ASD and schizophrenia show distinct developmental profiles in common genetic overlap with population-based social communication difficulties

Difficulties in social communication are part of the phenotypic overlap between autism spectrum disorders (ASD) and schizophrenia. Both conditions follow, however, distinct developmental patterns. Symptoms of ASD typically occur during early childhood, whereas most symptoms characteristic of schizophrenia do not appear before early adulthood. We investigated whether overlap in common genetic influences between these clinical conditions and impairments in social communication depends on the developmental stage of the assessed trait. Social communication difficulties were measured in typically-developing youth (Avon Longitudinal Study of Parents and Children, N⩽5553, longitudinal assessments at 8, 11, 14 and 17 years) using the Social Communication Disorder Checklist. Data on clinical ASD (PGC-ASD: 5305 cases, 5305 pseudo-controls; iPSYCH-ASD: 7783 cases, 11 359 controls) and schizophrenia (PGC-SCZ2: 34 241 cases, 45 604 controls, 1235 trios) were either obtained through the Psychiatric Genomics Consortium (PGC) or the Danish iPSYCH project. Overlap in genetic influences between ASD and social communication difficulties during development decreased with age, both in the PGC-ASD and the iPSYCH-ASD sample. Genetic overlap between schizophrenia and social communication difficulties, by contrast, persisted across age, as observed within two independent PGC-SCZ2 subsamples, and showed an increase in magnitude for traits assessed during later adolescence. ASD- and schizophrenia-related polygenic effects were unrelated to each other and changes in trait-disorder links reflect the heterogeneity of genetic factors influencing social communication difficulties during childhood versus later adolescence. Thus, both clinical ASD and schizophrenia share some genetic influences with impairments in social communication, but reveal distinct developmental profiles in their genetic links, consistent with the onset of clinical symptoms.


INTRODUCTION
The phenotypic overlap between autism spectrum disorder (ASD) and schizophrenia is complex and dates back to Kanner in 1943. 1 Individuals affected by either condition display deficits in the ability to initiate and maintain reciprocal interaction. 2 This includes impairments in social cognition 3,4 but also poor social competence 5 affecting verbal and nonverbal communication skills. Recent cross-disorder genetic analyses highlighted the continuity of psychiatric phenotypes beyond current diagnostic boundaries. 6 The nature of shared genetic influences between childhood neurodevelopmental disorders, such as ASD, and adult-onset psychiatric illnesses, like schizophrenia, however, remains less well understood.
ASD represent a group of neurodevelopmental conditions with a typical age of onset before the age of 3 years affecting~1 to 2% of children. 7,8 Core features include deficits in social interaction and communication, as well as highly restricted interests and/or stereotyped repetitive behaviours. 2 By contrast, schizophrenia is an adult-onset psychiatric illness with a typical first-time diagnosis between 16 and 30 years. The disorder has a lifetime prevalence of ~1% 9 and is characterised by hallucinations, delusions, disorganised speech or behaviour, apathy and lack of emotional reactivity. 2 Both ASD and schizophrenia are highly heritable 10,11 and recent studies have linked different types of genetic variation including common variants, 10,12,13 as well as rare inherited 14,15 and de novo variation 16,17 to risk of illness in both conditions. Contemporary research strongly supports a genetic overlap between ASD and schizophrenia for rare copy number variants 18 and rare de novo mutation events 16 with converging evidence for gene sets involved in synaptic function. 16 The role of shared common genetic risk between ASD and schizophrenia, however, is less clear. Common genetic influences account for 25 to 33% 19 of total liability to schizophrenia and up to 49% of total liability to ASD. 10,12 Despite this, the common genetic overlap between ASD and schizophrenia is small compared with the overlap between psychiatric adult-onset only disorders. 12,20 The framework of Research Domain Criteria (RDoC), including social communication difficulties, now actively facilitates the study of functional dimensions spanning the full range of human behaviour from normal to abnormal and across development. 21 Common disorders, due to their polygenic architecture, can be understood as quantitative traits. 22 For ASD, following the findings of earlier twin studies, 23,24 there is now molecular evidence for shared common genetic influences with social communication difficulties during childhood. 25 The genetic continuity of social interaction and communication deficits in schizophrenia has not yet been observed though it can be hypothesised that such common genetic links exist given the impairments in social cognition within first-degree relatives of schizophrenia patients. 3 Impaired abilities in social communication in affected children are heritable (twin-h 2 = 0.74) 26 and a large part of these genetic influences can be captured through common single-nucleotide polymorphisms (SNPs; SNP-h 2 ⩽ 0.45). 27 Beside some stable genetic influences, 28 genetic factors underlying social interaction impairments and social communication difficulties vary during development, 27,28 especially for common variation. 27 Thus, we hypothesise that also the genetic overlap between social communication difficulties and clinically recognised disorder may change during childhood and adolescence.
The primary aim of this study is to examine the nature of common polygenic influences in ASD and schizophrenia through their genetic overlap with phenotypic symptoms in the general population that are shared between both conditions, but differ according to developmental stage. We predict that if social communication difficulties are part of a common shared aetiology between ASD and schizophrenia, trait-disorder relationships for both conditions should follow similar patterns. Dissimilar patterns due to independent genetic influences would be expected for a non-shared genetic aetiology. Here, we report developmental profiles in common genetic overlap for both ASD and schizophrenia with respect to longitudinal measures of social communication difficulties within the general population. Analyses are based on the largest publicly available genome-wide data for ASD 29 and schizophrenia, 13 in addition to a large Danish ASD sample from the iPSYCH project and a deeply-phenotyped UK birth cohort, the Avon Longitudinal Study of Parents and Children (ALSPAC). 30,31

Genome-wide summary statistics
Population-based social communication difficulties. Genome-wide association studies (GWASs) were carried out in ALSPAC participants, a UK population-based longitudinal pregnancy-ascertained birth cohort (estimated birth date: 1991-1992). 30,31 Ethical approval was obtained from the ALSPAC Law-and-Ethics Committee (IRB00003312) and the Local Research-Ethics Committees, written informed consent was obtained from a parent or individual with parental responsibility and assent was obtained from child participants. ALSPAC children were genotyped using the Illumina HumanHap550 quad-chip and imputation was performed on 8237 children and 477,482 SNP genotypes using a 1000 Genomes reference (PhaseI_v3, http://www. 1000genomes.org/) 32

(Supplementary Methods).
Quantitative social communication problems in ALSPAC children were assessed with the 12-item Social Communication Disorder Checklist (SCDC; score-range: 0 to 24). 26 The SCDC is a brief screening instrument of social reciprocity and verbal/nonverbal communication (for example, 'Not aware of other people's feelings'), with high reliability and good validity, 26 which has been extensively investigated. 26,27,33 Higher SCDC scores reflect more social communication deficits and are positively skewed (Supplementary Figure 1). Mother-reported scores for children and adolescents were repeatedly measured at 8, 11, 14 and 17 years (Supplementary Table 1) and are inter-correlated (Spearman's ρ: 0.39 to 0.57, Supplementary Table 2). Information on phenotypic and genotypic data was available for 4175 to 5553 children (Table 1).
SCDC scores were residualised for sex, age and the two most significant ancestry-informative principal components 34 and then rank-transformed (Supplementary Figure 2). Transformed scores showed similar correlation patterns as untransformed scores (Pearson' r: 0.38 to 0.61, Supplementary Table 2).
Genome-wide single marker summary statistics were generated by regressing rank-transformed residuals on allele dosages using SNPTEST 35 (without genomic control-based correction 36 ).
Clinical ASD. The Psychiatric Genomics Consortium (PGC) has completed a genome-wide scan of 5305 ASD cases and their parents (PGC-ASD), all of European ancestry (2015 freeze; summary results at http://www.med.unc. edu/pgc/). An ASD diagnosis was confirmed using research standard diagnoses and expert clinical consensus diagnoses. 94.1% of all ASD cases had also a diagnosis of autism from the Autism Diagnostic Interview-Revised 37 and/or the Autism Diagnostic Observation Schedule. 38 Genomewide data were imputed to a 1000 Genomes reference (PhaseI_v3) and genetic association studied using a case and pseudo-control design. 29 This design is robust to population stratification as pseudo-controls are based on un-transmitted parental alleles, and thus cases and pseudo-controls are ancestrally matched. To replicate findings, we analysed ASD GWAS summary results in the Danish iPSYCH project (iPSYCH-ASD: 7783 ASD cases, 11 359 controls) using samples from the Danish Neonatal Screening Biobank hosted by Statens Serum Institute (Supplementary Methods). The iPSYCH-ASD project aims to genotype all Danish individuals with available DNA from bloodspots and an ASD diagnosis (International Classification of Diseases 39 ) in their medical record. iPSYCH-ASD has been genotyped using the Illumina Infinium PsychArray BeadChip and genotypes were imputed to a 1000 Genomes template (PhaseI_v3). This study has been approved by the Danish research ethical committee system.
Note that also a small number of ALSPAC children with clinical ASD (N ⩽ 36) has been included in this study (Supplementary Methods).
The studied population-based and clinical samples (Table 1) contain no sample overlap.
Other adult-onset disorders. To analyse the specificity of genetic overlap between SCDC scores and schizophrenia, we studied further adult-onset psychiatric disorders, such as major depressive disorder (MDD) and bipolar disorder (BIP; Supplementary Methods).

Statistical methods
Linkage disequilibrium (LD) score regression 41 was applied to estimate the cumulative effect of common SNPs on either variation in developmental SCDC scores or risk to disorder (SNP-h 2 ), using GWAS statistics and exploiting LD patterns in the genome. LD score correlation 20 analysis was carried out to estimate genetic correlations (r g ) between SCDC scores and clinical conditions, or among clinical conditions, that is, the extent to which two phenotypes share common genetic factors, based on GWAS statistics. All analyses were performed with LDSC software 20,41 using HapMap3 markers 42 (Supplementary Methods). Polygenic risk scores (PGS) 43,44 were analysed to estimate the explained phenotypic variance in social communication difficulties due to riskincreasing alleles for clinical disorder. Using a range of P-value thresholds (0.001oP T ⩽ 1), PGS for ASD (based on PGC-ASD), schizophrenia (based on PGC-SCZ2) and schizophrenia subsamples (based on PGC-SCZ1 and PGC-SCZ2i) were generated in ALSPAC (Table 1) using imputed genotypes (1000 Genomes, PhaseI_v3, INFO40.8). For this, common autosomal signals observed in clinical samples (with MAF40.01 in ALSPAC) were clumped (LD-r 2 40.25, ± 500 kb) consistent with current guidelines 45 using PLINK, 46 excluding duplicate SNPs (Supplementary Methods). Ranktransformed SCDC scores were regressed on Z-standardised PGS (Ordinary least square regression, R software Rv3.2.2, https://cran.r-project.org/), and the proportion of phenotypic variance explained by each PGS predictor reported as adjusted regression R 2 . Note that assuming an infinitely large clinical 'discovery' sample, the regression R 2 is equivalent to the product of r g squared and the heritability of the explained Z-standardised trait. 44 Mixed Poisson regression (R:lme4 library) was utilised to test the trend in common genetic overlap longitudinally using untransformed SCDC scores. Repeatedly assessed SCDC score counts were regressed on ASD-and schizophrenia-PGS with overdispersion being accounted for through the random error part. 47 Models included fixed effects for ASD-PGS and schizophrenia-PGS, sex, age at assessment, as well as random intercepts. Beta-coefficients for PGS quantify here the increase in natural-log SCDC scores for each increase in one standard deviation of PGS. Differences in sample-dropout across time were accounted for through bootstrapping, generating parametric 95%-bootstrap confidence intervals (N Bootstrap = 500). We have not estimated adjusted R 2 -related measures due to the difficulty of defining the residual variance for non-Gaussian responses, especially within a mixed model context. 48 Genome-wide Complex Trait Analysis (GCTA) 49,50 was utilised to estimate SNP-h 2 and genetic correlations among SCDC scores, as published previously, 27 for comparison only (Supplementary Methods).
Attrition analysis in ALSPAC studied the relationship between SCDCmissingness at each assessed age and PGS for clinical ASD and schizophrenia (Supplementary Methods).

SNP-heritabilities for social communication difficulties and psychiatric disorder
Genome-wide analyses of population-based SCDC scores at 8, 11, 14 and 17 years provided little evidence for bias in GWAS statistics due to population stratification. The estimated LDSC-h 2 intercepts were consistent with one, ranging from 0.988 (s.e. = 0.0067) to 1.009 (s.e. = 0.0070; Table 2). In subsequent analyses LDSC-h 2 intercepts were thus constrained to one, including LDSC correlation analyses.
Polygenic scores for risk-increasing alleles predicting social communication difficulties To provide an absolute measure of shared genetic influences between traits and clinically recognised conditions, we assessed the phenotypic variance in rank-transformed social communication difficulties due to risk-increasing alleles using polygenic scoring 43,44 (Table 1). Alleles more common in ASD cases than in pseudo-controls were only associated with variation in SCDC scores at 8 years (PGC-ASD: adjusted R 2 max = 0.13%, P min = 0.0042, Figure 2a, Supplementary Table 6). In contrast, alleles more often present in schizophrenia cases than controls explained predominantly variation in social communication difficulties at 17 years, based on risk alleles in both PGC-SCZ subsamples (PGC-SCZ1: adjusted R 2 max = 0.26%, P min = 0.00058; PGC-SCZ2i: adjusted R 2 max = 0.19%, P min = 0.0028; Supplementary Table 7) and the combined PGC-SCZ2 sample (adjusted R 2 max = 0.43%, P min = 0.000012, Figure 2b, Supplementary Table 6). Excluding ALSPAC children with a clinical ASD diagnosis had little influence on the reported changes in genetic effect (Supplementary Table 8). Importantly, adjustment of ASD-PGS and schizophrenia-PGS for each other did not affect the nature of these findings, suggesting the independence of ASDand schizophrenia-related polygenic influences (Supplementary Table 6).
To assess developmental trends in common genetic traitdisorder overlap, we modelled the effect of ASD-PGS and schizophrenia-PGS on untransformed SCDC scores longitudinally. Applying a mixed Poisson model, we found evidence for age-specific changes in genetic effects for both ASD-PGS and schizophrenia-PGS (Supplementary Table 9). For example, at P T o0.05 (Figure 3), a threshold shown to predict schizophrenia case-ness in independent samples, 13   Abbreviations: ALSPAC, Avon Longitudinal study of Parents and Children; GCTA, genome-wide complex trait analysis; h 2 , SNP heritability; LD, linkage disequilibrium; SCDC, Social Communication Disorder Checklist; y, age at assessment in years; λ GC , Genomic inflation factor. a Findings correspond closely to previously published estimates. 27 b LD score regression using an unconstrained intercept. c LD score regression constraining the intercept for the SNP-h 2 estimation to one. d Differences compared with the total sample N are due to the exclusion of individuals with a relatedness of ≥ 2.5%.
Developmental profiles in common genetic overlap B St Pourcain et al P = 0.000056). Similar developmental changes in genetic overlap were also found for other PGS thresholds (Supplementary Table 9).
Attrition in ALSPAC Analyses of SCDC-missingness in ALSPAC were carried out to investigate potential sources of bias (Supplementary Table 10).
Using for simplicity a PGS threshold of P T o0.05, there was little evidence for a relationship between sample-dropout and ASD-PGS, especially after adjustment for maternal educational level (age 8: odds ratio = 0.99 (s.e. = 0.03), P = 0.82), although there was support for an association with schizophrenia-PGS (age 17: odds ratio = 1.10 (s.e. = 0.03), P = 0.000050), consistent with previous studies. 52

DISCUSSION
This study provided evidence for shared common genetic overlap between social communication difficulties and both ASD and schizophrenia, but does not imply a shared genetic susceptibility between these clinical conditions. Instead, we identified distinct patterns in genetic trait-disorder relationships, largely consistent with the onset of clinical symptoms. Genetic links were driven by independent polygenetic influences and showed opposite trends in magnitude with progressing age of the population-based trait, as supported by longitudinal analyses. Genetic overlap with ASD was strongest for social communication difficulties during middle childhood (r g~3 3%), in line with recent cross-sectional studies, 25  adolescence (r g~1 8%). Complementary estimates were provided by polygenic scoring analyses. Up to 0.13% phenotypic variation in social communication difficulties could be explained by ASD risk-increasing alleles during childhood and up to 0.43% phenotypic variation by schizophrenia risk-increasing alleles during later adolescence, independently of each other. The genetic overlap with social communication difficulties during later adolescence was not observed for other adult-onset disorders, such as BIP, despite their strong genetic links with psychosis, 12 making unspecific age-related influences unlikely. Thus, our findings suggest that social communication impairments are part of the genetically influenced phenotypic spectrum of schizophrenia. Changes in genetic overlap over time need to be viewed within the context of cohort-specific sampling properties and clinical sample power. For instance, it is possible that the genetic overlap between schizophrenia and social communication difficulties has been underestimated, as SCDC-missingness, and more generally study non-participation, 52 has been related to common genetic risk for schizophrenia. In contrast, there was little evidence for a link between SCDC missingness and common ASD risk. In addition, mother-report of social communication difficulties may have contributed to enhanced variance sharing among population-based traits, and thus underestimated the true variation in child genetic effects. Finally, the studied clinical discovery sets differed in their inherent power. For example, the power 44 of ASD-PGS (PGC-ASD~5000 trios) was only 0.58 compared with 0.99 for schizophrenia-PGS (PGC-SCZ2, N~80 000), assuming a liability-scale SNP-h 2 of 0.23 for ASD and 0.25 for schizophrenia, a disease prevalence of 0.01, a type-I-error rate of 0.05 and a population-based target sample of 5000 individuals. Under attrition, such as a decrease of~1000 ALSPAC participants during later adolescence, the power of ASD-PGS would further drop (to 0.49), while the power of schizophrenia-PGS remains largely unaffected (0.99). Longitudinal analyses, adjusting for differences in population-based sample numbers across time through bootstrapping, suggested, however, that the observed developmental changes in polygenic risk effects are robust, even in the presence of sample dropout.
Our results have direct relevance for the definition of RDoC 21 within a developmental context. The lack of support for shared polygenic effects between ASD and schizophrenia, with respect to social communication impairments, is in agreement with recent studies. Molecular analyses of PGC samples reported modest correlations between ASD and schizophrenia 20,29 (r g~0 .20), confirmed within this study, and twin research suggested little genetic overlap between autistic traits and psychotic experiences. 53 The absence of shared aetiological factors strengthens furthermore positions suggesting that the exact nature of social deficits implicated within ASD and schizophrenia differs from each other. 54 Here we show that common genetic variation underlying complex disorders can be dissected through temporal changes in the genetic architecture of behavioural symptoms that are shared between disorders. Thus, a developmental analysis of genetic relationships between population-based and clinical samples can be informative with regard to the dimensional nature of psychiatric illness without discarding the aetiology of different disorders, a concern often raised with respect to RDoC. 21 The identification of distinct patterns in common genetic overlap between social communication difficulties and psychiatric illness is consistent with the presence of multiple distinct genetic influences contributing to variation in social communication behaviour during development. While genetic factors underlying SCDC scores across~3-year intervals are stable and shared by at least 80%, only~50% of common genetic influences are shared across 10 years intervals. 27 This may suggest developmental (but not rapid) changes in the phenotype capture by the SCDC with progressing age. For example, it is possible that the behavioural phenotypes influencing SCDC scores at age 8 or 11 years are, in terms of average composition, different from those influencing the scale at age 17. Social communication abilities comprise many components, such as social interaction, social cognition, pragmatic and language processing skills (http://www.asha.org/), some of which will vary during child and adolescent development, including changes in the social-cognitive understanding of