Polycomb-associated and Trithorax-associated developmental conditions—phenotypic convergence and heterogeneity

Chromatin remodelling is a fundamental epigenetic mechanism that orchestrates development and maintains the stability of differentiated tissues. Polycomb group (PcG) and Trithorax group (TrxG) multi-subunit complexes have key roles in chromatin remodelling, and act antagonistically to regulate chromatin accessibility by inducing the repression and activation of gene expression respectively [1, 2]. PcG complexes mediate transcriptional repression by promoting nucleosomal compaction, as well as deterring TrxG action [1, 3, 4]. TrxG complexes (which primarily include COMPASS methyltransferases and BAF (SWI/SNF) remodelling complexes) maintain active gene expression and oppose PcG action by relaxing nucleosomal compaction [1, 3, 4]. A tightly regulated balance of PcG and TrxG activity is thus necessary for the precise timing of gene expression during mammalian development; in particular, the PcG-TrxG system has a vital role in neurogenesis, by mediating the balance of neural progenitor cell differentiation and self-renewal [5].

Multiple developmental conditions arise from pathogenic variants in PcG and TrxG genes, forming a subset of Mendelian Disorders of Epigenetic Machinery (MDEMs) [6]. Whilst individually rare, PcG and TrxG-related conditions collectively account for ~8% of monogenic developmental disorders in the DECIPHER database [7]. These conditions are phenotypically heterogeneous, with variability within and across different genetic diagnoses. It has been suggested that while MDEMs are monogenic, they are conceptually similar to complex disorders, with epigenomic dysregulations contributing to an array of different phenotypes and severities via downstream pathways [8]. Patients with different PcG and TrxG-related conditions can share overlapping phenotypic characteristics: for example, broad similarities can be observed between BAF- and COMPASS-complex associated conditions [9,10,11,12,13,14]. It is plausible that phenotypic similarities across conditions are due to epigenomic dysregulation at shared genomic loci and at similar developmental stages, leading to downstream effects on gene expression that affect convergent developmental pathways. However, there are also phenotypic divergences within this group: for example, opposing features relating to growth have been observed amongst PcG-related conditions arising from variants in PRC1, PRC2 and PR-DUB complex genes [2, 5]. Phenotypic descriptions of each MDEM in isolation may not fully capture the extent of overlap or distinctiveness between conditions, and may underplay the extent of variability within each condition. A phenotypic evidence base that accounts for this complexity could improve post-diagnostic counselling and stimulate hypothesis-driven, clinically relevant research.

The current paper reports data-driven analysis, aiming to describe the phenotypic commonalities and heterogeneities of PcG and TrxG-related conditions. Our first objective was to identify phenotypes that occur at elevated frequencies across the population of patients with PcG and TrxG-associated conditions, in comparison to other monogenic conditions that are present in the DECIPHER dataset. If phenotypic enrichments exist for this broad group of conditions, this reinforces the evidence for pathway-level convergence in developmental mechanisms and clinical needs. Our second objective was to map gene-level heterogeneity within PcG and TrxG-related conditions. To do this, we investigated whether there are phenotypic clusters of PcG and TrxG-associated conditions when patient data are grouped by affected gene. If gene-driven clusters are present, this supports the recognition of clinical syndromes aligning with variants across several specific genes, and may point to shared epigenomic dysregulations. Our third objective was to map patient-level heterogeneity to explore whether there are phenotypic clusters of patients with PcG and TrxG-associated conditions; the results of this final analysis may support or challenge the alignment between MDEM-associated phenotype co-occurrences and variants in specific genes.

Gene list curation

Gene Ontology (GO) annotations were used to collate a comprehensive PcG and TrxG gene list (Supplementary Fig. 1A), following a similar approach to Ciptasari and van Bokhoven [15]. GO Cellular Component terms selected to represent PcG complexes were ‘PRC1 complex’ (GO:0035102), ‘ESC/E(Z) complex’ (synonymous with ‘PRC2 complex’; GO:0035098) and ‘PR-DUB complex’ (GO:0035517). To gather terms related to SWI/SNF and COMPASS complexes, several child terms of ‘SWI/SNF superfamily-type complex’ (GO:0070603) and three terms that correspond to the three subtypes of COMPASS complex were selected. After collating 114 genes using relevant GO terms, a further 21 genes that are likely to possess PcG/TrxG involvement were added based on literature reviews, and six genes that lack direct or specific PcG/TrxG involvement were removed (Supplementary Table 1A). This resulted in a final list of 129 PcG/TrxG genes (Supplementary Table 1B).

Cohort and phenotypes curation

The DECIPHER database is a repository of genotypic and Human Phenotype Ontology (HPO) standardised phenotypic data deposited by clinical geneticists and laboratory scientists [7, 16]. The DECIPHER open-access dataset was filtered using the pre-defined PcG/TrxG gene list, to identify all pathogenic or likely pathogenic sequence variants located in a PcG-related or TrxG-related gene (Supplementary Fig. 1B). 506 variants that were present across 499 patients were identified; importantly, all patients with more than one variant had variants present in the same gene. Patients with 0 HPO terms were removed (n = 37), resulting in a cohort of 462 patients with pathogenic or likely pathogenic variants across 38 PcG/TrxG genes (Supplementary Table 1C). 99 patients had PcG variants, 359 patients had TrxG variants, and 4 patients had variants in HCFC1 which has roles in both PcG (PR-DUB accessory subunit) and TrxG (SET1A/B-COMPASS complex subunit). All patients with autosomal variants were heterozygous (n = 447), while all patients with X-linked variants were hemizygous (n = 15). The majority (81%) of patients in this cohort had a de novo variant, while 13% had unknown inheritance and 6% had a maternally or paternally inherited variant. To obtain a comparison cohort, the remainder of the DECIPHER dataset was filtered to include unique patients with pathogenic or likely pathogenic sequence variants, and patients with 0 HPO terms were removed (n = 5070).

Group-level HPO enrichment analysis

To address the first study objective, the HPO terms reported for each patient were propagated using the Python package PyHPO, so that broader ‘top-level/parent’ phenotypes could be compared across groups, in addition to more specific ‘lower-level/child’ phenotypes [16]. The median number of reported HPO terms and top-level HPO terms for each patient in the PcG/TrxG and comparison cohorts were compared via Mann–Whitney U rank tests. We then compared the frequency of propagated HPO terms between the PcG/TrxG and comparison cohorts, using a two-proportions z-test with Benjamini-Hochberg correction for multiple testing (Fig. 1a). First, we examined whether any top-level HPO terms were enriched in the PcG/TrxG population: for this analysis the top-level term ‘Abnormality of the musculoskeletal system’ was split into the descendent terms ‘Abnormality of the skeletal system’, ‘Abnormality of the musculature’ and ‘Abnormality of connective tissue’, which resulted in a total of 25 top-level HPO terms (Supplementary Table 2). Second, we examined lower-level term enrichments.

A Method to compare HPO term frequencies between PcG/TrxG and comparison cohorts. B Method to identify clusters of PcG/TrxG genes showing phenotypic similarity. C Method to identify clusters of patients with PcG/TrxG gene variants showing phenotypic similarity. P/LP Pathogenic or likely pathogenic, IQR interquartile range, Q1 quartile 1, Q3 quartile 3, IC information content, BH Benjamini-Hochberg correction.

Full size image

Gene-level HPO analysis

Figure 1b details the approach taken to examine gene-associated phenotypes within PcG- and TrxG-related conditions. This involved grouping patients by genetic condition, identifying a set of representative HPO terms of each gene group, and computing semantic similarity scores, before performing cluster analysis to identify phenotypic similarity between gene groups. To identify representative HPO terms for each gene group, unique propagated HPO terms were collated, and depending on gene group size, HPO terms present in at least 2 patients or 20% of patients were retained (Supplementary Table 3). At this stage, each gene group contained a variable number of HPO terms, with some groups containing higher or lower number of terms. This could impact the cluster analysis outcomes by biasing the gene groups with similar numbers of HPO terms to artefactually cluster together. To reduce this bias, the number of HPO terms per group was approximately equalised. Based on the interquartile range of the number of terms across all gene groups, the cut-off for the minimum and maximum number of terms retained per gene group was determined (i.e., removing outliers). The HPO terms retained within the set limit were prioritised based on their information content (IC) [17, 18]. The IC of HPO terms is a measure of the rarity/ specificity of a term within a database, where the IC of term t is given by, IC(t) = −log p(t), and p(t) is the probability of occurrence of t. Thus, terms with higher IC reflect more specific and clinically informative terms. As such, in this analysis, terms with higher IC were prioritised to be retained within the IQR set limit of number of terms.

Next, the semantic similarity scores and pairwise gene group similarity scores for each pair of gene groups were computed, using the graph IC method followed by the funSimAvg method [18]. This approach takes into account all common ancestors of pairs of terms. The resulting similarity matrix was used as input for clustering. The method of cluster analysis chosen for this study was agglomerative hierarchical clustering using complete linkage; hierarchical clustering does not require a pre-specified number of clusters, which is useful for this study, as the number of different phenotypic profiles across the cohort was not known before clustering. Clustering was carried out using scikit-learn and SciPy packages in Python, and number of clusters was selected via visual inspection of the dendrogram [19]. To illustrate phenotype sharing between genes which leads to clustering, networks of semantic similarity were visualised using Cytoscape [20]. The method used to select HPO terms for network visualisation is detailed in Supplementary Table 4: in brief, terms were included if present in at least 4 gene groups, terms with listed descendent terms were excluded, and some terms were merged to provide clearer visualisation.

Patient-level HPO analysis

Figure 1c outlines the second approach for identifying phenotypic clusters, which involved directly computing the semantic similarity between patients, prior to performing hierarchical clustering. Raw HPO terms associated with each patient were collated. Modifier terms as well as terms with descendent terms were removed (i.e. parent term was removed if its child term is present). Similar to the gene-level pre-processing approach, the number of HPO terms associated with each patient was approximately equalised by identifying the upper outlier term frequency (Q3 + 1.5 × IQR) across all patients and using this as an upper limit with terms prioritised by highest IC. The semantic similarity between each pair of patients was then computed, before performing hierarchical clustering as described previously. Following clustering, HPO terms associated with each patient were propagated and the proportion of patients with each term in each cluster was compared to the proportion of patients with each term outside of each cluster, using Fisher’s exact test with a Benjamini-Hochberg correction. HPO terms that were significantly enriched (p_adj< 0.05) in each cluster were identified.

Group-level HPO enrichment analysis

The number of HPO terms per patient for the PcG- and TrxG-associated conditions and comparison cohorts are displayed in Supplementary Fig. 2. There is a significant distribution shift (p = 1.51e-10) towards a higher number of reported HPO terms for patients with PcG- and TrxG-associated conditions (median = 7) relative to the remainder of the DECIPHER dataset (median = 6). Similarly, using top-level HPO terms as a proxy for the number of organ systems affected, there is a significant shift (p = 1.51e-13) towards a higher number of top-level HPO terms across patients with PcG- and TrxG-associated conditions (median = 5) relative to the remainder of the DECIPHER dataset (median = 4). Together, these findings indicate an increased number of phenotypes per patient, and a larger breadth of affected organs/organ systems, in patients with PcG- and TrxG-associated conditions compared to other monogenic conditions.

Using propagated terms for each patient, the percentage occurrence of top-level HPO terms among patients with PcG and TrxG-associated conditions was compared to the percentage occurrence of each term across the remainder of the DECIPHER dataset. Five top-level terms were significantly increased at p_adj < 0.005 among patients with PcG and TrxG-associated conditions (Fig. 2, Supplementary Table 5, Supplementary Table 6): ‘Abnormality of the integument’ (p_adj= 2.04e-09), ‘Growth abnormality’ (p_adj= 5.44e-08), ‘Abnormality of head or neck’ (p_adj= 1.33e-06) ‘Abnormality of the digestive system’ (p_adj= 4.04e-06), ‘Abnormality of limbs’ (p_adj= 0.0012). ‘Abnormality of the nervous system’ was also marginally increased at p_adj = 0.019. No top-level terms were significantly reduced in the PcG/TrxG group. The distribution of each top-level term across each genetic condition is illustrated in Supplementary Fig. 3; it can be observed that a substantial proportion of patients with each PcG and TrxG-associated condition have abnormalities of the head or neck, integument, limbs and nervous system, while growth abnormalities and digestive system abnormalities are more unevenly distributed.

The proportion of each enriched (p_adj< 0.05) propagated HPO term among patients with PcG and TrxG-associated conditions is indicated by the y-axis, relative to the remainder of DECIPHER (x-axis). Terms with at least a ±3% change between the PcG/TrxG group and remainder of DECIPHER are shown. Significantly enriched top-level terms are labelled in bold; also labelled are the two terms with the largest positive percentage difference and the two terms with the largest negative difference between the groups.

Full size image

Supplementary Table 6 lists specific HPO terms which were significantly increased or decreased at p_adj < 0.05, and that have a percentage difference of at least ±3% between the PcG/TrxG group and comparison group. 66 descendent terms of the significantly enriched top-level terms were also significantly enriched in the PcG/TrxG group, and 2 descendent terms were significantly reduced (‘Seizure’ and ‘Abnormality of skull size’; Fig. 2). Five terms not descended from enriched top-level terms were significantly increased in the PcG/TrxG group (abnormalities of tone; abnormalities of respiratory tract/larynx; Supplementary Table 6). Of note, there are several specific HPO terms that are enriched in PcG/TrxG individuals with a prevalence within the case group of over 20%; these include abnormal skin or hair morphology, abnormal ocular or oral morphology, and abnormal hands or upper limbs. Additionally, 94% of patients with PcG/TrxG variants are reported to have nervous system abnormality, a modest elevation compared to other conditions (89%). ‘Global developmental delay’ (54% cases, 46% controls) and ‘Neurological speech impairment’ (25% cases, 19% controls) were also enriched, but terms relating to severity of developmental delay did not differ between groups. Furthermore, ‘Abnormal cerebral white matter morphology’ is an enriched phenotype, while seizures are reported at lower frequency (10% cases, 20% controls). In summary, despite the heterogeneity of PcG-related and TrxG-related conditions, there are detectable differences between this amalgamated population and the broader developmental disorders population: we observed differences in the median number of reported phenotypes, and differences across broad and specific phenotype frequencies.

Gene-level HPO analysis: semantic similarity, clustering and network analysis

Representative HPO terms for each genetic diagnosis are listed in Supplementary Table 3. The computed semantic similarity matrix was used to perform hierarchical clustering (Fig. 3A), resulting in 3 multi-gene clusters and 2 stand-alone gene groups. We examined clusters for potential bias arising from HPO term frequencies across each cluster (Supplementary Fig. 4A). Network analysis was applied to visualise the broad shared phenotypic characteristics of each cluster—for visualisation purposes, networks are presented separately for ‘Abnormalities of the nervous system’ (Fig. 3B) and descendent terms of ‘Head and neck abnormality’, ‘Abnormality of limbs’, ‘Abnormality of the integument’ and ‘Growth abnormality’ (Fig. 3C). Broadly, it can be observed that genes within cluster 1 share growth abnormalities (including terms relating to overgrowth and restricted growth) and atypical behaviour (including autistic behaviour); genes within cluster 3 share abnormalities of brain morphology, particularly microcephaly; genes within clusters 2 and 3 share limb abnormalities and facial dysmorphology. There was no clear pattern of PcG/TrxG complex membership aligning with phenotypic cluster membership i.e. COMPASS and PcG associated genes were distributed across all three clusters, while BAF complex genes were distributed across clusters 2 and 3.

A Dendrogram of gene-level clustering output. B Network illustration of gene relationships within clusters, based on nervous system abnormalities present in at least four gene groups. C Network illustration of gene relationships within clusters, based on abnormalities of physical development present in at least four gene groups. GDD Global developmental delay.

Full size image

Patient-level HPO analysis: phenotypic similarity, clustering and cluster comparisons

To identify clusters of patients with phenotypic similarity within the PcG- and TrxG-associated conditions population, irrespective of specific genetic diagnosis, hierarchical clustering was performed on the unpropagated HPO terms for each individual patient. Following inspection of the dendrogram (Fig. 4A), the population of patients was divided into 12 clusters, varying in size between 3 and 111 patients; this number of clusters was selected as it provides good insight into sub-clusters across the population, without separating larger clusters into smaller very similar clusters. Supplementary Fig. 4B shows the HPO term numbers across each cluster. The phenotypic profile of each of the 12 clusters was then analysed using propagated patient-level terms, with significantly enriched terms identified at p_adj < 0.05 (Supplementary Table 7). Figure 4C summarises the two most significant HPO terms for each cluster. Cluster 1 contains the largest number of patients (n = 111) and is enriched for a collection of neurodevelopmental abnormalities, particularly neurological speech impairment, in addition to abnormal muscle tone. Cluster 2 (n = 91) is characterised by various abnormal facial morphologies, and Cluster 3 (n = 48) has a higher proportion of patients with behavioural characteristics, including autistic behaviour. Cluster 4 (n = 48) is characterised by abnormalities of the eye, while Cluster 5 (n = 32) contains patients with growth abnormalities—both undergrowth (short stature, growth delay) and overgrowth (tall stature, overgrowth). Cluster 6 (n = 30) is associated with skull dysmorphologies, particularly microcephaly. The remaining clusters contained 10 or fewer patients and were each defined by more specific HPO terms.

A Patient-level dendrogram. B Within-cluster proportional occurrence of top-level terms that were increased across the population as a whole. C Within-cluster proportional occurrence of 2 representative HPO terms for each of the 12 clusters, which were identified using Fisher’s exact test.

Full size image

Supplementary Fig. 5 illustrates the distribution of cluster memberships for patients within each PcG- and TrxG gene diagnosis group. It can be observed that the majority of gene groups are phenotypically heterogeneous, and that phenotype clusters are distributed across genes. There are some exceptions, namely CHD8 (>50% patients within Cluster 11), SETD1A (>50% patients within Cluster 1) and BCOR (>50% patients within Cluster 12). However the absolute number of patients with each of these genetic diagnoses is low, hence this variation in proportional cluster membership may occur by chance.

This paper presents data-driven analyses of the phenotypic landscape of PcG- and TrxG-associated conditions. These analyses provide important insights into clinical convergences and divergences within this population, relevant to diagnosis, post-diagnostic counselling and management. Moreover, this study highlights opportunities and limitations of HPO computational methods for investigating disorder-associated pathways defined by convergent gene function.

We first assessed group-level differences in HPO term frequencies between PcG/TrxG-associated conditions and other monogenic developmental disorders. We identified an increased total number of HPO terms and an increased number of top-level HPO terms for patients with PcG/TrxG variants, reflecting the multi-system manifestation of these conditions; a broader study of MDEMs by Bjornsson [21] identified a similar trend, although it is important to note that our study will have included non-PcG/TrxG MDEMs in the control group, which may have weakened this effect. PcG/TrxG-associated conditions shared multi-system characteristics, with significantly increased frequency of abnormalities of growth, limbs, digestive system, integument and head or neck. In particular, the identification of enriched growth phenotypes is consistent with a previous MDEM investigation by Farhner and Bjornsson [6], which identified growth abnormalities as a consistent phenotypic finding. However, in our analysis head circumference differences were grouped under skeletal abnormalities, not growth abnormalities, due to the structure of the HPO. This data-driven separation of growth from head circumference enabled us to observe a divergence in results, whereby there was a reduced prevalence of abnormalities of head circumference in the PcG/TrxG versus comparison groups. More specific HPO terms reported in >20% of PcG/TrxG individuals include abnormal skin or hair morphology, abnormal ocular or oral morphology, and abnormal hands or upper limbs; these significant enrichments may be useful phenotypic markers for confirming pathogenicity of PcG/TrxG variants. On the other hand, occurrences of each specific characteristic affect a minority of the population, emphasising the variability of this population, and need to assess combinations of phenotypes rather than each in isolation.

Building on group-level analysis of phenotype enrichments, our next goal was to assess phenotypic variability within PcG/TrxG-related conditions. We observed three clusters of genes which share broad top-level terms but differ in frequency of lower-level terms. Genes within clusters show multi-phenotype similarity to their within-cluster neighbours, but sparse connectivity to genes in other clusters. The phenotypes driving cluster separation within this dataset are likely to be partially influenced by the reporting biases of DECIPHER. However, these results are potentially hypothesis-setting for mechanistic studies: for example, the high phenotypic similarity between conditions associated with BAF, COMPASS and PcG subunits may reflect convergent impacts on gene expression [8]. However we also observed contrasting phenotypic profiles associated with functionally similar genes (for example, KDM6A and KDM6B); these differences may be influenced by variant-specific, cell-specific or timing-specific effects.

The pooled phenotypic profile for each gene does not take into account patient-level heterogeneity [22]. Therefore, we carried out a second clustering analysis including all patients with at least four reported HPO terms. We identified 12 phenotypically-defined clusters, which did not align with genetic diagnoses. Smaller clusters are driven by low prevalence phenotypes, whereas larger clusters are characterised by broad and frequent terms, with more subtle between-cluster differences. Phenotype cluster memberships may be influenced by unanalysed factors, such as the age of patients at phenotyping, and variable reporting by clinical geneticists, developmental paediatricians or neurologists. Importantly, gene groups were distributed across multiple clusters, and the majority of clusters were genetically heterogeneous, suggesting that phenotyping is not overly influenced by prior expectations of gene-specific features.

The majority (94%) of patients with PcG/TrxG variants are reported to have nervous system abnormality, with the terms ‘Global developmental delay’ and ‘Neurological speech impairment’ enriched in this group. This reflects clinical experience whereby neurodevelopmental difficulties are predominant features for most individuals with MDEMs, and remain a high priority for patients and families across the lifespan. However, more specific terms such as those relating to severity of developmental delay, specific cognitive problems, speech and language difficulties, or social-behavioural characteristics did not differ between case and control groups, indicating extensive heterogeneity and lack of specificity. It is important to note that the reporting of cognitive and behavioural characteristics is sparse within DECIPHER, and HPO terminology is not designed for quantification of dimensional neurodevelopmental differences. Despite these limitations, neurobehavioural HPO terms did contribute to gene-level clustering. Specifically, autistic behaviour contributed to similarity of genes within cluster 1, suggesting that there is potential for some PcG/TrxG genes to be more strongly associated with neurobehavioural phenotypes than others. Moreover, patient-level analysis highlighted phenotypic cluster 3 in which both autistic behaviour and hyperactivity were significantly enriched (Supplementary Table 7), indicating that there are individuals across the PcG/TrxG spectrum for whom cognitive, social and behavioural difficulties are co-occurring features. Identifying predictive factors contributing to membership of this cluster is an important topic for future research. Given the prevalence and variability of neurodevelopmental phenotypes within the PcG/TrxG group, investigation of multi-level mechanisms linking epigenetic regulation to parameters of structural and functional brain development and emergent cognitive vulnerabilities is warranted.

Our results reinforce the complexity of MDEMs and the potential for early developmental mechanisms which cascade to system-level consequences that are not predictable by gene alone. However, limitations of our analyses include variable patient numbers across MDEMs in DECIPHER (more than 50% of the total group have one of four diagnoses), making it difficult to gauge and compare the phenotypic spectrum of every condition. There are disparities in phenotyping detail across the DECIPHER dataset—while some patients have an extensive list of fine-grained HPO terms, others have a single broad term, such as ‘Global developmental delay’; this may reflect the reality of a patient’s phenotype, or variations in phenotyping standards. However, we mitigated this bias by prioritising terms with higher IC (and therefore specificity), and application of a term equalisation step in our pre-processing to mitigate potential biases arising from disparities in HPO term numbers. Additionally, facial phenotypes and other specific morphological features can differentiate disorders from each other [23]. The fact that not every patient has a comprehensive list of specific phenotypes means that subtle dysmorphic differences will not contribute to our clustering analyses.

Our analyses focused on a subset of MDEMs, and future studies may be more informative by taking a more inclusive strategy to gene selection; we focused on the well-defined PcG and TrxG-associated complexes, while there is uncertainty as to what constitutes an ‘epigenetic’ protein involved indirectly in chromatin structural regulation. Despite this, there remains some ambiguity in our defined gene list: for example, while USP7 interacts with PRC1 variants, it may not strictly be considered a PcG protein [24]. Several genes have additional roles beyond the PcG-TrxG system, for example, KDM2B functions as a demethylase at several histone sites, as well as mediating ubiquitination [25]. There are also limitations arising from the classification of variants: this study relies on clinical reporting, and variants that may only partially explain phenotype were included. Additionally, this study did not take into account variant consequences; gain-of-function and loss-of-function variants may have different phenotypic consequences [26, 27].

Despite these limitations, a question arising from the current study is whether phenotypic clusters observed within the PcG/TrxG cohort, at gene level or at individual level, are matched to distinct molecular mechanisms. Firstly, we can ask whether gene-level clustering aligns with PcG/TrxG complex membership. We observed that several conditions associated with the BAF complex have high phenotypic similarity: for example, conditions associated with SMARCA2, SMARCB1, ARID1A, ARID1B and DPF2 share limb, digit, facial, hair and brain morphology abnormalities. However, these BAF complex genes are distributed across clusters 2 and 3, indicating that there are within-complex divergences as well as convergences. Moreover, BAF-associated conditions are phenotypically similar to COMPASS-associated conditions (including those associated with KMT2A, KMT2D and SETD1B), indicating cross-complex convergence. Alternatively, phenotype clusters could mirror downstream molecular mechanisms, an obvious possibility being genome-wide methylation status. DNA methylation (DNAm) profiles associated with PcG/TrxG complex subunits have been extensively investigated. For example, variants in PRC2 members EED, SUZ12 and EZH2 share a similar DNAm signature, and a similar clinical spectrum characterised by overgrowth and intellectual disability [28]. Similarly, there is a close resemblance between the DNAm signatures associated with pathogenic variants in ASXL2 and ASXL1 causing Shashi-Pena and Bohring-Optiz syndromes respectively [29], and between KDM6A and KMT2D variants causing Kabuki syndrome [30]. Ability to detect these convergences in the current dataset may be limited by low numbers for some genes (especially ASXL2, SUZ12 and EZH2, <5 patients per gene), and lack of phenotypic specificity in HPO reporting. However, our results indicate that future investigations integrating DNAm with phenotype may benefit from data-driven approaches.

An additional motivation for elucidating convergent clinical characteristics is that emerging treatment strategies may apply to a larger cohort beyond single genes. Ciptasari and van Bokhoven suggest that transcriptomic convergences across different MDEMs could be harnessed to develop effective interventions [15]. Drugs targeting epigenetic mechanisms or downstream gene expression have the potential to be efficacious in treating MDEMs [6]. HDAC inhibitors normalised H3K4 methylation and neurogenesis in a mouse model of Kabuki Syndrome [31]. CRISPR-based methods may also be applicable to MDEMs [32]. However, a major challenge is that therapies need to be administered at an appropriate developmental stage. Epigenetic interventions and CRISPR-based therapies risk introducing off-target changes [6]. The current study indicates that patient-relevant symptom domains are highly variable, hence target outcomes and therapeutic strategies will need to weigh-up potential benefits and risks on an individual basis.

The data-driven approach of this paper lays the foundation for future work disentangling MDEM-phenotype relationships at scale. Future phenotyping research and clinical initiatives should move beyond congenital characteristics to encompass health-related outcomes across the lifespan.