Abstract
Objective Genetic factors play a substantial role in the etiology of autism and its co-occurrence with other conditions and traits. The autism polygenic score, derived from the latest autism case-control meta-genome-wide association studies, captures some of the accumulated influence of common genetic variants on autism. We reviewed and meta-analyzed published studies that assessed the relationship between this autism polygenic score and autism diagnosis, and autistic, behavioral and neurobiological traits.
Method Systematically searching public databases, we identified 72 studies and > 750 outcome measures. Included studies received a quality assessment.
Results The majority of included studies were rated as good quality. The autism polygenic score was most strongly associated with autism diagnosis (meta-analytic r = .162, 95% CI .066 – .258). The autism polygenic score was also significantly associated with autistic traits but to a lesser degree than for autism (meta-analytic r = .042 (95% CI .004 – .081). Associations with other outcomes were inconsistent and meta-analytic effect sizes were generally small (median r = .03).
Conclusion We conclude that the current autism polygenic score is consistently associated with autism diagnostic status and autistic traits, but overlap between autism and other traits and conditions is not, from publications to date, explained significantly by the autism polygenic score. When compared to other mental conditions, autism is phenotypically and etiologically heterogeneous, which might drive the relatively modest associations observed with the autism polygenic score to date.
Introduction
Autism Spectrum Disorder (ASD; from here on autism as per Bottema-Beutel1) is a complex neurodevelopmental condition that is characterized by atypical social interaction and communication, repetitive behavior and focused interests, and sensory sensitivities. The prevalence of autism is estimated at 1-2% of the population worldwide2. Genetic factors substantially contribute to autism’s etiology. For example, a meta-analysis of twin and family studies indicated that autism is approximately 64 – 91% heritable3.
Autism shows considerable co-occurrence with developmental, psychiatric and somatic conditions and a range of other traits. The incidence of other psychiatric diagnoses in autistic individuals is estimated around 70%4. Autistic individuals1 also have increased liability for a range of physical conditions7. Shared genetic influences between autism, autistic traits and neurodevelopmental, psychiatric and physical conditions may explain part of this phenotypic co-occurrence8,9.
Genome-wide association studies (GWASs) assess the relationship between a large number of common genetic variants spread across the human genome and phenotypic outcomes of interest. The latest autism meta-GWAS by Grove et al. (2019)10 identified 5 independent common genetic loci that were significantly associated with autism and revealed a single nucleotide polymorphism (SNP)-based heritability, which reflects the proportion of phenotypic variance that is explained by additive effects of genomic variants included in the GWAS, of 11.8%. However, many common genetic loci beyond those that are genome-wide significant are involved in autism’s etiology10.
Polygenic scores use the summary statistics from GWAS to capture part of the combined influence of common genetic influences on an outcome. They reflect the effect of multiple genetic variants with small individual effects in a single value. They are calculated by summing together an individual’s common genetic variants, weighted by their effect size. Polygenic scores are one of the tools that can be employed to increase understanding of the genetic basis of complex traits such as autism, and overlap with other traits11, amongst many other popular methods in molecular genetic studies12. In addition to their scientific value, polygenic scores have potential clinical value in the future in addition to existing diagnostic tools13. However, at present, most psychiatric polygenic scores are not accurate for individuals, and bioethical issues relating to the application of polygenic scores need to be carefully planned13.
Although psychiatric phenotypes have long been among the most studied in genetics14, only few studies systematically studied reported polygenic score associations for a specific neuropsychiatric classification with related traits and conditions15–18. These studies report consistent associations between the ADHD polygenic score and ADHD diagnosis (OR range 1.22% to 1.76%), as well as other traits and diagnoses15,16,18 and between the schizophrenia polygenic score and other traits and disorders17.
The latest autism GWAS led to a wealth of studies assessing the relationship between the autism polygenic score and autism diagnosis, autistic traits, and a wide variety of other related traits. Here, we systematically reviewed and meta-analyzed all published studies that assessed the relationship between the autism polygenic scores and all available outcome measures, in order to understand whether the current autism polygenic score can explain phenotypic and genetic overlap between autism and co-occurring conditions and traits.
Methods
This study was pre-registered with PROSPERO Framework under reference number CRD42022307993. Primary analyses were carried out according to the preregistration. We performed additional analyses, of which details are described under ‘secondary analyses’.
Study Selection
We aimed to include all studies assessing the relationship between the autism polygenic score based on the most recent GWAS results10 and autism diagnostic status, autistic traits, and behavioral and neurobiological traits. A first systematic search of PubMed, Web of Science, Psychinfo and Scopus for published, peer-reviewed studies in the English language was performed on November 2, 2022 and a second search on January 6, 2023. The search terms per search engine are given in sTable 2, available online. Studies were excluded if: 1) The predictor was not an autism polygenic score based on the latest autism GWAS10; 2) The polygenic score was not based on genome-wide results, but included only a selection of SNPs; 3) The study was a review; 4) No direct associations between the autism polygenic score and an outcome variable were reported (e.g., polygenic transmission tests (pTDT), interaction effects, genetic correlations, multivariate associations). Abstracts were inspected by MdW. A second reviewer (TP) was involved in case of any doubt regarding inclusion. Figure 1 shows the full selection procedure.
Quality assessment
To evaluate the quality, validity and generalizability of our included studies, we reviewed each study against a set of quality assessment criteria based on Hayden et al. (2006, 2013)19,20. Criteria were largely similar to Ronald et al. (2019), with one addition in the control of confounders domain, where we stated that the polygenic score p-value threshold should be reported. Detailed descriptions of the quality assessment criteria are provided in the Supplementary Methods. All criteria were rated by reviewers MdW and MM. Since these raters had insufficient specialism in brain-related measures, these specific outcomes were reviewed by an expert (IL). In case of any inconsistencies, criteria were assessed by a third rater (TP). A quality assessment domain was considered to contain a bias if > 50% of the criteria in that domain were scored negatively.
Systematic review and meta-analyses
Data extraction
Results were extracted by authors MdW, and for brain measures by IL. Effect sizes that were not provided in the included papers were requested from corresponding authors via email. If authors did not respond to our request but effect sizes were displayed in Figures, we extracted the results using an online tool (https://automeris.io/WebPlotDigitizer/). All reported effect sizes were transformed to correlation coefficients.
If studies reported results for polygenic scores based on multiple GWAS p-value thresholds, we selected the threshold with the strongest association. Where studies reported multiple models, we selected the effect sizes from models that included covariates that most closely resembled those in other included papers and the requirements for the quality assessment (sex, age, socio-economic status (SES)).
Primary analyses
First, all included studies were systematically reviewed with the significance of individual effect sizes based on how they were reported in the respective studies.
Second, for the meta-analysis, all the extracted effect sizes were transformed into correlation coefficients. Since several studies assessed multiple effect sizes, and samples overlapped between studies, we applied three-level random effects meta-analyses per outcome category using the R package Metafor21. These models account for dependency between effect sizes by incorporating three variance components; sampling variance of extracted effect sizes (level 1), within-sample variance in effect sizes (level 2) and between sample variance (level 3) 22. We tested for overall significant heterogeneity, and heterogeneity within and between studies using the I2 statistic23. We assessed potential publication bias using funnel plots, and by performing and Rosenthal’s Fail Safe N test24.
Secondary analyses
In our systematic review, we additionally describe sex differences. We further performed a series of secondary meta-analyses: 1) differences in the association between the autism polygenic score and autism diagnosis for specific populations (Europe-based versus US-based), and 2) subgroup analyses for the associations between the autism polygenic score and ADHD, psychotic spectrum, eating disorders and self-harm and suicide ideation (in specific psychiatric classifications category).
Results
Study characteristics
Studies showed relatively large sample overlap: A total of 57 different samples were used across the 72 studies (pooled N = 720,087). The average percentage of males in the included study samples was 48.3% (25–31) omitted from this percentage due to missing or unclear sex information). Of the included studies 47.2% were performed on child or adolescent (<18 years) participants, 19.4% on adult (> 18 years) participants, 29.2% on a mixed sample of children and adults, and for 2,8% the sample age information was unavailable or unclear. Studies were most often based on participants living in Europe (63.8%), followed by Northern America (25.8%), Asia (5.6%), and Australia (1.4%). The remaining studies (6.9%) were based on participants from multiple continents. Genetic ancestry was mostly European (86.1%), with a limited number of papers including East Asian (5.6%) or trans-ancestry sample analyses (9.7%). Results can be considered to be based on participants of European ancestry, unless described otherwise. Individual study characteristics are presented in Table 1.
Category construction
Outcome categories were constructed through consultation and careful consideration between authors MdW, MM, TJP, SB, AR and AA. They were loosely based on codes from the International Classification of Diseases (ICD) and the World Health Organization International Classification of Functioning, Disability and Health (ICF) and resembled Ronald et al.’s (2021)18 category structure, although there were some dissimilarities (e.g., no addiction category, but a category for emotion recognition). This resulted in 11 categories, reflecting the literature that was identified: Autism diagnostic status, autistic traits, other specific psychiatric classifications, general psychopathology, brain measures, cognition and executive function, emotion recognition, early neurodevelopment, physical wellbeing, phenome-wide association studies (phe-WAS) and other. Note that some studies included outcome variables that fit into multiple categories.
Quality assessment
For 60 studies, 0 biases were detected. Nine studies had one bias, one had two, and one had three biases. We concluded that the quality of included studies is overall good and biases were not clustered in specific outcome categories. The quality assessments are presented in sTable 3, available online.
Systematic review and meta-analysis
A total of 760 outcome variables were included in the meta-analysis. Rosenthal’s failsafe N ranged from 102 for emotion recognition to 39,921 for autism diagnosis (sTable 4, available online), meaning that 102 to 39,921 null findings would be needed to negate significant findings. Funnel plot tests for asymmetry per outcome category based on standard error were not significant (p > .088) except for brain measures (p = .006) and autistic traits (p < .001), indicating a potential publication bias. Histograms and boxplots of effect sizes, and funnel plots of standard errors are presented in Supplementary Figures 1 – 3, available online. Detailed descriptions of sample characteristics, methods of calculating polygenic scores, GWAS p-value threshold, outcome measures, covariates, results, and multiple testing corrections per study and per outcome category are provided in Table 1. Meta-analytic pooled correlation coefficients ranged from .013 for early neurodevelopment to .162 for autism diagnosis. The number of studies in each category ranged from 3 (phe-was) to 20 (other), with a minimum of 2 and a maximum of 11 independent cohorts per category. A summary of the meta-analysis results for all outcome categories is presented in Table 2 and Figure 2.
Autism diagnosis
Systematic review
Nine studies assessed the association between the autism polygenic score and autism diagnosis. Of these, seven reported a significant positive association10,25,27,30,32,33,36. Two studies, based on smaller samples than the ones that reported significant associations, reported no significant association37,50. In addition, the autism polygenic score acted in a dose-dependent manner, where people in higher percentiles on average were more likely to be diagnosed with autism10,36.
Meta-analysis
Standardized correlation coefficients ranged from 0.030 to 0.346. One outcome was omitted from this range, as it differentiated whether the autism polygenic score distinguished autism diagnosis versus ADHD diagnosis27. The overall meta-analyzed effect size was r = .162 (95% CI .066 – .258), indicating a small positive association between the autism polygenic score and autism diagnosis. Total I2 was 99.6%, of which 28.4% between-study and 71.2% within-study. All meta-analysis results for the category of autism diagnosis, including secondary results are presented in Supplementary Figure 4.
Autistic traits
Systematic review
Eleven studies assessed the association between the autism polygenic score and autistic traits, either in the form of a total score or specific traits such as repetitive behavior, language and social communication. Four of these assessed total scores for autistic traits40,42,44,45, one of which was performed in a non-European sample. In children, the autism polygenic score was associated with overall autistic traits in a Japanese44 and Swedish sample45, but not in a United States (US) sample40. In adults, the autism polygenic score was not associated with overall autistic traits40,42.
Social behavior was the most commonly studied specific autistic trait, with seven studies assessing the association Four did not find a significant association31,38,39,45, and three did41,43,44, one of which was a Japanese sample. The autism polygenic score did not associate with restricted and repetitive behavior38,46, nor language and communication measures at various ages in childhood38,45,46, except for language difficulties at the age of 18 months38. In addition, the autism polygenic score was not associated with broad autistic traits as measured by the Broad Autism Phenotype Questionnaire (BAPQ) in children with autism or their parents, except for two significant positive associations with the pragmatic and total scale of the BAPQ in mothers of autistic children40.
Meta-analysis
One study on autistic traits was omitted from the meta-analysis effect sizes because effect sizes were not reported40. Correlation coefficients ranged from −.074 to .313. The overall meta-analyzed correlation coefficient was r = .04 (95% CI .004 – .081). Total I2 was 94.8%, of which 87.3% between-study and 7.5% within-study. Meta-analysis results for autistic traits are presented in Supplementary Figure 5.
Specific psychiatric classifications
Eighteen studies assessed the relationship between the autism polygenic score and specific psychiatric classifications (other than autism). Included studies assessed a range of classifications or traits : ADHD (n = 2)27,34, ADHD traits (n= 2)38,42, depression (n = 2)48,56, addiction (n = 1)49, suicidal ideation and/or suicide attempt (n = 3)51,59,97, eating disorders (n= 2)37,53, psychotic spectrum (n= 3)54,55,57 and nocturnal enuresis (n = 1)52.
One study showed a significant association with ADHD diagnosis27, whereas another did not34. The autism polygenic score was significantly associated with ADHD symptoms in an east Asian sample67, but not in two studies European ancestry samples38,42, except for one positive association with inattention and hyperactivity at the age of 8 years. The autism polygenic score did not associate with schizophrenia diagnosis in two East Asian samples57,58, and in a European ancestry sample it positively associated with some, but not all specific domains of the psychotic spectrum (e.g. auditory and visual hallucinations)55. Two studies found consistent positive associations with lifetime depression48,56 and one with depression or anxiety symptoms specifically during pregnancy48. No associations were found between the autism polygenic score and disordered eating behavior in two studies37,53, except for increased parent-reported Avoidant/restrictive food intake disorder (ARFID;. One study reported no associations between the autism polygenic score and cannabis use disorder49, except for an unexpected negative association in people diagnosed with DSM classifications other than schizophrenia. One study reported no association with nocturnal enuresis52. Lastly, one study reported that the autism polygenic score was not associated with the case-control status of any psychiatric classification50.
One study found that the autism polygenic score distinguished people with autism from those with ADHD27.
Meta-analysis
Correlation coefficients ranged from −.185 to .213 The overall meta-analyzed correlation coefficient was r = .044 (95% CI .016 – .073). Total I2 was 99.2%, of which 33.9% between-study and 65.4% within-study. Meta-analysis results for specific psychiatric classifications are presented in sFigure 6, and secondary results in sTable 5 and sFigure 7.
Other outcome categories
Associations between the autism polygenic score and other outcome categories were overall small (median r = .03) and inconsistent. Detailed descriptions of these associations are provided in Table 1, Figure 1 and the Supplementary Results and Supplementary Results, available online. Forest plots are presented in sFigures 8 – 13, available online.
Heterogeneity
We found considerable heterogeneity across all outcome categories (I2 ≥ 91.5%). To provide a more complete picture of the heterogeneity in the included studies, and because I2 is not a particularly suitable statistic when assessing very large sample sizes, all meta-analyses forest plots also report Q statistics98.
Discussion
Overall, our results showed that the autism polygenic score was most strongly and most consistently associated with autism diagnosis in independent cohorts. A higher polygenic score translates into an increased likelihood of being diagnosed with autism10,36. The association of the autism polygenic score with autistic traits was significant but smaller. Associations with outcomes other than autism diagnosis and autistic traits were inconsistent and meta-analytic effect sizes were generally small. Here we discuss our results, specifically zooming in on autism diagnosis, autistic traits and other psychiatric classifications because these categories contained the most consistent results. We discuss implications and provide recommendations for future directions.
The lack of consistent associations with outcomes other than autism diagnosis and autistic traits is in contrast with a vast body of literature supporting phenotypic associations between autism and neurodevelopmental, psychiatric, and physical conditions and other traits, as well as genetic associations found in twin studies. They are also in contrast with studies that used alternative genetic methodologies such as genome-wide restricted maximum likelihood (GREML), genome-wide complex trait analysis (GCTA), and linkage disequilibrium score (LDSC) regression. A difference between polygenic scores and these methods is that the latter captures the total influence of measured SNPs whereas polygenic scores do not, potentially resulting in less power to detect genetic associations (explanations of these methods are described in Martin et al.12). Our results also highlight disparities between the autism polygenic score and the ADHD polygenic score, with the latter showing stronger (e.g., OR between 1.22% and 1.76% for ADHD diagnosis) and more reliable, consistent associations. This is despite comparable heritability estimates and GWAS sample sizes for ADHD and autism, and also both being a neurodevelopmental condition18.
The overarching category of specific psychiatric classifications comprised numerous classifications and had an overall significant but small association with the autism polygenic score. However, results varied markedly for different classifications. Our secondary analyses to dissecting the overarching category revealed a significant association with the psychotic spectrum, but not with ADHD. These findings were in contrast with some past studies in this area10,99–102.
The inconsistency of our results and their lack of alignment with previous findings indicates that the current autism polygenic score may lack the statistical power to adequately capture shared polygenic effects. In line with this notion, two included studies found that the autism polygenic score acts largely independent from psychiatric family history36,61, despite both factors being expected to reflect a genetic predisposition to autism. This finding extends beyond autism to multiple psychiatric classifications103,104, indicating that current psychiatric polygenic scores may overall insufficiently capture genetic propensity towards psychiatric traits. As the predictive power of polygenic scores depends heavily on the statistical power of the discovery GWAS, we underscore the importance of ongoing efforts to increase GWAS sample sizes.
Our study is not without limitations. First, by focusing on univariate associations and excluding genetic measures and methods other than polygenic scores, we may have overlooked important insights into autism genetics. We refer the reader to Akingbuwa et al. (2022)99 for results that are based on other methods than univariate associations99. Second, we were not able to include all the associations in our meta-analyses because some studies did not report all effect sizes in text. Similarly, we did not include all outcome measures assessed through phenome-wide association studies in our meta-analysis, as this was infeasible due to the number of studies. Lastly, we point out the considerable heterogeneity in our included studies, as is reflected in the high I2 statistics. A high I2 value could be due to low within-sample estimation error, which is likely to occur with large sample sizes such as the ones included in our study105. However, we acknowledge that this heterogeneity may also be due to the construction of our outcome variable categories. However, even in the categories that are largely similar in their outcomes (such as autism diagnosis), heterogeneity is high. This heterogeneity remained substantial even when dissecting outcome categories into more homogeneous subcategories. This heterogeneity may have added noise to our estimates. Future studies should aim to reduce such heterogeneity to enhance our understanding of autism’s polygenic nature.
A specific source of heterogeneity may be the substantial variation in polygenic score calculation, both in terms of software and the selection of the GWAS p-value thresholds. While the majority of studies apply methods that select the polygenic score that yields the strongest associations, others use a pre-defined threshold or employ principal component analyses to construct a polygenic score from multiple thresholds. Similarly, a diverse set of polygenic score calculation programs was applied. Several recent reports address this issue and advocate for standardized guidelines and reporting practices106,107, which might improve the comparability of polygenic score findings.
Despite the marked diversity in methods between studies, we observed limited diversity in the participant characteristics. The majority of studies were based on people of European ancestry, resulting in our current understanding of the autism polygenic score being almost solely based on people of European ancestry. This observation is concerning, especially considering the increasing autism prevalence rates in people from non-European backgrounds108. Similarly, few included studies addressed understudied subgroups of autism, such as those with intellectual disability, or people with autism that are nonverbal. Both scientifically, and ethically, it would be of great value to radically improve inclusive research practices109.
We propose that GWASs, and subsequently polygenic scores, based on individual autistic traits instead of the overarching autism diagnosis might enhance our understanding of autism genetics110. Early twin studies revealed that the three domains of autistic traits (social, communication, and restricted/repetitive behavior) only modestly correlate genetically, a finding that suggests largely independent genetic effects may affect different autistic traits111,112. Yet molecular genetics have not consistently pursued this research direction96. Although some molecular genetic studies were performed on specific autistic traits, large-scale approaches are still lacking46,113.
Despite its limitations, the current autism polygenic score comprises part of a bigger picture, and should ideally, as suggested by several recent papers, be integrated with other genetic factors such as common and rare variants, psychiatric family history, and sex, in a comprehensive approach to studying autism25,36,61,114. For example, Grove et al. (2019)10 found that a combined polygenic score of autism and correlated traits, such as major depressive disorder and schizophrenia, improved the prediction of autism diagnostic status. Rare genetic variants are not captured by typical polygenic scores, but have been suggested as factors of high influence on autism diagnostic status115. In line, Antaki et al. (2022)114 showed that a genetic score containing a combination of common and rare genetic factors significantly increased statistical power over separate common and rare variant scores. Schendel et al. (2022)36 argue that psychiatric family history and polygenic scores only minimally correlate and should therefore be considered separate, mostly uncorrelated, measures of autism’s genetic etiology. Lastly, although we do not find significant sex differences in the association of the autism polygenic score with outcome variables, there is building evidence for a different genetic makeup for men and women with autism. Autistic women on average have a higher number of genetic mutations116, higher mean polygenic load87,114 and more rare copy number variants (CNV) load117. However, a potentially male-biased clinical view on the manifestation of autism might explain diagnostic differences, and subsequently observed genetic differences118
In conclusion, we show that the autism polygenic score as a standalone variable associates consistently with autism diagnostic status and dimensional autistic traits, but is unable to capture the spectrum’s complex phenotypical and etiological differences and its genetic overlap with other traits and conditions. We propose considering the current autism polygenic score as a complementary measure in research, with improvements needed for a more robust understanding of autism’s polygenic underpinnings.
Data Availability
All data extracted for this study are available upon reasonable request to the authors.
Financial support
This work was supported by ZonMw (TP, SB and MdW: grant number 60-63600-98-834)
Disclosures
Authors declare they have nothing to disclose.
Acknowledgements
We thank our EGAL (Environmental and Genetic Influences on Autism across the Lifespan) panel for their valuable input.
Footnotes
References
- 1.↵
- 2.↵
- 3.↵
- 4.↵
- 5.↵
- 6.↵
- 7.↵
- 8.↵
- 9.↵
- 10.↵
- 11.↵
- 12.↵
- 13.↵
- 14.↵
- 15.↵
- 16.↵
- 17.↵
- 18.↵
- 19.↵
- 20.↵
- 21.↵
- 22.↵
- 23.↵
- 24.↵
- 25.↵
- 26.
- 27.↵
- 28.
- 29.
- 30.↵
- 31.↵
- 32.↵
- 33.↵
- 34.↵
- 35.
- 36.↵
- 37.↵
- 38.↵
- 39.↵
- 40.↵
- 41.↵
- 42.↵
- 43.↵
- 44.↵
- 45.↵
- 46.↵
- 47.
- 48.↵
- 49.↵
- 50.↵
- 51.↵
- 52.↵
- 53.↵
- 54.↵
- 55.↵
- 56.↵
- 57.↵
- 58.↵
- 59.↵
- 60.
- 61.↵
- 62.
- 63.
- 64.
- 65.
- 66.
- 67.↵
- 68.
- 69.
- 70.
- 71.
- 72.
- 73.
- 74.
- 75.
- 76.
- 77.
- 78.
- 79.
- 80.
- 81.
- 82.
- 83.
- 84.
- 85.
- 86.
- 87.↵
- 88.
- 89.
- 90.
- 91.
- 92.
- 93.
- 94.
- 95.
- 96.↵
- 97.↵
- 98.↵
- 99.↵
- 100.
- 101.
- 102.↵
- 103.↵
- 104.↵
- 105.↵
- 106.↵
- 107.↵
- 108.↵
- 109.↵
- 110.↵
- 111.↵
- 112.↵
- 113.↵
- 114.↵
- 115.↵
- 116.↵
- 117.↵
- 118.↵