The Genetic Control of Stoichiometry Underlying Autism

Abstract

Autism is a common and complex neurologic disorder whose scientific underpinnings have begun to be established in the past decade. The essence of this breakthrough has been a focus on families, where genetic analyses are strongest, versus large-scale, case-control studies. Autism genetics has progressed in parallel with technology, from analyses of copy number variation to whole-exome sequencing (WES) and whole-genome sequencing (WGS). Gene mutations causing complete loss of function account for perhaps one-third of cases, largely detected through WES. This limitation has increased interest in understanding the regulatory variants of genes that contribute in more subtle ways to the disorder. Strategies combining biochemical analysis of gene regulation, WGS analysis of the noncoding genome, and machine learning have begun to succeed. The emerging picture is that careful control of the amounts of transcription, mRNA, and proteins made by key brain genes—stoichiometry—plays a critical role in defining the clinical features of autism.

INTRODUCTION

Autism refers to a variety of clinical features observed in children in the first three years of life who develop difficulties with direct social interactions along with a variety of manifestations, including language dysfunction, repetitive behaviors, and restricted interests and activities, together termed autism spectrum disorder (ASD). It is therefore not surprising that, in parallel to the many clinical manifestations of those given this diagnosis, there may be many predisposing factors.

Similar phenomena abound in neurology. For example, epilepsy is a general diagnostic term with a multitude of predisposing conditions, and genetic understanding of epilepsy has identified mutations in a variety of channel proteins involved in the balance between excitation and inhibition, as well as in proteins involved in many different aspects of cell metabolism, and these genetic insights are driving rational therapeutics (Dhindsa & Goldstein 2015). Similarly, the set of dementias includes many disorders with many etiologies. Rapid-onset early dementia in Creutzfeldt-Jakob disease has an entirely different pathophysiology than the more gradual onset of dementia associated with APOE4 homozygosity. Common to all such general diagnostic categories are several features: They are clinical definitions with key common clinical manifestations that group them together—intellectual disability presenting in youth, seizures, or loss of cognitive function with age—even as clinicians recognize that there exist distinct and sometimes overlapping subsets of individuals within each category. It is important that basic scientists appreciate that such distinctions may underlie a plethora of different etiologies that manifest clinically as the same disease.

This complexity is worth emphasizing for several reasons. First, strategies that do not recognize the inherent diversity underlying categorical clinical terms such as autism have struggled to identify pathophysiologic causes for disease in individual cases. Efforts to take large cohorts of individuals with a general diagnosis of autism and extrapolate a common set of etiologic factors for what in reality are different physiologic disorders can lead to confusion. Genome-wide association study (GWAS)-focused autism studies struggled to identify common genetic features in what we now recognize as a multitude of different underlying genetic disorders unified by a common clinical diagnostic term.

Autism does have a genetic basis, with concordance rates of at least 90% in monozygotic twins with ASD, compared to 10% in dizygotic twins (Tick et al. 2016). Large-scale population-based epidemiologic studies, analyzing over 22,000 ASD individuals in five countries, estimated overall heritability of ASD as at least 80% (Bai et al. 2019). This review presents the current understanding of the molecular genetics of autism, framing an overarching model in which a wide variety of critical regulatory, synaptic, and other neuronal proteins underlie the common phenotypes of ASD, and explores the idea that one theme tying these disparate proteins together is tight stoichiometric control of their expression.

BREAKTHROUGHS IN AUTISM GENOMICS

Early studies to assess genetic causes of autism attempted to leverage the paradigm used in analysis of single nucleotide polymorphisms (SNPs) in case-control studies that used GWAS (Klein et al. 2005). This strategy was based on the recognition that common SNP variants reside in blocks and that linkage disequilibrium of such blocks can lead to insight into disease heritability. Such studies were more readily accessible than family studies in which genetics are much more tightly controlled but which are more difficult to assemble at scale.

GWAS searching for common but pleomorphic disorders such as schizophrenia (Pardiñas et al. 2018) have identified gene-associated loci through analysis of common SNP variants (for example, those present at ≥1% in the population), suggesting the possibility that this set of disorders results from the collection of many such variants in sporadic cases of this common disease. Nonetheless, it has been difficult to validate and make use of such genetic findings clinically, and this same problem has been true for studies of autism.

One approach to pursuing such studies has been to build larger sets of variants by increasing sample size and by looking at less common (rare) variants. Statistical power analysis argued that studying 25,000 cases of autism and comparing, via GWAS analysis, 100,000 controls would yield specific risk alleles (Pardiñas et al. 2018, Zuk et al. 2014). While entirely logical, this approach was confounded by the heterogeneity in the clinical features of autism as well as requisite a priori assumptions about the impact genetics would have on fecundity and selection coefficients. Autism is a spectrum of disorders and can range from those who are severely disabled to those who can function well, sometimes exceptionally, in the workplace and family. Consistent with this complexity, linkage studies did not unambiguously identify specific genes that caused ASD (Autism Spectr. Disord. Work. Group Psychiatr. Genom. Consort. 2017, Weiss et al. 2009), although larger GWAS (as applied to the recent collection of over 18,000 ASD cases and 28,000 controls) continue to be of interest (Grove et al. 2019).

A different approach to the study of autism came from bypassing the challenges and opportunities of GWAS and instead focusing on tightly genetically controlled sets of individual families. These studies began with analysis of copy number variants (CNVs) in families (Sebat et al. 2007) and then were refined to studies of whole-exome sequencing (WES) of trios or quads—genetic analysis of four individual family members, one affected and three not. Such analyses of an affected child (proband), a sibling, and their parents provided a tight-knit group of genetically related individuals from whom scientists could compare de novo genetic variants in probands.

It is worth noting that getting a firm grip on specific genes implicated in ASD, through studies of either small numbers of familial ASD (necessarily small because the disease itself can have an effect on fecundity) or larger numbers of de novo (simplex) ASD cases, has offered the potential to have important feed-forward effects in developing a comprehensive view of autism. By analogy, early linkage studies of the genetics of breast cancer failed to identify risk genes due to the complexity of the genetics and environmental factors that impact their manifestations. A foothold on the problem was accomplished by Mary-Claire King and colleagues (Hall et al. 1990) by focusing on families with a high incidence of breast cancer, and even there such studies required further refinement to look specifically at those with early-onset disease. This intuition, following basic tenets of a geneticist, allowed the identification of the BRCA1 gene and fed forward to the current recognition that a series of well over a dozen genetic risks all relating to mutations in DNA repair pathways underlie a great deal of breast cancer genetic risk.

Copy Number Variants Underlying ASD

The results of initial studies to identify ASD-related genes underlying ASD began with analysis of CNVs in 165 affected families. These included those defined as both simplex autism cases and multiplex cases with more than one affected sibling (Hall et al. 1990, Sebat et al. 2007).

These results were astounding and had two important consequences. First, while CNVs were not the only causative variants to be found in ASD individuals, the early focus on gene copy number introduced stoichiometry as a key variable to consider in framing disease pathogenesis. In the pioneering CNV study it immediately became evident that most of the de novo CNVs associated with ASD (12 out of 15) were deletions (Sebat et al. 2007), suggesting that a more common mechanism (although not necessarily cause, as discussed below) for CNV-associated stoichiometric variation in ASD was likely associated with loss of function. Such studies could then be tied to specific gene mutations within these regions. For example, studies of Angleman’s syndrome identified a CNV variant resulting from maternal deletion of chromosome 15q11-q13 as being associated with ASD (Cook et al. 1998). Subsequent analysis indicated that the same phenotype could be caused by duplication or triplication of the same region (Baker et al. 1994, Bundey et al. 1994, Schinzel et al. 1994) as well as point mutations of the UBE3A gene within this region (Fang et al. 1999, Kishino et al. 1997, Matsuura et al. 1997, Sutcliffe et al. 1997).

A number of additional studies have underscored the role of stoichiometric balance in ASD pathogenesis. Together, they more clearly define the remarkable observation that either too much or too little of the same gene product could lead to the same clinical ASD outcome. For example, Rett syndrome, a neurologic disorder associated with ASD, is classically defined as haploinsufficiency in the MECP2 gene in females (Amir et al. 1999). However, neurodevelopmental phenotypes are also found in males with MECP2 duplications (Van Esch et al. 2005). Another instance of this stoichiometry phenomenon is seen in Smith-Magenis syndrome, associated with either microdeletion of 17p11.2 (Van Esch et al. 2005) and ASD or de novo 17p11.2 duplication (Potocki et al. 2000). Subsequently, work identified mutations in the RAI1 gene as the dosage-sensitive gene within the 17p11.2 region (Carmona-Mora & Walz 2010).

A major consequence of the success of CNV analysis of ASD families was the robust conclusion that genetics could lead to the identification of single genes that were likely to contribute to ASD. The consequent increased impetus to study ASD genetics pushed forward attempts to undertake more sophisticated genomic studies, and technologic advances moved from copy number studies to WES studies.

Whole-Exome Sequence Analysis in ASD

Whole-exome analysis was pioneered by Hannon and colleagues (Hodges et al. 2007), who recognized that the ability to inexpensively synthesize mass-scale oligonucleotides could produce enough probes to bind to all transcribed exons. The strength of this bind and capture approach was primarily that all coding sequences in nearly all transcribed genes could be directly sequenced. Many such overlapping molecules from captured exons allowed great sequencing depth (complexity) within any individual locus, providing statistical robustness in the ability to call true variants (relative to low abundance technical sequencing errors).

The limitations of WES strategies were severalfold and in part presaged the drive to undertake whole-genome sequencing (WGS) studies (discussed below). The requirement to hybridize and capture individual fragments on an arrayed platform—literally thousands of hybridization reactions done in parallel—meant that the ensuing binding and wash conditions were an average best-case set of conditions for an average oligonucleotide capture. Variants in which GC content varied, for example, would be expected to be variable in their capture from being either too easy (AT-rich) or too difficult to release after binding and wash from capture oligos. Moreover, the length of captured overlapping RNA fragments obscured detection of small insertions and deletions, while the nonlinear nature of PCR amplification obscured analysis of CNVs.

Nonetheless, the strengths of WES combined with earlier methods were evident in the first such studies applied to ASD. These strengths still relied on the design of quad analysis, as early efforts applying WES in trios did not yield significant findings of ASD-related genes. In contrast, WES analysis of quads, published by the Wigler group (Iossifov et al. 2012, 2014) and independently by the groups of Eichler (Krumm et al. 2014, O’Roak et al. 2012), State (Sanders et al. 2012), and Daly (De Rubeis et al. 2014), made huge strides in the understanding of ASD genetics. What started as the identification of small numbers of loss-of-function de novo variants that were proband specific quickly became important and generalizable findings. The reproducibility of finding de novo variants in the same genes by these groups, and the subsequent replication of variants in these same genes in larger WES studies, confirmed with statistical robustness the emerging understanding that mutations in biologically coherent groups of genes relating to neuronal function underlay the development of autism in many cases.

Whole-Genome Sequencing: De Novo Variants in Noncoding RNA

While WES could readily identify many loss-of-function de novo variants, discovery of small insertions or deletions (e.g., of individual exons) or missense mutations that disrupted open reading frames leading to either nonsense-mediated decay (NMD) or loss of messenger RNA (mRNA) levels was more difficult. These limitations became apparent after WES of 10,000 quads available through the Simons Simplex Collection (SSC)—as increasing numbers of individuals were sequenced, the yield of findings began to plateau at ~25–30% of all children with de novo ASD. While the results from WES provided the breakthrough needed to begin to demarcate the functions encoded by proteins that led to susceptibility to ASD, the work also left open the question of what variants were responsible for the ~70% or more of probands that remained undiagnosed (Iossifov et al. 2014, Yuen et al. 2017).

Systematic identification of regulatory variants that could either increase or decrease the levels of gene expression of ASD-associated genes was beyond the reach of WES. Such variants were known to generally preside in noncoding regions of genes. While on occasion such variants could be picked up by WES, the capture design and depth were not able to detect them in either a consistent or comprehensive manner. This obstacle could in theory be overcome by WGS, in which coding and noncoding variants are detected with equal sensitivity. Strategies to identify even small CNVs could be enhanced by WGS relative to WES, and in fact an enrichment of such CNVs has been reported in analysis of WGS data (Fang et al. 2016, Turner et al. 2017).

However, WGS raised the issue of how to systematically interpret the impact of noncoding variants. It was anticipated that there are many variants per individual that are difficult to interpret and that even causative variants would have smaller effect sizes than missense variants in coding sequence.

Interpreting single-nucleotide variants from WGS raised big issues for investigators. Functional variants in introns, for example, could lead to aberrant splicing and switches in exon inclusion that could lead to such consequences as truncated proteins, premature stop codons and NMD, or de novo coding exons that are pathogenic. For example, NMD (Eom et al. 2013) or de novo exons (Cummings et al. 2017) are underrepresented or absent in genomic data sets. More generally, there were not established methods to systematically annotate noncoding functional variants such as, for example, those that impacted the nucleotide-binding sites of specific regulatory factors that in turn would cause RNA misregulation.

At the same time, the idea that noncoding variants might contribute to ASD gained strength from the general observation that regulatory regions are critical to rapid evolutionary change in humans (Indjeian et al. 2016, McLean et al. 2011). Moreover, it was becoming increasingly clear that regulatory sequences were likely to mediate proper stoichiometry in the context of complex disease (Liu et al. 2019). These observations were pertinent to findings that ASD probands on average harbor more noncoding de novo variants, in aggregate, than probands (Liu et al. 2019, Turner et al. 2017).

Understandably, given the large number of noncoding variants and difficulty interpreting them biologically, initial efforts to apply different analytical frameworks for such analyses of WGS data from small numbers of families present in data sets such as MSSNG, or from even larger data sets (519 families) from the SSC, were unable to demonstrate specific noncoding variants that were enriched in probands versus siblings (Werling et al. 2018, Yuen et al. 2017).

Applying Deep Learning Tools to Whole-Genome Sequencing

Machine learning tools were initially applied successfully to identify ASD-associated genetic changes with neural development, underscoring that a key feature of autism is alterations in genes expressed in and hence involved in early brain development (Krishnan et al. 2016), but these tools were not designed to analyze noncoding variants. More recently, machine learning was applied to larger WGS cohorts [up to 1,902 SSC families (Werling et al. 2018, An et al. 2018)] to compare rare and de novo noncoding variants. Using a risk score based on analysis of five large groups of transcription-associated annotations (e.g., functional annotations of histone acetylation, or GENCODE-defined promoters), researchers presented evidence for a contribution of noncoding variants in ASD probands in regions upstream (~750–2000 base pairs) of transcription start sites (An et al. 2018), suggesting that de novo mutations in promoter regions might contribute to ASD. However, after correction for multiple hypotheses (An et al. 2018), no single noncoding category was significant, and the use of a risk score could not identify specific variants with functional changes in transcription of specific genes. Similarly, while targeted and machine learning analysis of additional WGS in 493 multiplex families with autism was successful in identifying many deleterious protein-coding variants, analysis of noncoding variants identified only a few noncoding promoter deletions (in DLG2 and NR3C2), and no global enrichment in promoters was found (Ruzzo et al. 2019).

More recently, an important step forward in interpreting WGS data from the SSC came from a combination of leveraging biochemical demarcation of regulatory variants with deep learning strategies. This more integrated strategy focused on noncoding variants, pairing a deep convolutional-network-based framework with biochemical data demarcating regions of noncoding genome bound by DNA or RNA regulatory proteins. The latter included 2,000 regulatory profiles of histone marks, transcription factor binding sites, and chromatin accessibility identified by epigenetic or chromatin IP studies, as well as large sets of precise RNA regulatory sites identified by 230 crosslinking immunoprecipitation (CLIP) [a method for covalently crosslinking RNA-binding proteins (RBPs) to their RNA targets in situ in tissues (Ule et al. 2003)] biochemistry studies (Zhou et al. 2019).

This approach led to remarkable results, identifying an increased burden of de novo mutations disrupting both DNA-binding proteins and RBPs in ASD probands compared to unaffected siblings. Interestingly, the set of genes found to be under selective pressure (“loss of function intolerant” by the ExAC study; see discussion below) showed a significant concentration of noncoding RBP regulatory-disrupting mutations in ASD probands. Similar results were found for enrichment of de novo mutations disrupting DNA-binding proteins, and the loss of function–intolerant genes identified for RNA- and DNA-binding proteins were convergent, supporting the conclusion that observed noncoding regulatory events played causal roles in ASD. Importantly, both sets of results were independent of any selection for subsets of genes demarcated a priori using risk scores and were significant after multiple-hypothesis correction (Zhou et al. 2019), addressing issues raised in earlier WGS machine learning studies (An et al. 2018, Werling et al. 2018).

The biological implications of this strategy (Zhou et al. 2019) were important in finding that noncoding regulatory variants implicated gene regulation of genes and transcripts involved in synaptic transmission, neuronal development, and autism and were consistent with previous studies (Iossifov et al. 2014, Krishnan et al. 2016, Packer 2016). Moreover, they pointed the way to a means for annotating specific noncoding variants more precisely as disease associated. For example, the 57 genomic variants predicted to have high impact in transcriptional regulation and enriched in probands could be demonstrated to have differential transcriptional activity relative to wild-type (WT) sequences in cell-based assays (Zhou et al. 2019). In addition, a noncoding mutation, outside of canonical splice sites, was found to disrupt splicing of SMEK1 transcripts (encoding a protein that regulates neurogenesis through Wnt signaling; the fraction of exon inclusion [delta-I; (Ule et al. 2005)] relative to WT was highly significant [≃ 0.4 (Zhou et al. 2019)]). Moreover, reanalysis of the gene lists analyzed in Werling et al. (2018) with this new strategy revealed that noncoding regulatory mutations that were predicted to have posttranscriptional (RBP) regulatory effects in probands (i.e., large effect size between proband and sibling) were specifically enriched in ASD susceptibility genes (Zhou et al. 2019, figure 1c).

Together, these studies suggest a pathway forward for annotating noncoding variants in WGS data from ASD patients, a significant issue given the finding that only a minority of ASD individuals sequenced have been found to harbor coding mutations. While the selection coefficient for individual noncoding variants may be smaller than for coding variants, it is worthwhile noting that they also may account for more subtle differences in ASD phenotypes. For example, de novo noncoding variants were found to be associated with lower IQ in a subset of ASD individuals with RNA-binding mutations in introns flanking alternatively spliced exons (Zhou et al. 2019). At the same time, while these studies demonstrate a path forward by which WGS can identify rare regulatory variants with large effects, complex traits such as those seen in ASD are likely to be impacted by many genes contributing incrementally to any individual’s phenotype, and the actions and effects of any one trans-regulatory factor–DNA/RNA interaction may well be relatively small (see discussion in Liu et al. 2019).

ASD SUSCEPTIBILITY GENES

Genomic analysis has led to the identification of a vast number of ASD susceptibility genes. In general, their identification has relied on careful clinical descriptions combined with genomic sequencing, most often through analysis of quads. Some interesting exceptions to this successful strategy have emerged from more targeted studies. Careful clinical phenotyping of congenital syndromic disorders led to recognition that they may be associated with ASD phenotypes. Examples include Angelman syndrome (Khatri & Man 2019), the fragile X syndrome (FXS) (Niu et al. 2017), tuberous sclerosis (Krueger & Bear 2011), Rett syndrome (Moretti & Zoghbi 2006), Coffin-Siris syndrome 1 (CSS1) (Santen et al. 2012), and Asperger’s syndrome (Sebat et al. 2007). An unusual report of a syndrome involved a single sperm donor, in which an impacted recipient mother gave birth to two boys with ASD; crowdsourcing then traced a total of 12 offspring from the same donor sperm, all of whom had ASD (Cha 2019). This sperm was given to the geneticist Steve Scherer, who is undertaking genomic analysis.

Genomic studies have identified an expanding number of ASD susceptibility genes, with high confidence estimates of 53 genes in 2015 (Sanders et al. 2015), 102 genes in 2018 (Satterstrom et al. 2020), and 106 by mid-2019 (Iakoucheva et al. 2019), with some arguing that there may be up to 1,000 susceptibility genes (Ramaswami & Geschwind 2018). More may be identified in the future by analysis of whole-genome sequence, allowing an understanding of regulatory variants (see above), as well as by better clinical understanding of genes that have complex phenotypes that both contribute to and mask overt ASD. The latter refers to the fact that ASD is a neurodevelopmental disorder of variable severity presenting years postnatally, such that variants that could give rise to clinical autism may present earlier with severe clinical symptoms, precluding an ASD diagnosis.

Nearly all of the genes identified through WES as ASD susceptibility genes are associated as haploinsufficiencies (Iakoucheva et al. 2019). While there is some ascertainment bias, given the approaches used (e.g., searching for gene-disrupting/loss-of-function mutants), this nonetheless underscores the concept that stoichiometric control of key neuronal proteins is necessary for intellectual function. This point is further underscored in consideration of genes found associated with ASD susceptibility through study of CNVs (see also the above section titled Copy Number Variants Underlying ASD). Some genes in this category are linked to ASD through either overexpression or underexpression of the same gene. For example, while NRXN1 is one of the most well-studied ASD susceptibility genes associated with loss of function, by either CNV or likely gene-disrupting mutations, duplications of NRXN1 have also been associated with ASD (Wiśniowiecka-Kowalnik et al. 2010). Both deletion and duplication of 16p11.2 (Hippolyte et al. 2016, Pizzo et al. 2019, Rosenfeld et al. 2010) are associated with ASD. Similarly, duplication of 22p13, a region encoding SHANK3, and deletions of SHANK3 are clinically associated with intellectual disability and ASD (Jiang & Ehlers 2013, Wilson et al. 2003). Moreover, point mutations in SHANK3 are associated with ASD (Durand et al. 2007, Ruzzo et al. 2019), restoration of SHANK3 function in SHANK3-deleted neurons restores synaptic deficits (Shcheglovitov et al. 2013), and similar findings of reversible autism-like defects were observed in mice harboring a Shank3 mutation (Speed et al. 2015, Mei et al. 2016), although these mice studies are complex and need additional confirmation (Speed et al. 2019). Interestingly, Sebat and colleagues (Iakoucheva et al. 2019, Malhotra & Sebat 2012) have noted that such over- and underexpression of the same gene regions associated with ASD (1q21.1 and 16p11.2) can also be associated with either macrocephaly or microcephaly.

Given the early action and pleiomorphy of ASD genes, there is some evidence that the full set of susceptibility genes may be overlooked because their phenotypes with respect to ASD may be obscured in some instances by more severe phenotypes. An illustration of how ASD susceptibility genes could thus be masked is given by considering mutations associated with the chromatin remodeling protein ARID1B. Haploinsufficiency of ARID1B has been recognized as underlying the congenital disorder CSS1, and in some children this results in intellectual disability and ASD (Santen et al. 2012). However, it has subsequently been recognized that newborns presenting with ARID1B mutations may present with a wide spectrum of clinical symptoms, including life-threatening malformations of major organs [microcephaly (10%) and severe cardiac defects (ventricular septal defects, coarctation of the aorta; 8%) (Mannino et al. 2018)] that may preempt an ASD diagnosis that would otherwise manifest later in life [for example, after surgical repair of previously devastating cardiac anomalies (Bean Jaworski et al. 2017)]. Analogous gene variants can cause severe perinatal defects but may also, in milder clinical manifestations or after early aggressive medical/surgical management, remain to be identified and hence contribute to greater understanding of developmental pathways associated with ASD.

To date, two major sets of genes that encode quite distinct functions have been identified by CNV analysis and WES. These sets can be described as genes encoding proteins involved in chromatin regulation and those encoding proteins involved in synaptic function, particularly relating to the balance between excitation and inhibition. In both instances, the theme of regulation of stoichiometric levels of gene control per se is evident in ASD and consistent with the overall model that the fine-tuning of neuronal activity and gene expression may be fundamental in allowing normal neuronal development and ultimately cognitive function.

Gene Regulation and ASD Susceptibility

Proteins that regulate transcription or chromatin state are vastly overrepresented in the known set of ASD susceptibility genes, together accounting for nearly one-third of the currently identified genes (Iakoucheva et al. 2019). Transcriptional regulators and their binding partners include both general factors and those associated with specific neuronal functions, such as ADNP, which may be regulated by and associated with vasointestinal peptide signaling (Sragovich et al. 2019).

In addition to transcription factors, an even larger set of ASD susceptibility genes encode chromatin modifiers. These again include general factors, some of which are implicated in regulating neuronal function, such as BRD4 (see below). While many functions of these proteins regarding specific actions in neuronal gene regulation are not yet known, some again may be more specifically tied to regulation of neuronal genes. For example, PHF21, believed to be the gene within 11p11.2 deletions (Potocki-Shaffer syndrome) that mediates intellectual disability (Kim et al. 2012), is a histone deacetylation/demethylating protein that is also part of the LSD1-CoREST corepressor complex thought to repress expression of neuronal genes (Porter et al. 2018).

RNA regulation and ASD susceptibility.

A second aspect of gene regulation intersecting with ASD susceptibility is the identification of large numbers of RNA regulatory genes. These fall into various categories, including factors involved in splicing (see below) and other aspects of RNA regulation (3 UTR binding, RNA localization; CELF4, ELAVL3), translational regulation [EIF3G, FMRP (see below)], and microRNA regulation and mRNA turnover (TNRC6B).

Understanding the role of RBPs in ASD has benefited from the ability to identify, with nucleotide resolution, their direct RNA-binding sites using CLIP methods (Ule et al. 2018). In ASD, a number of the networks of regulated RNAs identified show remarkable overlap with ASD susceptibility genes. The foremost among this group is FMRP, discussed below. Networks regulated by 3′UTR–binding factor CELF4 have also been found to overlap ASD susceptibility gene networks, specifically those involved in neuronal excitation (Porter et al. 2018, Wagnon et al. 2012).

Splicing factor networks regulated by RBPs also show a significant relationship to ASD (Gandal et al. 2018, Voineagu et al. 2011). When the targets of RBFOX splicing factors were determined by CLIP, some regulated exons were found in ASD susceptibility genes, including NLGN3 (Zhang et al. 2008), and subsequent studies found that over 10% of RBFOX-regulated alternative exons are regulated from among 48 ASD susceptibility genes, including SHANK3 (Weyn-Vanhentenryck et al. 2014). These findings are complemented by additional analysis of splicing (Fogel et al. 2012) and cytoplasmic transcripts regulated by RBFOX1 and implicated in ASD, including cytoplasmic transcripts encoding ASD-related proteins affecting synaptic function and calcium signaling (Lee et al. 2016). Together, these studies are consistent with the strong link between RBFOX proteins in autism, observed in both studies of copy number variation in ASD (Gambin et al. 2017, Sebat et al. 2007) and the original case report of a 160-kb deletion in the RBFOX1 gene (Martin et al. 2007).

Several studies have implicated the NOVA splicing factors (Buckanovich et al. 1993, Yang et al. 1998) in both RBFOX and ASD biology. NOVA expression is extremely restricted to neurons in the central nervous system, an observation that first suggested that the mammalian brain regulates RNA in uniquely specified ways (Darnell 2013). Several NOVA-regulated precursor mRNAs harboring YCAY clusters were originally described in cellular and in vitro splicing assays and in Nova-null mice (Dredge & Darnell 2003; Dredge et al. 2001, 2005; Jensen et al. 2000; Musunuru & Darnell 2001; Saito et al. 2019; Ule et al. 2006), which established NOVA as the first tissue-specific alternative splicing factor in vertebrates. Subsequent unbiased, genome-wide approaches afforded by CLIP, including pioneering exon-junction splicing microarrays (Ule et al. 2005), HITS-CLIP (Licatalosi et al. 2008), and bioinformatic analyses (Ule et al. 2006, Zhang et al. 2010), identified a robust set of ~800 target alternative exons (Ule et al. 2018).

Together, these studies allowed bioinformatic analyses that revealed robust overlap with ASD susceptibility genes. Bioinformatic analysis of target transcripts, based largely on analysis of HITS-CLIP data, tied NOVA to both RBFOX proteins and ASD (Zhang et al. 2010). Moreover, a subset of NOVA-regulated transcripts identified by Bayesian network analysis were found to harbor orphan conserved elements not bound by NOVA proteins themselves. Further analysis revealed that these were RBFOX-binding elements, and NOVA and RBFOX proteins were found to synergize in regulating splicing cooperatively (Zhang et al. 2010). Independently, NOVA targets were examined for potential relationships to neurologic disease, including ASD. Among 358 NOVA target transcripts identified by network analysis, 88 were implicated in genetic disorders, including mental retardation, epilepsy, and autism. Interestingly, the latter were encoded by the RBFOX1 gene as well as a host of other ASD susceptibility genes (AUTS2, CADM1, GRIK2, NRXN1, SLC4A10, etc.). Together, 8.5% of transcripts coordinately regulated by both NOVA and RBFOX proteins are implicated in autism (Zhang et al. 2010).

Notably, very specific delineations of misspliced transcripts are beginning to be identified as correlated and potentially causally related to ASD (reviewed in Quesnel-Vallières et al. 2019). Conserved neuronal microexons, highly conserved exons that are believed to modulate protein-protein interactions (Scheckel & Darnell 2015), have been found to be frequently misregulated in the ASD brain, again often in transcripts associated with neuronal vesicle transport (Irimia et al. 2014), and more recently in transcripts encoding translational elongation factors (Gonatopoulos-Pournatzis et al. 2020). Several RBFOX transcripts show dysregulated splicing in the ASD brain (Voineagu et al. 2011). An assessment of allelic imbalance in proband versus siblings has implicated small nucleolar RNA–targeted splicing changes in ASD-related target genes (Lee et al. 2019), and a specific error in splicing of the CPEB4 transcript, which encodes a polyA-binding protein linked to deadenylation, has been associated with ASD (Parras et al. 2018). Finally, since RNA regulatory elements identified by CLIP have primarily been identified in noncoding regions of transcripts (with FMRP binding elements a notable exception), these studies have led the way into new approaches to understand ASD susceptibility variants in noncoding regions (see discussion above and Zhou et al. 2019).

Excitatory/inhibitory balance and ASD susceptibility.

Perhaps the earliest considerations regarding the imbalance in neuronal excitation/inhibition in ASD came from hypotheses based on early genetic studies of autism and observations from mouse models (Rubenstein & Merzenich 2003), combined with early studies on FXS. FXS is directly relevant to autism, given the extremely high prevalence of ASD in children with FXS. Early studies by Bear, Huber, and colleagues (Bear et al. 2004; Huber et al. 2000, 2002) studying the synaptic pathology in Fmr1-null neurons demonstrated that FMRP plays a critical role in the regulation of a key receptor for the excitatory neurotransmitter glutamate, specifically the metabotropic glutamate receptor mGluR5. These studies led to the mGluR hypothesis of FXS in which imbalance of glutamate-dependent excitation/inhibition underlay the development of intellectual disability in FXS (Bear et al. 2004; Huber et al. 2000, 2002; Parras et al. 2018; Zhang et al. 2009). Moreover, as mGluR5 mRNA was found to be directly bound to the translational regulatory protein FMRP (Darnell et al. 2011), these studies tied synaptic translation to long-term depression in hippocampal CA1 neurons in FXS animal models.

Subsequently, the many genes that have been identified as ASD susceptibility genes (as well as FMRP targets) have obvious direct and indirect (downstream signaling) functions in excitatory or inhibitory neurotransmission. These include both susceptibility genes themselves and direct and indirect protein-protein interaction networks, as recently reviewed in detail by Wall and colleagues (Ruzzo et al. 2019; see also Darnell et al. 2011). One unexplained observation is that large genes are disproportionately represented in ASD studies (Sebat et al. 2007) and as FMRP targets (Darnell et al. 2011). Another and perhaps more salient connection is that ASD susceptibility genes and genes encoding FMRP target RNAs are under high selective pressure (Darnell et al. 2011, Iossifov et al. 2012). This concept has since been applied more generally, with the development of loss of intolerance scores, to successfully narrow lists of candidate genes in other complex neurogenetic disorders such as epilepsy (Petrovski et al. 2013).

The protein products of genes directly identified as ASD susceptibility loci include 98 categorized by analysis of whole-genome sequences from 2,308 individuals (Ruzzo et al. 2019) and what turns out to be a largely overlapping set of 106 ASD susceptibility genes comprehensively assembled by Sebat and colleagues (Iakoucheva et al. 2019), and the reader is referred to both of those reviews for details and discussion. Many of these ASD susceptibility genes encode inhibitory (GABABR2, GABABR3, SLC6A1) or excitatory (GRIA2, GRIN2B) neurotransmitter receptors or transporters and ion channels that impact neuronal sensitivity, including potassium (KCNMA1, KCNQ3), calcium (CACNA1E, CACNA2D3), and sodium (SCN1A, SCN2A) channels and allied proteins (e.g., NRXN1, involved in calcium channel function, and ANK2, involved in ion channel localization). Together, these observations support the identification of excitatory/inhibitory imbalance as one means of leading to an ASD phenotype.

ASD and Epilepsy

The relationship between ASD and the balance between neuronal excitation and inhibition is underscored by considering the relationship between ASD and epilepsy (Lee et al. 2015). Up to 20% of children with ASD have epilepsy, and this relationship was noted early by the association of tuberous sclerosis with ASD (Walsh et al. 2008). Tuberous sclerosis is a congenital neurodevelopmental disorder associated with heterotopias of ectopic groups of neurons in the brain, and these are believed to be epileptic foci. These correlations in tuberous sclerosis complicate conclusions regarding the cause-effect relationship between epilepsy and ASD, a caveat that in general (although see discussion of BCKDK below) complicates interpretation of the coincidence of epilepsy and autism.

Specific genetic variants associated with ion channel function and the maintenance of excitatory/inhibitory balance have been associated with ASD. For example, variants in the potassium channel inward rectifier gene KCNQ3 have recently been associated with ASD. Interestingly, molecular analysis of these missense mutations indicate that they result in a gain of function (Sands et al. 2019); conversely, loss-of-function mutations in KCNQ2 or KCNQ3 can also cause neurodevelopmental problems, including epilepsy and global neurologic disability (Miceli et al. 2014, Sands et al. 2019). This observation emphasizes the delicate nature of KCNQ3 (and other potassium channels related to ASD, e.g., KCNMA1) stoichiometric control and its relationship to ASD and epilepsy.

Indeed, the frequent concurrence of epilepsy in ASD patients may directly relate to RNA regulation. Human and mouse studies have demonstrated that haploinsufficiency of RNA regulatory proteins is frequently associated with epilepsy, as seen with FMRP (Lee et al. 2015), NOVA1 and NOVA2 (Eom et al. 2013), ELAVL4 (Ince-Dunn et al. 2012), RBFOX proteins (Bhalla et al. 2004; Lal et al. 2013a,b), and CELF4 (Wagnon et al. 2012). The relationship between RBP stoichiometry and the nature of excitatory/inhibitory balance in ASD is particularly relevant given the findings of Zhou et al. (2019), discussed above, demonstrating that noncoding variants that affect RBP binding are enriched in ASD probands and ASD susceptibility genes.

Another relationship exists between dysregulation of metabolic pathways, ASD, and epilepsy. In one particularly interesting example, recessive gene-disrupting mutations in BCKDK, the gene encoding branched-chain ketoacid dehydrogenase kinase, were identified in families with autism, epilepsy, and intellectual disability (Novarino et al. 2012). Interestingly, mouse models (Bckdk-null mice) recapitulated neurobehavioral defects that were ameliorated with dietary supplements, suggesting both a potential treatment for these patients and a dissociation of epilepsy from the developmental aspects of intellectual disability in this instance of ASD. The gene encoding the HDL-binding protein HDLBP is downregulated in ASD (Satterstrom et al. 2020), and there is a relationship between epilepsy, metabolic defects, and verbal learning defects, specifically including lower HDL levels (Hermann et al. 2017).

Finally, a smaller but significant number of ASD susceptibility genes encode proteins involved in actin or microtubule biology (see, for example, Ruzzo et al. 2019), and there is emerging interest in the relationship between epilepsy and microtubule dynamics (Xu et al. 2016). This intersection is not well articulated in experimental science other than through observations of actions of pharmacologic agents that act on microtubules, which may exacerbate or ameliorate human epilepsy. In ASD, such a relationship may converge on the need for neurons to organize processes in development and to do so in rapid time scales, for example, to reorganize microtubular arrays in response to synaptic plasticity stress. This question may be developed further, given the finding that ASD susceptibility genes include proteins associated with microtubule dynamics such as MAP1A (Myers et al. 2011), NCKAP1 (Zweier et al. 2019), and DYNC1H1 (Ding et al. 2016), which is essential for retrograde axonal transport, as well as several that encode proteins [KATNAL2, SPAST (Matthews et al. 2017)] that are believed to act to sever microtubules and promote rapid reorganization of the neuronal cytoskeleton.

AUTISM AND THE STOICHIOMETRY OF BRAIN REGULATION

A general concept of gene balance underlying variations in human phenotype and disease was put forth over a decade ago based on considerations of the consequences of gene duplication (Conrad & Antonarakis 2007). More specifically in ASD studies, gene dosage effects leading to cognitive development tie the control of neuronal plasticity to the control of protein levels mediated by transcriptional and translational controls. In part, this stoichiometric control seems likely to include those proteins encoding the balance between neuronal excitation and inhibition, a point emphasized above by the observation that 20% of ASD patients have epilepsy. More generally, the idea of gene balance relates to the concept of homeostatic plasticity, discussed below and considered in great detail by Turrigiano and colleagues (Turrigiano & Nelson 2004, Turrigiano et al. 1998).

Autism and the Fragile X Syndrome

An aspect of stoichiometric misregulation predisposing to ASD—related to but distinct from variants in chromatin or synaptic function–associated genes—came from the study of FXS. Approximately 50% of children with FXS are diagnosed with ASD. Nearly all children with FXS harbor a triplet repeat in the 5′ UTR of the fragile X gene, FMR1 (Santoro et al. 2012). This leads to hypermethylation of the DNA locus and a loss of function through transcriptional silencing of the FMR1 gene. The FMR1-encoded protein, FMRP, was recognized to harbor two KH-type RNA-binding domains, a spacer region, and a third RGG-type RNA-binding domain.

Precisely what mechanism might underlie the connection between FXS and ASD first came to light by identifying FMRP’s RNA targets in the brain. FMRP was recognized as similar in overall architecture to the KH domain containing NOVA neuron-specific RBPs, leading to the application of CLIP—first developed for analysis of Nova function (Ule et al. 2003)—to FMRP. This identified 842 robust mRNAs bound by FMRP in the mouse brain (Darnell et al. 2011), a list now expanded to more than 1,000 transcripts (Van Driesche et al. 2019).

Remarkably, these studies revealed that the ranked list of FMRP-bound brain transcripts fell into several major gene ontology categories that closely overlapped the nascent set of ASD targets (Darnell et al. 2011). This observation was replicated independently as more ASD targets were identified by WES analysis of quads (Iossifov et al. 2012). At the top of the FMRP mRNA-bound target list were synaptic proteins and chromatin modifiers. These studies also led to the discovery that FMRP target and ASD risk genes were under a high degree of selective pressure, implying that they encoded key cellular functions (Iossifov et al. 2012). This concept, subsequently termed loss-of-function intolerance, has been applied more widely in genomics (Petrovski et al. 2013), including, for example, to amyotrophic lateral sclerosis (Cirulli et al. 2015). The variant intolerance evident in FMRP and ASD joint targets underscores how, when mutated, such genes may lead to severe neurologic phenotypes, including not only intellectual disability and ASD but potentially other neurologic disorders such as schizophrenia (Fromer et al. 2014).

FMRP CLIP not only delivered a biochemically robust list of FMRP targets but revealed what is likely the major function of the FMRP protein. FMRP CLIP results differed from those previously seen with most other RBPs such as NOVA, which bound to discrete RNA elements through sequence-specific interactions mediated by the KH-type RNA-binding domains. In vitro, FMRP was found to bind complex RNA elements (G-quadruplex and RNA pseudoknots), revealed by RNA selection, biochemistry (Darnell et al. 2001, 2005), and structural studies (Phan et al. 2011, Vasilyev et al. 2015). In vivo, while FMRP CLIP demonstrated binding directly to a specific subset of brain transcripts, within those transcripts its primary binding appeared indiscriminate, covering entire coding sequences (Darnell et al. 2011). This observation, together with the finding that the vast majority of FMRP protein within the brain is present on polyribosome fractions of sucrose gradients (Darnell et al. 2011, Stefani et al. 2004), led to biochemical studies of the role of FMRP in translational regulation, an issue that had previously been raised by in vitro (nonstoichiometric) FMRP studies (Laggerbauer et al. 2001).

Assessment of a role for FMRP in translation done in WT and Fmr1-null brains revealed several points. FMRP RNA binding was reliant on its KH domains, as mutations in the protein’s KH1 (Myrick et al. 2014) or KH2 (Zang et al. 2009) domains (mutations that reproduced those seen in two clinical rare cases of FXS in which the FMR1 gene did not harbor triplet repeats) abrogated polysome association and RNA binding. Moreover, it was found that FMRP binding to polyribosomes on its target mRNAs was associated with both binding to ribosomes (observed by electron microscopy) and stalled ribosomes (observed in biochemical assays) (Darnell et al. 2011). These findings were supported by additional in vitro studies showing an increased translation rate in the absence of FMRP (Udagawa et al. 2013). The findings suggested that FMRP binding to specific mRNAs (via still undiscovered specificity determinants) leads to reversible ribosomal elongation arrest, thereby inhibiting translation of those transcripts.

Additional direct actions of FMRP have since been described, including association with CAPRIN1, in vitro assembly of protein-protein (and likely RNA) complexes into insoluble cell fractions (El Fatimy et al. 2012, Kim et al. 2019), and undefined (direct and/or indirect) actions of FMRP on elongation factors and translational initiation that may act to balance FMRP actions on ribosomal elongation (Das Sharma et al. 2019). To date, these studies together point to a model in which FMRP loss in FXS (and by proxy ASD) results from a primary defect in translational control of specific target mRNAs.

Since some of the FMRP targets themselves mediate translational controls, one net takeaway from studies of FMRP function is that FMRP may act at multiple levels to regulate the stoichiometry of proteins involved in synaptic function and chromatin modification. This is consistent with the observation made above that FMRP acts to inhibit the synthesis of dosage-sensitive gene products in neurons. Several such targets, including some whose loss of function has been associated with ASD, are NRXN1, NLGN3, SHANK3, TSC2, and CENTG2, consistent with the idea that stoichiometric control of key neuronal transcripts is important in autism.

Interestingly, the number of splice site variants (as well as other nonsense mutations) present in FMRP targets was also found to be significantly decreased (Iossifov et al. 2012). Thus, there appears to be a strong purifying selection in humans against splicing errors or nonsense mutations in FMRP, suggesting the importance of RNA-protein stoichiometry for key autism-related transcripts, and the potential for overlap between splicing and translational regulatory networks. A potential relationship between FMRP-dependent translational control and RBFOX/NOVA-dependent splicing control in autism remains largely unexplored.

Stoichiometric Control of Synaptic Excitatory/Inhibitory Balance and Homeostatic Plasticity

A summary of all genes identified to date in ASD suggests subgroups of disorders—for example, those involved in excitatory/inhibitory balance, and those involved in regulation of transcription. A unifying feature of these two sets is that they themselves are regulated by the translational control protein FMRP.

The genes involved in these biological balances, between neuronal excitability and molecular output of the genome, may be considered as a group to be involved in maintaining normal stoichiometric output of these proteins. Each of these domains in turn relates to ASD in more specific ways. Fifteen percent of males with FXS have epilepsy, and 25% of FXS and ASD children have concurrent epilepsy (Niu et al. 2017). While this remains a poorly understood general phenomenon, the observations fit well with the critical need to control excitation/inhibition balance, a key underlying feature of both epilepsy and ASD.

The need to control the translation of excitatory/inhibitory proteins for cognitive function fits with our understanding of Hebbian synaptic control and synaptic learning, in which coincidence detection requires translation during a brief (~20–60 min) sensitive window. There is great interest in the molecular neuroscience community in whether this window of translational control may in turn relate to the local regulation of translation of synaptic mRNAs (Hafner et al. 2019). Indeed, it will be of interest to understand whether such regulation by local translational regulators such as FMRP may play such a role.

A local synaptic mechanism for tight regulation of the stoichiometry of synaptic proteins does not preclude the need to maintain general homeostasis in the neuron. This concept, termed homeostatic plasticity, has been considered in great detail in studies of monocular deprivation and plasticity in the visual system by Turrigiano and colleagues (Miska et al. 2018). Interestingly, preliminary studies suggest that the synaptic scaffolding protein SHANK3, which itself is a stoichiometric target in ASD (see above), is also necessary for homeostatic plasticity in the brain (Hafner et al. 2019, Tatavarty et al. 2018).

Moreover, SHANK3 variants can lead to either excitatory or inhibitory imbalance in different neuronal populations (Monteiro & Feng 2017, Van Driesche et al. 2019). While this argument could reflect a general requirement for SHANK3 (and other ASD targets) in homeostatic plasticity across whole neurons or neuronal circuits (Hafner et al. 2019, Tatavarty et al. 2018), they clearly reflect the critical nature of proper stoichiometric control of SHANK3 in ASD.

Given the role of SHANK3 and other synaptic proteins in local synaptic control and their relationship to Hebbian function and local synaptic plasticity, their stoichiometric control is again likely to reflect excitatory/inhibitory imbalance at the level of individual synapses. Indeed, SHANK3 and many other synaptic proteins are direct targets of FMRP (Darnell et al. 2011), linking FMRP’s role in local synaptic translational controls to its role in the neuronal cell body to circuit wide homeostatic control.

Stoichiometric Control of Transcriptional Regulation in ASD and Homeostatic Plasticity

Proper stoichiometric control of the level of chromatin regulators is critical in ASD, given their role both as genetic susceptibility genes and as targets of translational overexpression in the absence of FMRP. Indeed, FMRP loss leads to overexpression of one such regulator, the bromo-domain protein BRD4 (Korb et al. 2017, Monteiro & Feng 2017, Van Driesche et al. 2019), and missense mutations in BRD4 [for example, P113Q (Iossifov et al. 2012)] have been seen in ASD probands. Mutations leading to decrease of BRD4 levels lead to downstream dysregulation of synaptic proteins, memory defects, and impacts on seizures. Interestingly, correcting stoichiometric imbalance of BRD4 overexpression in Fmr1-null mice, using the BRD4-specific inhibitor JQ1, corrects corresponding synaptic and physiologic abnormalities (Korb et al. 2015, 2017).

FMRP as General Molecular Organizer

Three arguments may be made regarding FMRP as a central regulator of stoichiometric control in autism. First, there is an undisputed relationship between the two disorders; ~60% of males with FXS meet diagnostic criteria for ASD (Niu et al. 2017). Second, FMRP plays a role in the control of overall gene expression—most clearly through translational regulation of transcriptional controls. Such a relationship between factors impacting translational state, through regulation of FMRP stoichiometry itself and/or its impact on gene expression, for example, through modulation of chromatin state and transcription factors, is suggestive of a global role of FMRP on transcriptional regulation (Korb et al. 2017). This suggests a potential relationship between FMRP and neuronal homeostasis.

The third activity of FMRP relates to its role in the synapse. Most of this literature (but not all) relates to the presence, regulation, and activity of FMRP locally in the cell dendrite. Neuronal dendrites harbor ribosomes at the bases of spines and beneath the synapse (Ostroff et al. 2002), harbor a discrete but sizable number of transcripts (Glock et al. 2017, Tushev et al. 2018), and harbor the FMRP protein itself (Bassell & Warren 2008), as do axon termini (Akins et al. 2017, Christie et al. 2009), and FMRP may act in both compartments. Functionally, FMRP acts to stall ribosomal elongation (Darnell et al. 2011). Taken together, these observations suggest that synaptic activity could have unique (and localized) actions on translation and the stoichiometry of key synaptic proteins.

For example, FMRP binds to the mRNA encoding several dendritic neurotransmitter receptors, including key receptors such as the metabotropic glutamate receptor and the NMDA receptor, which are present at low molar stoichiometry within individual dendrites [i.e., ~20 molecules (Sheng & Hoogenraad 2007)]. Regulated translation of even two extra receptor molecules would yield a 10% increase in receptor stoichiometry, which in turn could have a long-term impact on the function of that synapse. Such a model would be consistent with a number of proposed means by which FMRP could be regulated, including those involving Hebbian mechanisms that lead to calcium influx, potentially triggering FMRP dephosphorylation (Lee et al. 2011) and acute and transient changes in its ability to regulate translation (Ceman et al. 2003, Muddashetty et al. 2011). Since FMRP’s action to bind mRNA and inhibit ribosomal elongation is reversible in experimental settings (Darnell et al. 2011), these observations underscore a model for FMRP action in normal and ASD synapses in which local Hebbian input leads to stoichiometric variation in protein molarity and synaptic function.

An overarching model (Figure 1) would place FMRP as a sensor of neuronal activity (for example, through calcium-dependent coincidence detection) in both the synapse and cell soma, impacting its ability to drive translation-dependent synaptic activity and transcriptional state. The readout of its activity would be a direct action on protein stoichiometry, both within the synapse (as one example, via modulation of NMDA receptors) and globally (on transcription state, as one example, acting on BRD4 levels); Figure 1 relates additional aspects of this model. An interesting aspect of such a model is that it provides a means to contribute simultaneously to regulation of both localized protein synthesis within specific synaptic compartments and global gene regulation, contributing toward a potential mechanism governing both synaptic learning and homeostatic plasticity within the neuron. Such a model integrating mechanisms that may jointly impact the neurobiology of learning and homeostatic plasticity in the neuron reflects the growing understanding of how genomics contributes to our understanding of neurologic disorders of intellectual function of the complex sort observed in ASD.

Figure 1.

A model for stoichiometric integration of neuronal activities in autism. Autism spectrum disorder (ASD) susceptibility genes fall into three major classes under different levels of stoichiometric control. Each has a connection to ASD and FMRP: synaptic proteins and chromatin regulators are both major ASD gene susceptibility targets and major FMRP targets, while the loss of FMRP in fragile X syndrome leads to ASD in about half of affected children. This model connecting FMRP, intellectual function, and ASD proposes that FMRP integrates protein expression and stoichiometry in two ways. First (➊), within the synapse, FMRP integrates local signals (such as Ca²⁺-mediated signals) to modulate translation (Darnell et al. 2011) and local stoichiometry of proteins involved in synaptic plasticity and protein degradation. Second (➋), within the cell soma, FMRP integrates the sum of dendritic signaling globally, modulating RNA translation to regulate the stoichiometry of nuclear proteins—chromatin modifiers and transcription factors (➌)—to effect global homeostasis in the neuron. This model is supported by the recognition that local dendritic translation plays an important role in protein synthesis–dependent synaptic plasticity (Kang & Schuman 1996). Such local control of synaptic protein stoichiometry can thereby be differentially regulated in independent synapses within a single neuron (via translational control, by FMRP as a local regulator of ribosomal elongation, and by other factors). Notably, joint synaptic-related, FMRP-regulated target messenger RNAs (mRNAs) and ASD risk alleles also include proteins involved in microtubule/actin dynamics and protein turnover (not pictured), including those involved in protein degradation pathways [e.g., the ubiquitin ligase Ube3A; reviewed in Kasherman et al. (2020)]. A connection with Hebbian plasticity (involving NMDA receptor–mediated Ca²⁺ signaling) and FMRP regulation is consistent with data suggesting that the ability of FMRP to suppress translation may be regulated by Ca²⁺-dependent phosphatases (see main text). Globally, there is a need for homeostasis within the whole neuron, as it simultaneously balances thousands of synaptic excitatory/inhibitory signals without triggering neuronal electrical instability. This concept reflects the need for homeostatic plasticity, as articulated by Turrigiano and colleagues (Turrigiano & Nelson 2004; Miska et al. 2018), with particular connection to ASD exemplified through studies of SHANK3 (Tatavarty et al. 2018). A molecular mechanism in such neuronal homeostatic control, acting over different timescales ranging from milliseconds to days (Abraham & Williams 2003), is proposed here as the net feedback between translational control of synaptic activity locally, cell soma regulation of mRNAs encoding chromatin regulators and transcription factors (➌), and additional joint FMRP-regulated targets and ASD risk alleles, including proteins involved in RNA binding/regulation (not pictured; ELAVL3, EIF3G, TNRC6B; see text and Iakoucheva et al. 2019). Again, these transcripts encoding gene regulatory functions are enriched as FMRP targets (Darnell et al. 2011, Iossifov et al. 2012) and are implicated in ASD (Satterstrom et al. 2020; see text). An untested hypothesis is that mRNAs encoding different functions—synaptic vs. nuclear factors—are differentially localized in the dendrite and soma. In sum, this model suggests that net signaling from synapses may be integrated in the soma in a fast timescale by electrical integration at the axon hillock and at a slower timescale by translational regulation (e.g., as mediated by FMRP, via a balance between phosphorylation/dephosphorylation status), leading to stoichiometric homeostasis of synaptic proteins and nuclear transcriptional state.