We genotyped the International Schizophrenia Consortium (ISC) case-control sample for up to ~1 million single nucleotide polymorphisms (SNPs) augmented by imputed common HapMap SNPs. In the GWAS (λ_GC=1.09; Table S1, Figure S1–S3), the most associated genotyped SNP (P = 3.4×10⁻⁷) was located in the first intron of myosin XVIIIB (MYO18B) on chromosome 22. The second strongest association comprised over 450 SNPs on chromosome 6p spanning the MHC (Figure 1). There is some evidence for between-site heterogeneity in both allele frequencies and odds ratios (Table 1). We observed associations consistent with previous reports in the 22q11.2 deletion region and ZNF804A³ (Table S2, Figure S2 and Section S5).

Open in a separate window

Figure 1

Association results across the MHC region

Results are shown as –log10 (P-value) for genotyped SNPs. The most associated SNP is shown as a blue diamond. The colour of the remaining markers reflects r² with rs3130375, light pink, r² > 0.1, red, r² > 0.8. The recombination rate from the CEU HapMap (second Y-axis) is plotted in light blue.

The best imputed SNP, which reached genome-wide significance (rs3130297, P = 4.79×10⁻⁸, T allele, OR=0.747, MAF=0.114, 32.3Mb), was also in the MHC, 7kb from NOTCH4, a gene with previously reported associations with SCZ⁴. We imputed classical human leukocyte antigen (HLA) alleles; 6 were significant at P < 10⁻³, found on the ancestral European haplotype⁵ (Table 1, Table S3, Section S3). However, it was not possible to ascribe the association to a specific HLA allele, haplotype or region (Table S3, Figure S4).

We exchanged GWAS summary results with the Molecular Genetics of Schizophrenia (MGS) and SGENE consortia for genotyped SNPs with P < 10⁻³. There were 8,014 cases and 19,080 controls of European descent in the combined sample (see companion manuscripts, Section S7). Our top genotyped MHC SNP (rs3130375) had P = 0.086 and P = 0.14 in MGS and SGENE. Considering combined results for genotyped and imputed SNPs across the MHC region more broadly, rs13194053 had a genome-wide significant combined P = 9.5×10⁻⁹ (ISC, MGS and SGENE P = 3×10⁻⁴, 1×10⁻² and 1×10⁻⁴ respectively; C allele OR = 0.82, 0.88 and 0.78) and was in LD with rs3130375 (r²=0.35 in HapMap). Across the region 11 other SNPs had P < 10⁻⁷ at 27.1 – 27.3Mb and 32.7Mb (Table S5).

Our second approach was to evaluate whether common variants play an important role en masse, directly testing the classic theory of polygenic inheritance⁶, applied to SCZ by Gottesman & Shields⁷. While our GWAS analysis did not identify a large number of strongly associated loci, there could still be many – potentially thousands – of very small individual effects that collectively account for a substantial proportion of variation in risk. We summarized variation across many nominally-associated loci into quantitative scores and related the scores to disease state in independent samples⁸. Although variants of small effect (e.g. genotypic relative risk, GRR=1.05) are unlikely to achieve even nominally significant p-values, increasing proportions will be detected at increasingly liberal significance thresholds (p_T), for example p_T < 0.1 or p_T < 0.5. Using such thresholds, we defined large sets of “score alleles” in a discovery sample, in order to generate aggregate risk scores for individuals in independent target samples. We use the term score, instead of risk, as we cannot differentiate the minority of true risk alleles from unassociated variants.

We performed the score analyses on a reduced set of SNPs to facilitate analysis and interpretation. After filtering on MAF, genotyping rate and LD (independent of association with SCZ) we obtained a subset of 74,062 autosomal SNPs in approximate linkage equilibrium (Table S5, S6). In each discovery sample, we selected sets of score alleles at different association test p_T thresholds. For each individual in the target sample, we calculated the number of score alleles they possessed, each weighted by the log odds ratios from the discovery sample. To assess whether the aggregate scores reflect SCZ risk, we tested for a higher mean score in target cases compared to controls (Sections S9–S11, Table S7).

We selected males (2176 cases, 1642 controls) and females (1146 cases, 1945 controls) to form arbitrary discovery and target samples (Table S8). Score alleles designated in the discovery sample were significantly enriched among target cases and the effect was larger for increasingly liberal p_T thresholds. The score based on all SNPs with male discovery p_T < 0.5 (N=37,655 SNPs) was highly correlated with SCZ in target females (P=9.4×10⁻¹⁹), explaining ~3% of the variance (Nagelkerke’s pseudo R² from logistic regression), with higher scores in cases. The results were not driven by only a few highly associated regions (Section S12).

We eliminated several possible confounders, with emphasis on subtle population stratification (Table S9–S15). Defining score alleles in British Isles samples and testing in samples from Sweden, Portugal and Bulgaria, and vice versa, we observed a similar pattern of results. It is unlikely that the same substructure is overrepresented in the corresponding phenotype class when discovery and target samples are from distinct populations. The effect is also stronger for SNPs within annotated genes (Table S16).

We used independent GWAS samples a) to replicate the polygenic component, b) to examine whether this component is shared with BPD⁹ and c) to demonstrate specificity by considering non-psychiatric diseases. We used the entire ISC for the discovery sample, considering the five most informative p_T thresholds from the intra-ISC analyses. The independent target samples were the MGS European-American (MGS-EA), the MGS African-American (MGS-AA) and the UK sample described by O’Donovan et al³. The ISC-derived score was highly associated with disease in both European SCZ samples (Figure 2, Figure S6 and Table S17). The MGS-EA had a significantly higher mean p_T < 0.5 score in cases (P = 2×10⁻²⁸; R² = 3.2%), as did the smaller O’Donovan sample (P = 5×10⁻¹¹; R² = 2.3%). Aggregate differences in allele frequencies and patterns of LD between Europeans and African-Americans are expected to lead to an attenuated effect. Still, MGS-AA cases carried more of the European-derived score alleles than MGS-AA controls (P = 0.008; R² = 0.4%).

Open in a separate window

Figure 2

Replication of the ISC-derived polygenic component in independent SCZ and BPD samples

Variance explained in the target samples based on scores derived in the entire ISC for five significance thresholds (p_T < 0.1, 0.2, 0.3, 0.4 and 0.5, plotted left to right in each study). The y-axis indicates Nagelkerke’s pseudo-R²; the number above each set of bars is the P-value for the p_T < 0.5 target sample analysis. Numbers for cases/controls: MGS-EA 2687 / 2656; MGS-AA 1287 / 973; O’Donovan 479 / 2938; STEP-BD 955 / 1498; WTCCC 1829 / 2935; CAD 1926 / 2935; CD 1748 / 2935; HT 1952 / 2935; RA 1860 / 2935 ; T1D 1963 / 2935 ; T2D 1924 / 2935.

The ISC-derived score alleles were also associated with BPD in two independent samples. Both STEP-BD¹⁰ and WTCCC¹¹ had higher mean p_T < 0.5 scores in cases (P = 7×10⁻⁹, R² = 1.9%; and P = 1×10⁻¹², R² = 1.4% respectively) indicating a substantial, shared genetic component.

To test disease specificity, we selected all six non-psychiatric WTCCC samples (coronary artery disease, Crohn’s disease, hypertension, rheumatoid arthritis, type I diabetes and type II diabetes). Controls are shared among the WTCCC case samples, including BPD. In contrast to SCZ and BPD, there was no association (p>0.05) between the ISC-derived SCZ scores and these non-psychiatric diseases, for any p_T threshold.

We next investigated the genetic models consistent with our data. The total additive genetic variance (V_A) reflects the number of causal alleles as well as their frequency and effect size distributions. However, the variance explained by the markers that tag these causal alleles (V_M) will be attenuated, reflecting the average extent of LD between marker and causal allele. In our target samples, the variance explained by the observed score alleles, V_S, will be further attenuated by sampling variation and p_T threshold, such that V_S ≤ V_M ≤ V_A.

We used simulation to estimate possible values for V_M and V_A, by identifying models that produced profiles of V_S across p_T threshold that were similar to those observed in ISC data, as indexed by the target sample R². Under a variety of genetic models, we simulated discovery and target datasets of comparable sample size to the ISC. Based on the empirical allele frequency distribution, we simulated marker SNPs, varying the proportion that were in LD with causal variants, for which we varied allele frequency (uniform, U-shaped) and effect size distributions (fixed GRRs, exponential GRRs or fixed variance explained) as well as the extent of LD (Section S16).

From a broad range of models, a subset produced results consistent with the ISC data (Figures 3 and S7). Among these, all led to similar estimates of V_M (mean 34%, range 32 to 36%). In models in which the causal alleles were imperfectly tagged (r²<1) estimates of V_A can be considerably larger. Therefore, our estimate that common polygenic variation accounts for one third of the total variation in SCZ risk is a lower bound for the true value, which could be much higher. Figure 3b shows seven examples from the range of consistent models, detailed in Table S18.

Open in a separate window

Figure 3

Observed and simulated profiles of target sample variance explained

Panel A shows the observed variance explained (R², black line). Panel B shows a subset of models that produced results consistent with the observed data. All yielded similar estimates of the total variance explained by the SNPs that tag the causal variants, V_M, with a mean value of 34%. The seven models were: (% SNPs, Mean GRR/variance explained (V) per causal allele, LD, frequency model) M₁: 6.25%, GRR=1.05, r² =1, empirical; M₂: 25%, GRR=1.025, r² =1, empirical; M₃: 12%, GRR=1.05, r² <1, uniform; M₄: 32%, GRR=1.04, r² <1, U-shaped; M₅: 11%, V=0.00006, r² =1, empirical; M₆: 25%, GRR(Exponential)=1.025, r² <1, uniform; M₇: 100%, GRR(Exponential)=1.012, r² <1, uniform. Panel C shows four inconsistent models with fewer variants of larger effect.

The simulated models consistent with our observed results all implied a substantial number of common variants, whereas models that invoked only a few common variants of large effect or only rare variants were not able to account for our findings. For example, if V_M ≈ 34% arose from only 100 common causal alleles, with GRRs at the tagging marker between ~1.2–1.5, the majority would be detected at p_T<0.01, and so the variance explained would decline, not increase, as more SNPs were added (Figure 3c, Table S19). It is possible that an observed GRR of ~1.05 could represent a large effect of a weakly tagged rare variant, e.g. a 10-fold effect of a 1/10,000 variant in complete LD (D’ = 1, but low r²) with a genotyped SNP. However, as this would only hold for low frequency markers (MAF < ~0.1), we stratified our analysis by score allele frequency (Figure 4a). For simulated models in which all causal variants were of low frequency (<0.05), a stratified analysis revealed the expected, skewed distribution (Figure 4c, Section S17), which was more pronounced for rarer causal alleles, e.g. 1/1,000 (data not shown). In contrast, models in which causal alleles followed a uniform frequency distribution provided a closer fit to our data (Figure 4b; although note some enrichment in the 2^nd quintile, of ~13–30% score alleles). Moreover, rare variants are likely to be population specific and if recurrent, in LD with different common alleles within and between populations. As such, they could not account for the observation of disease variation that is largely shared across our different populations.

Open in a separate window

Figure 4

Analysis stratified by score allele frequency

Panel A shows the observed data for the ISC/MGS-EA comparison. The y-axis is the target sample pseudo R², scaled within each figure as a proportion of the maximum value observed. The two other plots show results for simulated data: the common variant model, with a uniform frequency distribution for causal risk-increasing alleles (panel B) and a multiple rare variant (MRV) model, in which the collective frequency of rare variants at a locus that all reside on the same haplotypic background with respect to the genotyped SNP was bounded at a maximum of 5% (panel C).

Decreased reproductive fitness in SCZ¹² suggests that risk alleles of large to moderate effect will be under negative selection and therefore very rare^13,14. This is not inconsistent with our results, since the common variants indexed by our polygenic score will not be subjected to strong selection, by virtue of their very small individual effect sizes. Our results do not exclude important contributions of rare variants for SCZ¹³, since rare variants are expected as part of the allele frequency/effect size spectrum of a polygenic model. We and others recently reported higher genome-wide rates of rare copy number variants in SCZ^15,16,17. However, our results imply that medical sequencing and studies of structural variation to identify rare, highly penetrant variants will not alone fully characterize the genetic risk factors.

In conclusion, our molecular genetic data strongly support a polygenic basis to SCZ that a) involves common SNPs, b) explains at least a third of the total variation in liability, c) is substantially shared with BPD, and d) is largely not shared with several non-psychiatric diseases. We also identified variants in the MHC region that received support in two independent studies, although the population specificity and extensive LD will make follow-up challenging.

A highly polygenic model suggests that genetically influenced individual differences across domains of brain development and function may form a diathesis for major psychiatric illness, perhaps as multiple growth and metabolic pathways influence human height¹⁸. Our results may also reflect heterogeneity, such that some patients have aetiologically distinct diseases. The shared genetic liability between SCZ and BPD, previously suggested by clinical and genetic epidemiology^9,19, opens up the possibility of genetically-based refinements in diagnosis. However, the scores derived here have little value for individual risk prediction, meaning that application to clinical genetic testing for SCZ would be unwarranted. In the future, measures of polygenic burden, along with known risk loci and non-genetic factors such as season of birth, life stress, obstetrical complications, viral infections and epigenetics, could open new avenues for studying gene-gene and gene-environment interplay.

Increasing the discovery sample size should substantially refine the polygenic scores derived here. The variance explained by the observed score increases from ~3% to over 20% in extended simulations of 20,000 case/control pair, as will soon be available via international meta-analytic efforts such as the Psychiatric GWAS Consortium^20–22 (Section S18, Figure S8). In addition, analyses that focus on gene pathways, clinical features and non-additivity may increase the variance captured by the score and identify genes or biological systems, that are either shared by, or unique to, SCZ and BPD.

We identified fewer unambiguously associated variants than studies of some non-psychiatric diseases of comparable size²³. Nonetheless, for other diseases replicated variants typically account for only a modest fraction of risk. The nature of this “missing heritability” is a general problem now faced by complex disease geneticists²⁴. For SCZ, our data point to a genetic architecture that includes many common variants of small effect. The extent to which similar models characterize genetic variation within and across other complex diseases remains to be investigated.