Prioritization of neurodevelopmental disease genes by discovery of new mutations

1) Recurrently mutated genes

One of the frequently used concepts in considering possible ‘new disease genes’ responsible for a given neurodevelopmental phenotype is the recurrence of de novo mutations in the same gene as well as the absence of such mutations in healthy controls. This rule follows the precedent established for the discovery of pathogenic de novo copy number variants (CNVs) during the last decade with the highest priority given to recurrent mutations that lead to a complete loss of function of one of the parental copies of the gene. Up to ten independent reports of de novo mutations in SCN2A and nine independent reports of de novo mutations in SCN1A and STXBP1 have been described (Tables 2 and 3). Strikingly de novo mutations in those genes are found, to date, exclusively in probands but never in controls. Simulation data suggest that at least two but certainly three or more recurrent de novo loss-of-function (LoF) events are unlikely to occur by chance, making such genes outstanding candidates (Table 2)⁷.

The frequency of recurrence is dependent on the extent of locus heterogeneity associated with each disease. Leveraging observed recurrences of de novo mutations (Fig. 1), diseases such simplex autism and SCZ are thought to arise from >500 genes, while studies of severe ID and epileptic encephalopathies (infantile spasms and Lennox-Gastaux subtypes) suggest lower heterogeneity. Such estimates should be considered only rough approximations at this point since they are highly dependent upon the fraction of de novo mutations that are in fact pathogenic as well as ascertainment biases in sample collection (Fig. 1). In this regard it is interesting that several of these top-scoring candidate genes have been observed in patients with epilepsy¹¹, ID^1,2, and ASD^3-7. This may not be surprising in light of the comorbidity of these diseases and if one accepts that the level of heterogeneity is lower for epilepsy-related disorders than for ASD, SCZ, or ID. In such a scenario, sequencing of even a modest number of epilepsy cases delivers recurrent mutations more frequently than more broadly defined DD (developmental delay) or ASD^11,18. One study of SCZ, for example, highlighted four recurrently mutated genes, but maybe more remarkable the same study identified overlapping genes with ASD when focusing on prenatally expressed genes⁹. The stronger overlap between ASD, ID and epilepsy and yet limited overlap with SCZ (Table 2) could be largely because only a subset of the disease stems from a neurodevelopmental origin.

Perhaps, the most striking examples are recurrent identical de novo mutations within the same gene. Across the various studies, such identical recurrences have already been observed for six genes (ALG13, KCNQ3, SCN1A, CUX2, DUSP15 and SCN2A; Table 4). Such events are exceedingly unlikely with estimates of identical recurrences in ALG13 and SCN2A calculated at p = 7.77 × 10⁻¹² and p = 1.14 × 10⁻⁹, respectively, in the case of epilepsy¹¹. Most of these estimates of significance, however, assume a random mutation process. Yet mutational hotspots certainly exist and recurrence of the same mutation cannot be taken as proof positive of an association.

The significance of the majority of de novo mutations remains unclear. For most genes, only a single de novo mutation has been identified. Nevertheless, based on the observation that de novo LoF mutations occur 2–3 times more frequently in ASD probands when compared to unaffected siblings, it is now estimated that a large fraction of these singletons will be relevant to disease etiology. When considering all studies in aggregate, de novo LoF mutations are observed significantly more in cases than in controls (Table 1; Fisher’s exact test: p-value = 0.0062). The interpretation of recurrent missense mutations, however, represents a greater challenge. Sanders et al.⁴ estimated that four missense de novo mutations in the same genes would be required in simplex autism to exceed a chance finding; this was based on a cohort size of up to 2,000 with an estimated locus heterogeneity of 1,000 ASD risk loci. Given the extreme locus heterogeneity of diseases such as ASD and ID, other strategies have been adopted to prioritize likely causal genes. High-throughput targeted multiplexed resequencing technologies, such as molecular inversion probes, have been employed to screen ~50 candidate genes in thousands of patients and controls^3,18. Such approaches are scalable, inexpensive (<$1 per gene per sample), sensitive, and specific increasing by an order of magnitude the number of patients that can be screened. The strategy was particularly useful in discovering a burden of de novo LoF mutation of CHD8 associated with ASD²⁰. The relatively ease in detecting de novo mutations allows rapid identification of potential candidate genes; the number of cases that are required to make these findings statistically significant (Fig. 1) can be lower for the genes that are mutated exclusively in a large number of patients when compared to standard case-control studies²¹.

2) Prior evidence of overlap with pathogenic CNVs

Another strategy has been to compare patterns of CNVs in patient and control populations to prioritize genes (Fig. 2)²². Extensive CNV morbidity maps have been developed for tens of thousands of children with autism, ID and epilepsy helping to define pathogenic regions of dosage imbalance in the human genome^23-26. Overlapping deletions in such collections occasionally refine the smallest region of overlap highlighting a modest number of candidate genes. Recurrent de novo point mutations in a gene within such a region with CNV burden significantly increases the likelihood that LoF of the gene is responsible for a phenotype. O’Roak et al.³ and Rauch et al.² each discovered, for example, LoF point mutations for SETBP1—a gene where a significant enrichment for deletion CNVs has been seen in patients with overlapping neurodevelopmental phenotypes but not in controls. Similar patterns have recently been observed for DYRK1A and MBD5 (Fig. 2), including reports of balanced but gene-disrupting chromosomal translocations²⁷. Such information has been used to compute haploinsufficiency scores^2,28 to strengthen the case of ’causality’ of de novo LoF mutations (Fig. 2).

3) Position of the mutation within the protein

The bulk of de novo mutations discovered from exome sequencing projects are missense mutations, with more than 60% in probands and controls (Table 1). Distinguishing pathogenic signal from the background of benign mutations is an active area of research. For genes for which previous CNVs or LoF mutations were described, ‘severe’ missense mutations are also likely to result in a dosage effect. However, there are examples, usually from clinically well-defined neurodevelopmental syndromes, showing that missense mutations result in different outcomes either based on the protein domain they affect, the position in the gene (e.g., the amino- or carboxy-terminus of the resulting protein)²⁹, or their potential to modify the normal function of the protein. Examples for the latter include gain-of-function (GoF) mutations and LoF mutations in the same gene that result in different phenotypic outcomes^30,31. SETBP1 GoF in a degron sequence (ubiquitination motif) results in the rare but well-defined Schinzel-Giedion syndrome³² while deletions or LoF mutations may result in a distinct and milder phenotype comprising ASD/ID with speech delay and other features^2,3,33,34. Other examples include different phenotypic effects dependent on the location of the mutation; early truncating and missense mutations in NOTCH2 are known to cause Alagille syndrome³⁵ while truncating events restricted to the last exon escape nonsense mediated decay (NMD) and result in Hajdu-Cheney syndrome^36,37.

4) Mutational burden among healthy individuals

One approach to prioritizing missense mutations leverages evolutionary conservation by assigning ‘mutability scores’ per gene or even at the base pair level. O’Roak et al.³, for example, established an evolutionary mutation score per human gene based on the human-chimpanzee divergence and the size of a gene. Similarly the Epi4k Consortium used a gene-specific mutation rate based on a per-base score³⁸; this score was, however, not based on human-chimpanzee evolution but made use of human-specific polymorphism data. A recent publication used rare variant data from healthy individuals and offers a new integrative annotation tool for noncoding variants³⁹. The wealth of available control exome sequence data can also be used to estimate the (rare) variant load per gene (and distribution). For example, the analysis of data generated from sequencing 6,500 ‘control’ exomes as part of the ESP6500 has been used to define the load of LoF mutations per gene² and to prioritize >4,000 human genes that are most intolerant to variation¹¹. Another approach uses random mutation modeling⁴⁰ to calculate the likelihood that observed (de novo) mutations have a damaging effect; similar prioritizations are provided by tools that score individual mutation severity (SIFT, PolyPhen2, MutationTaster, MutPred, CONDEL, etc.), some of which can be adapted to a gene-based prioritization score from genome-wide data⁴¹. These population data provide a powerful unbiased approach to hone in on genes that are likely to be among the most penetrant because of the complete absence of disruptive variation in the general population (e.g., CHD8 or DYRK1A). A critical aspect of such analyses is the reliability of a particular gene model. Most human genes show evidence of alternative splice forms—many of which have no known function. Apparent hotspots of mutation for a particular exon (often 5′ or 3′) in both cases and controls may suggest mis-annotation, the presence of a processed pseudogene, or an alternative nonfunctional splice form.

5) Pathway enrichment and links to cancer biology

Another popular approach to suss out the most important gene candidates for further characterization has been to identify specific biological networks of genes enriched in cases as compared to controls. Although this approach cannot be used unequivocally to define causality, membership of a specific gene in a particular protein-protein interaction (PPI) or co-expression network may increase the likelihood of its association with disease. Numerous studies have reported significant enrichment of both de novo CNV and SNV mutations in particular pathways^3,4,42,43. O’Roak et al., for example, reported a significant enrichment of de novo disruptive autism mutations among proteins associated with chromatin remodeling, beta-catenin and WNT signaling—a finding that was replicated in a follow-up resequencing study of more than 2,400 probands. One instance, in which membership of a new candidate gene within a PPI network led to the discovery of an autism gene, is the recent example of ADNP. A single ADNP LoF mutation was initially observed from exome sequencing studies. Although the gene did not reach statistical significance when comparing cases and controls²⁰, it was strongly implicated in the PPI network originally defined by O’Roak et al. Targeted resequencing experiments combined with clinical exome sequencing identified several additional cases with de novo mutations and remarkably similar phenotypes representing a new SWI-SNF-related autism syndrome (Fig. 3)⁴⁴. Notably, many of the genes implicated in the beta-catenin pathway have also been described as mutated in patients with ID¹ but not in patients with SCZ. Similarly, an enrichment of genes interacting with FMRP—the gene responsible for Fragile X syndrome—has been reported with de novo mutations in ASD⁵, epilepsy¹¹ and, most recently, SCZ^10,45. Whether this observation is due to the relative high degree of cases that also presented with comorbid ID remains to be determined.

In addition to PPI networks, studies of co-expression have shown enrichment for specific spatio-temporal patterns of expression. A study of co-expressed genes affected by de novo mutations reported an enrichment in fetal prefrontal cortical network in SCZ⁸, which is in line with the finding by Xu et al.⁹ that genes with higher expression in early fetal life have significant contribution to SCZ by de novo mutations. Similarly, Willsey et al.⁴⁶ working with a few high-confidence sets of ‘ASD genes’ as seeds reported a convergence of deep layer cortical projection neurons (layers 5 and 6) in mid-fetal development. Another analysis using a larger set of ASD and ID risk genes suggested translational regulation by FMRP and an enrichment in superficial cortical layers⁴³. Implicit in these types of analyses is the notion that while more than 1,000 genes may be responsible for ASD or ID, in the end the genes will converge on a few highly enriched networks of related genes. It is possible that molecular therapies targeted to the network at a specific stage of development as opposed to the individual gene may be beneficial to specific groups of patients.

Related to this, it is intriguing that several recurring genes and pathways that have been implicated in neurodevelopmental disease have also been associated with different forms of cancer (Fig. 4)⁴⁷. While clear-cut examples like the mutation of tumor suppressor genes, PTEN (Cowden syndrome) or ARID1B (Coffin-Siris syndrome), and neurodevelopmental disease have been extensively reviewed⁴⁸, more recent exome sequencing data from patients with neurodevelopmental disease suggests potentially new links. The most striking observation here is the identical point mutations reported to cause cancer when mutated somatically and (severe) neurodevelopmental syndromes when mutated in the germline. Examples include the identical mutations in SETBP1³², ASXL1⁴⁹, and EZH2⁵⁰, as well as several genes of the RAS-MAP-kinase pathway associated with parental-age effect Mendelian disorders⁵¹ (Table 5). It is important to stress that this is an observation at an individual gene level and should not be translated to an epidemiological link, i.e., this cannot be generalized to speculate that patients with neurodevelopmental disorders in these specific genes will all be at a higher risk for certain cancer types. Instead, it is likely that this convergence represents a selection of genes that play a fundamental role in cell biology (e.g., cell proliferation and/or membership in multi-subunit complexes associated with chromatin remodeling). There is also the distinct possibility of pleiotropy; i.e., genes and pathways have completely unrelated functions explaining developmental defects and cancer independently. Therefore, de novo mutations in those genes can result in different outcomes depending on timing, genetic background, and cellular context. Nevertheless, there may be advantages to integrating sequence data from patients with neurodevelopmental disease and massive sequencing programs devoted to the discovery of somatic mutations within tumors, e.g., the International Cancer Genome Project⁵². It is possible that these intersections will help to further prioritize genes important in both cellular development and neurodevelopment.

Phenotypic similarity of recurrent de novo mutations

Statistical support of recurrent mutations is not the sole arbiter in determining pathogenicity of particular mutations and genes. In particular, it is important to consider the phenotypic presentation and overlap of the individuals with the same presumptive underlying genetic lesion. In this regard, we note that many of the initial studies are likely enriching for the most severe cases since ascertainment is clinical as opposed to population based. As a result, initial estimates of penetrance may be overestimated and phenotypic heterogeneity underestimated. Nevertheless, identification of clinically recognizable syndromes or sets of phenotypic features has historically been used to strengthen the case of a particular gene’s involvement. In the past, gene discovery was usually driven by detailed description of a particular syndrome (e.g., Fragile X or Rett syndrome) followed by a systematic hunt for the mutated gene. Recognition of clinical subtypes, however, is now beginning to occur after mutation and gene discovery. In the case of PACS1⁵³, for example, identical de novo mutations result in patients with a strikingly similar phenotypic outcome (Fig. 3). Such clinical discernment now, more than ever, requires the expertise of the clinician.

It should be noted, however, that not all genes when mutated will show a phenotypic convergence but rather may be much more variable in their phenotypic presentation. For example, mutations in ARID1B can either lead to isolated ID⁵⁴ or syndromic forms of ID with a recognizable phenotype known as Coffin-Siris syndrome^48,55. The type of mutation may be critical in this regard. It is noteworthy that patients with LoF mutations of SCN2A described in autism cohorts⁴ do not present with epilepsy in contrast to multiple recurrent missense mutations identified among epileptic encephalopathies, or more specifically infantile spasms or Lennox-Gastault subtype. Similarly, de novo missense mutations in CHD2, SETD5, and SLC6A1 have been reported in patients with ASD yet frameshift mutations in the same genes are seen in patients with ID but without ASD features². There are several considerations regarding genotype-phenotype correlations.

1. Some of the classically defined neurodevelopmental clinically defined syndromes may present with broader (or milder) phenotypes as defined by initial clinical case reports². There is evidence that the genetic background upon which these mutations occur significantly influences phenotypic outcome^56-58.

2. ‘Genotype-first’ approaches using current genomic technologies followed by ‘reverse phenotyping’ are beginning to define more subtle syndromes that are still opaque within large umbrella cohorts such as ASD or ID¹³. Some examples include macrocephalic subtypes of ASD/ID caused by mutations in PTEN and CHD8^20,59 and DD/ID and epilepsy caused by de novo mutations in SCN2A^1,2,18.

3. After discovery of potential causative mutations, more detailed and standardized phenotyping assessments are necessary to eliminate disease ascertainment biases. Since patients with a specific mutation will be individually rare, greater coordination, including patient recontact, will need to occur across clinical research centers.

Detailed phenotypic characterization of patients is an important first step in modeling mutations in other organisms. Indeed, additional support for a gene’s involvement in disease often is provided by related pathologies in these model organisms and may be used to rapidly prioritize genes for further study as well as to provide additional insight into function. In many cases, mouse models⁶⁰ or Drosophila mutant lines⁶¹ already exist and neurologic phenotypes have been, at least, partially documented. For example, recurrent LoF mutations of ADNP (activity dependent neuroprotective peptide) were recently described in patients with autism and ID³. Heterozygous knockout mice show a neuronal glial pathology associated with reduced cognitive function⁶² and this phenotype was recognized in mouse models prior to the association in human disease. Similarly, heterozygous deletions or mutations of DYRK1A in humans^20,63, mice⁶⁴ and fruitflies⁶⁵ all show a phenotype of reduced brain volume associated with microcephaly. In this regard, it is interesting that the DYRK1A LoF were the last to be documented with the models preexisting the discovery of human genetic diseases. With new resources like the Zebrafish Mutation Project⁶⁶ and the International Knockout Mouse Consortium⁶⁰, more systematic and high-throughput genome-wide approaches for model organisms may be achieved in particular for LoF mutations.

Limitations and future directions

Despite the great success of recent exome studies, it is important to note that most of the analyses, to date, have been restricted to the protein-coding portion of the genome—a very small fraction (1.5%) of all human genetic variation. Furthermore, the definition of the protein-coding portion is far from perfect. Portions of the reference genome^67,68 and gene annotation⁶⁹ are incomplete especially in relation to isoforms specifically expressed in the brain. Regulatory variation and its impact are currently ignored. Even though genome sequencing costs have reduced, discovery and interpretation of genetic variation remain significant hurdles. Unlike protein-coding sequencing, defining the functional regions and the type of mutations that will abrogate such function remain active areas of research. Nevertheless, the genes where dosage imbalance have been found to strongly associate with disease represent a logical starting point to begin to interrogate regulatory mutations as well as epigenetic effects that may have a similar effect. Targeted resequencing of the entire genomic loci as well as full genome sequencing will undoubtedly discover new mutations and further improve our understanding of the phenotype-genotype relationship. Despite the recent emphasis on de novo mutations, their contribution to disease can only be understood in the context of the full spectrum of genetic variation of each individual^25,57.

Even for the protein-coding component, sequence discovery is incomplete with 5–10% of the exons either being missed or insufficiently captured to call genetic variation. The bias is particularly pronounced for genes mapping to high GC content regions of the genome where as many as 20–30% of the exons may be insufficiently covered. The sequencing technology also introduces biases against certain types and classes of mutation. The discovery of indels is largely regarded as incomplete because of difficulties associated with mapping short sequence reads in low complexity regions⁷⁰. Although there have been recent methods to call smaller CNVs, validation experiments indicating specificity and sensitivity are still far from ideal^71,72. The development of new sequencing chemistries and platforms that can cheaply and in a high-throughput manner access these regions of the genome should remain a high priority.

There is another level of reduced sensitivity, related to the timing of de novo mutations. It is increasingly recognized that postzygotic de novo mutations, i.e., mutations in somatic state, may play an important role in neurodevelopmental disorders⁷³. While the importance of somatic de novo mutations has been recognized for many years in the field of cancer genetics^52,74-76, we are only starting to appreciate its prevalence in neurodevelopmental disease^77,78. Several exome studies report that individual de novo mutations likely occurred postzygotically with estimates ranging from 1–2% of new mutations based on analysis of DNA derived from blood. O’Roak et al.³, for example, have shown that ~4% (9/209 cases) of de novo mutations likely occurred postzygotically. Mosaic mutations have been observed as a more general theme for CNVs, but not yet linked systematically to disease⁷⁹. More sensitive technologies^80,81 are required to identify lower-level mosaics as well as access to clinically more relevant tissue types. For defined disorders with isolated neurologic involvement, this may never be possible if the mosaicism is restricted to neuronal subtypes in the brain.

Next to technical hurdles, there is the daunting prospect of the extreme locus heterogeneity of these diseases. This raises the distinct possibility that a recurrent de novo mutation in a second patient will never be seen again in the same clinic. This can be partially overcome by developing new models for data sharing (e.g., de novo variant databases) and generating larger sample collections of patients (>50,000) that may be screened in follow-up targeted resequencing experiments. This requires a shift toward a more integrated and collaborative model of clinical and basic research. Successful models of clinical lab cooperation and standardization have already been established for the exchange of CNV data, e.g., the ISCA (International Standards for Cytogenomic Arrays) Consortium, and there is momentum to do the same for exome and genome sequence datasets, e.g., the ICCG (International Collaboration for Clinical Genomics)²⁴. The sheer size of the dataset (multiple Petabytes), ever-changing advances in sequencing technology, and the importance of standardized call sets, however, pose major challenges moving forward.

Although sporadic mutations have been the focus of this review, the importance of inherited mutations should not be underestimated. There is in fact compelling evidence that such variation contributes significantly to these diseases^45,72,82,83. While specific gene effects are much more difficult to tease apart in the general population owing to the genetic heterogeneity of these diseases⁴⁵, other approaches such as studies of consanguineous families have identified numerous candidate risk genes under a recessive disease model^84-86. It should also be noted that the effect size and penetrance for many of the recurrent de novo mutations is not yet known. For autism, de novo mutations have been thought to collectively increase risk 10– to 20–fold for up to 20% of cases with disease. It is likely that in some cases a rare variant will be necessary but not sufficient to confer phenotype requiring that both inherited and de novo mutations be jointly considered in order to understand their impact as has been noted for some CNV risk variants^57,87. Understanding the gender bias, which is particularly pronounced for ASD and ID, will require integrating inherited and de novo mutations from both the X chromosome and autosomes. Data from CNVs as well as SNVs suggest that the carrier burden of males and females differ significantly^42,72,88,89. The evidence suggests that females are more likely to be carriers of deleterious mutations and, therefore, protected against such diseases perhaps sex dependent modifiers map genetic architecture to diagnostic boundaries differently between the sexes.

Since many of the genes linked to disease will likely have minimal functional annotation in the human genome, it will be necessary to perform systematic studies to understand their specific role in neurodevelopment. The sheer volume of high-impact genes will probably necessitate large-scale model organism knockouts in Drosophila, zebrafish, and mouse^60,66, industrial-level development of iPSC cell lines and neuronal differentiation protocols⁹⁰, as well as massive screens using mass spectrometry to identify protein-interacting partners. All of these approaches have their own limitations. For example, it is an open question how well knockouts will model neurodevelopmental diseases such as ASD or ID since most of the known effects in humans occur in the heterozygous state and most knockout phenotypes are typically studied as homozygous LoF. Many phenotypic aspects of complex neuropsychiatric and neurobehavioral disease will not be amenable to model systems further limiting such functional approaches. Notwithstanding these challenges, it is a golden age of ‘neurogene’ discovery which promises to improve not only our understanding of disease but provide fundamental insight into the biology of human brain development.