Genomics, convergent neuroscience and progress in understanding autism spectrum disorder

Abstract

More than a hundred genes have been identified that, when disrupted, impart large risk for autism spectrum disorder (ASD). Current knowledge about the encoded proteins — although incomplete — points to a very wide range of developmentally dynamic and diverse biological processes. Moreover, the core symptoms of ASD involve distinctly human characteristics, presenting challenges to interpreting evolutionarily distant model systems. Indeed, despite a decade of striking progress in gene discovery, an actionable understanding of pathobiology remains elusive. Increasingly, convergent neuroscience approaches have been recognized as an important complement to traditional uses of genetics to illuminate the biology of human disorders. These methods seek to identify intersection among molecular-level, cellular-level and circuit-level functions across multiple risk genes and have highlighted developing excitatory neurons in the human mid-gestational prefrontal cortex as an important pathobiological nexus in ASD. In addition, neurogenesis, chromatin modification and synaptic function have emerged as key potential mediators of genetic vulnerability. The continued expansion of foundational ‘omics’ data sets, the application of higher-throughput model systems and incorporating developmental trajectories and sex differences into future analyses will refine and extend these results. Ultimately, a systems-level understanding of ASD genetic risk holds promise for clarifying pathobiology and advancing therapeutics.

Autism spectrum disorder (ASD) constitutes a group of aetiologically and symptomatically heterogeneous, early-onset developmental disorders that are defined by core deficits in social communication and the presence of repetitive, stereotyped behaviours¹. Although these features are the diagnostic sine quo non, individuals come to clinical attention representing a very wide range of severity and often with co-occurring signs and symptoms that can include (but are not limited to) intellectual disability, anxiety, depression, sleep disturbance, epilepsy, specialized talents (also known as savant skills), delayed motor development and neurological soft signs. The inherent breadth of the presentation has led to an often-stated clinical pearl that “if you have seen one person with ASD, you have seen one person with ASD”. Indeed, the notion of an autism spectrum is a central feature of the diagnostic nosology.

However, despite the wide variety of clinical manifestations and the absence of reliable diagnostic biomarkers², ASD persistently ranks at the very top of the list of neuropsychiatric disorders with regard to relative genetic contribution^3–20. In fact, over the past 15 years, despite an extraordinary degree of aetiological and clinical heterogeneity, the pursuit of specific genetic contributors to ASD has been remarkably successful. Similarly to that of other relatively common disorders, the genetic architecture of ASD is now confirmed to include a wide range of variation with regard to type, frequency and effect size. However, the greatest promise for insight into the pathobiology of ASD presently is the result of striking progress in gene discovery based on the identification of large-effect, rare (often de novo) damaging variants in the coding portion of the genome^21–34.

The success in characterizing many rare large-effect mutations and concomitant gene discovery would reasonably be expected to set the stage for the rapid identification of key biological features and potential treatment targets. However, the extensive locus heterogeneity and biological pleiotropy of ASD risk genes, the developmental dynamism of the human brain along with its structural and functional complexity, the wide range of clinical outcomes associated with even large-effect genetic risks, the presence of marked sexual dimorphism and the inherent limitations of relying on evolutionarily distant model system to illuminate distinctly, if not uniquely, human characteristics have posed formidable barriers to translating genetic insights into an actionable understanding of underlying mechanisms.

However, as gene identification has progressed over the past decade, these successes have also ushered in a generation of ‘convergent neuroscience’ analyses aimed at illuminating core aspects of ASD pathology. By convergent neuroscience, we refer here to analyses that address the overlap, or intersection, of diverse ASD genetic risks with respect to molecular-level, cellular-level and circuit-level function as well as across multiple dimensions of analysis including anatomical localization and developmental timing^35–40. These conceptual approaches originated with the earliest successful efforts at ‘idiopathic’ ASD gene discovery⁴¹ and have often relied on Gene Ontology terms to offer initial insights, although contemporary approaches have begun to integrate other data, such as gene expression patterns from the human brain and maps of interactions between proteins and/or genes.

These studies have proven valuable in generating hypotheses about functional commonalities among ASD risk genes, providing consistent evidence for chromatin modification, synapse structure and function, RNA-binding proteins and several other points of mechanistic overlap^29,42–44. They have also generated hypotheses about spatial-level, temporal-level and cellular-level convergence of ASD genetic risk, repeatedly implicating mid-gestational developmental stages, the frontal cortex and developing cortical excitatory neurons. However, in evaluating the current state of the field with regard to the understanding of mechanism, it is important to consider the potential for cryptic biases in these types of analyses related to the selection and the sources of the underlying data. Nonetheless, with the accelerating development of foundational large-scale biological data sets, particularly those that derive from direct experimental evidence and include human data and developmental trajectories, convergence neuroscience analyses hold increasing promise for identifying core biological features and offer avenues to constrain key experimental parameters in investigations of the pathological sequelae of individual genes and mutations — a key challenge for studies of the developing brain that confront an extraordinarily complex and developmentally dynamic biological landscape³⁷. In addition to offering relatively hypothesis-free insights into genetic risks and related pathobiology, newer convergent approaches relying on highly parallelized modelling of rare variants carrying large risk for ASD also have the potential to identify previously unappreciated, yet widely applicable, mechanisms of resilience^39,40,45, a potentially promising avenue for the development of treatments impacting broad populations of individuals who do not necessarily carry mutations in the same gene or genes. In this Review, we summarize progress in the genetics and genomics of ASD, assess the state of the field with respect to findings from analyses employing a convergent neuroscience framework and consider the implications of progress in genomics and convergent neuroscience for the future of translational research.

Gene and locus discovery in ASD

Whole exome-based discovery.

The past two decades have been a period of striking progress in the identification of specific genes and loci carrying individually large risks for ‘idiopathic’ ASD (BOX 1). The first successful efforts at isolating bona fide risk genes can be traced to just after the turn of the millennium, with the identification of heterozygous damaging and putative loss-of-function mutations in the genes NLGN3 and NLGN4, through traditional cytogenetic mapping strategies⁴⁶ and parametric linkage analysis⁴⁷ followed by DNA sequencing. Soon thereafter, the rapid evolution of genomic technologies allowed for cost-effective, comprehensive scanning of the genome for rare and de novo mutations at increasingly high resolution (BOX 2). The critical observations that de novo structural and sequence variants contribute writ large to common forms of ASD^21,30,31,34, together with the insight that even a small number of exceedingly rare de novo variants mapping to the same gene or genomic interval in unrelated individuals offers sufficient statistical power to reliably identify risk loci^{22,24,33,48–51} and genes^{23–30,32,42} with limited confounds due to ancestry, constituted a critical turning point for the field, setting the stage for systematic and reliable gene discovery (FIG. 1). At present, microarray and whole-exome sequencing (WES) studies focused on rare variants have yielded convincing statistical support for the association of about a dozen copy number variant (CNV) loci²⁴ and more than 100 genes²³ (FIG. 1a), with this list rapidly growing (for example, see REF.⁵²) and many more of each predicted to exist^{22,23,28,30,33,53}.

Box 1 |. Idiopathic versus syndromic ASD.

There is debate about the definition of ‘syndromic’ forms of autism spectrum disorder (ASD) versus ‘idiopathic’ presentations. Syndromic generally refers to rare and severe conditions with a highly characteristic presentation that may have co-occurring ASD — for example, fragile X syndrome, Angelman syndrome, Rett syndrome, tuberous sclerosis, neurofibromatosis and PTEN hamartoma tumour syndrome. In practice, syndromic ASD tends to have a greater contribution of features apart from social communication deficits and repetitive behaviours, shows greater penetrance and tends towards monogenic causality. By contrast, idiopathic refers to cases that are not obviously accompanied by characteristic physical features or a pathognomic natural history. Although there may be considerable overlap in symptoms among these groups, and an increasing appreciation of intersecting biology, the history of gene identification efforts has diverged at the extremes of these distinctions. For example, there was substantial progress in the identification of syndromic loci and genes leveraging classic genetic methods, before the sequencing of the human genome and the genomic era^227–247. By contrast, reliable, systematic gene discovery in the larger group of individuals who commonly present clinically with ASD was highly dependent on the development of high-throughput genomic technologies focused on rare and de novo mutations, as discussed in the main text.

As gene discovery has progressed, and larger groups of affected individuals with rare variants in the same gene have been characterized, the boundaries between syndromic and non-syndromic have become less clear. Careful retrospective clinical analyses of patients whose ASD is apparently idiopathic and who share a mutated ASD gene or chromosomal segment have not infrequently led to the recognition of distinctive physical features and co-morbidities suggesting the presence of previously unappreciated syndromes^248–254. Therefore, at present, syndromic versus idiopathic ASD may be most productively thought of existing on a continuum both with regard to relative contribution of core features of the ASD diagnosis as well as with regard to the reliability of genotype–phenotype correlations.

Box 2 |. Major methods to identify ASD risk genes and loci.

Rare variant-based approaches (allele frequency <1%)

Whole-exome sequencing (WES).

This approach aims to identify rare, especially de novo, coding variants recurring in the same gene in unrelated patients more often than expected by chance (based on theoretical mutation rates, unaffected siblings or population frequency). These approaches have historically focused on ‘simplex’ families (unaffected parents with one child with autism), which tend to be enriched for de novo variants. However, rare inherited variants and case–control data from simplex, multiplex or unknown family types have increasingly contributed to gene discovery, especially as databases accumulate information on population variant frequency and the ability to stratify the most deleterious variants improves^58–60. As these databases grow, the definition of rare is beginning to shift to <0.1% or even <0.01%. Genes identified by this approach tend to carry large effect sizes (FIG. 1).

Whole-genome sequencing (WGS).

Conceptually similar to WES, these analyses expand to include non-coding variants. Identifying non-coding loci requires very large cohort sizes and is challenging due to incomplete characterization of regulatory loci (that is, functional boundaries, target gene), the relative inability to prioritize deleterious variants (especially compared with coding variants) and the tendency for these effects to be restricted to particular (yet currently unknown) developmental epochs, tissues and/or cell types. Accordingly, specific non-coding loci have not yet been identified at genome-wide significant thresholds.

Copy number variant (CNV) detection.

CNVs are deletions or duplications of DNA (also known as structural variants). The approach to detecting CNVs is conceptually similar to WES and WGS: identify rare and/or de novo CNVs from genotyping or sequencing data, and then search for recurrent variants overlapping the same locus in multiple unaffected individuals. Associated loci tend to overlap genes and carry large effects (FIG. 1). When combined with WES and WGS, small coding CNVs, especially deletions, can help identify individual risk genes²⁴.

Common variant-based approach (allele frequency >1%)

Genome-wide association.

This approach relies on common variants (also known as single-nucleotide polymorphisms (SNPs)), occurring at a frequency greater than 1% in the population. Variants are generally identified with genotyping arrays. Cases and controls are compared to identify SNPs with an allele that occurs more frequently in cases. Very large cohorts are required to identify risk loci as they carry small effect sizes (FIG. 1). Only five genome-wide significant loci for autism spectrum disorder (ASD) have been identified to date⁹² but this number is expected to grow as sample sizes increase.

Fig. 1 |. Relationship between effect size and allele frequency for loci discovered in autism spectrum disorder or schizophrenia.

a,b | Systematic studies of rare and common variants have identified risk loci for autism spectrum disorder (ASD) and for schizophrenia, but the trajectory of discovery differs between these conditions. In ASD, the majority of loci discovered to date are genes and have been identified by exome-wide sequencing studies of rare, generally de novo, coding variants (P < 2.5 × 10⁻⁶; red dots; for ASD, 26 genes²³; for schizophrenia, 10 genes⁶⁶), whereas in schizophrenia the majority are single-nucleotide polymorphisms (SNPs) and have been discovered based on genome-wide association studies (GWAS) of common variants (P < 5 × 10⁻⁸; teal dots; for ASD, 5 SNPs⁹²; for schizophrenia, 270 SNPs¹³⁰, of which only the 132 SNPs with relative risk >1 are shown). None of these genes (red) or SNPs (teal) overlap between ASD and schizophrenia. Both disorders have a similar number of loci associated based on rare, generally de novo, copy number variants (CNVs; blue dots; for ASD, permutation-based false discovery rate (FDR) ≤ 0.02, 10 loci²⁴; for schizophrenia, Benjamini–Hochberg FDR ≤ 0.02, 8 loci²²²). In contrast to rare coding variants and common variants, most of the top CNVs overlap between ASD and schizophrenia. In general, rare variants carry substantially higher relative risks than common variants. c | De novo damaging variants in the top 26 ASD risk genes tend to carry higher relative risk than de novo damaging variants within the top 7 schizophrenia risk genes (de novo damaging variants have been identified in only 7 out of the 10 schizophrenia risk genes), underscoring the particularly large contribution of rare variants to ASD. d | Utilizing a conservative FDR threshold instead of an exome-wide significant P value (FDR ≤ 0.01 versus P < 2.5 × 10⁻⁶) highlights 47 high-confidence ASD risk genes²³, with varying relative risks. In general, the most strongly associated genes carry the largest relative risks. See Supplementary Table 1 for a complete list of gene names, FDRs and effect sizes; see REF.²²³ for a broader list of ASD risk genes as determined by FDR. Note that we defined relative risk as the ratio of the frequency of a given variant in cases versus unaffected controls. Dot size is proportional to relative risk. Shaded area represents 95% confidence interval of the locally weighted least squares regression. Location on x axis in parts a,b is based on frequency in unaffected controls. As there are no associations in ASD for alleles with frequencies between 0.001 and 0.1, no curve is shown in this interval in part a. Damaging variants consist of protein-truncating variants (frameshift, canonical splice-acceptor, canonical splice-donor and nonsense variants) and missense 3 (Mis3) variants (those predicted to be probably damaging by PolyPhen-2 (REF.⁶²)). To estimate the per gene relative risk for de novo damaging variants, we defined the frequency in cases as the total number of de novo damaging variants observed in a given gene^23,66 and the frequency in controls based on estimates of the number of mutations expected per generation²³. All 26 genes identified in ASD have multiple de novo damaging variants in probands²³. However, only seven of ten genes identified in schizophrenia have de novo variants (the rest were identified by rare damaging variants in case–control data only)⁶⁶. Therefore, we estimated relative risk for de novo damaging variants when possible (red dots) and for rare damaging variants when necessary (orange dots). To estimate relative risk for rare damaging variants (case–control, schizophrenia only), we defined the frequency of each allele in cases based on reported frequency for schizophrenia⁶⁶ and the frequency in controls based on the frequency observed in the non-psychiatric subset of the gnomAD v2 data set⁵⁸. To estimate relative risk for rare CNVs, we defined the frequency in cases based on reported frequencies for ASD²⁴ and for schizophrenia²²² and the frequency in controls based on frequency observed in the Database of Genomic Variants (DGV)²²⁴ (ASD) or directly from controls (schizophrenia)²²². We estimated the 3q29(del) frequency in schizophrenia as 0.00005 because no 3q29(dels) were observed in 20,227 controls. The 15q13.3(del) is referred to as 15q13.2–13.3 in REF.²⁴ and we estimated its frequency as 0.000092 based on structural variants in gnomAD²²⁵ because of inconsistent frequencies reported in smaller studies in DGV. We estimated relative risk for common variants based on reported odds ratios for ASD⁹² or for schizophrenia¹³⁰. Part b adapted from REF.⁶⁶, CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).

The initial progress in ASD gene discovery relied heavily on the identification of rare heterozygous de novo mutations leading to putative loss of function of the encoded protein (also referred to as protein-truncating variants). However, missense de novo mutations³⁰ as well as rare inherited and somatic mutations (also known as mosaic mutations) have now also been demonstrated to carry risk^{23,24,29,53–57}. In addition, the dramatic expansion of databases of control exome and genome sequences has enabled the identification of ultra-rare genetic variation, the characterization of evolutionarily constrained genes (that is, those likely intolerant to variation) and the prioritization of missense variants most likely to be deleterious (for example, as determined by the MPC or ‘missense badness, PolyPhen-2 and constraint’ score)^58–62, all of which have demonstrated utility in parsing risk for ASD and, consequently, the potential to improve gene discovery^{23,29,55,60,63–65}. Indeed, characterizing ultra-rare variants among highly conserved genes in case–control data alone has emerged as an efficient, viable alternative to family-based studies designed to identify rare de novo mutations (for example, FIG. 1b; also REF.⁶⁶).

Likewise, there have been long-standing efforts to identify rare recessive variants and compound heterozygous variants contributing to ASD. Based on the strong track record of success relying on homozygosity mapping of consanguineous pedigrees in a wide range of severe neurological and neurodevelopmental disorders (NDDs), including brain malformation syndromes, severe intellectual disability and epilepsy^67–69, it has been somewhat surprising that the yield of novel genes from similar efforts have been relatively limited for the ASD phenotype, although there have been some notable exceptions^70–74. Additionally, several more recent studies have demonstrated an overall modest contribution of recessive inheritance, including compound heterozygous damaging mutations in non-consanguineous families^74–77.

Studies have also supported the contribution of rare somatic variation to ASD, suggesting that these variants may contribute to risk in approximately 5–22% of simplex cases^{56,57,78–83}. However, gene discovery based on somatic variants alone has been underpowered. Additionally, because these studies generally use data derived from the DNA of clinically accessible tissues such as peripheral blood and saliva^57,79–82, brain-specific somatic mutations will, by their nature, be missed. Identifying somatic variants within brain tissue from individuals with ASD is an active area of investigation⁵⁶, but the obstacles to this as a systematic approach to identifying novel ASD risk genes remain formidable.

Collectively, rare coding variants identified from WES not only have been a particularly valuable resource for biological studies but also have been shown to contribute to a substantial minority of individuals presenting for clinical evaluation. The estimates of clinical contribution range from a low of about 10% to a high of approximately one half of the clinical ASD population, depending in part on the ascertainment criteria. For example, increasing parental age, low IQ relative to mid-parental expectation, dysmorphology, epilepsy, congenital heart disease, female sex and simplex status — along with a larger number of unaffected siblings — are all known to lead to higher positive yields in genetic testing^{21,22,24,28–31,33,44,64,84–89}.

Whole-genome sequencing.

To date, the contribution of whole-genome sequencing (WGS) studies to identifying risk genes and loci for ASD has been limited compared with the approaches noted above. Capturing the full complement of variation across the entire genome, including both sequence and structural mutations, clearly represents the future of genomic analyses for ASD and, indeed, all of medicine. However, several factors limit its current utility: broadly, that the analysis and storage of WGS data require extensive computational infrastructure and confer substantial costs; the ‘search space’ for WGS is 100-fold greater than for WES, yet is less well characterized; the ability to interpret the consequences of non-coding variation is far more challenging than assigning functional significance to canonical coding mutations; and, as a group, rare non-coding variation is estimated to carry relatively more modest risks for ASD as compared to rare damaging coding mutations⁹⁰. Consequently, the cohort sizes needed to identify associated loci with confidence are estimated to be an order of magnitude or more larger than for contemporary WES studies, and, currently, no published investigations have yet realized sufficient numbers to identify, definitively, individual non-coding risk variants (BOX 2).

Common variant-based discovery.

Although most of the ASD risk loci confirmed to date have come from studies of rare coding variants of large effect, there is strong evidence that common variants of small effect carry the majority of population risk^4,5. Despite this, initial ASD genome-wide association studies (GWAS) yielded disappointing results owing to insufficient statistical power⁹¹. However, more recent studies, involving more than 18,000 cases and almost 28,000 controls, have successfully identified five common risk alleles of modest effect (BOX 2; FIG. 1a), meeting rigorous statistical criteria for association⁹². Undoubtedly, this number will continue to grow as case–control cohorts further increase in size⁹¹. Moreover, despite the relatively small number of loci implicated to date, polygenic risk scores, which are a means to quantify cumulative common genetic risks, have already demonstrated a modest ability to differentiate cases versus controls⁹². Additionally, polygenic risk scores appear to differ by sex^93,94, with a higher genetic load observed in females, providing evidence for a ‘female protective effect’ (FPE; discussed below) — a phenomenon that has also been widely observed with respect to rare variants^{22,24,27,28,30,32,33,95}. There has also been some success in identifying shared common genetic risks that cross a wide range of apparently disparate psychiatric diagnoses⁹⁶.

There is also evidence that among individuals carrying large-effect rare risk variants, common alleles and polygenic inheritance likely contribute to idiopathic ASD^97,98. Indeed, the question of the extent to which polygenic ‘background’ may shape the impact of rare mutations and help define an individual’s clinical presentation and natural history has been long-standing⁹⁹. The potential for very large, well-powered cohort studies in ASD makes this an increasingly tractable area of investigation^93,100. Similarly, as the yield of SNPs from ASD GWAS grows, important questions regarding the genetic architecture underlying various groups of individuals with autism, including those with so-called high-functioning autism, may well be clarified. Finally, the discovery of a much larger set of genome-wide significant common alleles will undoubtedly empower studies of convergence similar to those currently successfully leveraging rare, large-effect mutations.

The number of ASD-associated genes.

Contemporary association analyses of common, genetically complex and heterogeneous phenotypes such as ASD rely heavily on the application of rigorous statistical methods that effectively protect against false positive findings, which have historically bedevilled psychiatric genetic studies. Over the past decade, two broad approaches to the analysis of rare variants, using either a Bayesian or a frequentist framework, have emerged, providing the field with robust and consistent results. For WES studies specifically, it has become commonplace to require a discovery P value of less than or equal to 2.5 × 10^–6 — effectively correcting for the multiple comparisons among the approximately 20,000 genes comprising the human exome. Widely used alternative methods estimate false discovery rates (FDRs) per gene using a Bayesian approach, assigning significant association typically to genes showing FDR ≤ 0.1 and reserving a ‘highest-confidence’ designation to those with FDR ≤ 0.01. By contrast, GWAS arrived relatively quickly at a single consensus standard and threshold, namely a discovery P value of less than 5 × 10⁻⁸ followed by the finding of nominal significance in a replication cohort¹⁰¹. Importantly, the fact that there remain alternative accepted statistical approaches to WES studies of rare variants contributes to some degree of variability in gene ‘accounting’. For example, the recent omnibus analysis by the Autism Sequencing Consortium²³ is commonly cited as identifying 102 ASD risk genes, based on the Bayesian FDR threshold of 0.1 noted above. Using the alternative accepted exome-wide significant P value threshold (2.5 × 10⁻⁶) yields 26 risk genes (FIG. 1a) and restricting to the highest-confidence genes based on FDR ≤ 0.01 identifies 47 risk genes (FIG. 1d).

Additional definitional issues contribute to a range of views on what constitutes the definitive list of bona fide risk genes. For example, some widely used databases (for example, SFARI Gene^102,103) include genes discovered in studies of syndromic forms of ASD. Such approaches have the advantage of being more complete than those relying solely on large-scale association studies restricted to ‘idiopathic’ clinical samples (BOX 1). However, consensus approaches also introduce a degree of subjectivity and potential ascertainment bias that may be largely avoided by focusing only on genes that exceed predetermined statistical thresholds^23,24. The task of building a singular high-confidence ASD risk gene list is similarly made more challenging by studies that cross diagnostic categories. For instance, although approaches that combine ASD and other NDD cohorts increase the yield of gene discovery, they also complicate our understanding of the relative risk that each identified gene carries for ASD versus other NDDs. Further, results from studies that rely on targeted sequencing can be difficult to integrate with findings from WES or WGS, given the challenges of arriving at a single applicable statistical approach or threshold.

ASD versus other NDDs.

With recent success in gene identification across multiple NDDs and psychiatric disorders, the question of overlap of genetic risks among apparently distinct clinical syndromes has emerged as a key area of inquiry, for both conceptual and pragmatic reasons. Some of the areas currently of greatest interest in this regard include the degree to which ASD risks are distinguishable from the risks for schizophrenia or from intellectual disability. Both questions have historical relevance as well: autism was initially conceptualized clinically as a form of childhood-onset schizophrenia^104,105, but conventional wisdom subsequently shifted and the notion was formally rejected after the second edition of the Diagnostic and Statistical Manual of Mental Disorders¹⁰⁶. Similarly, it is now clear that early diagnostic criteria for autism tended to enrich for individuals with co-existing intellectual disability, resulting in a long-standing overestimate of the prevalence of intellectual disability in the population with ASD¹⁰⁷.

The question of whether and to what degree there is specificity to ASD genetic risk versus other developmental phenotypes remains a controversial topic. Successive generations of systematic CNV studies have shown that identical structural variations can contribute to multiple distinct diagnostic outcomes, including, but not limited to, epilepsy, schizophrenia, bipolar disorder, intellectual disability, attention deficit hyperactivity disorder, Tourette disorder and ASD (see also FIG. 1a,b). Moreover, as noted above, the yield of rare risk mutations from high-throughput sequencing studies of ASD cohorts has consistently been found to increase as IQ declines. Not surprisingly, these kinds of observations have fuelled debates over whether these diagnoses are truly separable at the genetic level^{6,92,108–112}. For CNVs, this variability in phenotypic outcomes has been particularly dramatic. From over a decade of progress in studies of ASD, schizophrenia, epilepsy and intellectual disability, it has become increasingly difficult to identify any structural variant associated with one of these diagnoses that is not also associated with some risk for one or more of the others — and the list of additional clinical outcomes showing overlap continues to grow^113–116. Although there is some evidence for differing relative representations of clinical manifestations for individual CNVs, including for duplications versus deletions of the same genomic interval²⁴, overall, the data have consistently pointed to the absence of diagnostic specificity^{6,108,110–112} (FIG. 1a,b).

These findings, not surprisingly, led to concern that the pursuit of biological mechanisms related to any given CNV would lack the ability to inform development of ASD therapeutics, as well as to speculation that individual genes of large effect would pose similar challenges. A related concern was, and remains, that the wide range of potential clinical outcomes from any given variant would pose serious obstacles to the design and execution of early intervention trials, including those targeting individuals with large-effect genetic risks. For example, it is not obvious how one would assess efficacy in a group of high genetic-risk individuals when the clinical course might evolve over weeks, months or years and include symptoms that vary dramatically from one person to the next, irrespective of treatment.

As both rare and common variant studies have progressed, a more nuanced view of the genetic architecture and loci contributing to ASD versus other neurodevelopmental syndromes is beginning to emerge. For example, in comparison with schizophrenia, much smaller cohorts have been required to identify large-effect risk genes in ASD. Conversely, much larger sample sizes have been required to identify the first genome-wide significant common variants in ASD versus in schizophrenia^92,117–119. Consistent with these observations, rare variants show, on average, a larger relative risk in ASD than in schizophrenia (FIG. 1a–c). Moreover, in a recent study of schizophrenia focused on identifying ultra-rare coding variants, none of the 10 genes meeting the most rigorous cut-off for statistical significance (P ≤ 2.5 × 10^–6) overlaps with the 26 genes identified at the same threshold in prior studies of ASD^23,66 (FIG. 1a,b). Moreover, an examination of a broader set of genes from this recent study that fell just below a threshold of FDR ≤ 0.1 for schizophrenia risk demonstrates that only 3 out of these 34 probable schizophrenia risk genes overlap with the 102 high-confidence ASD risk genes identified based on the same statistical cut-off^23,66. In a similar vein, recent detailed investigations of diverse mutations in a high-confidence ASD risk gene have shown a stronger relationship between individual alleles and specific phenotypic outcomes than would be anticipated based on the findings from genotype–phenotype studies of high-confidence CNVs¹²⁰. Moreover, given strong evidence that rare and common alleles combine to contribute to both disorders^93,100 and the current state of locus discovery overall — with a paucity of common alleles identified in ASD and rare alleles in schizophrenia — definitive conclusions about the presence or absence of specificity between ASD and schizophrenia genetic risks are likely premature.

There has also been long-standing interest in the overlap, both clinically and aetiologically, of ASD and intellectual disability. As noted, early diagnostic criteria for autism were biased towards the ascertainment of individuals with concomitant intellectual impairment. However, as the wider spectrum of ASD clinical manifestations has been increasingly appreciated, the percentage of individuals with ASD and without intellectual disability has increased markedly and now reflects the majority of those who carry an ASD diagnosis. Nonetheless, the frequent co-occurrence clinically of intellectual disability and ASD has prompted an ongoing debate over the separability of social versus intellectual functioning and has raised key questions as to whether recent gene discovery efforts in ASD cohorts have the potential to offer insight into the core biology of social disability^110–112. There has been a long-standing hypothesis that social ability is continuously distributed in the population and that higher IQs buffer behaviourally against autism symptoms^121,122. One contemporary interpretation of this idea is that the recent spate of large-effect mutations identified in WES studies primarily lead to lower IQ, which, in turn, limits an individual’s ability to compensate for social impairment mediated by common alleles. Along these lines, multiple analyses have noted that the rate of rare de novo mutations in ASD cohorts increases as IQ declines^33,44,64,88 and some have concluded that rare de novo sequence mutations are over-represented only in individuals with IQ below the mean⁶⁴.

By contrast, multiple studies have also shown that rare sequence variants and CNVs contributing large amounts of risk are found across the full range of IQ^23,24,33. Moreover, one of these studies evaluated ASD genes with regard to whether they showed evidence for being either ‘ASD-predominant’ or ‘ASD/NDD-predominant’ and found evidence for differing distributions of phenotypes corresponding to specific genes²³. Although the authors of the study did not conclude that any genes exclusively carried risk for ASD alone, their analysis was consistent with numerous observations suggesting that the rare variant genetics of human social and intellectual functioning are separable to some degree. For example, one study showed that duplications versus deletions of the identical Williams syndrome region on chromosome 7q11.23 leads, on average, to similar levels of intellectual impairment but also, typically, to strikingly divergent social phenotypes³³.

As progress in both common and rare variant studies paints a more complete picture of the allelic architecture of ASD, intellectual disability, schizophrenia and other psychiatric disorders, multiple related questions will become increasingly tractable. Studies examining the correlation of specific genes with differing distributions of diagnoses will have increased power to draw definitive conclusions. Moreover, the relationships between diagnosis, variant class (for example, protein-truncating variants versus missense variants¹²³), direction of effect (for example, duplication versus deletion and loss of function versus gain of function¹²⁰) and patient sex are likely to provide key insights into the mediators and moderators of whatever specificity is found to be present. Finally, although there is understandable consternation regarding the possibility that genetic risks for ASD, intellectual disability, epilepsy and other NDDs are inseparable, the high degree of comorbidity that is also observed, including for individuals with large-effect rare mutations in the highest-confidence ASD genes, may turn out to facilitate early therapeutic trials. For example, phenotypes apart from social functioning, such as seizures, motor development and overall intellectual functioning, are presently more amenable to quantification and to reliable longitudinal assessment, particularly early in development.

Future work in ASD genetics.

For those who recall the early days of gene discovery efforts in psychiatric disorders writ large, and in ASD in particular, the past 15 years of progress has been nothing short of spectacular. However, considerable opportunities remain: the methods that have been most successful in elaborating the growing list of risk genes introduce some biases owing, in part, to differential detection of types of variants. For example, the sex chromosomes have been largely unexplored to date, owing to complexities in accurate variant calling at scale. In addition, their distinctive inheritance patterns and selection pressures preclude the use of current metrics (for example, the probability of being loss-of-function intolerant) that have accelerated gene discovery on the autosomes. It is also likely that the relative sparsity of female samples in existing research cohorts, especially from simplex families, has limited the ability to detect sex-specific or sex-biased risk associated with specific large-effect variants. Similarly, missense mutations, gain-of-function mutations, non-canonical splice site mutations, somatic mutations and the universe of non-coding variants remain far more difficult to interpret at scale than canonical loss-of-function mutations. Recessive variants, inherited variants and variants with low probability of being loss-of-function intolerant also remain difficult to identify, although network-based approaches have begun to fill in these gaps, by identifying putative novel risk genes based on proximity to strongly associated risk genes^54,124,125.

Similarly, efforts to identify genotype–phenotype relationships have so far been met with limited success^24,126 but this is a highly promising area for future investigation. Hypotheses regarding the utility of endophenotypes to increase the power of gene discovery efforts have not panned out¹²⁷. Thankfully, however, the well-known limitations of categorical psychiatric diagnoses combined with the extreme clinical heterogeneity of ASD have not thwarted the identification of dozens of large-effect genes and the first set of common risk alleles. The combination of increasingly large cohorts combined with the ability to work from a well-established list of definitive genes and mutations promises to provide insight into questions, such as the relative specificity of genetic risks, that have long been vexing to investigators. As foundational network data characterizing the relationship between diverse risk genes and the proteins they encode are generated, so-called ‘network–phenotype’ analyses may increase power to extract meaningful relationships between risk variants co-localizing in network space and patient phenotypes^37,124,128. That being said, despite the abundance of genes identified based on rare and de novo mutations, there is a strong rationale to continue to pursue common variant analyses. Although studying individual small-effect non-coding alleles may not be as direct a path to exploring core biology, they account for a substantial proportion of population risk and the interplay of common and rare variant risk is now a tractable area of inquiry that could have profound importance^124,129,130. The ability to evaluate the combined contributions of variation across the frequency and effect spectrum promises to yield critical insights into natural history and treatment response, and potentially to contribute to diagnostic strategies especially aimed at very early detection.

Progress in convergent neuroscience

The progress made to date in gene and locus discovery in ASD has set the stage for understanding the role of these genes and loci in human brain development and function, and for clarifying how disruptions can lead to the emergence of social impairments. However, the extensive biological and locus heterogeneity revealed by contemporary genomic analyses has, simultaneously, led to concern that any individual gene or mutation, and consequently any related therapy, might be relevant only for a fleetingly small number of individuals with autism. At the same time, from the earliest successful studies, there has also been evidence that a larger number of apparently functionally diverse ASD risk genes tend to converge on a smaller number of biological pathways, developmental stages, brain regions and cell types. Indeed, steady progress has been made in specifying and defining these points of biological overlap through reliance on relatively hypothesis-free approaches and on the expanding armamentarium of developmentally contextualized neurobiological data resources, spanning levels of analysis, construct-valid models and human brain tissue.

The problem of pleiotropy.

A major challenge in translating successful gene discovery in ASD into a reliable understanding of pathobiological mechanisms — and the identification of related treatment targets — has been the pleiotropy of genes, that is, their tendency to be involved in different roles throughout brain development and function, potentially dependent on, for example, cell type, brain region and/or developmental stage^35,37,38. Consequently, although an identified risk gene may have a well-studied and well-characterized role in a biologically plausible process, it does not necessarily follow that either this is the only function that the gene plays or it is necessarily the role that contributes to the emergence of ASD.

The challenge of pleiotropy is, of course, also highly relevant in the study of animal models. Recapitulating a mutation in a human ASD risk gene in, for example, rodents, has become routine, as has the ability to characterize a very wide range of neurobiological sequelae, including in awake behaving animals. However, the observation of even the most plausibly ASD-relevant phenotype in a construct-valid mutant does not, on its own, provide reassurance that the observation is pointing to causal mechanisms in the human syndrome — a problem that is compounded by the well-established challenges of relying on face-valid behavioural phenotypes in evolutionarily distant species to illuminate psychiatric syndromes³⁵. Similarly, even for a single gene, limitations in the ability to assay molecular and cellular phenotypes in a hypothesis-free manner across the entire brain at high resolution in more complex models introduces inherent analytical biases. In this regard, having strong evidence regarding when and where to look for ASD-related pathobiology is potentially invaluable in sorting through an immensely complex and dynamic biology. These types of considerations have motivated convergent neuroscience approaches over the past several years. These studies assess risk genes in parallel, in a hypothesis-free manner, to identify and confirm shared developmental epochs, brain regions, specific cell types and biological pathways associated with human vulnerability to social impairment as a prelude to resolving the pleiotropy of ASD risk genes³⁷.

Incomplete gene ontologies.

Convergent neuroscience approaches are based on the hypothesis that subsets of risk genes for a given disorder, such as ASD, will point to shared vulnerability somewhere along the path from gene to complex behaviour. This logic is reflected in the earliest sequencing studies in ASD as successful gene discovery was immediately followed by a search for functional overlap. Many of the first genes associated with ASD based on targeted sequencing studies were known to encode proteins localized to the synapse (for example, NRXN1, SHANK3, NLGN3 and NLGN4)^46,131,132. Not surprisingly, as a result, the field focused on the structure and function of the synapse as a convergent point of ASD-relevant biology^41,133,134. As systematic gene and locus discovery proceeded, prima facie evidence for convergence was also found for genes involved in chromatin modification and transcriptional regulation^{27,28,30,33,43}. As risk gene lists have grown and the vast degree of allelic and locus heterogeneity has been revealed, enrichment analyses (for example, Gene Ontology enrichment analysis and gene set enrichment analysis) have been invoked to identify statistical over-representation of molecular functions, cellular components, biological processes and other gene lists from various pathway databases^43,135–140. These analyses have repeatedly highlighted gene targets of FMRP and the β-catenin pathways^28,34,42 and confirmed significant enrichment of Gene Ontology terms associated with chromatin modification and synaptic function^{23,24,29,41,43,133,134}. However, it is important to note that the underlying data that support Gene Ontology terms and other annotations may offer an incomplete and biased view of gene function and that these approaches are generally unable to account for pleiotropy and/or missing data^37,141. The annotations underlying these terms are often created by aggregating existing data, a process that can skew towards categories of function and, potentially, cell types and developmental time points that are more commonly or readily studied, while minimizing or omitting others^142,143. For example, given several decades of effort aimed at characterizing neurons and a particular emphasis on illuminating the structure and function of the synapse, it is not surprising that ontologies tend to be relatively enriched for related annotation terms. These types of inherent, and often cryptic, ascertainment biases can cascade into self-reinforcing, but potentially incomplete, biological interpretations.

Indeed, recent analyses of some of these same genes annotated at the synapse have shown experimental evidence of functional convergence that precedes synapse formation — for example, in neurogenesis^40,144. Of course, these genes probably have multiple roles during brain development with their observed function dependent, in part, on spatiotemporal parameters. In addition, it is possible that both processes are relevant to ASD and interrelated, with synapse dysfunction following from earlier developmental derangements. To address the inherent limitations of Gene Ontology analyses, recent work has increasingly focused on the development of highly parallelized, systematic, experimental characterizations of relevant genes and/or proteins across developmental stages, tissues and cell types. Similarly, data sets developed through the manipulation of known ASD genes within intact biological systems also offer a basis for convergence analysis that mitigates some of the limitations of ontology-based approaches. These efforts are described in more detail below.

Examining neurotypical brains.

Several groups have taken advantage of the growing list of reproducibly associated risk genes to identify anatomical regions, cell types and/or developmental stages of the typically developing brain of particular relevance to ASD. Conceptually, these studies are based on two underlying hypotheses: first, genes expressed in a highly coordinated spatiotemporal fashion probably share functions; and, second, identifying points of significant overlap among a group of disparate ASD risk genes in the context of typical development is a means of triangulating key parameters associated with core pathobiology. These hypotheses highlight the importance of the nature of the input data: these analyses are dependent on gene lists derived in a systematic, unbiased fashion. For example, the potential for circular reasoning introduced by investigating convergence among a set of ‘candidate’ genes identified based, in part or in whole, on biological plausibility is obvious. For this reason, the emergence of widely accepted statistical thresholds in genome-wide and exome-wide studies of idiopathic ASD cohorts to identify gene lists has been a key to the development of meaningful convergence analyses.

Some of the earliest efforts in this vein used network-based approaches to analyse data from the BrainSpan Atlas of the Developing Human Brain, which generated transcriptomic data from early gestation to late adult stages across multiple anatomical regions of ‘neurotypical’ human brains¹⁴⁵. Several studies leveraged these data to develop spatio-temporally defined co-expression networks and then examined their relationship to a set of ASD genes meeting specified statistical thresholds for association. The observation of a statistically significant over-representation of risk genes mapping with co-expression networks or of greater than expected expression correlation between ASD genes within networks was considered evidence for convergent vulnerability. Three groups published the first examples of these approaches leveraging somewhat different gene lists and approaches to network generation. Strikingly, despite these differences and the sparse lists of associated genes at the time, all three groups identified evidence for convergence in prenatal development^25,146,147, with two resolving this further to mid-gestational prefrontal cortex (PFC) and developing excitatory neurons^25,146. More recent studies — using similar or divergent methods, rapidly expanding gene lists and more comprehensive foundational transcriptomic resources — have consistently replicated these findings and, additionally, provided evidence for involvement of the cerebellum and striatum^{126,148–153} (FIG. 2a).

Fig. 2 |. Emerging patterns of ASD convergence identified from gene expression data.

Numerous studies investigating human gene expression data sets for convergence related to autism spectrum disorder (ASD) risk have been conducted, spanning so-called ‘neurotypical’ (control) brain samples (parts a–c) as well as post-mortem brain samples from individuals with ASD (parts c,d). These data sets encompass bulk tissue from major regions of the brain (parts a,d), bulk tissue from more finely dissected regions of the developing cortex (part b) and single cells from the cortex (part c). Neurotypical data tend to include prenatal and postnatal samples across the entire spectrum of human brain development whereas patient-derived data include postnatal samples only (ASD diagnosis requires behavioural assessment). Generally, approaches utilizing neurotypical expression data seek to identify convergence by searching for ‘enrichment’ of ASD risk genes in particular developmental epochs, brain regions and/or cell types. In this case, enrichment is indicated by the ‘strength’ of co-expression of ASD risk genes or by ‘high’ or ‘specific’ expression of ASD risk genes. By contrast, approaches utilizing patient expression data generally aim to identify convergence by characterizing systematic differences between cases and matched controls (that is, genes consistently differentially expressed in brain tissue from individuals with ASD). Although data and approaches differ substantially across studies, the developing frontal cortex and excitatory neurons are recurring points of enrichment. a | Studies investigating spatiotemporal convergence of ASD genetic risk in gene expression data from bulk samples of neurotypical brains have consistently highlighted mid-gestational frontal and parietal cortices^{23,25,126,129,146,148,149,153}. We summarize the extent of replication by color intensity, with the strongest replication indicated by dark red and the weakest by light red. These studies varied widely in granularity of developmental periods and brain regions studied, and therefore we considered studies that examined two or fewer regional (for example, grouping all cortical regions, studying only one brain region) or temporal (for example, prenatal versus postnatal) contexts to provide ‘limited’ evidence. Strong replication denotes a significant finding in at least three ‘non-limited’ studies, moderate replication indicates a significant finding in two non-limited studies and weak replication corresponds to a significant finding in one non-limited study and at least one limited study. We define early gestation as 8–10 post-conception weeks (PCW), early mid-gestation as 10–16 PCW, late mid-gestation as 16–24 PCW and late gestation as 24–38 PCW. b | Based on findings summarized in part a, several studies have examined laminar convergence of ASD genetic risk in neurotypical early mid-gestational frontal cortex. An initial study assessing preservation of spatiotemporal co-expression networks by layer implicated the inner cortical plate²⁵ whereas a contemporaneous study assessing specificity of expression of ASD gene lists by layer indicated a relative lack of layer-specific enrichment¹⁴⁶. A later study incorporating orthogonal gene interaction data and assessing ASD gene network connectivity by layer replicated the inner cortical plate finding and newly implicated the inner subventricular zone and subplate⁴⁰. c | Analyses of single-cell gene expression data from neurotypical controls and individuals with ASD have consistently highlighted excitatory neurons. In neurotypical mid-gestational cortex, ASD risk genes tend to be highly and/or specifically expressed in developing excitatory and inhibitory neurons (progenitors and maturing neurons)^{23,54,151,157,159}. Analyses of single-cell gene expression data from (postnatal) cortex of individuals with ASD have similarly identified consistent differences between cases and controls in excitatory neurons, as well as potential differences in microglia and oligodendrocytes^166,167. d | Analyses of bulk samples of brain tissue from individuals with ASD compared with those from matched neurotypical controls support the major findings from parts a–c. These studies implicate frontal cortex, excitatory and inhibitory neurons, and microglia, and additionally highlight differences in genes associated with astrocytes and in expression modules related to synaptic function and inflammation^{162–165,226}. CBC, cerebellar cortex; DFC, dorsal frontal cortex; iCP, inner cortical plate; iSVZ, inner subventricular zone; ITC, inferior temporal cortex; OFC, orbital frontal cortex; oSVZ, outer subventricular zone; PFC, prefrontal cortex; URL, upper rhombic lip. Some of the brain images in parts a,b and d are adapted with permission from REF.¹⁹⁹

Based on the aforementioned reproducible association of the developing mid-gestational frontal cortex, analyses of the more fine-grained BrainSpan prenatal laser micro-dissected data set¹⁵⁴ have similarly been used to identify points of laminar convergence for ASD risk genes in the developing cortex, through the analysis of the preservation of network connections by layer²⁵, the over-representation of associated gene lists¹⁴⁶ or the degree of connectivity of ASD-associated interaction networks⁴⁰ (FIG. 2b). Early analyses implicated the inner cortical plate (iCP)^25,146, and a recent analysis using more than 100 ASD genes and a comprehensive database of molecular interactions confirmed this finding and further identified the inner subventricular zone (iSVZ) and subplate⁴⁰. Developmentally, a key process occurring in mid-gestational cortex is neurogenesis, with neural progenitors actively undergoing cell division and differentiation in the subventricular zone, leading to the generation of the postmitotic excitatory neurons of the cortex (migrating through the subplate and into the cortical plate) (FIG. 2b). The findings from the biological analyses of these systems are also consistent with findings from orthogonal enrichment analyses^{23,25,29,146,148}, as well as emerging evidence from model system studies, as noted in later sections.

Recent advances in single-cell RNA sequencing (scRNA-seq) and single-nucleus RNA sequencing technology have enabled the generation of high-resolution expression data at the level of single cells or nuclei from the typically developing prenatal cortex^{151,155–158}. Analyses of scRNA-seq data tend to first identify the cell types present based on clustering patterns of individual cells and expression of cell type-specific marker genes, and then assess each cell type for concurrent (that is, co-expression), high or specific expression of ASD risk genes. These studies have again implicated excitatory progenitor cells and maturing excitatory cortical neurons, and have additionally implicated inhibitory progenitor cells and maturing inhibitory neurons of the cortex^{23,54,151,157,159} (FIG. 2c, left).

Finally, these observations of convergent expression of ASD risk genes in excitatory and inhibitory neurons and the developing cortex have also been reproduced in analyses of expression data from wild type model organisms. In one of the earliest studies, analysis of layer-specific expression data from developing mouse brains supported the idea that ASD-associated genetic risk converges in developing deep layer cortical excitatory neurons²⁵. A contemporaneous study leveraging gene expression data from adult primate cortex¹⁶⁰ observed similar enrichment in excitatory neurons, although the signal was stronger for more superficial layers¹⁴⁶. Two additional studies in mice generated and analysed a data set with gene expression profiles for 25 distinct CNS cell types and observed enrichment of ASD-associated genes in cortical excitatory neurons, as well as inhibitory neurons of the cortex and striatum^126,148,161. A recent study using Xenopus tropicalis observed that ten ASD risk genes chosen solely based on the strength of statistical evidence for association are co-expressed in the developing telencephalon at time points that correspond to human mid-gestational periods in the frontal cortex⁴⁰. These non-human studies suggest that expression convergence among ASD genes is at least partially conserved through evolution, although this needs to be examined in more detail and across additional models.

Comparative post-mortem studies.

An alternative approach to identifying convergence of ASD risk involves comparing the molecular profiles of post-mortem brain tissue from individuals with ASD and from neurotypical controls. One of the first published studies of this type profiled the transcriptomes of ASD and typically developing brains across the PFC, temporal cortex and cerebellum, and observed that regional expression differences identified in control brains were attenuated in the frontal and temporal lobes of individuals with an ASD diagnosis¹⁶² (FIG. 2d). This observation was replicated in a subsequent PsychENCODE study with larger cohort sizes¹⁶³. These studies and others have also identified downregulation of neuron and synapse-related genes and upregulation of microglia and immune-related genes in the brains of individuals with ASD compared with controls^162–166 (FIG. 2d).

scRNA-seq has provided some additional granularity to the observed differences between brains of individuals with ASD and neurotypical brains (FIG. 2c, right). Experiments leveraging data from scRNA-seq to deconvolve bulk RNA-sequencing (RNA-seq) expression data have shown that the majority of expression variation across brain tissue samples is attributable to varying proportions of basic cell types, and that that the cortex of individuals with ASD may have higher fractions of excitatory neurons, microglia and astrocytes, and lower fractions of oligodendrocytes¹⁶⁷. This suggests that prior observations, from comparative bulk RNA-seq studies, of alterations in synapse-related and immune-related genes may derive from changes in cell representation rather than from large alterations in gene expression within given cell types^162–164.

Overall, studies that compare the transcriptome of ASD and neurotypical brains have repeatedly noted that there may be altered cortical patterning and changes in relative abundance and/or function of excitatory neurons and microglia (FIG. 2d). As noted above, these findings are largely congruent with data from studies searching for convergence of ASD risk gene expression in neurotypical brain tissue, in that both implicate the frontal cortex and excitatory neurons in ASD risk. However, they differ in that neurotypical data implicate inhibitory neurons but not microglia^{23,25,54,126,129,146,148,151,153,157,159} (FIG. 2c). There are multiple potential explanations for these divergent findings, including that they may reflect some of the inherent practical limitations of studies involving comparisons of affected versus unaffected post-mortem tissue. For example, brain bank resources from individuals with ASD remain quite limited and, therefore, most studies involve very small sample sizes and have attendant challenges with statistical power. Additionally, those samples that are available are often highly heterogeneous in characteristics such as age and sex, as well as with regard to important covariates such as medication history, post-mortem interval, agonal events and/or time of collection^{165,168–170}. Moreover, as the diagnosis of ASD is based on behaviour, these studies involve postnatal samples only, generally from patients 2 years of age or older, raising the critical issue of how to relate postnatal changes in cases versus controls to pathobiology of a developmental disorder with strong evidence for a nexus of prenatal risk, such as ASD. Given the natural history, with symptoms emerging typically in the first 1–2 years of life, as well as the evidence for the relevance of prenatal events in syndrome aetiology, it is plausible and, indeed, likely that observations in postnatal tissue may capture events downstream of primary pathobiology and may be offering insights into ‘effect’ rather than ‘cause’. Although it would be ideal to perform differential expression analyses in prenatal affected versus neurotypical tissue, these types of experiments pose formidable logistical challenges, including as a result of ASD being a clinical and behavioural diagnosis that can only formally be made postnatally. Although one might consider experiments leveraging prenatal post-mortem tissue carrying large-effect ASD-associated mutations without relying on clinical diagnosis, the observation of the wide range of potential diagnostic and phenotypic outcomes from even the largest-effect risk genes would place limits on the ability to interpret findings with respect to ASD pathology specifically.

Patient-derived model systems.

Functional and transcriptional profiling has also been conducted on human induced pluripotent stem cell (iPSC)-derived monolayer models of cell types from the developing brain as well as in 3D models of human brain development (also known as organoids and spheroids)¹⁷¹. Over the past few years, there have been several studies comparing neurons generated from iPSCs derived from donors with ASD versus those from neurotypical individuals. Some recurrent findings in iPSC-derived neurons from patients with ASD and macrocephaly, compared with controls, include increased cellular proliferation, accelerated neuronal differentiation, impaired synapse development and decreased spontaneous and synchronous neuronal activity^172–174. Interestingly, a recent study showed that iPSCs derived from individuals with ASD but without macrocephaly exhibited impairments in neurogenesis compared with those from neurotypical individuals¹⁷⁵. Thus far, systematic profiling of iPSCs derived from individuals with idiopathic autism have been limited by small sample sizes (on the order of ten individuals carrying an ASD diagnosis) and varied timing and methods of profiling cellular phenotypes^172–178. Given the aetiological heterogeneity of the syndrome, the presence of comorbid phenotypes (such as macrocephaly), the diverse genetic backgrounds of probands and controls, and the variability in derivation of iPSC lines and downstream cell types, it is not yet clear whether these findings will replicate in larger, better-powered analyses.

Transcriptomic and epigenetic profiling of iPSC-derived telencephalic organoids has identified ASD-associated co-regulated gene modules related to synaptic function that overlap with those previously identified in studies using post-mortem brain tissue^146,163,178. Recently, organoid models of developing forebrain generated from individuals with ASD and from neurotypical controls were profiled for chromatin accessibility and gene expression over time¹⁷⁹. Overall, ASD genetic risk mapped to glial progenitor cells, mid-stage excitatory neurons and late-stage excitatory and inhibitory neurons. These results are therefore largely consistent with the aforementioned findings from human brain. It is important to note, however, that numerous distinctions exist between organoids and primary tissue, including limited spatial organization, impaired cell-type maturation, the absence of cell types found in primary tissue and the fact that organoids may be significantly impacted by factors related to oxygen diffusion and stress^180–182. As organoid technology continues to develop (for example, see REFS^183–185), it will be an increasingly promising avenue for a more granular understanding of the molecular-level and cell type-level changes taking place during early neurodevelopment, including with respect to ASD.

Parallelized modelling of genetic risk.

The efforts to identify where and when ASD pathology arises have already provided important insights into underlying biology and now offer opportunities for parallel investigations of ASD risk genes that constrain or define key experimental parameters including cell types and developmental stages of interest. In addition, in recent years, advances in CRISPR-based technology have facilitated the parallelized study of multiple ASD-associated genes in vitro and in vivo to identify points of functional convergence. Initial in vitro studies have characterized the effect of perturbing roughly a dozen ASD-associated risk genes in neurons derived from human iPSCs or embryonic neuronal precursor cells^186,187. These initial studies suggest that perturbation of ASD genes results in impaired neuronal differentiation and reduced spontaneous synaptic activity, and highlight the feasibility of assessing the functional consequence of multiple ASD genes in parallel. Going forward, advances in pooled CRISPR screening will enable hypothesis-naive functional screening of larger numbers of genes with a wide range of read-outs, including proliferation, differentiation and cellular function as well as gene expression and chromatin accessibility at single-cell resolution^188–195. Patient-derived iPSC model systems still tend to suffer from small sample sizes in addition to variability among iPSC lines and heterogeneity in the derived cell populations¹⁹⁶. In this regard, systems with isogenic or comparatively homogeneous genetic backgrounds are an important addition to the experimental armamentarium, as they moderate some of the challenges of interpreting subtle biological effects across varied genetic backgrounds^{39,40,197 —} although the generalizability of findings in isogenic backgrounds needs further investigation, as does the effect of sex (see below).

Studying ASD risk genes in vivo in intact biological systems is critical to model dynamic neurodevelopmental contexts, circuit-level function and higher-order phenotypes. However, generating animal models of individual ASD genes is labour and time-intensive. Consequently, historically, studying multiple ASD genes in parallel to identify points of functional convergence has been challenging. Increasingly, these challenges can be mitigated through the use of pooled screening approaches or using model organisms that allow for scalable arrayed screening. In a recent study, an in vivo pooled CRISPR–scRNA-seq system (in vivo Perturb-seq) that combines genetic editing in the developing mouse brain with postnatal scRNA-seq was used to identify the cell type-specific functional consequences of perturbing 35 ASD risk genes during brain development³⁹. The authors observed overlapping gene expression changes linked to perturbation of small subsets of ASD genes (up to four) in different cell types, hinting at cell type-specific functional convergence. Importantly, this study highlights the feasibility of pooled in vivo CRISPR screens with single-cell profiling to interrogate many risk genes for functional convergence across multiple cell types. However, there were no clear convergent findings involving large numbers of genes. Larger-scale experiments with greater numbers of cells and multiple time points across brain development may be better powered to identify convergent gene modules within specific cell types.

Our group recently leveraged CRISPR alongside the X. tropicalis model system to study the effects of loss-of-function mutations in ten ASD risk genes in parallel in an arrayed format⁴⁰. Strikingly, perturbation of all ten ASD risk genes altered telencephalic size and impaired neurogenesis. This convergent neurogenesis phenotype, spanning ASD genes with apparently disparate cellular functions, was conserved in human 2D and 3D in vitro models of brain development. Overall, these observations are in line with prior findings that ASD risk converges in mid-gestational telencephalic excitatory neurons, and provide strong evidence for impaired neurogenesis as a point of vulnerability, as has been previously hypothesized¹⁹⁸. These findings underscore the power of Xenopus for productive functional genomics studies, in particular highlighting the ability of medium-throughput parallelized in vivo studies to identify convergent mechanisms that can then be prioritized for validation and extension in human systems¹⁹⁹.

Sex differences and genetic risk

Although, in general, the field is enjoying success in elaborating points of functional and spatiotemporal convergence among a rapidly expanding list of autosomal ASD genetic risks, far less is understood about the nature of the contribution of sex to ASD aetiology. A strong male sex bias has consistently been observed for ASD, but this remains largely unexplained²⁰⁰. The male to female ratio has been consistently estimated to be around 4:1 (REF.²⁰¹), although estimates of male predominance for individuals with autism with normal to high IQ tend to be even higher. Similarly, although the findings have been roughly consistent across studies, it is also clear that these estimates are vulnerable to several potential confounds beyond IQ, including ascertainment methods and potential diagnostic bias^200–202. Still, studies that have endeavoured to account for these issues have, nonetheless, found consistent evidence for male bias in ASD risk^95,202,203. One putative avenue of explanation for these findings is the FPE, evidence for which derives mostly from genetic studies that show enrichment of risk in females versus males — that is, female probands tend to carry a relatively increased burden of rare damaging de novo and transmitted mutations, including CNVs and sequence variants^{23,24,29,31,33,44}, as well as rare complete knockouts⁷⁶. Additionally, females — including unaffected mothers of children with autism — also appear to be enriched for common variant risk^93,94. Similarly, unaffected female siblings of children diagnosed with ASD appear to carry an increased rate of rare variants^{22–24,27,28,30,32,33,95} and appear to inherit more polygenic risk^93,94, as compared with male siblings. These observations support the hypothesis that females are more resilient to genetic risks, resulting in the observation of higher thresholds for females to become symptomatic and meet diagnostic criteria. Interestingly, to date, there has been no evidence of sex-biased risk for individual autosomal genes²⁴ and rare genetic risks mapping to the sex chromosomes have been relatively unexplored owing to technical and methodological barriers.

In contrast to the types of data noted above, family-based and twin-based studies of recurrence rates have provided less clear evidence for the FPE. Younger siblings of females diagnosed with ASD may be more likely to receive an ASD diagnosis than younger siblings of males diagnosed with ASD^94,204; however, there are conflicting data in this regard and these analyses are constrained by limited statistical power owing to the relatively small female cohort sizes^95,153,203.

There is also the possibility of male-specific ‘risk’, which is not mutually exclusive with the FPE model. For example, males with ASD have an increased rate of hemizygous rare loss-of-function variants on chromosome X⁷⁶. However, the sex chromosomes, in general, appear to be a minor factor in determining risk versus resilience^31,205,206, although they also tend to be omitted from large-scale genetic studies owing to technical difficulties and complexities in accurate variant calling due to confounding sequence homology between sex chromosomes and differences in ploidy by sex. Therefore, it is possible that there are as yet undiscovered factors on the sex chromosomes.

These findings suggest that sex differences are a critical aspect of the biology of ASD and a potentially promising avenue for the development of broadly applicable therapeutics, especially in light of the possibility that unravelling underlying mechanisms may illuminate resilience as well as risk factors. The lack of mechanistic understanding of sex bias is probably a consequence of the relative paucity of knowledge regarding differential molecular and cellular mechanisms in male and female brain development, inherent limitations in commonly employed model systems for interrogating these differences and the lack of any immediately obvious explanation for sex bias in otherwise highly successful gene discovery efforts. Identifying robust differences in ‘neurotypical’ male and female brain development will probably facilitate understanding of risk for many psychiatric disorders beyond ASD, as many show some degree of sex bias²⁰⁷. Rich developmental human brain omics data sets from both males and females (see TABLE 1 for a list of currently available gene expression data) will be a critical resource in efforts to identify these differences.

Table 1 |.

Available human developmental brain expression data sets for studying sex differences

Resource	Brain regions	Developmental age	Sample size (female, male)	Technology	Number of cells	Year
Colantuoni et al.218	PFC	12–18 PCW, birth–78 years	269 (92, 177)	Microarray	NA	2011
BrainSpan microarray145	16 (12 cortical)	5 PCW–82 years	57 (26, 31)	Microarray	NA	2011
BrainSpan LMD154	~300 cortical regions (LMD of 6 layers)	15–21 PCW	4 (3, 1)	Microarray	NA	2014
BrainSpan RNA-seqa	16 (12 cortical)	8 PCW–40years	42 (19, 23)	BulkRNA-seq	NA	2013
O’Brien et al.219	Entire brain hemisphere	12–19 PCW	120 (50, 70)	BulkRNA-seq	NA	2018
Li et al. (PsychENCODE)151	16 regions (11 cortical)	8 PCW–40years	41 (18, 23)	BulkRNA-seq	NA	2018
Walker et al.152	Unspecified cortex	14–21 PCW	201 (84, 117)	BulkRNA-seq	NA	2019
Werling et al. (BrainVar)153	1 region (DPFC)	6 PCW–20 years	176 (72, 104)	BulkRNA-seq	NA	2020
Pollen et al.155	Unspecified cortex (MD of 2 layers)	14–16 PCW	3b(3, 0)	scRNA-seq	393	2015
Nowakowski et al.156	2 cortical regions (MD of 2 layers), MGE	5–37 PCW	48b (21, 27)	scRNA-seq	4,261	2017
Zhong et al.220	PFC	6–24 PCW	12b (9, 3)	scRNA-seq	2,309	2018
Fan et al.159	22 regions (20 cortical)	22–23 PCW	3 (2, 1)	scRNA-seq	4,213	2018
Li et al. (PsychENCODE)151	Cortex (fronto-parietal)	5–20 PCW	9 (3, 6)	scRNA-seq	1,195	2018
Polioudakis et al.157	Unspecified cortex (MD of 2 groups of cortical layers)	15–16 PCW	4 (3, 1)	scRNA-seq	33,976	2019
Bhaduri et al.182	7 regions (all cortical)	4–20 PCW	5b(4,1)	scRNA-seq	189,409	2020
Eze et al.158	18 regions (3 cortical)	4–8 PCW	10b (7, 3)	scRNA-seq	289,000	2021
Trevino et al.221	Unspecified cortex	16–24 PCW	4b(3, 1)	scRNA-seq	57,868	2021

DPFC, dorsolateral prefrontal cortex; LMD, laser micro-dissection; MD, micro-dissection; MGE, medial ganglionic eminence; NA, not applicable; PCW, post-conception weeks; PFC, prefrontal cortex; RNA-seq, RNA sequencing; scRNA-seq, single-cell RNA sequencing.

^aBrainSpan RNA-seq data can be accessed at http://brainspan.org/static/download.html.

^bSex inferred by X-inactive specific transcript expression.

In considering points of intersection between ASD risk factors and biological sex differences, there have been some recent data pointing to the potential contribution of the sex hormones, testosterone and oestrogen, to ASD sex bias^95,208–210. Indeed, two independent in vivo resilience screens identified oestrogen as a protective factor in the context of modelled ASD genetic risks^40,45, implying that oestrogen could be an endogenous protective factor if it is active at higher levels in females than in males. However, testosterone excess has also been hypothesized to be a risk factor for ASD^209,211,212. Further research into the precise manner by which sex hormones intersect with ASD risk is needed, especially characterizing developmental differences in sex hormone levels in the brain and whether they result in sex differences in brain development and function. With the recent discovery that oestrogen inhibits sonic hedgehog (SHH) signalling endogenously in the developing brain⁴⁰, critical future areas of research will include determining whether SHH signalling is sexually dimorphic during prenatal development and whether it is a convergent pathway in ASD risk. SHH has key roles throughout brain development, spanning patterning, proliferation, arealization and differentiation^213–215, suggesting that there may be many points of potential contribution to sexually dimorphic outcomes just within this single pathway.

Perspective

We are at a highly opportune moment in the effort to identify core pathobiological processes underlying ASD. There is a long and growing list of reliably associated large-effect genes accompanied by an expanding armamentarium of multidimensional omics data sets, higher-throughput model systems and CRISPR-based genetic tools that are empowering studies focused on convergence. Although Gene Ontology enrichment analyses highlighting chromatin modification and synapse function have been a readily accessible resource for developing evidence of shared biology, newer hypothesis-free system biological approaches are important complements to these types of studies and have already provided strong evidence for a nexus of ASD risk, particularly in developing excitatory neurons in the mid-gestational frontal cortex.

Of course, critical challenges remain. First, it is not clear how to span the gap from molecular and cellular mechanisms of convergence to differences in the clinical presentation of ASD. Animal models provide an opportunity to observe and manipulate behaviours, as well as the ability to identify underlying circuit-level changes. However, the limitations of anthropomorphizing and relying on face-valid behaviours for complex neuropsychiatric phenotypes are well known³⁵. Second, despite the long-observed male predominance in ASD, the field still lacks a detailed understanding of how this phenomenon is generated and therefore, going forward, it will be critical to investigate sexual dimorphism in brain development and to intersect this with genetic risk.

Despite these challenges, there are already ongoing efforts to exploit successful gene discovery directly to develop novel therapies³⁵. These include gene replacement, gene editing and antisense oligonucleotide strategies. Particularly with regard to the most common forms of syndromic ASD, these initiatives are gaining momentum and clinical trials are on the horizon. Similar efforts for the most highly penetrant high-confidence ASD genes are unlikely to be far behind. The pursuit of convergence neuroscience approaches can contribute to these and related efforts, in part, by clarifying when and where vulnerability to ASD risk emerges, and therefore can help answer key remaining questions regarding when interventions addressing ASD would need to be delivered and where in the brain^37,38. At the same time, points of convergence, with regard to risk and resilience, have the potential to support more traditional therapeutics development by identifying biological mechanisms and targets that are ‘downstream’ of individual genes and applicable to more than one gene at a time. These points of convergence may also have diagnostic utility. Also, there are critical practical and ethical issues that still confront the design and execution of clinical trials in ASD, particularly early in development. Nonetheless, progress in other areas of medicine, including with regard to early-onset neurodegenerative disorders^216,217, combined with the progress in multiple areas noted above points to several viable paths towards more effective, personalized and rational treatments for ASD, particularly for individuals with severe autism.

Autism spectrum disorder.

(ASD). A group of developmental disorders characterized by deficits in social communication and social interaction, and restrictive and repetitive patterns of behaviour, interests or activities.

Intellectual disability.

A developmental disorder characterized by deficits in intellectual functioning (including reasoning, problem-solving and academic learning) and adaptive functioning (including communication and independent living).

Genetic architecture.

The characteristics of genetic variation that contribute to a specific phenotype, including variant frequency, effect size and their interaction with each other and the environment.

Locus heterogeneity.

Variants at different gene loci result in a similar phenotype, individually (for monogenic diseases) or in combination (for complex traits).

Pleiotropy.

A phenomenon in which a single gene contributes to multiple processes or phenotypic traits.

Convergent neuroscience.

Studies that address the overlap, or intersection, of genetic risk for a psychiatric disorder with respect to molecular-level, cellular-level and circuit-level function as well as across multiple dimensions of analysis, including anatomical localization and developmental timing.

Gene Ontology terms.

A hierarchical set of terms that aim to define the universe of possible descriptors a gene can have, including properties such as molecular function, cellular component and biological process.

Penetrance.

The probability that an individual with a given genotype will exhibit the associated phenotype.

Resilience.

The capacity of an individual to have a ‘better than expected outcome’ (for example, being unaffected despite having multiple autism spectrum disorder-associated genetic variants).

Protein-truncating variants.

(Also referred to as loss-of-function or likely gene disrupting variants). Sequence variants that are predicted to shorten the protein-coding sequence of a gene — usually to the point of resulting in a non-functional protein or no protein at all due to nonsense-mediated decay — including nonsense, frameshift and essential splice site variants.

Somatic mutations.

(Also known as mosaic mutations). Variants that are present in fewer than 100% of the cells of an individual (for example, present in brain cells but not present in germ-line cells), generally because the mutation occurred after fertilization.

Compound heterozygous variants.

Variants similar to recessive variants, in that both alleles of a gene are mutated; however, in this case, the two alleles are different.

Neurodevelopmental disorders.

(NDDs). A group of conditions (including autism spectrum disorder, attention deficit hyperactivity disorder, tic disorders and intellectual disability) characterized by developmental onset, atypical brain development and resultant impairments in cognition, communication, behaviour and/or motor skills.

Polygenic risk scores.

Estimates of an individual’s genetic predisposition for a disorder or disease based on the collective effects of many genetic variants.

SFARI Gene.

An evolving database compiled by the Simons Foundation Autism research initiative (SFARI) research program that includes a list of genes and copy number variants (CNVs) associated with autism spectrum disorder curated from human studies and animal models.

Missense variants.

Single base-pair coding variants that result in altered amino acids, for which several metrics have been developed to predict the functional consequence, including the PolyPhen-2 (Polymorphism Phenotyping v2) and MPC (missense badness, PolyPhen-2 and constraint) metrics, for which a categorization of PolyPhen-2 missense 3 (Mis3) or an MPC score ≥2 reflects a probably damaging variant.

Gene Ontology enrichment analysis.

A statistical approach that assesses whether specific Gene Ontology terms are statistically over-represented in a (generally large) set of genes or proteins (for example, among autism spectrum disorder risk genes).

Gene set enrichment analysis.

Like Gene Ontology enrichment analysis, this statistical approach assesses whether specific Gene Ontology terms are statistically over-represented in a large set of genes or proteins; however, unlike Gene Ontology enrichment analysis, gene set enrichment analysis also incorporates the rank of genes within the set into the statistical test (for example, differentially expressed genes ranked based on the fold-change between cases and controls).

Pathway databases.

Databases that annotate genes with an ontological approach to capture functional relationships, including molecular interactions, regulation and phenotype associations; common pathway databases include Gene Ontologies (GO) and Kyoto encyclopedia of Genes and Genomes (KeGG).

Neurogenesis.

The process by which new neurons are generated, in humans most active during gestational weeks 10–25.

BrainSpan Atlas of the Developing Human Brain.

A foundational resource that includes bulk transcriptome profiling of up to 16 cortical and subcortical structures across human brain development (prenatal to adult).

Neurotypical.

Description of individuals with apparently typical intellectual and cognitive development (for example, not affected by a neurodevelopmental or psychiatric disorder).

PsychENCODE.

A consortium-based project that aims to produce multidimensional genomic data from human post-mortem brain tissue from neurotypical and patient donors to begin to functionally characterize risk variants in model systems, with an initial focus on autism spectrum disorder, bipolar disorder and schizophrenia.

Induced pluripotent stem cell.

(iPSC). A pluripotent cell obtained by reprogramming somatic cells through ectopic expression of defined pluripotency factors and/or treatment with small molecules.

Organoids.

(Also known as spheroids) 3D structures, derived in vitro from primary tissue, embryonic stem cells or induced pluripotent stem cells (iPSCs), that self-organize and recapitulate aspects of organ development, anatomy, cellular composition, physiology and function.

Gene modules.

Typically, a set of genes with similar expression profiles that are inferred to be functionally related and co-regulated, and that can be identified through various module-detection methods (for example, gene co-expression).

Functional convergence.

A situation in which seemingly disparate entities are associated with an overlapping function, which can occur at various levels of investigation, including molecular, cell taxonomic, morphological, neural circuit or phenotype level.

Sex differences.

Differences between individuals of different sex in the same species, often including secondary sex characteristics, size and behavioural or cognitive traits.