Published online February 28, 2013 http://dx.doi.org/10.1016/S0140-6736(12)62129-1
varying degrees of overlap in familial and genetic liabilityfor pairs of these disorders. For example, some ﬁndings
from family and twin studies support diagnostic boun-daries between schizophrenia and bipolar disorder andbipolar disorder and major depressive disorder, but alsosuggest correlations in familial and genetic liabilities.
Several molecular variants confer risk of both schizo-phrenia and bipolar disorder.
Autism was once knownas childhood schizophrenia and the two disorders werenot clearly diﬀerentiated until the 1970s. Findings fromthe past few years have emphasised phenotypic andgenetic overlap between autism spectrum disorders andschizophrenia,
including identiﬁcation of copy numbervariants conferring risk of both.
Findings from family,twin, and molecular studies
suggest some geneticoverlap between autism spectrum disorder and attentiondeﬁcit-hyperactivity disorder.In this ﬁrst report from the PGC Cross-Disorder Group,we analyse data on genome-wide single-nucleotide poly-morphism (SNP) for the ﬁve PGC disorders to answer twoquestions. First, what information emerges when all ﬁvedisorders are examined in one GWAS? When risk iscorrelated across disorders, pooled analyses will be betterpowered than individual-disorder analyses to detect riskloci. Second, what are the cross-disorder eﬀects of variantsalready identiﬁed as being associated with a speciﬁcpsychiatric disorder in previous PGC analyses? We aimedto examine the genetic relation between the ﬁve psychiatricdisorders with the expectation that ﬁndings will ultimatelyinform psychiatric nosology, identify potential neuro-biological mechanisms predisposing to speciﬁc clinicalpresentations, and generate new models for preventionand treatment.
Samples and genotypes
The sample for these analyses consisted of cases,controls, and family-based samples assembled forprevious genome-wide PGC mega-analyses of individual-level data.
Cases and controls were not related. Forthe family-based samples, we matched alleles transmittedto aﬀected oﬀspring (trio cases) with untransmittedalleles (pseudocontrols). We estimated the identity-by-descent relation for all pairs of individuals to identify anyduplicate individuals in the component datasets. Whenduplicates were detected, one member of each set wasretained. We then randomly allocated these individuals,with a random number generator, to a disorder case-control dataset. Sample sizes diﬀer from previous reportsbecause of this allocation of overlapping individuals. Allpatients were of European ancestory and met criteriafrom the DSM third edition revised or fourth edition forthe primary disorder of interest.To ensure comparability between samples, raw geno-type and phenotype data for each study were uploadedto a central server and processed through the samequality control, imputation, and analysis process(appendix).
We analysed imputed SNP dosages from1 250 922 autosomal SNPs.
In the primary analysis, we combined eﬀects of eachdisease analysis by a meta-analytic approach that applieda weighted
in which weights equalled theinverse of the regression coeﬃ cient’s standard error.This strategy assumed a ﬁxed-eﬀects model, with weightsindicating the sample size of the disease-speciﬁc studies.In a second analytical approach, we did a ﬁve-degree-of-freedom test by summing the χ² values for eachindividual disease meta-analysis. Unlike our primaryanalysis, this model did not assume that all diseases hadthe same direction of eﬀect and could detect allelic eﬀectsthat increase risk for some diseases and decrease risk forothers. The appendix describes statistical methods andresults, including the handling of trios and populationstratiﬁcation. We also examined loci that previouslyachieved genome-wide signiﬁcance in PGC meta-analyses of schizophrenia and bipolar disorder.
To characterise the speciﬁcity of the allelic eﬀects forour main ﬁndings, we examined the association evidencein three ways: we generated forest plots of the disorderbeta coeﬃ cients with 95% CIs; we calculated a hetero-geneity p value for the disorder-speciﬁc eﬀects con-tributing to the overall statistics for meta-analyticassociation; and we undertook a multinomial logisticregression procedure with model selection
for eachmain SNP for all ﬁve disorders to assess the pattern of phenotypic eﬀects (appendix pp 8–11). To compare the ﬁtof various models of genotype–phenotype associations,we applied established goodness-of-ﬁt metrics (theBayesian information criteria and the Akaike informationcriteria). We report the best-ﬁtting model by Bayesiancriteria and show results of both metrics for a range of models (appendix pp 38–45, 51–61).To examine shared polygenic risk at an aggregate levelbetween pairs of diagnoses, we used risk-score proﬁling aspreviously described.
For each pair, we selected onedisorder as a discovery dataset and the other as a targetdataset and calculated the proportion of variance in thetarget set explained by risk scores from the discovery setwith a range of statistical cutoﬀs for SNP inclusion in thescore (appendix p 13). To assess the role of speciﬁcbiological systems in the pathogenesis of the ﬁve disorders,we did pathway and eQTL analyses. Pathway analysis wasby interval-based enrichment analysis (INRICH) for thefull dataset consisting of linkage disequilibrium segmentscontaining signals with association p<10
³ in the primarymeta-analysis. INRICH accounts for potential genomicconfounding factors, such as variable gene and pathwaysizes, SNP density, linkage disequilibrium, and physicalclustering of biologically related genes (appendix pp 14–16).We did eQTL enrichment analysis
to assess whetherSNPs associated with ﬁve psychiatric disorders wereenriched for regulatory SNPs in post-mortem brain tissue