Genes with high penetrance for syndromic and non-syndromic autism typically function within the nucleus and regulate gene expression

A comprehensive list of forms of ID were accessed from the Mendelian Inheritance in Man (MIM) database [15]. Only conditions with known molecular basis were collected. Keywords for initial search comprised “intellectual disability”, “mental retardation”, “mentally retarded”, “global developmental delay”, “severe developmental delay”, and “profound developmental delay” (for a full listing of OMIM numbers, gene/locus numbers, and associated data, see Additional file 1). Conditions were then curated and removed if: (1) the ID was not a primary feature but was variably expressed; (2) the ID displayed onset later than three years of age; (3) the condition tended to be lethal in infancy or early childhood; (4) the condition had a known yet complex genetic etiology, e.g., large recombination events that include two or more genes (with the exception of chromosome 2p16.3 deletion syndrome which has been directly linked to NRXN1 mutations); (5) autism was a defining symptom for diagnosis, as in the cases of certain “susceptibility” genes; (6) only one or two cases were noted in the literature; (7) mutations occurred in only a single family; (8) the condition was a chromosome instability syndrome, leading to variable features due to the accumulation of different mutations; or (9) a condition contained an unconfirmed or potentially spurious mapping as indicated by a “?” before the disease name. This led to a final list of 465 different forms of ID and 434 unique genes (some genes whose functions are unknown were also removed from analyses although are still included in the main list. In addition, a small selection of genes was not recognized by gene ontology and therefore was not included within those analyses. Therefore, gene lists sometimes varied minimally between analyses).

Associated genes were assigned to one of five categories according to the information available regarding their comorbidities with autism and epilepsy. This information was initially derived from the MIM database and was subsequently confirmed through thorough literature review. Genes were additionally cross-referenced with Pinto et al. [16], whose supplemental material includes a large list of conditions associated with autism and ID. The categories ultimately were: (1) ID with highly comorbid autism (HCA) (N = 72 conditions, 71 genes), (2) ID with variable autism (VarAut) (N = 139 conditions, 124 genes), (3) ID with highly comorbid epilepsy (HCE) (N = 88 conditions, 86 genes), (4) ID with variable epilepsy (VarEp) (N = 84 conditions, 78 genes), and (5) ID without autism or epilepsy (ID only) (N = 82 conditions, 75 genes) (Fig. 1a).

If a gene was redundant across two or more IDs, those with occurrences of autism were preferentially given preference and placed within HCA or VarAut, followed by epilepsy (HCE, VarEp), and so forth. Conditions were assigned the above comorbidity groups if: (1) there was even minor evidence of overlapping comorbidity reported on MIM or in the broader literature; (2) if, in the case of autism, the associated gene was included under “syndromic”, “high confidence”, “strong candidate”, or “suggestive evidence” headings in the SFARI Gene database or rated 9+ within the AutismKB database; (3) there were no reported instances of epilepsy/seizures but abnormal epileptiform activity had been noted; or (4) the gene was included in the epilepsy gene database, CarpeDB [17‐19]. In instances in which a single MIM condition had multiple associated genes, all genes were placed in the same category provided any of those were not already in a superseding category.

For the purposes of dividing all autism-associated IDs into either the HCA or VarAut groups, any condition that had ≥20 % autism comorbidity rating according to intensive literature review was placed within the HCA category; meanwhile, all other IDs with at least two unrelated case studies involving autism or autistic symptomology were placed within VarAut (see Additional file 1 and Additional file 2 for comorbidity references). For the purposes of studying genetic penetrance, HCA and HCE conditions were annotated according to their inheritance patterns (dominant, recessive) and broken into their respective pattern groups for further analysis.

In order to assess associated gene product functions, we analyzed differences in sample frequency for gene ontology (GO) terms according to group across categories of biological process, cellular component, and molecular function [20]. A listing of significantly enriched terms were initially accessed for each gene group, followed by the removal of redundant parent terms and those unusable for direct annotation. Absolute frequencies were then compared for each of these significant terms across all groups. The prop.test() function in the statistical computing software, R, was used for most statistical analyses. All pairs of proportions were compared using a Chi-square test of two proportions with one degree of freedom. A false discovery rate adjustment was applied to account for multiple comparisons.

For more intensive examination, UniProt/Swiss-Prot and Entrez Gene summaries of gene products’ molecular functions were used as the basis to determine whether a given gene was considered a nuclear epigenetic regulator (transcription factor/repressor, methylator, ubiquitinase, chromatin remodeler, etc.) [21, 22]. The same statistical analyses as used in the main GO experiment were also used here. For the purposes of studying rates of nuclear epigenetic regulators across dominant x recessive subgroups, a proportions comparison was used, without need for correction for multiple comparisons.

In addition, because a large minority of the HCA genes are not currently included or rated within the SFARI Gene Database, we compared the SFARI-only HCA group to the HCA non-SFARI group in terms of number of nuclear epigenetic regulators to ensure that no significant differences existed. This was performed using a two-tailed proportions comparison.

Finally, a thorough literature search was performed to determine which conditions within HCE and VarEp were considered neurodegenerative. HCE and VarEp rates of neurodegeneration were compared using a two-tailed proportions comparison.

For the protein-protein interaction experiment, each gene group was loaded individually into String 10 alongside a selection of proteins representative of the core PPI network (WNT = CTNNB1, SHH = PTCHD1, NCOR = NCOR1, SWI/SNF = SMARCA1, NOTCH = NOTCH1, ERK1/2 = FGF8, TGF-β/BMP = SMAD4), and confidence data were analyzed [23]. Both the percentage of experimental genes connected with the core PPI and the number of intermediary nodes that lay between the genes of interest and their nearest core PPI neighbors were assessed according to group. For the former analysis, a proportions test with correction for multiple comparisons was used; meanwhile, for the latter, an ANOVA was utilized.

We studied the comorbidities of a substantial list of IDs with known molecular origins derived from the MIM database. Of the 465 forms of IDs collected, we found that 45 % (N = 211) were comorbid with autism in at least a minority of cases. Meanwhile, 15 % (N = 72) were highly comorbid with autism, co-occurring in ≥20 % of reported cases. Some of these include conditions well known for autism association, such as FXS and TSC, however, also included conditions less well known, such as non-photosensitive trichothiodystrophy and Mowat-Wilson Syndrome. Meanwhile, 44 % (N = 204) of IDs were highly comorbid with epilepsy, while an additional 32 % (N = 151) exhibited variable rates of epilepsy. In fact, 55 % (N = 116) of autism-related conditions listed epilepsy as a common feature, which was consistent across both HCA and VarAut groups, reinforcing ideas of their shared etiologies (Table 1).

We went on to study the genetic etiologies of our conditions of interest, investigating GO term associations in the areas of biological process, molecular function, and cellular component (see Table 2 for gene list by category). Both autism groups, HCA and VarAut, exhibited enrichment in nervous system development compared to HCE and ID only (p = 0.0035–0.052, see Additional file 3 for full statistical results). However, despite differences in this overarching parent term, our comorbidity groups did not differ significantly in the child terms neurogenesis (p = 0.1218–0.8594), neuron differentiation (p = 0.3118–1.00), neuron projection development (p = 0.2608–1.00), and synaptic transmission (p = 0.5075–0.7988). HCA was mildly enriched in the regulation of synaptic structure or activity compared to HCE (p = 0.0477) and ID only (p = 0.0453), but not compared to the variable groups. In summary, both autism groups exhibit stronger nervous system enrichment than either HCE or ID only, suggesting that gene product involvement in nervous system development may characterize a significant subset of autism risk genes (Fig. 1b).

HCA was also particularly enriched in anatomical structure development compared to all groups (p = 0.0021–0.0419) except VarAut (p = 0.2084), indicating the genes’ probable roles in structural morphogenesis (Fig. 1b). However, above all else, HCA was typified by regulation of gene expression (p = 0.000–0.0404) and was involved in regulation of DNA-templated transcription (p = 0.000–0.0406) and chromatin binding (p = 0.0018–0.0326). Matching its functional enrichment, HCA was strongly enriched within the nucleus (p = 0.0002–0.0009) in contrast to all other groups and was also modestly enriched at the chromosome compared to HCE (p = 0.0392) and ID only (p = 0.0416) (Fig. 1c).

UniProt/Swiss-Prot and Entrez Gene analysis further revealed that HCA gene expression regulation was largely carried out through nuclear epigenetic means, such as transcription factors and repressors, methylation regulators, ubiquitin ligases, and other chromatin remodelers, which comprised over half of that gene group, a substantial increase compared to all other comorbidity groups (p = 0.000–0.0004) (Fig. 2a). In addition, 45 % of the genes in HCA are not currently included or rated within the SFARI database, yet even with their removal, the SFARI-only HCA group did not differ from those not included within the database in terms of their functional enrichment (p = 0.4501, Z = 0.8, Diff = −0.1436, 0.3236).

VarAut exhibited similar though more modest trends in functional enrichment as seen within HCA, such as regulation of gene expression (p = 0.0101) and regulation of DNA-templated transcription (p = 0.0034), although this was only apparent compared to HCE, the latter which tended to house particularly low enrichment in all of these terms.

HCE, VarEp, and ID Only did not show consistent differences in enrichments in biological processes, with the exception of HCE functional enrichment in lipid metabolic processes compared to all groups (p = 0.0092–0.0486) except VarEp (p = 0.1135). Compartmental enrichments for the non-autism groups were also minimal, with the exception of ID only enrichment within the Golgi membrane compared to all groups (p = 0.0298) except VarEp (p = 0.3132).

For more in-depth analysis, the HCA and HCE comorbidity groups were each divided in two according to their patterns of inheritance (dominant vs. recessive) and were compared against one another for significant GO term enrichments. Significant functional enrichments differentiated both sets of dominant and recessive groups. HCA dominant genes, for instance, were comparatively more enriched than HCA recessive genes in anatomical structure development (p = 0.0340), nervous system development (p = 0.0126), cell differentiation (p = 0.0395), regulation of gene expression (p = 0.0033), regulation of DNA-templated transcription (p = 0.0048), and chromosome organization (p = 0.0232) (Fig. 2b). This suggests that many of the significant GO terms associated with the larger HCA group are primarily driven by this dominantly inherited subgroup.

Meanwhile, the HCE recessive gene group neared significant enrichment in the term lipid metabolic process compared to HCE recessive (p = 0.0561). In addition, though they likewise did not reach significance, the recessive group was also comparatively enriched in the endoplasmic reticulum as well as the endoplasmic reticulum membrane (p = 0.0913). This suggests that disturbances to protein trafficking through the cell, particularly the endoplasmic reticulum, may be a risk factor for recessive forms of epilepsy.

Dominant HCE genes, in contrast, were enriched in terms related to structural constituent of cytoskeleton (p = 0.0018), transmembrane transporter complex (p = 0.0003), potassium ion transmembrane transporter (p = 0.0018), protein complex (p < 0.0001), and myelin sheath (p = 0.007) (Fig. 2c). The functional significance of these associations is not currently well understood.

Upon further study of our gene groups, we find that neither of our autism groups differ significantly from one another in their connectivity to the core PPI network, either in the number of proteins that connect to the core PPI (p = 0.1053) nor in the number of intermediary nodes that lay between our proteins and their nearest core PPI neighbor (network “tightness”) (p = 0.6098) (Fig. 3c, d). In addition, HCA does not differ from HCE in terms of the number of proteins that connect with the network (p = 0.9151), but they do vary according to the tightness of the networks surrounding the core (p < 0.0001). Meanwhile, HCA and VarAut exhibit a larger core network than both VarEp (p = 0.0001–0.0011) and ID only (p = 0.0001–0.0003), but the level of network tightness does not differ significantly. Overall, our autism groups exhibit a larger, tighter protein network surrounding the core PPI compared to all other groups.

The core PPI network is a single extensive protein network integral for patterning of the central nervous system (e.g., dorsoventral patterning) as well as later processes of neural maturation and ongoing plasticity. This network includes morphogens such as Wingless Integration Site (WNT), NOTCH, nuclear receptor corepressor (NCOR), SWItch/Sucrose Non-Fermentable (SWI/SNF), Sonic Hedgehog (SHH), transforming growth factor-β (TGF-β), bone morphogenetic proteins (BMP), and extracellular signal-related kinase 1 and 2 (ERK1/2)—a functional module with considerable overlap with recent reports by Hormozdiari et al. [24] pinpointing a particularly enriched protein network in autism that centers around WNT, NOTCH, SWI/SNF, and NCOR [25].

Why the VarEp group diverged from HCE somewhat in the String analysis, as well as various GO term enrichments, is currently unknown. However, the HCE group (N = 88) contained a significantly greater number of neurodegenerative conditions compared to VarEp (N = 84) (z = 2.5, p = 0.0119, Diff = 0.0375, 0.3025), suggesting that the prevalence of neurodegeneration could underlie divergence in their respective etiologies.

Despite that the rare conditions reported here within the HCA group display very high rates of autism comorbidity, genes associated with close to half of these conditions are not currently included within the SFARI Gene Database, are included but unscored, or are scored as “6” indicating that the “evidence does not support a role in [autism spectrum disorders]” (Table 1) [19]. In addition, 69 % of HCA genes that are scored in SFARI but not considered “syndromic” do in fact present with multiple physical dysmorphia, such as complex craniofacial malformations, indicating that with further investigation, these genes should and are likely to be subsumed under this umbrella category.

Some of the non-SFARI conditions that are used in this study are based on small numbers of patients and likely require further in depth diagnostics to confirm the presence of autism symptoms in these conditions at a higher-than-expected rate. In addition, a number of the studies is also hampered by poor diagnostic methodology, an issue that has plagued many of the earlier genetics studies, whereas today’s gold standards include the use of the ADOS and ADI-R as well as ID comparison groups for the purposes of research. Most tellingly, however, the removal of these conditions from the HCA group did not change our results; genes with high penetrance for syndromic and nonsyndromic autism are typically localized to the nucleus and are involved in transcription regulation. These results in particular stress the importance of grouping risk genes according to penetrance when possible, because this information was already present within the SFARI database, though is currently lost amidst the other genetic data.

Because genes associated with HCA are overrepresented within the nucleus and tend to directly regulate transcription, this suggests that mutation penetrance for autism may be strongly linked with regulatory, as opposed to enzymatic, transduction, and structural, cellular networks [26]. Transcription factors and regulators are the most common examples within this group; however, other epigenetic regulators, such as heterochromatin remodelers, ubiquitin ligases, and methylation regulators, were also overrepresented. It may be for their phenotypic penetrance that dominantly inherited conditions were so common in the HCA comorbidity group.

In addition, the presence of gene subgroups within HCA that share considerable overlap with a module for autism risk reported by Hormozdiari et al. [24] suggests that a significant portion of these cases, as well as those in VarAut, are rooted in disturbances to patterning of the CNS and ongoing deviations in neural maturation and plasticity. All of these morphogenetic pathways share considerable crosstalk that is foundational to dorsoventral and rostrocaudal patterning, planar cell polarity, locomotion, neuritogenesis, and finally synaptogenesis and plasticity [25, 27‐29]. If disturbed, they are likely to affect all stages of neuronal development, spanning from the most foundational to the most nuanced. Further work at the cellular and tissue levels will be required to investigate whether disturbances to patterning may play roles in these conditions and how such patterning defects, alongside later impairments to neuronal development and plasticity, underlie the behavioral and neurological phenotypes.

In support of this, previous work investigating high-risk autism-related genes has suggested that disturbances to neural maturation may be a common theme to autism [30]. Our present results indicate that epigenetic dysregulation could inappropriately suppress or prematurely promote the expression of gene products, leading to chronological changes in the typical developmental process and ultimately to gross structural, microstructural, and physiological perturbations.

For example, the fragile X mental retardation protein (FMRP) associated with FXS normally aids in suppression of translation, thereby controlling timing of neural differentiation. Instead, when the FMR1 gene is mutated leading to decreased production of FMRP, neurogenesis occurs prematurely [31]. These neurons also fail to express mature markers in a timely fashion (e.g., GAD67), a disparity likely resulting in poor maturation and circuit integration of adult neurons, and absolute or relative macrocephaly, periventricular heterotopias, and volumetric increase in periventricular white matter, further evidence of a pathological heterochrony and disturbances to patterning [6, 7, 32].

In contrast to our autism, variable epilepsy, and ID only groups, ID with highly comorbid epilepsy exhibited particularly low enrichment in nervous system-specific processes, was more often involved in lipid metabolism, and, compared to conditions with variable epilepsy, had higher rates of neurodegeneration. While this evidence is tantalizing, further work is needed to determine whether ID and epilepsy related to neurodegenerative processes follow a different etiological path than those related to general nervous system development as may be seen in ID with autism or variable epilepsies.

In contrast to the other groups, it was clear that the HCA group is surprisingly homogeneous, suggesting that risk for autism lies within a specific and very definable set of molecular events that confer greater risk the further downstream these elements are affected, i.e., at the level of the gene and its product. This likewise suggests that the further upstream a particular risk factor or environmental effector, the more variable the penetrance for autism due to the number of elements that may intercede and alter events, e.g., feedback inhibition. This is strongly suggested by the divergent compartmental enrichments seen in HCA vs. VarAut, in which the former is highly enriched for the nucleus while the latter is mildly enriched throughout numerous cellular compartments and within cell projections in particular. Ultimately, risk is a threshold effect and a risk factor must be closely upstream of its target (e.g., in the case of epigenetic regulators) or, if further upstream, then it must be capable of avoiding feedback inhibition in order to reach threshold in a consistent highly penetrant fashion (e.g., in the case of select sodium channel mutations).

On a similar note, factors that are comparatively less penetrant yet still confer measurable risk suggest the presence of additional variables, e.g., polygenic effects, environmental agencies, etc., in the determination of their etiologies. Given the nature of genetic selection, common gene variants that provide variable autism risk (i.e., common disease-common variant) are more likely to explain a wider breadth of cases than the rare, often de novo, mutations that confer higher penetrance for the phenotype. Although interestingly, a recent study by Alvarez-Mora et al. [33] reported that in a subset of high-functioning cases they studied, over 50 % (6/11) of the identified rare potentially deleterious single nucleotide variants (SNV) occurred within the HCA genes reported here, suggesting that these genes may be targets with variable penetrance dependent upon the specific type of mutation. Sanders et al. [34] have found that highly penetrant deleterious SNVs tend to affect the same genes that are also targets of small copy number variants (CNV) in autism, such as occurs in the monogenic conditions studied here. Meanwhile, individual genes that comprise larger CNVs each confer comparatively lesser risk. It is possible that if rare SNVs tend to overlap HCA genes, less penetrant SNVs (e.g., common variants) may overlap genes typically comprising larger CNVs and reflect polygenic risk.

In the future, we may find that the genetics of autism tends to diverge according to levels of severity, with rare mutations (e.g., small and large CNVs, highly deleterious rare SNVs) responsible for a significant portion of low-functioning individuals with intellectual disability while other rare SNVs and common variants, perhaps even polygenic and/or environmentally driven, play important roles in a larger portion of the moderate-to-high-functioning ranges. This hypothesis is not entirely unlike that proposed by Folstein [35] in which she suggests that autistic individuals with profound ID, complex dysmorphic features, or specific genetic conditions represent phenotypes that are clinically unique compared to the idiopathic autism reported by Kanner. In this case, however, we are suggesting that the genetics, though not necessarily the overall biology, diverges between the two.

Aside from the limitations mentioned above concerning questions of the diagnostic reliability of some of the HCA conditions, additional shortcomings of this study involve the availability of information regarding what are typically rare conditions and potential underreporting regarding comorbidities such as autism and epilepsy. There are, for instance, a number of VarAut conditions in which case studies or small group studies reporting autism incidence are available but no larger studies have been performed in order to provide better estimates of co-occurrence. Examples include conditions such as Succinic Semialdehyde Dehydrogenase Deficiency (#271980), Autosomal Dominant Mental Retardation 21 (#615502), and Dihydropyrimidine Dehydrogenase Deficiency (#274270) to name just a few that are likely worthy of more intensive study in relation to autism. Therefore, it is highly likely that some of the conditions presented within this study have been mis-categorized. In order to limit that occurrence to an absolute minimum, various genetic databases were used in conjunction with phenotypic data.

In addition, the fact that we limited our study of autism and epilepsy to forms of ID subsequently limits the potential scope of applicability of our results, although in doing so we were able to estimate comorbidity rates. We therefore hope that future research may elucidate which of the results presented here are applicable to the broader autism spectrum or whether these data solely define a subgroup of autism.

Because genetic mutations are infrequently identical across different individuals with a single form of ID, it is possible that some cases of autism or epilepsy within our variable groups were not due to mutations involving the primary gene associated with the monogenic condition but were instead due to confounding effects of other genes, such as may be seen in larger chromosomal rearrangements. However, most of the results presented here exhibit strong functional patterns and therefore while individual IDs may ultimately be mis-categorized, we are confident that the conclusions regarding the larger groups are relatively sound.