 0.895. The pairwise distance distribution of C is further compared to the control and cases distribution. If more than 50% of distribution C is below the threshold, we consider the given cohort to present similar facial gestalt compared to the random selection.

GestaltMatcher analysis (comparison to selected, disorder-specific cohorts)

We built a pairwise distribution based on the distance between each combination of two cohorts to investigate the similarities between ZNF699, Cornelia de Lange Syndrome (CdLS), and Coffin-Siris Syndrome (CSS) cohorts. This was done using a similar experimental setting as the one introduced above, but instead of calculating the distance within the same cohort, here we calculated the pairwise distance between two different cohorts (disorders or genes).

ROC analysis was performed to identify the threshold that optimally distinguishes whether two cohorts represent the same or different syndromes. The control distributions and resulting threshold were fine-tuned using a five-fold cross-validation, resulting in a threshold of c = 0.9176. If the distance distribution between two cohorts is 50% above the threshold, it was considered evidence that thetwo cohorts represent different syndromes. This approach was tested on 66 syndromes not deriving C in ROC analysis. The method correctly identified cohorts from the same syndrome in 80.30% of the comparisons and cohorts from different syndromes in 91.98%. Ultimately, we compared face descriptors from five ZNF699 affected individuals to those of 319 CdLS affected individuals and 121 CSS affected individuals, separately.

Genetic testing and reporting of sequence variants

The genomic DNA of the proband in each family was analysed by research or clinical exome sequencing. All ZNF699 variants were verified by Sanger sequencing. All variants were described using the reference transcript NM_198535.3 (Mane Select) in reference genome assembly GRCh37. All variants were classified according to the American College of Medical Genetics and Genomics and the Association for Molecular Pathology (ACMG-AMP) classification system and ClinGene Sequence Variant Interpretation general recommendations for using the ACMG-AMP criteria [14, 15]. More specifically PVS1 was used as defined in the PVS1 decision tree (Fig. 1) in Tayoun et al. [16], PM2 was only used at supporting level and PM3 was used as recommended (https://clinicalgenome.org/site/assets/files/3717/svi_proposal_for_pm3_criterion_-_version_1.pdf).

Fig. 1. Spectrum of pathogenic variants across the ZNF699 protein domains.

a Krüppel-associated box (KRAB) domain and 16 zinc fingers (ZF). Domain architectures of human ZNF699 protein was given by Uniprot Q32M78. The numbers showed to the left of the figure indicate which amino acid are defining the start and stop of the protein, the start and stop of the KRAB domain, and the start of the 16 zinc finger domains (Uniprot Q32M78). The seven novel variants are highlighted in bold. Four of the 15 variants (27%) cluster in the sixth C2H2 ZF domain. A non-standard variant nomenclature was used due to space constraints. b 3D detail of the sixth C2H2 ZF domain using the predicted structure of the ZNF699 protein (AF-Q32M78-F1) modeled by AlphaFold2 and visualised in PyMol v2.5.5. The two missense variants affect Cys340 and His353 (in bold, variants in these positions are found in affected individuals), both participating in the stabilization of the ZF structure. The zinc ion is coordinated by two conserved cysteine residues at one end of the β sheet and with two conserved histidine residues at the α-helix C-terminus. AlphaFold structures do not predict the positions of any non-protein components found in experimental structures (such as zinc ions), therefore it is not included in the figure.

For PP1 we used the recommendations described in Jarvik and Browning, (2016) [17]. For the use of PP3/BP4 for missense variants, we used a REVEL cutoff of 0.7 and 0.4, respectively.

Methylation cohort

DNA was extracted from peripheral blood of 14 individuals, nine with ZNF699 pathogenic variants and clinical features consistent with DEGCAGS (age range 0.2-21 years; seven males, two females), and five parents. All were initially included in episignature discovery to obtain a general overview of DNA methylation profiles across samples. The cohort was then divided into two subsets, one with biallelic and the other of heterozygous genotypes. The biallelic samples (n = 9) were used to select probes and construct the classification model while heterozygous subjects were used as a testing set (Supplementary Table S2).

DNA methylation analysis

DNA methylation analysis was carried out using the Illumina Infinium Methylation EPIC bead chip arrays (San Diego, CA, USA) according to manufacturer’s protocols. DNA methylation analysis and episignature discovery were performed using EpiSign software platform in reference to the EpiSign Knowledge Database (EKD) following the previously described protocols [18, 19]. The analysis of methylated and unmethylated signal intensities was performed using R program (version 4.2.0). The minfi package (version 4.2.0) was used to normalize methylation data using the Illumina normalization method with background correction [20]. Probes with detection p value > 0.01, and those located on chromosomes X and Y were removed from the data set. Also, probes containing single nucleotide polymorphisms (SNPs) at or near the CpG interrogation or single nucleotide extension sites and those know to cross-react with chromosomal locations other than their target regions were excluded from the analyses. Additionally, samples with >5% failed probes and those imposing the batch effect were eliminated from the data set. In total, 675,766 probes remained for subsequent analyses. Principal component analysis was carried out to examine the structure of batches and to identify outliers.

Selection of matched controls

Controls were randomly selected from EKD at the London Health Sciences Center [19]. Controls were matched to cases based on age, sex, and array type using MatchIt package (version 4.5.1). Matching trials were repeated to achieve consistency across all analyses. The principal component analysis was employed to ensure none of the matched controls were outliers. A control to case ratio of 7:1 was chosen for the analysis.

DNA methylation profiling of DEGCAGS

For each probe, the methylation levels (β-values) were calculated as the ratio of methylated signal intensity to the sum of methylated and unmethylated signal intensities. The β-values were then logit transformed into M-values using formula log₂(β/(1-β)). The M-values were employed to fit a multivariate linear regression model using the limma package (version 3.52.4) [21]. The blood cell type composition was estimated using the algorithm developed by Houseman et al., and incorporated in the model as a confounding variable [22]. The resulting p values were moderated using the eBayes function of the limma package and corrected for multiple testing according to the Benjamini and Hochberg (BH) method [23]. Three measurements were applied to select probes. Firstly, the methylation differences between cases and controls were multiplied by the negative value of the log-transformed p values, and 1000 probes with the highest product were selected. Secondly, a ROC curve analysis was performed to select 200 probes with the highest area under the ROC curve (AUC). Lastly, Pearson’s correlation coefficient was computed between probes to verify that there are no highly correlated pairs among the selected probes. No probes were identified to be highly correlated. The selected set of 200 probes was then used to perform hierarchical clustering using Ward’s method on Euclidean distance by the gplots package in R. In addition, multidimensional scaling (MDS) was carried out by scaling of the pair-wise Euclidean distances between samples.

Construction of a classification model

A binary support vector machine (SVM) classification model was constructed using the e1071 R package (version 1.7-13), as previously described [18, 19]. This classifier generates Methylation Variant Pathogenicity (MVP) score, which is a metric to determine whether a sample has a DNA methylation pattern similar to the target cohort. The MVP scores range from 0 to 1, where lower scores indicate resemblance to the control group, and values nearing 1 indicate a DNA methylation profile akin to DEGCAGS subjects. The SVM classifier was constructed using training cases, matched controls, 75% of other controls and 75% of 56 other neurodevelopmental disorders from the EKD. The model was then tested using the remaining 25% of other controls and 25% of other neurodevelopmental disorders that are part of the EpiSign v3 classifier within the EKD. This was repeated four times in total so each sample sample could be tested once.

Cross validation

The reproducibility of the model was evaluated using nine rounds of leave-one-out cross-validation. In each round, probes were selected and the model was trained using eight DEGCAGS cases, and the remaining one sample was used for testing. The differentiation of the testing sample was then visualized by generating heatmap and MDS plots.

Functional annotation and comparison of DEGCAGS to other episignature disorders

Functional annotation was performed following the procedure previously described by Levy et al. [24]. The DEGCAGS affected individuals’ samples were trained against controls matched for age and sex to identify differentially methylated probes (DMPs). These controls were selected from a pool of healthy individuals not carrying any of the known episignature. The number of shared DMPs were compared between DEGCAGS and 56 other neurodevelopmental disorders from the EKD using the heatmap. The similarity of different cohorts was further evaluated using a clustering method. A tree diagram was then constructed to evaluate the similarities between cohorts. The median methylation level of each DMP was computed across all samples with the same condition, forming a beta value matrix, followed by computing Euclidean distances between cohorts and clustering these distances using Ward’s method. For each cohort, the top 500 DMPs, ranked by p values, were included in the analysis [24]. For each cohort, the top 500 DMPs, ranked by p values, were included in the analysis. The relationships between cohorts were visualized through a tree and leaf plot generated using the TreeAndLeaf package (version 1.8.0) in R.

Identification of the differentially methylated regions of DEGCAGS

Differentially methylated regions (DMRs) were detected using the DMRcate package (version 2.10.0) [25]. The DMRs were defined as regions that included at least five significantly different CpGs within a 1 kb range, with a minimum mean methylation difference of 10% and a Fisher’s multiple comparison p value less than 0.01. We then used the annotatr package of R (version 1.22.0) to annotate the DMRs to CpG islands (CGIs) and genes [26]. Annotations hg19_cpgs, hg19_basicgenes, hg19_genes_intergenic, and hg19_genes_intronexonboundaries were employed using AnnotationHub (version 3.2.2). In addition to CpG islands, the CGIs annotations included 0 to 2 kb on either side of CGIs (CGI shores), 2 to 4 kb on either side of CGIs (CGI shelves), and other genomic regions (inter-CGI regions). The gene annotations encompassed gene body (including untranslated regions (5ʹ-UTR and 3ʹ-UTR), exons, introns and exon/intron boundaries), promoters (up to 1 kb upstream of the transcription start site (TSS)) and promoter+ (1–5 kb upstream of the TSS). Gene ontology (GO) enrichment analysis of the identified DMRs was conducted using the missMethyl package in R (version 1.30.0) [27].

Comparison of biological and epigenetic age

Because methylation analysis is age-sensitive and biological age affects the individual epigenome [28], we compared the predicted age using methylation data (Watermelon package in R) to the reported chronological age. Chronological age data were not available for the five samples of the carrier parents (Supplementary Table 2). We used Horvath clock method from Watermelon R package to predict the ages [29]. This method uses 353 CpG probes that are present on both 450k and EPIC arrays. We compared the list of these probes with 2101 probes located on the identified DMRs. Only one probe (chr6: 25652602 located within SCGN gene, MIM 609202) was in common between these two probe sets. Interestingly, SCGN gene methylation markers are among key probes to estimate the human age.

Predicted three-dimensional protein model of ZNF699

Currently, no experimentally determined protein structure of ZNF699 is available. Therefore, we used the predicted three-dimensional protein model of ZNF699 (AF-Q32M78-F1) by AlphaFold2 for protein modeling, further visualised in PyMol v2.5.5 [30, 31]. The four amino acids of the sixth C2H2 ZF domain was in a high confidence region (90 > pLDDT > 70).

Results

Cohort description

This cohort comprises 30 affected individuals from 23 families (20 males and 10 females; 12 not previously described and 18 collected from the existing literature; 26 with reported clinical data) (Supplementary Table S2). Affected individuals are of Middle Eastern, European, Asian and Hispanic descent and span an age range from 1 month to 21 years. The mean age at diagnosis was 4.9 years. Three affected individuals passed away due to heart failure (Tetralogy of Fallot), Dengue fever and sepsis, at age 6 months, 9 months and 8 years, respectively.

Molecular spectrum and variant interpretation

Fifteen different variants were identified in 30 individuals (two individuals with two variants in compound-heterozygous state and 28 individuals in homozygous state) and one variant was detected in affected individuals of different ethnic origin (c.339del, p.(Cys113Trpfs*11)). Out of the 14 variants, seven were not previously described in DEGCAGS (NM_198535.3): c.421_424del p.(Glu141Profs*15), c.339del p.(Cys113Trpfs*11), c.918_1001del p.(Cys318_Ser345del), c.1014_1097del p.(Ser346_Ser373del), c.175+1 G > A p.?, c.1039del p.(Ser347Profs*5) and c.1019 G > A p.(Cys340Tyr). The variants were located either within the KRAB domain, the C2H2 ZF domain, or in the region between (Fig. 1a). Five variants were classified as Variants of Uncertain Significance (VUS) but four of them were associated with a DEGCAGS DNA methylation episignature (homozygous or compound heterozygous) (Supplementary Table S2). Thirteen variants are LoF or presumed LoF. The only two missense variants affect Cys340 and His353, both important in binding and coordinating the zinc ion (Zn²⁺) which stabilizes the fold of the sixth C2H2 ZF domain (Fig. 1b). The individual reported by Ali SM et al., (2023) [32] was not included in this cohort because of the unclear significance of the published variant (VUS) NM_198535.3(ZNF699): c.1379 C > T p.(Ser460Leu). The variant’s MAF in GnomAD v.2.1.1 is 0.02% (popmax; Admixed American), and in GnomAD v4.1.0, 0,69% (popmax: Middle Eastern). There is one homozygote in GnomADv.4.1.0, age 55-60. In addition, Ser460 is not well conserved in vertebrates and has a low REVEL score of 0.07 (scale 0–1) with 2x VUS in ClinVar. Methylation profiling of the reported VUS is not available.

Clinical characteristics

The reported phenotypic data on 26 ZNF699 affected individuals is summarized in Fig. 2, Human Phenotype Ontology (HPO)-based phenotypic analysis can be found in Supplementary Table S1 and detailed phenotypic information in Supplementary Table S8. The most consistent clinical features include facial dysmorphism (n = 26; 100%), variable global developmental delay and/or intellectual disability (25/26; 96.2%), skeletal (and dental) abnormalities (22/26; 85%) and an array of neurological findings (hypotonia: 15/26; 58%, CNS structural and myelination abnormalities: 17/26; 65%, movement abnormalities 4/26; 15% and Supplementary Video). In addition, variable patterns of skin and adnexa abnormalities (19/26; 73%), gastrointestinal (16/26; 62%), and genitourinary (16/26; 62%) anomalies were frequently observed. Sensorineural hearing loss, myopia, or hypermetropia were detected in 50% (14/26) of individuals. Immunodeficiency with recurrent infections, growth retardation, anemia and pancytopenia were reported in 11/26 (42%). Some features exhibit age-related presentation. This includes an array of rare skin and adnexa abnormalities, with an exceptionally distinctive finding of selective, premature graying of the eyebrows in one of the two adult individuals in our cohort (Fig. 3a). Premature graying of hair is reported at as early as 11 years of age. The skeletal abnormalities are variable in age of detection (prenatal to adulthood) and in their forms (shortening of long bones, radial ray abnormalities, delayed bone age, epiphyseal dysplasia affecting mainly the hip joints, and lordosis with vertebral body height reduction) (D.W., R.M.L. and S.D.H. pers. obs. and Fig. 3l-p). One of the individuals of this cohort was initially recruited for genetic testing with the suspicion of Diamond-Blackfan Anemia (DBA), an often-syndromic form of congenital hypoplastic anemia caused by pathogenic variants in ribosomal genes or genes involved in ribosomal biogenesis. ZNF699 is thought to interact with RNA polymerase II, and the latter has recently been shown to be involved in ribosomal biogenesis [33]. However, none of the other individuals for whom detailed hematologic analyses (MCV and erythrocyte adenosine deaminase activity) were possible, had consistent hematologic findings.

Fig. 2. ZNF699-related clinical spectrum.

A Overview. B Neurological and (C) Skin and adnexa findings. This is a visual representation of collected clinical data on a very diverse cohort recruited in the diagnostic setting, from 12 different countries including 2 developing countries. Given the rarity of the condition, the extended, variable, and often age-dependent morbidity, and the different healthcare settings, a distinction between not observed (no) or not available (n/a) could not always be objectively assessed. We chose thus to group, via HPO-terms, all clinical/phenotypic occurrences which we could document and group, separately, no/na as a distinct group that could not be objectively assessed.

Fig. 3. The ZNF699 pattern of recognizable dysmorphic features.

a, b Craniofacial dysmorphic features of Patient 1, 17 years old and (c), (d) Patient 8, 8 years old: low anterior and posterior hairline, frontal upsweep of hair, full, thick, highly-arched eyebrows, mildly low-set ears, hooded eyelids, blepharophimosis, acute glabellar/nasal angle, pear-shaped nose with bulbous nose/tip and nasal skin induration, macrostomia, thick vermillion border of lips, retrognathia, open-mouth appearance; (e) GestaltMatcher analysis compares ZNF699 sampling with the general distribution of the GestaltMatcher database (GMDB). The blue distribution was a pairwise distance distribution sampled from the affected individuals with the same disorder, and the red distribution was sampled from the random affected individuals in GMDB. In the orange distribution, 96% of the ZNF699 sampling falls below the threshold (0.915). It indicates that ZNF699 affected individuals share a moderate degree of similar facial gestalt; Patient 1 (f) and Patient 8 (g) distinctive posture with narrow trunk, upper limb contractures, lordosis and wide-based standing position; digital anomalies Patient 8 (h), (k) and Patient 1 (i), (j) brachydactyly, polydactyly, thumb hypoplasia, 2^nd-3^rd toe syndactyly; X-rays Patient 8, 9 years old (l) delayed bone age, hypoplastic terminal phalanges I-V; Patient 1, 20 months of age (m) left hand with shortened metacarpals I-V and clinodactyly of the thumb, right hand with shortened metacarpals I-V, bifid proximal phalange and duplicated distal phalange of right thumb; Patient 8, 14 years 4 months (n) bilateral dysplastic iliac wings and epiphyseal dysplasia of the femoral heads and (o) accentuated lordotic curve. Patient 1, 12 years 5 months, MRI of the abdomen and spine (p) T12-L3 cover plate impression and partial vertebral body height reduction. Partially insufficient vertebral arch closure in the thoracic-lumbar transition and the sacrum area.

Gestalt analysis

In the GestaltMatcher analysis, 96% of the ZNF699 sampling falls below the threshold which indicates that ZNF699 affected individuals share a similar facial gestalt (Fig. 3e). Comparison between ZNF699 and CdLS affected individuals showed that 52% of the random sampling between ZNF699 and CdLS falls below the threshold indicating that the two groups share a moderate degree of gestalt similarity (Supplementary Fig. S6a). On the other hand, although only 10% of ZNF699 and CSS comparison falls below the threshold, the overall distribution is very close to the threshold, which indicates that ZNF699 and CSS share a certain degree of gestalt similarity (Supplementary Fig. S6b).

Identification of an episignature unique for DEGCAGS

We initially compared the methylation β-values between all cases with variants in ZNF699 and controls. The MDS plot clearly differentiated between the ZNF699 biallelic and heterozygous individuals (Supplementary Fig. S1). The classifier was then constructed using the biallelic samples while the heterozygous samples were used as testing set. The methylation levels were compared between cases and controls with a case-to-control ratio of 1:7. In total, 200 differentially methylated CpG probes were selected for the discovery of DEGCAGS episignature. The selected probes were predominantly hypermethylated with 90% exhibiting a positive mean difference in methylation (Supplementary Table S4). A distinct DNA methylation pattern was observed for biallelic cases compared to controls, demonstrating that the selected probes were able to distinguish cases from controls (Fig. 4a). A clear separation was also observed between biallelic cases, and the controls in the MDS plot (Fig. 4b), further supporting the robustness of the identified episignature. Moreover, the heterozygous samples form a distinct cluster – separate to both the controls and samples with biallelic variants (Fig. 4c).

Fig. 4. Episignature of DEGCAGS cohort.

a Euclidean hierarchical clustering (heatmap): each column presents a single DEGCAGS case or control, and each row represents one of 200 probes selected for episignature discovery. A clear separation was observed between biallelic cases (homozygous, red and compound heterozygous, orange), and controls (blue). b Multidimensional scaling (MDS) plot presents the differentiation between homozygous cases (red), compound heterozygous (orange), and controls (blue). c Support vector machine (SVM) classifier: the model was trained by comparing the biallelic cases against matched controls, 75% of other controls, and 75% of the other 56 neurodevelopmental disorders (blue circles). The remaining 25% of controls and 25% of the other 56 neurodevelopmental disorders were used for testing (gray circles).

Construction of a classification model

The SVM classifier was then constructed using 200 selected probes. The model was trained using nine biallelic cases, 56 matched controls, 75% of other controls, and 75% of cases of 56 other neurodevelopmental disorders from the EKD. The classification model was then tested using the remaining 25% of other controls and 25% of other neurodevelopmental disorders from the EKD. All the biallelic ZNF699 samples presented high MVP scores in the SVM plot while all heterozygous samples and cases from other neurodevelopmental disorders provided MVP scores close to zero, confirming the 100% sensitivity and specificity of the classifier in distinguishing biallelic ZNF699 cases, at least among healthy individuals and the tested 56 disorders (Fig. 4c).

Cross-validation

The reproducibility of the classifier was assessed using nine rounds of leave-one-out cross-validation (Supplementary Fig. S2), and every round resulted in all testing DEGCAGS cases being correctly clustered alongside the training cases in both heatmap and MDS plots, demonstrating the robustness of the DEGCAGS episignature.

Comparison of the DEGCAGS methylation profile with other neurodevelopmental disorders

The DMPs were identified by comparing the DEGCAGS cases with matched controls from the EKD. A total of 34,870 DMPs were identified in the DEGCAGS cohort, compared to the range of 279 to 151,848 probes reported in other neurodevelopmental disorders [24]. The probe overlap between DEGCAGS and the 56 disorders in the EpiSign v3 classifier was evaluated (Supplementary Fig. S3). The list of EpiSign™ disorders and their abbreviations are presented in Supplementary Table S5. The DEGCAGS cohort shared the highest percentage of overlapping probes (42%) with the ‘BAFopathy’ episignature. The mean methylation differences were also compared between the DEGCAGS and other cohorts, revealing a predominant pattern of hyper-methylation changes in DEGCAGS cases (Supplementary Fig. S4a). The clustering analysis of the top 500 DMPs from each cohort were employed to examine the similarities between cohorts. In the tree and leaf plot, the DEGCAGS was clustered alongside the Hunter-McAlpine craniosynostosis syndrome (HMA) (MIM# 601379) (Supplementary Fig. S5).

Advanced epigenetic age

Interestingly, all predicted ages of the ZNF699 biallelic samples analysed in our cohort were higher than the reported ages (when available) at the time of sample collection (nine individuals). This gap was particularly significant in three individuals (predicted/chronologic ages in years: 35.2/16, 36.6/5, and 20.3/4). The overall correlation between the predicted and chronological ages was ~0.75. (Supplementary Table 2).

Differentially methylated regions (DMRS)

A total of 210 DMRs were detected in the DEGCAGS cohort using the DMRcate algorithm (Supplementary Table S6). The majority of the identified DMRs (>98%) were hypermethylation events, and annotation analysis revealed that more than 88% were located within CpG islands (Supplementary Fig. S4b). Additionally, the gene annotation showed that most of DMRs are found within promoters (~58%) and CDs (~19%) (Supplementary Fig. S4c). The GO enrichment analysis identified 13 significant GO terms (False Discovery Rate <0.01), the majority of which were associated with MIR21, CALB1, MPL, DPP6, and FAAH genes (Supplementary Table S7). Significant biological processes linked to the identified genes include cell adhesion, cell-cell adhesion, mRNA trans-splicing via spliceosome, mRNA trans-splicing spliced-leader addition, formation of quadruple SL/U4/U5/U6 snRNP, and melatonin receptor activity.

Discussion

Disruption of several KZNFs is known to cause various forms of neurodevelopmental disorders (NDDs) such as intellectual disabilities and psychiatric disorders, with occasional structural brain abnormalities [34]. We show that loss of ZNF699 causes a distinctive syndromic form of ID and multiple malformations, with unique findings that exhibit age-dependent presentation such as movement abnormalities (4/26, 15%, Supplementary video). Affected individuals can develop skeletal abnormalities that affect their posture and become more obvious around puberty, characterized by lumbar lordosis, contracture of large joints, and wide-based standing position (Fig. 3f, g); as well as very distinctive, selective premature graying of the eyebrows (Fig. 3a) and a recognizable facial gestalt (Fig. 3a–e).

Our group recently provided some evidence of the utility of episignature as a diagnostic tool [8]. Larger scale projects (e.g., EpiSignCAN) are also on ongoing to provide more information on clinical epigenomics [6]. These studies are necessary to assess the diagnostic yield and health system impact as either a first-line test or in unresolved cases post–genomic assessment. Then, the development of clinical guidelines for use and application of clinical epigenomic technologies can be warranted. The DEGCAGS episignature allowed the reclassification of ZNF699 VUS (3/9, 30%, homozygous or compound heterozygous) to Likely Pathogenic (Supplementary Table S2). We propose Episign^TM as a screening and diagnostic tool in individuals with a dysmorphic NDD and craniofacial hypertrichosis. The comparison of the DEGCAGS episignature with other NDDs showed a partial overlap with the BAFopathy episignature. Craniofacial hypetrichosis was reported as a distinct phenotypic feature in >50% of ZNF699 individuals (Supplementary Table S1) and was reminiscent of that observed in individuals with Cornelia De Lange (MIM Phenotypic Series - PS122470), Coffin-Siris (MIM Phenotypic Series - PS156200, PS135900) or other, dysmorphic NDDs with a defined episignature. On this basis, we chose to specifically target the GM analysis for comparison with CdLS and CSS which also confirmed a certain degree of gestalt similarity. GestaltMatcher may also aid variant classification providing evidence using the Bayesian framework of the ACMG classification guidelines [35]. Yet, photographs from a larger number of affected individuals are required to achieve this. This analysis can be performed in the future once more families with affected ZNF699 individuals consent to share representative facial photographs. Still, our results showed that the facial features of ZNF699 affected individuals are very similar to each other and distinctively different to a random selection.

Another frequent clinical finding in this DEGCAGS cohort is the premature graying of the hair, which can be seen in several, very rare, genetic disorders, including those characterized by premature aging or progeroid phenotype [36]. However, none of the patients we assessed exhibit other features of premature aging [37]. Interestingly, all predicted ages of the ZNF699 biallelic samples analysed in our cohort were higher than the reported ages at the time of sample collection (nine individuals) (Supplementary Table 2).

Based on the methylation data, the samples heterozygous for a ZNF699 variant form a distinct cluster – separate to both the controls and samples with biallelic variants (Fig. 4c) which suggests a level of cellular consequence for the heterozygous state too. Although clinical phenotyping of heterozygotes (presumed) parents was beyond the scope of this study, we believe that clinical phenotyping of the parents of affected individuals for any ZNF699-related clinical effects, is an important aspect to follow-up in the future.

The DMRs annotation and subsequent enrichment analysis revealed several genes that are associated with biological processes implicated in other NDDs with clinical features overlapping with DEGCAGS. One of DMRs identified on chromosome 17 contains the MIR21 gene, which plays key roles in several mechanisms implicated in the pathogenesis of CNS disorders [38]. Another DMR located on chromosome 17 overlaps with TIMP2 gene. This gene contributes to the maintenance of tissue homeostasis, and has been suggested to support normal cardiovascular, immune and cognitive functions [39]. The DMR located on chromosome 1 contains the MPL gene, which is involved in the production of the thrombopoietin receptor protein which promotes the growth and proliferation of megakaryocytes and production of platelets [40]. These roles might be involved in the pathogenesis of some of the features observed in DEGCAGS syndrome, such as hematopoietic abnormalities (anemia or pancytopenia) and immunodeficiency.

The annotated Drosophila homolog of the ZNF699 gene (35% homology) has been associated with the development of ethanol tolerance in fruit flies [41]. While an early study in humans suggested an association between several markers within the ZNF699 locus and alcohol dependence in humans, a follow-up study did not replicate those findings except in one subgroup [42]. However, given the known association of KRAB-ZNFs with NDDs and psychiatric disorders, we examined the DEGCAGS-related DMRs for overlap with genes implicated in substance use disorders. Out of the 210 ZNF699-LoF-related DMRs (>5 DM-probes in a cluster), 173 (82%) were in genes. Of those gene-related DMRs, seven (4%) were in genes previously linked to substance use disorders in GWAS studies: GRIK2, CALB1, and STXBP5L linked to tobacco; DPP6, GABRG1, and TRAK1 to alcohol; and FAAH to polysubstance use disorder. The DMR located on chromosome 7 involves DPP6, a gene implicated in neuronal excitability dysfunction and pathogenesis of dementia [43], with possible involvement in NDDs [44]. The DMR identified on chromosome 8 overlaps with CALB1, a gene that has been associated with epigenetic progression of seizure-related cognitive deficits and autism [45]. The FAAH gene, which is located on one of the DMRs on chromosome 1, was associated with epilepsy, ADHD, anxiety, and depression [46]. The DMR located on chromosome 5 includes several members of the protocadherin gene family (PCDHGA1, PCDHGA2, PCDHGA3, PCDHGB1, PCDHGA4, PCDHGB2 and PCDHGA5), which are involved in neurodevelopment and neuropsychiatric diseases [47]. In addition, alterations in DNA methylation patterns within these genes have been linked to prenatal alcohol exposure and fetal alcohol spectrum disorder [48]. While these findings suggest potential interactions between ZNF699 and other genes, further functional experiments are necessary to gain a deeper understanding of the target pathways involved in the pathogenesis of DEGCAGS syndrome.

Author contributions

Conceptualization: DW, BS and SDH; Data Curation: KK, MH, IA, T-ChHs, HL., AA and DW; Formal Analysis: KK, DW, IA, and T-ChHs; Funding acquisition: FH, TJS, BS and SDH; Investigation: KK, DW, IA, T-ChHs, JP, RML, LD, MD, HL, TL, KHS, B.M.S.Al-M, RGYAl-Ob, DM, MR, MB, MSS, AA, HTG, LP, ShSh, HD, SB, FL, AS, TK, TJ-Sch, NAS, RR, MAL, JK, JS, HH, GVP, FH, SBS, RM, TWY, and SDH; Methodology: KK, T-ChHs, MAL, BS and SDH; Project Administration: HL, RM, MH, DW, KK, IA and SDH; Software: KK, IA, T-ChHs, BS and PK; Supervision: DW, RML, SBS, RM, TY, PK, BS and SDH; Validation: KK, IA, T-ChHs; Visualization: KK, DW, IA, RML, LD, T-ChHs and SDH; Writing-original draft: KK, DW, T-ChHs, IA and SDH; Writing-review and editing: KK, DW, T-ChHs, AA, BS and SDH.

Funding

This study was funded by the Norwegian National Advisory Unit on Rare Disorders (grant #43066 to SDH), the government of Canada through Genome Canada and the Ontario Genomics Institute (OGI-188 to BS), grants from the National Institutes of Health (DK068306) and the Begg Family Foundation (both to FH) and grants by the Interdisciplinary Center for Clinical Research Erlangen (J70) and the Deutsche Forschungsgemeinschaft (281319475) (both to TJS). Sequencing and analysis of [GAZ_431/ Patient 16] were provided by the Broad Institute of MIT and Harvard Center for Mendelian Genomics (Broad CMG) and were funded by the National Human Genome Research Institute (NHGRI) grants UM1HG008900 (with additional support from the National Eye Institute, and the National Heart, Lung and Blood Institute), and R01HG009141 (to LP). All funders played no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript.

Data availability

Data and materials can be supplied upon request. All seven novel ZNF699 variants were deposited in ClinVar, accession numbers SCV004565351 – SCV004565357. Some of the datasets used in this study are publicly available and may be obtained from gene expression omnibus (GEO) using the following accession numbers. GEO: GSE116992, GSE66552, GSE74432, GSE97362, GSE116300, GSE95040, GSE104451, GSE125367, GSE55491, GSE108423, GSE116300, GSE 89353, GSE52588, GSE42861, GSE85210, GSE87571, GSE87648, GSE99863, and GSE35069. These include DNA methylation data from individuals with Kabuki syndrome, Sotos syndrome, CHARGE syndrome, immunodeficiency-centromeric instability-facial anomalies (ICF) syndrome, Williams-Beuren syndrome, Chr7q11.23 duplication syndrome, BAFopathies, Down syndrome, a large cohort of unresolved subjects with developmental delays and congenital abnormalities, and also several large cohorts of DNA methylation data from the general population. Remaining data are not available due to institutional or REB restrictions. EpiSign is proprietary, trademarked analytical software owned by EpiSign Inc. and parts of it are based on the methods and publicly available software that are referenced in the methods section.

Competing interests

BS is a shareholder in EpiSign Inc, company involved in commercialization of EpiSignTM technology.

Ethical approval

This study was performed according to the Declaration of Helsinki and approved by the Western Norway Regional Ethics Committee (REC 604007) and the Western Ontario University Research Ethics Board (REB 106302). Written informed consent for the publication of photographs, videos and medical information was obtained from parents/legal guardians.

Supplementary information

The online version contains supplementary material available at 10.1038/s41431-024-01702-y.