Common schizophrenia alleles are enriched in mutation-intolerant genes and in regions under strong background selection

Abstract

Schizophrenia is a debilitating psychiatric condition often associated with poor quality of life and decreased life expectancy. Lack of progress in improving treatment outcomes has been attributed to limited knowledge of the underlying biology, although large-scale genomic studies have begun to provide insights. We report a new genome-wide association study of schizophrenia (11,260 cases and 24,542 controls), and through meta-analysis with existing data we identify 50 novel associated loci and 145 loci in total. Through integrating genomic fine-mapping with brain expression and chromosome conformation data, we identify candidate causal genes within 33 loci. We also show for the first time that the common variant association signal is highly enriched among genes that are under strong selective pressures. These findings provide new insights into the biology and genetic architecture of schizophrenia, highlight the importance of mutation-intolerant genes and suggest a mechanism by which common risk variants persist in the population.

Schizophrenia is characterized by psychosis and negative symptoms such as social and emotional withdrawal. While onset of psychosis typically does not occur until late adolescence or early adulthood, there is strong evidence from clinical and epidemiological studies that schizophrenia reflects a disturbance of neurodevelopment¹. It confers substantial mortality and morbidity, with a mean reduction in life expectancy of 15-30 years^2,3. Although recovery is possible, most patients have poor social and functional outcomes⁴. No substantial improvements in outcomes have emerged since the advent of antipsychotic medication in the mid-twentieth century, a fact that has been attributed to a lack of knowledge of pathophysiology¹.

Schizophrenia is both highly heritable and polygenic, with risk ascribed to variants spanning the full spectrum of population frequencies^5–7. The relative contributions of alleles of various frequencies are not fully resolved, but recent studies estimate that common alleles, captured by genome-wide association study (GWAS) arrays, explain between one-third and one-half of the genetic variance in liability⁸. There has been a long-standing debate, from an evolutionary standpoint, as to how common risk alleles persist in the population, particularly given the early mortality and decreased fecundity associated with schizophrenia⁹. Various hypotheses have been proposed, including compensatory advantage (balancing selection), whereby schizophrenia-associated alleles confer reproductive advantages in particular contexts^10,11; hitchhiking, whereby risk-associated alleles are maintained by their linkage to positively selected alleles¹²; and contrasting theories that attribute these effects to rare variants and gene-environment interaction¹³. Addressing these competing hypotheses is now tractable given advances from recent studies of common genetic variation in schizophrenia.

The largest published schizophrenia GWAS, that from the Schizophrenia Working Group of the Psychiatric Genomics Consortium (PGC), identified 108 genome-wide significant loci and unequivocally demonstrated the value of increasing sample sizes for discovery in schizophrenia genetics research⁵. Here we report a large, phenotypically homogeneous GWAS of schizophrenia that, when combined with previously published data, identifies new facets of genetic architecture and biology and demonstrates that the evolutionary process of background selection contributes to the persistence of common risk alleles in the population.

Results

GWAS and meta-analysis

We obtained genome-wide genotype information for schizophrenia cases from the UK (the CLOZUK sample), which we combined with control datasets obtained from public repositories or through collaboration. The final sample size was 11,260 cases and 24,542 controls (5,220 cases and 18,823 controls not in previous schizophrenia GWAS; Methods and Supplementary Figs. 1 and 2). At a genome-wide level, the association statistics indicated that the common variant architecture in the CLOZUK sample was highly correlated with that in an independent sample of 29,415 cases and 40,101 controls from the PGC (genetic correlation = 0.954± 0.030; P = 6.63 × 10⁻²²⁷), and this was further confirmed by polygenic risk score and trend test analyses across the datasets at a range of association P-value thresholds (Methods and Supplementary Tables 1 and 2).

Meta-analysis of the CLOZUK and independent PGC datasets, excluding related and overlapping samples (total of 40,675 cases and 64,643 controls; Supplementary Fig. 3) identified 179 independent genome-wide significant SNPs (P < 5 × 10⁸; Supplementary Table 3) mapping to 145 independent loci (Fig. 1, Methods and Supplementary Table 4). The 145 associated loci included 93 of those that were genome-wide significant in the study of the PGC, the majority of which showed a strengthened association (Supplementary Fig. 4 and Supplementary Table 5). This does not imply that the remaining 15 PGC loci were false positives; rather, this reflects the expected inflation of effect sizes for genome-wide significant SNPs in incompletely powered studies and, as we demonstrate, is consistent with all 108 PGC loci representing true positives (Supplementary Note). Of the 52 loci not identified by the PGC, 2 have been reported as genome-wide significant in other studies: the locus at ZEB2¹⁴ and a locus on chromosome 8 (38.0-38.3 Mb)¹⁵.

Fig. 1. Manhattan plot of schizophrenia GWAS associations.

Associations are shown from the meta-analysis of CLOZUK and an independent PGC dataset (n = 105,318; 40,675 cases and 64,643 controls). The 145 genome-wide significant loci are highlighted in green. The red horizontal line indicates the genome-wide statistical significance threshold (P = 5×10⁻⁸).

In further independent samples (5,662 cases and 154,224 controls), 43 of the 50 genome-wide significant index SNPs showed the same pattern of allelic association, a level that far surpassed chance (P = 1.05 × 10⁻⁷). Despite the modest number of cases in these samples, 18 of the 50 index alleles reached nominal significance (P < 0.05), which again is implausible by chance (P = 1.46 × 10⁻¹¹). None demonstrated evidence for heterogeneity of effect (Methods and Supplementary Table 6).

Mutation-intolerant genes

Recent studies have shown that mutation-intolerant genes capture much of the rare variant architecture of neurodevelopmental disorders such as autism, intellectual disability and developmental delay, as well as schizophrenia^16–19. Here we show that, for schizophrenia, this also holds for common variation. Using gene set analysis in MAGMA²⁰, loss-of-function (LoF)- intolerant genes (n = 3,230) as defined by the Exome Aggregation Consortium (ExAC)²¹ using their gene-level constraint metric (pLI ≥ 0.9), were enriched for common variant associations with schizophrenia in comparison with all other annotated genes (P = 4.1 × 10⁻¹⁶).

It has been shown that pLI is correlated with gene expression across tissues, including brain²¹, which raises the possibility that the enrichment for LoF-intolerant genes in schizophrenia may reflect enrichment for signal in genes expressed in the brain. However, LoF-intolerant gene set enrichment was robust to the inclusion of both ‘brain-expressed’ (n = 10,360) and ‘brain-specific’ (n = 2,647) gene sets¹⁹ as covariates in the analysis (P = 1.89 × 10⁻¹⁰) or to controlling for FPKM gene expression values in brain²² (P = 1.03 × 10⁻¹⁴).

It has been suggested that clustering of risk alleles in mutationintolerant genes is a hallmark of early-onset traits under natural selection^23,24. However, LoF-intolerant genes are known to be enriched for SNPs identified as genome-wide significant in GWAS (as listed in the NHGRI-EBI GWAS Catalog²⁵) and for broad categories of disorders²¹. To examine whether our finding is a property of polygenic disorders in general, we obtained summary genetic data from a late-onset neuropsychiatric disorder (Alzheimer’s disease), a non-psychiatric disorder (type 2 diabetes) and a psychological trait (neuroticism), each of which has been shown to be under minimal selective pressure (Methods). These other phenotypes showed at best a weak signal for enrichment of the LoF-intolerant gene set in the MAGMA analysis, with the signal not comparable to that seen in schizophrenia (Alzheimer’s disease, P = 0.008; type 2 diabetes, P = 0.016; neuroticism, P = 0.066).

To quantify the contribution of SNPs within LoF-intolerant genes to schizophrenia SNP-based heritability (h²_SNP), we used partitioned linkage disequilibrium score regression (LDSR)²⁶ (Supplementary Table 7). Overall, genic SNPs accounted for 64% of h²_SNP, a 1.23-fold enrichment proportional to their SNP content (P = 5.93 × 10⁻¹⁴). Consistent with the analysis using MAGMA, h²_SNP was enriched in LoF-intolerant genes (2.01-fold; P = 2.78 × 10⁻²⁴), which explained 30% of all h²_SNP (equating to 47% of all genic h²_SNP). In contrast, genes classed as not LoF intolerant (pLI < 0.9) were significantly depleted for h²_SNP relative to their SnP content (0.90-fold; P = 5.86 × 10⁻³), although in absolute terms SNPs in these genes accounted for 34% of h²_SNP A finer-scale analysis of the relationship between LoF intolerance scores and enrichment for association showed that enrichment was restricted to genes with a pLI score above 0.9, precisely those defined as ‘LoF intolerant’ (Supplementary Fig. 5).

Common risk alleles in regions under background selection

Our finding that LoF-intolerant genes are enriched for common risk variants raises the question of how such alleles are found at common frequencies in the population. While the contribution of ultra-rare variation in functionally important genes to disorders associated with low fecundity can be accounted for by de novo mutation^16,19,27, this cannot explain the persistence of common alleles. To address this question, we used partitioned LDSR to test the relationship between schizophrenia-associated alleles and SNP-based signatures of natural selection. These included measures of positive selection, background selection and Neanderthal introgression. We examined the heritability of SNPs after thresholding them at extreme values for these metrics (top 2%, 1% and 0.5%), including in the baseline model annotation sets such as LoF-intolerant genes and genomic regions with extreme LD patterns (Methods).

We observed strong evidence for schizophrenia h²_SNP enrichment in SNPs under strong background selection (BGS), which was consistent across all the thresholds we examined (Table 1). We also found a significant depletion of h²_SNP in SNPs subject to positive selection as indexed by the CLR statistic. These two results are mutually consistent, as calculation of the CLR statistic explicitly controls for the effect of BGS²⁸. This suggests that SNPs under positive selection, but under weak or no BGS, are depleted for association with schizophrenia. No significant relationship between h²_SNP and other positive selection or Neanderthal introgression measures was found after correction for multiple testing (Table 1). An LDSR analysis treating BGS measures as a quantitative trait rather than as a binary one confirmed that the relationship between BGS and schizophrenia association was not due to the imposition of arbitrary thresholds to define strong BGS (P = 7.73 × 10⁻¹¹). We also note that the τ_c statistic of the LDSC model was significant for BGS, in both the binary (P = 0.041) and quantitative (P = 0.023) analyses (Supplementary Table 8). The τ_c statistic indicates the enrichment of BGS after controlling for all other annotations in the model (including LoF-intolerant genes)²⁶ and thus represents a robust and conservative test for BGS enrichment.

Table 1.

Heritability analysis of natural selection metrics

		Top 2% of scores (genome wide)		Top 1% of scores (genome wide)		Top 0.5% of scores (genome wide)
Metric	Ref.	Enrichment	2-sided P value	Enrichment	2-sided P value	Enrichment	2-sided P value
Background selection (B statistic)	[29]	1.801	0.001	2.341	9.90×10−4	2.365	0.002
Positive selection (CLR)	[28]	0.408	6.53×10−5	0.173	5.80×10−7	0.259	0.016
Positive selection (CMS)	[88]	0.054	0.001	−0.037	0.006	−0.039	0.007
Positive selection (XP-EEH)	[87]	0.621	0.342	0.383	0.303	0.125	0.268
Positive selection (iHS)	[86]	0.973	0.946	0.980	0.974	1.633	0.557
Neanderthal posterior probability (LA)	[89]	0.807	0.347	0.800	0.462	0.858	0.745

Partitioned LDSR regression results for SNPs thresholded by extreme values (defined as top percentiles versus all other SNPs) of each natural selection metric. All tests have been adjusted for 58 ‘baseline’ annotations, which include categories such as LoF intolerant, recombination coldspot and conserved (Methods). Enrichment values below 1 indicate a depletion of h²_SNP in an annotation category (less contribution than expected for a given number of SNPs). Negative enrichments should be considered zero (no contribution to h²_SNP by these SNPs). Bold values indicate results surviving correction after adjusting for all tests (Bonferroni a = 0.05/18 = 0.0028).

The above analyses account for a possible confounding relationship between LoF intolerance and BGS. To illustrate this more clearly, we binned the BGS intensities into four categories of increasing score and classified SNPs in these bins according to whether they were in LoF-intolerant genes, ‘all other’ gene sets or a non-genic set (Supplementary Fig. 6). Note that the lower boundary of the top bin (BGS intensity > 0.75) corresponds approximately to the top 2% BGS threshold in Table 1 and is equivalent to a reduction in effective population size estimated at each SNP of 75% or more²⁹. We found significant heritability enrichment across all BGS intensity intervals in LoF-intolerant genes that increased progressively with higher intensity scores. Notably, we also found heritability enrichment for SNPs under BGS pressure in genes that were not LoF intolerant, restricted to the highest BGS intensity bin. Indeed, the highest BGS intensity bin in non-LoF-intolerant genes was enriched for heritability at a level roughly equivalent to that for all LoF-intolerant genes. These findings point to BGS and LoF intolerance as making at least partially independent contributions to heritability enrichment in schizophrenia. In contrast, none of the phenotypes we selected on the basis of their minimal impact on fecundity (Alzheimer’s disease, type 2 diabetes and neuroticism) showed significant BGS enrichment for heritability either when using the BGS τ_c statistic of the LDSR model (minimum P > 0.22; Supplementary Table 8) or when specifically testing regions of high BGS intensity in genes that were tolerant (pLI < 0.9) of functional mutations (minimum P > 0.40).

Systems genomics

Using MAGMA, we undertook a primary analysis of 134 central nervous system (CNS)-related gene sets we have previously shown capture the excess copy number variation (CNV) burden in schizophrenia³⁰. In a GWAS context, we now show that, collectively, this group of gene sets captures a disproportionately high fraction of h²_SNP (30% of total heritability, enrichment = 1.63, P = 8.57 × 10⁻¹³, 46% of genic heritability; Supplementary Table 7). Of the 134 sets, 54 were nominally significant, of which 12 survived multiple-testing correction (family-wise error rate (FWER) P < 0.05; Supplementary Table 9), with no notable association for gene sets such as the ARC protein complex and the NMDAR protein network, that we have previously implicated in rare variant studies^30,31. Stepwise conditional analysis, adjusting sequentially for the more strongly associated gene sets, resulted in six gene sets that were independently associated with schizophrenia (Table 2 and Supplementary Data). These extended from low-level molecular and subcellular processes to broad behavioral phenotypes. The most strongly associated gene set constituted the targets of the fragile X mental retardation protein (FMRP)³². FMRP is a neuronal RNA-binding protein that interacts with polyribosomal mRNAs (the 842 target transcripts of this gene set³²) and is thought to act by inhibiting translation of target mRNAs, including many transcripts of pre- and postsynaptic proteins. The FMRP target set has been shown to be enriched for rare mutational burden in exome sequencing studies of de novo variation in autism³³ and intellectual disability³¹. In schizophrenia, it has also been shown to be nominally significantly enriched for association signal in sequencing studies^8,31 and GWAS^5,8, but has only inconsistently been associated in studies of CNV^30,34. Here we provide the strongest evidence thus far for enrichment of this gene set in schizophrenia.

Table 2.

Functional gene set analysis highlights six independent gene sets associated with schizophrenia

Gene set	Number of genes	Enrichment P value (FWER)a	Conditional P valueb
Targets of FMRP32	798	1×10−5	1.9×10−8
Abnormal behavior (MP:0004924)	1,939	1.8×10−4	1.4×10−5
5-HT2C receptor complex37	16	0.029	0.001
Abnormal nervous system electrophysiology (MP:0002272)	201	0.003	0.002
Voltage-gated calcium channel complexes36	196	0.011	0.016
Abnormal long-term potentiation (MP:0002207)	142	0.030	0.031

MP refers to Mammalian Phenotype Ontology terms of the MGI³⁵, from which gene sets were derived. FMRP, fragile X mental retardation protein.

^aWestfall-Young family-wise error rate, as implemented in MAGMA²⁰.

^bFrom stepwise conditional analysis that adjusts sequentially for ‘stronger’ associated gene sets.

We highlight another five gene sets that are independently associated with schizophrenia. Three of these derive from the Mouse Genome Informatics (MGI) database³⁵ and relate to behavioral and neurophysiological correlates of learning: abnormal behavior (MP:0004924), abnormal nervous system electrophysiology (MP:0002272) and abnormal long-term potentiation (MP:0002207). We note that two of these gene sets (MP:0004924 and MP:0002207) were among the five most enriched of the 134 gene sets tested in a recent schizophrenia CNV analysis³⁰. The remaining two independently associated genes sets were voltage-gated calcium channel complexes³⁶ and the 5-HT_2C receptor complex³⁷. The calcium channel finding confirms extensive evidence from common and rare variant studies implicating calcium channel genes in schizophrenia^5,8, including a new GWAS locus in CACNA1D identified in our meta-analysis. While there is less convergent evidence in support of the involvement of the 5-HT_2C receptor complex in schizophrenia, the fact that we identify independent association for this gene set implicates these genes in schizophrenia pathophysiology and potentially rejuvenates a previous avenue of 5-HT_2C ligand therapeutic endeavor in schizophrenia research³⁸. However, we interpret this result with caution given the small size of this gene set and the fact that a number of its genes encode synaptic proteins that are structurally related to other receptor complexes³⁷, not only 5-HT_2C.

Systems genomics and mutation-intolerant genes

The LoF-intolerant genes and the six conditionally independent (‘significant’) CNS-related gene sets together account for 39% of schizophrenia SNP-based heritability (P = 5.07 × 10⁻²⁶), equating to 61% of genic heritability (Fig. 2a and Supplementary Table 7). This is likely to be an underestimation of the true effect of these gene sets, as distal non-genic regulatory elements (not included in this analysis) will add to the heritability explained by these genes. In examining the relationship between the LoF-intolerant and CNS-related gene sets (Fig. 2a), genes belonging to both categories were the most highly enriched (2.6-fold, P = 7.90 × 10⁻¹⁵), although LoF-intolerant genes that were not annotated to our significant CNS gene sets still displayed enrichment for SNP-based heritability (1.74-fold, P = 9.77 × 10⁻¹⁰), while genes that were in the significant CNS gene sets but had pLI <0.9 showed more modest enrichment (1.39-fold, P = 6.05 × 10⁻⁴). Notably, genes outside these categories were depleted in heritability relative to their SNP content (enrichment = 0.79, P = 1.82 × 10⁻⁷).

Fig. 2. Partitioned heritability analysis of gene sets in schizophrenia.

a, Heritability of genomic partitions and the six conditionally independent (‘significant’) gene sets (Table 2). The radius of each segment indicates the degree of enrichment, while the arc (angle of each slice) indicates the percentage of total SNP-based heritability explained. No relative enrichment (enrichment = 1) is shown by the dashed red line (and depletion equates to enrichment <1, inside red line). b, Heritability of the significant CNS gene sets dissected by their overlap with LoF-intolerant genes. Whiskers represent heritability or enrichment standard errors. Asterisks indicate the significance of each heritability enrichment (*P≤0.05, **P≤0.01, ***P≤0.001).

This general pattern remained when we focused on the six significant CNS gene sets individually, in that the enrichment in these gene sets derived primarily from their intersection with LoF-intolerant genes (Fig. 2b). Indeed, only the targets of FMRP showed significant enrichment for SNPs in genes that were not LoF intolerant (2.06-fold, P = 4.23 × 10⁻⁵).

Data-driven gene set analysis

To set the systems genomics results in context and to ensure that we were not missing enrichment in other gene sets by our hypothesis-driven approach, we undertook a purely data-driven analysis of a larger comprehensive annotation of gene sets from multiple public databases, totaling 6,677 gene sets (Methods and Supplementary Table 10). Six gene sets survived FWER correction for the full 6,677 gene sets and showed independence through conditional analyses. The LoF-intolerant gene set was the most strongly enriched, followed by the two most strongly associated functional gene sets we had specified in our hypothesis-driven CNS gene set analysis (FMRP targets and MGI abnormal behavior genes). The other three sets were calcium ion import (GO:0070509), membrane depolarization during action potential (GO:0086010) and synaptic transmission (GO:0007268). These are highly overlapping with the independently associated sets from our primary CNS systems genomics analysis. Indeed, if we repeat the data-driven comprehensive gene set analysis while adjusting for the six independently associated CNS gene sets, the only surviving enrichment term is the LoF-intolerant genes. These results are consistent with those from CNV analysis³⁰ in that they do not support annotations other than those related to CNS function and demonstrate that hypothesis-based analysis to maximize power does not substantially impact the overall pattern of results.

Identifying likely candidates within associated loci

To identify SNPs and genes that might be causally linked to the genome-wide significant associations, we used FINEMAP³⁹ to identify credibly causal alleles (those with a cumulative posterior probability for a locus of at least 95%) and functionally annotated these alleles using ANNOVAR⁴⁰. This identified 6,105 credible SNPs across 144 genome-wide significant loci, excluding the major histocompatibility complex (MHC) region (Methods and Supplementary Table 11). From these, we defined a highly credible set of SNPs (n = 25) as those that were more likely to explain the associations than all other SNPs combined (i.e., with a FINEMAP posterior probability greater than 0.5). Of these, 14 mapped to genes on the basis of putative functionality (exonic SNPs that cause nonsynonymous or splice variations or promoter SNPs; n = 6) or mapped to regions identified as likely regulatory elements (n = 8) through chromosome conformation analysis performed in tissue from the developing brain using Hi-C⁴¹ physical interactions (Methods and Supplementary Table 12). One of the implicated alleles was a nonsynonymous variant in the manganese and zinc transporter gene SLC39A8. Nonsynonymous variants in this gene, which lead to SLC39A8 deficiency, have been associated with severe neurodevelopmental disorders putatively through impaired manganese transport and glycosylation⁴², highlighting a mechanism of therapeutic potential for schizophrenia.

We also applied Summary-data-based Mendelian Randomization (SMR) analysis⁴³ to the data in concert with dorsolateral prefrontal cortex expression quantitative trait locus (eQTL) data from the CommonMind Consortium⁴⁴, aiming to identify variants that might be causally linked through expression changes in specific genes (Methods and Supplementary Table 13). After applying a conservative threshold (P_HEIDI > 0.05) that prioritized colocalized signals due to a single causal variant⁴³, we identified 22 candidates at 19 loci with false discovery rate (FDR) P < 0.05.

In total, the combination of FINEMAP, Hi-C and SMR analyses assigned potentially causal genes at 33 genome-wide significant loci and implicated a single gene at 27 of these loci. However, the analyses intersect for only a single gene, ZNF823, indicating the need for more comprehensive functional genomic annotations in CNS- relevant tissues.

Discussion

In the largest genetic study of schizophrenia thus far, we explore the genomic architecture of and the evolutionary pressures on common variants associated with the disorder. Our study provides the first evidence linking common variation in LoF-intolerant genes to risk of developing schizophrenia and demonstrates that these genes account for a substantial proportion (30%) of the SNP-based heritability for schizophrenia. Systems genomics analysis highlights six gene sets that are independently associated with schizophrenia and point to molecular, physiological and behavioral pathways involved in schizophrenia pathogenesis.

Given that mutation intolerance is due to high selection pressure^21,23,24, our finding that schizophrenia risk variants that persist at common allele frequencies are enriched in LoF-intolerant genes might appear counterintuitive. However, new evidence presented here suggests that this can be reconciled by BGS, which is a consequence of purifying selection in regions of low recombination^45,46. In such regions, recurrent selection against deleterious variants causes haplotypes to be removed from the gene pool, which reduces genetic diversity in a manner equivalent to a reduction in effective population size⁴⁷. This in turn impairs the efficiency of the selection process, allowing alleles with small deleterious effects to rise in frequency by drift⁴⁸. Such a consequence of purifying selection has been shown to be compatible with the genomic architecture of complex human traits⁴⁹ and to influence phenotypes in model organisms⁵⁰. We have explicitly modeled this effect (both theoretically and via simulations; Supplementary Note) and provide strong evidence for the feasibility of this effect as explanatory for the effect sizes seen for common alleles in schizophrenia.

We did not find enrichment for any measure of positive selection or Neanderthal introgression. A recent study explained a negative correlation between schizophrenia associations and metrics indicative of a Neanderthal selective sweep as evidence for positive selection or polygenic adaptation in schizophrenia¹². We do not find any significant correlation in our model, which addresses the contribution of BGS, and hence our results are not consistent with large contributions of positive selection to the genetic architecture of schizophrenia (Table 1). Indeed, positive selection is not widespread in humans, as reported by other studies that explicitly considered or accounted for BGS^28,51. Polygenic adaptation, the co-occurrence of many subtle allele frequency shifts at loci influencing complex traits⁵², remains an intriguing possibility but has not been implicated in psychiatric phenotypes, including schizophrenia, in recent analyses^53,54. In contrast, BGS has been proposed as a mechanism driving human-Neanderthal incompatibilities, as regions with stronger estimated BGS have lower estimated Neanderthal introgression⁵⁵. We therefore conclude that the bulk of the BGS signal we obtain is unlikely to be influenced by positive selection²⁹, challenging theories of the selective advantage of schizophrenia risk alleles to explain the high population frequencies of these alleles.

Methods

Methods, including statements of data availability and any associated accession codes and references, are available at https://doi.org/10.1038/s41588-018-0059-2.

Methods

GWAS and reporting of independently associated regions

Details of sample collection and genotype quality control are given in the Supplementary Note. The CLOZUK schizophrenia GWAS was performed using logistic regression with imputation probabilities (‘dosages’) adjusted for 11 principal-component analysis (PCA) covariates. These covariates were chosen as those nominally significant (P < 0.05) in a logistic regression for association with the phenotype⁵⁶. To avoid overburdening the GWAS power by adding too many covariates to the regression model⁵⁷, only the first 20 principal components were considered and tested for inclusion, as higher numbers only become useful for the analysis of populations that bear strong signatures of complex admixture⁵⁸. The final set of covariates included the first five principal components (as recommended for most GWAS approaches⁵⁹) and principal components 6, 9, 11, 12, 13 and 19. Quantile–quantile and Manhattan plots are shown in Supplementary Figs. 7 and 8.

To identify independent signals among the regression results, signals were amalgamated into putative associated loci using the same two-step strategy and parameters as PGC (Supplementary Table 14). In this procedure, regular LD clumping is performed (r² = 0.1, P < 1×10⁻⁴; window size < 3 Mb) to obtain independent index SNPs. Afterward, loci are defined for each index SNP as the genomic region that contains all other imputed SNPs within the region with r² ≥ 0.6. To avoid inflating the number of signals in gene-dense regions or in those with complex LD, all loci within 250 kb of each other were annealed.

Meta-analysis with PGC

A total of 6,040 cases and 5,719 controls from CLOZUK were included in the recent PGC study⁵. We reanalyzed the PGC data after excluding all these cases and controls, obtaining a sample termed ‘INDEPENDENT PGC’ (29,415 cases and 40,101 controls). Adding the summary statistics from this independent sample to the CLOZUK GWAS results allowed for a combined analysis of 40,675 cases and 64,643 controls (without duplicates or related samples). This meta-analysis was performed using the fixed-effects procedure in METAL⁶⁰ with weights derived from standard errors. For consistency with the PGC analysis, additional filters (INFO > 0.6 and MAF > 0.01) were applied to the CLOZUK and INDEPENDENT PGC summary statistics, leaving 8 million markers in the final meta-analysis results. Quantile–quantile and Manhattan plots are shown in Supplementary Fig. 3 and Fig. 2. The same procedure as above was used to report independent loci from this analysis (Supplementary Tables 3 and 4). As raw PGC genotypes were not available for the LD clumping procedure, phase 3 of the 1000 Genomes Project (1KGPp3) was used as a reference.

Replication of new GWAS loci

To validate the association signals from the CLOZUK + PGC meta-analysis, we amalgamated data contributed by other schizophrenia genetics consortia (total of 5,762 cases and 154,224 controls; details in the Supplementary Note). We sought GWAS summary statistic data for the index SNPs from the 50 new genome-wide significant loci (Supplementary Table 4). These summary statistics were subjected to meta-analysis in METAL using the fixed-effects procedure to obtain replication and heterogeneity statistics (Supplementary Table 6).

Estimation and assessment of a polygenic signal

Association signals caused by the vast polygenicity underlying complex traits can be hard to distinguish from confounders related to sample relatedness and population stratification. To effectively disentangle this issue, we used the software LD Score v1.0 to analyze the summary statistics of our association analyses and estimate the contribution of confounding biases to our results by LDSR⁶¹. An LD reference was generated from 1KGPp3 after restricting this dataset to strictly unrelated individuals and retaining only markers with MAF >0.01. To improve accuracy, the summary statistics used as input were refined by discarding all indels and restricting SNPs to those with INFO >0.9 and MAF >0.01, a total of 5.16 million SNPs. The resulting LD score intercept for the CLOZUK GWAS was 1.085 ± 0.010, which compared to a mean χ² of 1.417 indicates a polygenic contribution of at least 80%. For the CLOZUK + PGC meta-analysis, the LD score intercept was 1.075 ± 0.014 (mean χ² = 1.960), which supports more than 90% of the signal being driven by polygenic architecture. Both of these figures are in line with those for other well-powered GWAS of complex human traits⁶⁴, including schizophrenia⁵. This analysis was also used to calculate SNP-based heritability (h²_SNP) for our three datasets (CLOZUK, INDEPENDENT PGC and the CLOZUK + PGC meta-analysis), which we transformed to a liability scale using a population prevalence of 1% (registry-based lifetime prevalence⁶²). For reference and compatibility with epidemiological studies of schizophrenia, prevalence estimates of 0.7% (lifetime morbid risk⁶³) and 0.4% (point prevalence⁶³, more akin to treatment-resistant schizophrenia prevalence (appropriate for CLOZUK)) were used for additional liability-scale h²_SNP calculations (Supplementary Table 15).

The LDSR framework allowed us to compare the genetic architecture of CLOZUK and INDEPENDENT PGC, by calculating the correlation of their summary statistics⁶⁴. A genetic correlation coefficient of 0.954±0.030 was obtained, with a P value of 6.63×10⁻²²⁷. We also examined the independent SNPs that reached a genome-wide significant level in the INDEPENDENT PGC dataset, of which there were 76 after excluding the extended major histocompatibility complex (xMHC) region. In the CLOZUK sample, 76% (n = 57) of these genome-wide significant SNPs were nominally significant (P < 0.05). Using binomial sign tests based on clumped subsets of SNPs⁶⁵, we found that all but 1 (98.6%) of these 76 genome-wide significant SNPs were associated with the same direction of effect in the CLOZUK sample, a result highly unlikely to reflect chance (P = 2.04 × 10⁻²¹; Supplementary Table 1). Moreover, of the 1,160 SNPs with an association P value less than 1 × 10⁻⁴ in the INDEPENDENT PGC sample, 82% showed enrichment in the CLOZUK cases (P = 3.44 × 10⁻¹¹³), confirming that very large numbers of true associations will be discovered among these SNPs with increased sample sizes. Additionally, the new sample introduced in this study (CLOZUK2) was compared by the same methods with the PGC dataset and showed results consistent with the full CLOZUK analysis, providing molecular validation of this sample as a schizophrenia sample (Supplementary Table 1).

We went on to conduct polygenic risk score analysis. Polygenic scores for CLOZUK were generated from INDEPENDENT PGC as a training set, using the same parameters for risk profile score (RPS) analysis in PGC⁵, arriving at a high-confidence set of SNPs for RPS estimation by removing the xMHC region and indels, and applying INFO > 0.9 and MAF > 0.1 cutoffs. Scores were generated from the autosomal imputation dosage data, using a range of P-value thresholds for SNP inclusion⁶⁶ (5 × 10⁻⁸, 1 × 10⁻⁵, 0.001, 0.05 and 0.5). In this way, we can assess the presence of a progressively increasing signal-to-noise ratio in relation to the number of markers included⁶⁷. As in the PGC study, we found the best P-value threshold for discrimination to be 0.05 and report highly significant polygenic overlap between the INDEPENDENT PGC and CLOZUK samples (P < 1 × 10⁻³⁰⁰, Nagelkerke r² = 0.12; Supplementary Table 2), confirming the validity of combining the datasets. For comparison with other studies, we also report polygenic variance on the liability scale⁶⁸, which amounted to 5.7% for CLOZUK at the 0.05 P-value threshold (Supplementary Table 2). As in the PGC study, the limited r² and area under the receiver operating characteristic curve (AUROC) obtained by this analysis restrict the current clinical utility of these scores in schizophrenia.

Gene set analysis

To assess the enrichment of sets of functionally related genes, we used MAGMA v1.03²⁰ on the CLOZUK + PGC meta-analysis summary statistics. From these, we excluded the xMHC region for its complex LD and the X chromosome given its smaller sample size. In the resulting data, gene-wide P values were calculated by combining the P values of all SNPs inside genes after accounting for LD and outliers. This was performed allowing for a window of 35 kb upstream and 10 kb downstream of each gene to capture the signal of nearby SNPs that could fall in regulatory regions^69,70. Next, we calculated competitive gene set P values on the gene-wide P values after accounting for gene size, gene set density and LD between genes. For multiple-testing correction in each gene set collection, an FWER⁷¹ was computed using 100,000 resamplings.

We performed sequential analyses using the following approaches:

1. LoF-intolerant genes. We tested the enrichment of the LoF-intolerant genes described by ExAC²¹. This set comprised all genes defined in the ExAC database as having a probability of LoF intolerance (pLI) statistic higher than 90%. Although these genes do not form part of cohesive biological processes or phenotypes, they have previously been found to be highly expressed across tissues and developmental stages²¹. Also, they are enriched for hub proteins⁷², which makes them interesting candidates for involvement in the ‘evolutionary canalization’ processes that have been proposed to lead to pleiotropic, complex disorders⁷³.

2. CNS-related genes. These gene sets were compiled in our recent study³⁰ and include 134 gene sets related to different aspects of CNS function and development. These include, among others, gene sets that have been implicated in schizophrenia by at least two independent large-scale sequencing studies^8,31: targets of FMRP³², constituents of the N-methyl-D-aspartate receptor (NMDAR⁷⁴) and activity-regulated cytoskeleton-associated protein complexes (ARCs^75,76), as well as CNS and behavioral gene sets from MGI database version 6³⁵.

3. Genes identified by data-driven analysis. The final systems genomic analysis was designed as an ‘agnostic’ approach, with the aim of integrating a large number of gene sets from different public sources, not necessarily conceptually related to psychiatric disorders, as this has been successful elsewhere^70,77. We conducted this analysis to test whether additional gene sets were associated in addition to those from the 134 CNS-related gene sets. For this, first we merged together the LoF-intolerant gene set and the 134 sets in the CNS-related collection. Second, we selected additional gene set sources to encompass a comprehensive collection of biochemical pathways and gene regulatory interaction networks: 2,693 gene sets with direct experimental evidence and a size of 10–200 genes⁷⁰ were extracted from Gene Ontology (GO⁷⁸) database release 01/02/2016; 1,787 gene sets were extracted from the fourth ontology level of MGI database version 6³⁵; 1,585 gene sets were extracted from REACTOME⁷⁹ version 55; 290 gene sets were extracted from KEGG⁸⁰ release 04/2015; and 187 gene sets were extracted from OMIM⁸¹ release 01/02/2016.

The total number of gene sets included was 6,677.

Detailed results of the analyses of the CNS-related and data-driven collection are given in Supplementary Tables 9 and 10. Reported numbers of genes in each gene set are those with available data in the meta-analysis. This may differ from the original gene set description, as some genic regions had null or poor SNP coverage. Following the data-driven gene set analysis as described, we also conducted analysis adjusting for our CNS-related gene sets to determine whether the data-driven analysis was contributing additional findings.

Partitioned heritability analysis of gene sets

It is known that the power of a gene set analysis is closely related to the total heritability of the phenotype and the specific heritability attributable to the tested gene set⁸². To assess the heritability explained by the genes carried forward after the main gene set analysis, LD Score was again used to compute a partitioned heritability estimate of CLOZUK + PGC using the gene sets as SNP annotations. As in the MAGMA analysis, the xMHC region was excluded from the summary statistics. These were also trimmed to contain no indels and only markers with INFO >0.9 and MAF >0.01, for a total of 4.64 million SNPs. As a recognized caveat of this procedure is that model misspecification can inflate the partitioned heritability estimates²⁶, all gene sets were annotated twice: once using their exact genomic coordinates (extracted from the NCBI RefSeq database⁸³) and another time with putative regulatory regions taken into account using the same upstream/downstream windows as in the MAGMA analyses. Additionally, all SNPs not directly covered by our gene sets of interest were explicitly included into other annotations (‘non-genic’, ‘genic but not LoF intolerant’) on the basis of their genomic location. Finally, the ‘baseline’ set of 53 annotations from Finucane et al.²⁶, which recapitulates important molecular properties such as presence of enhancers or phylogenetic conservation, was also incorporated in the model. All of these annotations were then tested jointly for heritability enrichment. We note that using exact genic coordinates or adding regulatory regions made little difference to the estimated enrichment of our gene sets; thus, throughout the manuscript, we report the latter for consistency with the gene set analyses (Fig. 2 and Supplementary Table 8).

Natural selection analyses

We aimed to explore the hypothesis that some form of natural selection is linked to the maintenance of common genetic risk in schizophrenia^12,84,85. To do this, for all SNPs included in the CLOZUK + PGC meta-analysis summary statistics, we obtained four different genome-wide metrics of positive selection (iHS⁸⁶, XP-EEH⁸⁷, CMS88 and CLR²⁸), one of background selection (B statistic²⁹, postprocessed by Huber et al.²⁸) and one of Neanderthal introgression (average posterior probability LA⁸⁹). The use of different statistics is motivated by the fact that each of them is tailored to detect a particular selective process that acted on a particular timeframe (see Vitti et al.⁵¹ for a review). For example, iHS and CMS are based on the inference of abnormally long haplotypes and thus are better powered to detect recent selective sweeps that occurred during the last ~30,000 years⁸⁸, such as those linked to lactose tolerance or pathogen response⁹⁰. On the other hand, CLR incorporates information about the spatial pattern of genomic variability (the site frequency spectrum⁹¹) and corrects explicitly for evidence of BGS, thus being able to detect signals from 60,000 to 240,000 years ago²⁸. The B statistic uses phylogenetic information from other primates (chimpanzee, gorilla, orangutan and rhesus macaque) to infer the reduction in allelic diversity that exists in humans as a consequence of purifying selection on linked sites over evolutionary time frames⁹². As the effects of background selection on large genomic regions can mimic those of positive selection⁴⁶, it is possible that the B statistic might amalgamate both, although the rather large diversity reduction that it infers for the human genome as a whole suggests that any bias due to positive selection is likely to be minor⁹³. Finally, XP-EEH is a haplotype-based statistic that compares two population samples, and its power is thus increased for alleles that have suffered differential selective pressures since those populations diverged⁹⁰. Although methodologically different, LA has a similar rationale by comparing human and Neanderthal genomes⁸⁹, to infer the probability of each human haplotype having been the result of an admixture event with Neanderthals.

For this work, CLR, CMS, the B statistic and LA were retrieved directly from their published references and lifted over to GRCh37 genomic coordinates if required using the Ensembl LiftOver tool^94,95. As the available genome-wide measures of iHS and XP-EEH were based on HapMap 3 data⁹⁶, both statistics were recalculated with the HAPBIN⁹⁷ software directly on the EUR superpopulation of the 1KGPp3 dataset, with the AFR superpopulation used as the second population for XP-EEH. Taking advantage of the fine-scale genomic resolution of these statistics (between 1-10 kb), all SNP positions present in CLOZUK + PGC were assigned a value for each measure, either directly (if the position existed in the lifted-over data) or by linear interpolation. To simplify the interpretation of our results, all measures were transformed before further analyses to a common scale, in which larger values indicate stronger effect of selection or increased probability of introgression. For example, the BGS B statistic, for which values of zero indicate the strongest effect (see Charlesworth⁴⁵ for its theoretical derivation), was included in all our analyses as 1 - B, which we termed ‘BGS intensity’.

Heritability enrichment of these statistics was tested by the LD Score partitioned heritability procedure. We derived binary annotations from the natural selection metrics by dichotomizing at extreme cutoffs defined by the top 2%, 1% and 0.5% of the values of each metric in the full set of SNPs. This approach is widely used in evolutionary genomics, owing to the difficulty of setting specific thresholds to define regions under selection^28,51. Consistent with the previously described LDSR partitioned heritability protocol, enrichment was estimated with all binary annotations included in a model with multiple categories that represent important genomic features. This model included the 3 main categories of our set-based analysis (‘non-genic’, ‘genic’ and ‘LoF intolerant’), 2 categories based on genomic regions with outlying LD patterns (recombination hotspots and coldspots)⁹⁸ and the 53 ‘baseline’ categories of Finucane et al.²⁶.

We then derived the τ_c coefficient²⁶ (and associated P value) of the significantly enriched natural selection annotations (i.e., the background selection metric), This represents the enrichment of an annotation over and above the enrichment of all other annotations, which is a conservative approach, as most of the categories in our model are partially overlapping. To increase our power and for additional validation, we noted that LD Score allows testing of the full range of quantitative metrics, in an extension of the partitioned heritability framework. Results of this analysis are reported in Supplementary Table 8.

Analysis of other phenotypes

To explore the specificity of our natural selection results, we retrieved data from other well-powered GWAS of complex traits. We selected three phenotypes for which (i) the genome-wide summary statistic data were publicly available, (ii) the sample size was larger than 50,000 individuals, (iii) the phenotype has minimal impact on fecundity^99–101 (and hence the traits behave as neutral or approximately neutral to selection) and (iv) summary statistics were considered adequate for LD Score analysis based on baseline z scores >4^26,102 (Supplementary Table 8). The phenotypes chosen were Alzheimer’s disease¹⁰³, neu roticism¹⁰⁴ and type 2 diabetes¹⁰⁵. For the LD Score analyses, as the public release of these statistics did not include imputation INFO scores at the time of this study, we restricted the set of SNPs to those included in the HapMap 3 project⁹⁶, as recommended⁶¹. To facilitate comparison with the schizophrenia results, we also restricted our schizophrenia summary statistic data to these SNPs and repeated the analyses above using BGS as a binary (top 2%) and quantitative trait.

We also employed MAGMA on the summary statistics of these additional phenotypes to examine whether the LoF-intolerant gene set enrichment displayed specificity to schizophrenia, after excluding the xMHC and APOE regions.

Fine-mapping, Hi-C and SMR

Accurately locating causal genes (‘fine-mapping’) for complex disorders is a challenge to GWAS and usually requires multiple approaches¹⁰⁵. To highlight credibly causal variants, we used FINEMAP v1.1³⁹ at each of the 145 identified loci (Supplementary Table 3), selecting variants with a cumulative posterior probability of 95%. These were then annotated with ANNOVAR⁴⁰ release 2016Feb1 (Supplementary Table 11). We mapped the SNPs with a FINEMAP posterior probability higher than 0.5 to the developing brain Hi-C data generated by Won et al.⁴¹, following the methodology described therein, which allowed us to implicate genes by chromatin interactions instead of solely chromosomal position (Supplementary Table 12). We compiled results from the eQTL analysis of the CommonMind Consortium post-mortem brain tissues⁴⁴. This included 15,782 genes, which were curated to remove any genes with FPKM = 0 across >10% of individuals. All the SNPs from the meta-analysis data were mapped to the eQTL data using rs numbers, position and allele matching. Both datasets were analyzed together using SMR⁴³, which resulted in 4,276 genes showing eQTLs with overlapping SNPs and genome-wide significant P values (Supplementary Table 13).

URLs

CLOZUK + PGC2 meta-analysis summary statistics, http://walters.psycm.cf.ac.uk/; CRESTAR Consortium, http://www.crestar-project.eu/; Wellcome Trust Case Control Consortium, http://www.wtccc.org.uk/; People of the British Isles project, http://www.peopleofthebritishisles.org/; Mouse Genome Informatics (MGI), http://www.informatics.jax.org/; Psychiatric Genomics Consortium, http://www.med.unc.edu/pgc/; 1000 Genomes IBD segment sharing within and between populations, http://ftp.1000genomes.ebi.ac.uk/vol1/ftp/release/20130502/supporting/ibd_by_pair/.

Life Sciences Reporting Summary

Further information on experimental design is available in the Life Sciences Reporting Summary.

Data availability

The gene content of the CNS-related gene sets that survived conditional analysis (significant) is given in MAGMA format in the Supplementary Data. Summary statistics from the CLOZUK + PGC2 GWAS are available for download (see URLs).

GERAD1 Consortium

Denise Harold^47,48, Rebecca Sims⁴⁷, Amy Gerrish⁴⁷, Jade Chapman⁴⁷, Valentina Escott-Price¹, Richard Abraham⁴⁷, Paul Hollingworth⁴⁷, Jaspreet Pahwa⁴⁷, Nicola Denning⁴⁷, Charlene Thomas⁴⁷, Sarah Taylor⁴⁷, John Powell⁴⁹, Petroula Proitsi⁴⁹, Michelle Lupton⁴⁹, Simon Lovestone^49,50, Peter Passmore⁵¹, David Craig⁵¹, Bernadette McGuinness⁵¹, Janet Johnston⁵¹, Stephen Todd⁵¹, Wolfgang Maier⁵², Frank Jessen⁵², Reiner Heun⁵², Britta Schurmann^52,53, Alfredo Ramirez⁵², Tim Becker⁵⁴, Christine Herold⁵⁴, André Lacour⁵⁴, Dmitriy Drichel⁵⁴, Markus Nothen⁵⁵, Alison Goate⁵⁶, Carlos Cruchaga⁵⁶, Petra Nowotny⁵⁶, John C. Morris⁵⁶, Kevin Mayo⁵⁶, Peter Holmans¹, Michael O’Donovan¹, Michael Owen¹ and Julie Williams⁴⁷

CRESTAR Consortium

Evanthia Achilla⁵⁷, Esben Agerbo^21,22, Cathy L. Barr⁵⁸, Theresa Wimberly Böttger⁵⁹, Gerome Breen^7,8, Dan Cohen⁶⁰, David A. Collier^7,44, Sarah Curran^61,62, Emma Dempster⁶³, Danai Dima⁷, Ramon Sabes-Figuera⁵⁷, Robert J. Flanagan⁶⁴, Sophia Frangou⁶⁵, Josef Frank⁶⁶, Christiane Gasse^59,67, Fiona Gaughran⁴, Ina Giegling⁴⁵, Jakob Grove^21,23,24,32, Eilis Hannon⁶³, Annette M. Hartmann⁴⁵, Barbara Heißerer⁶⁸, Marinka Helthuis⁶⁹, Henriette Thisted Horsdal⁵⁹, Oddur Ingimarsson⁷⁰, Karel Jollie⁶⁹, James L. Kennedy⁷¹, Ole Köhler³³, Bettina Konte⁴⁵, Maren Lang⁶⁶, Sophie E. Legge¹, Cathryn Lewis⁷, James MacCabe⁴, Anil K. Malhotra⁷², Paul McCrone⁵⁷, Sandra M. Meier⁵⁹, Jonathan Mill^7,63, Ole Mors^21,34, Preben Bo Mortensen^21,22,23, Markus M. Nöthen⁵⁵, Michael C. O’Donovan¹, Michael J. Owen¹, Antonio F. Pardiñas¹, Carsten B. Pedersen^21,22, Marcella Rietschel⁶⁶, Dan Rujescu^45,46, Ameli Schwalber⁶⁸, Engilbert Sigurdsson⁷⁰, Holger J. Sørensen³⁵, Benjamin Spencer⁷³, Hreinn Stefansson³⁹, Henrik Støvring⁶⁷, Jana Strohmaier⁶⁶, Patrick Sullivan^74,75, Evangelos Vassos⁷, Moira Verbelen⁷, James T. R. Walters¹ and Thomas Werge^21,41,42

⁴⁷MRC Centre for Neuropsychiatric Genetics and Genomics, Neurosciences and Mental Health Research Institute, Department of Psychological Medicine and Neurology, School of Medicine, Cardiff University, Cardiff, UK. ⁴⁸Neuropsychiatric Genetics Group, Department of Psychiatry, Trinity Centre for Health Sciences, St James’s Hospital, Dublin, Ireland. ⁴⁹Institute of Psychiatry, Department of Neuroscience, King’s College London, London, UK. ⁵⁰Department of Psychiatry, University of Oxford, Oxford, UK. ⁵¹Ageing Group, Centre for Public Health, School of Medicine, Dentistry and Biomedical Sciences, Queen’s University, Belfast, UK. ⁵²Department of Psychiatry, University of Bonn, Bonn, Germany. ⁵³Institute for Molecular Psychiatry, University of Bonn, Bonn, Germany. ⁵⁴Deutsches Zentrum für Neurodegenerative Erkrankungen (DZNE), Bonn, Germany. ⁵⁵Department of Genomics, Life & Brain Center, University of Bonn, Bonn, Germany. ⁵⁶Departments of Psychiatry, Neurology and Genetics, Washington University School of Medicine, St. Louis, MO, USA. ⁵⁷Centre for Economics of Mental and Physical Health, Health Service and Population Research Department, Institute of Psychiatry, King’s College London, London, UK. ⁵⁸Toronto Western Research Institute, University Health Network, Toronto, Ontario, Canada. ⁵⁹National Centre for Register-Based Research, Department of Economics and Business, School of Business and Social Sciences, Aarhus University, Aarhus, Denmark. ⁶⁰Department of Community Mental Health, Mental Health Organization North-Holland North, Heerhugowaard, the Netherlands. ⁶¹Department of Child and Adolescent Psychiatry, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, London, UK. ⁶²Brighton and Sussex Medical School, University of Sussex, Brighton, UK. ⁶³University of Exeter Medical School, RILD, University of Exeter, Exeter, UK. ⁶⁴Toxicology Unit, Department of Clinical Biochemistry, King’s College Hospital NHS Foundation Trust, London, UK. ⁶⁵Clinical Neurosciences Studies Center, Department of Psychiatry, Icahn School of Medicine at Mount Sinai, New York, NY, USA. ⁶⁶Department of Genetic Epidemiology in Psychiatry, Central Institute of Mental Health, Medical Faculty Mannheim/Heidelberg University, Mannheim, Germany. ⁶⁷Centre for Integrated Register-based Research, CIRRAU, Aarhus University, Aarhus, Denmark. ⁶⁸Concentris Research Management, Fürstenfeldbruck, Germany. ⁶⁹Leyden Delta, Nijmegen, the Netherlands. ⁷⁰Department of Psychiatry, Landspitali University Hospital, Reykjavik, Iceland. ⁷¹Centre for Addiction and Mental Health, Toronto, Ontario, Canada. ⁷²Division of Psychiatry Research, Zucker Hillside Hospital, Northwell Health System, Glen Oaks, NY, USA. ⁷³Department of Psychological Medicine, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, London, UK. ⁷⁴Center for Psychiatric Genomics, Department of Genetics, University of North Carolina, Chapel Hill, NC, USA. ⁷⁵Department of Medical Epidemiology and Biostatistics, Karolinska Institutet, Stockholm, Sweden.