23andMe sample.

The GWAS summary statistics were obtained from a subset of 23andMe participants. 23andMe uses a survey designed to collect a number of phenotypes, including the personality traits reported here, and the sample has been described previously for other phenotypes31,32. We included only participants (N = 59,225) who showed >97% European ancestry as determined by analyzing local ancestry and comparing to three HapMap2 populations33. Relatedness between participants was examined by a segmental identity-by-descent (IBD) method34 to ensure that only unrelated individuals (sharing less than 700 cM IBD) were included in the sample. All participants included in the analyses provided informed consent and answered surveys online according to a human subject research protocol, which was reviewed and approved by Ethical and Independent Review Services, a private institutional review board accredited by the Association for the Accreditation of Human Research Protection Programs.

Additionally, we obtained independent replication results of GWAS from the 23andMe replication sample. This sample included ∼39,500 participants (N = 39,452 for conscientiousness, 39,484 for extraversion and 39,488 for neuroticism) who met the inclusion criteria described above.

Genetics of Personality Consortium (GPC) sample.

The Genetics of Personality Consortium (GPC) is a large collaboration of GWAS for personality. Summary statistics of the GPC data used in the current study included the first meta-analysis of GWAS (GPC-1)6 for three traits (agreeableness, conscientiousness and openness) and the second meta-analysis of GWAS (GPC-2) for neuroticism and extraversion7,35,36. The results of 10 discovery cohorts for GPC-1 and 29 discovery cohorts for GPC-2 are available in the public domain, and consist of 17,375 and 63,661 participants, respectively, with European ancestry across Europe, Australia and United States. These studies were performed with oversight from local ethic committees, and all participants provided informed consent6,7,35,36.

UK Biobank sample.

UK Biobank is a large prospective cohort of more than 502,000 participants (aged 40–69 years)3 with genetic data and a wide range of phenotypic data, including social, cognitive, personality (neuroticism trait), life style, and physical health measures collected at baseline recruitment from 2006 to 2010. We used a subsample of this cohort for neuroticism replication. Exclusion criteria included UK Biobank genomic analysis exclusions, relatedness, gender mismatch, non-white UK ancestry and failure of quality control of UK BiLEVE genotyping3, resulting in a sample of 91,370 individuals. Association analysis was conducted using linear regression under a model of additive allelic effects with sex, age, array and the first eight PCs as covariates3. Informed consent was obtained from all participants, and the study was approved by the UK National Health Service National Research Ethics Service3.

deCODE sample.

Icelandic participants (N = 7,137 for extraversion, 7,136 for neuroticism and 7,129 for conscientiousness) were enrolled in various ongoing deCODE studies administering the Neuroticism–Extraversion–Openness Five-Factor Inventory (NEO-FFI) measure of the Big Five personality traits37,38. All deCODE studies were approved by the appropriate bioethics and data-protection authorities, and all subjects donating blood provided informed consent. The personal identities of participants from whom phenotype information and biological samples were obtained were encrypted by a third-party system overseen by the Icelandic Data Protection Authority39. A generalized form of linear regression that accounts for relatedness between individuals was used to test the correlation between normalized NEO-FFI trait scores and genotypes.

Personality assessment.

In the 23andMe sample, individuals completed a web-based implementation of the Big Five Inventory (BFI)40,41 that includes 44 questions. Scores for agreeableness, conscientiousness, extraversion, neuroticism and openness were computed using 8 to 10 items per factor40.

In GPC-1, scores of personality traits were based on the 60-item NEO-FFI with 12 items per factor6,37. In GPC-2, harmonization of measures for neuroticism and extraversion across 9 inventories and 29 cohorts was performed by applying Item Response Theory (IRT) to avoid personality scores being influenced by the number of items and the specific inventory. Because the personality measures were not assessed similarly across GPC-2 cohorts, the harmonized or calibrated scores of personality are more comparable, thereby increasing power for meta-analysis of GWAS using fixed-effect models7,35,36. As described in the main text, we found high genetic correlations between 23andMe and GPC samples, suggesting a highly consistent pattern of associations despite the discrepancy in questionnaires (Supplementary Fig. 1).

In the UK Biobank sample, neuroticism was scored between 0 and 12 using the 12 items of the Eysenck Personality Questionnaire, Revised Short Form (EPQ-R-S)42 with high reliability and concurrent validity42.

In the deCODE sample, NEO-FFI personality trait scores37,38 were adjusted for sex and age at measurement and were then normalized to a standard normal distribution using quantile normalization.

Regional association and annotation plot.

The regional plot of chromosome 8p (Fig. 2) was constructed by a web-interface tool, LocusZoom43. The bottom panel displays gene symbol and location within the region derived from UCSC Genome Browser human hg19 assembly. The regional and annotation plots for other significant SNPs are also shown in Supplementary Figure 4.

Distributions and correlations for personality scores in the 23andMe sample.

Quantile–quantile (QQ) plots of covariate-adjusted personality scores to examine normality are shown in Supplementary Figure 5. The distributions at the top tail deviate from normality owing to the limited range of the scores, and those at the bottom tail deviate due to the limited range (for neuroticism and extraversion) and/or extreme values. This violation of the normality assumption can be influential for genetic variants with very low minor allele frequencies (e.g., rare variants)44. However, this did not affect our results because our GWAS and LD Score regression9 include only common variants.

Pearson correlations, unadjusted and after adjusting for the covariates (age, sex and top five principal components (PCs) for population structure correction45), were used to assess phenotypic correlations among the five traits (Supplementary Table 3).

Genotyping and imputation.

In the 23andMe sample, DNA extraction and genotyping were performed on saliva samples by National Genetics Institute (NGI), a CLIA-licensed clinical laboratory and a subsidiary of Laboratory Corporation of America. Samples were genotyped on one of four genotyping platforms. The V1 and V2 platforms were variants of the Illumina HumanHap550+ BeadChip, including about 25,000 custom SNPs selected by 23andMe, with a total of about 560,000 SNPs. The V3 platform was based on the Illumina OmniExpress+ BeadChip, with custom content to improve the overlap with 23andMe's V2 array, with a total of about 950,000 SNPs. The 23andMe's V4 platform in current use is a fully custom array, including a lower redundancy subset of V2 and V3 SNPs with additional coverage of lower-frequency coding variation, and about 570,000 SNPs. Samples that failed to reach a 98.5% call rate were reanalyzed. As part of 23andMe standard practice, individuals whose analyses failed repeatedly were contacted and asked to provide a new sample.

23andMe participant genotype data were imputed using the 1000 Genomes Project phase 1 version 3 reference panel46. The phasing and imputation for each genotyping platform were separated. First, chromosomal segments of no more than 10,000 genotyped SNPs, with overlaps of 200 SNPs, were phased using Beagle (version 3.3.1)47. Then, each phased segment was imputed against all-ethnicity 1000 Genomes Project haplotypes (excluding monomorphic and singleton sites) using a high-performance version of Minimac48 for 5 rounds and 200 states to estimate parameters. SNPs were filtered by procedures including Hardy–Weinberg equilibrium P < 10−20 (stringent threshold for large sample size), call rate < 95% and allele frequencies apparently different from European 1000 Genomes Project reference data. A total of 13,341,935 SNPs was retained after filtering and excluding chromosome X, Y and mitochondria. We focused on autosomal SNPs, which are available for 23andMe, GPC and UK Biobank samples.

Genotyping in cohorts of GPC-1 (ref. 6) and GPC-2 (refs. 7,35) was conducted on Illumina or Affymetrix platforms. Quality control of genotype data was examined in each cohort independently, including checks for European ancestry, sex inconsistencies, Mendelian errors, high genome-wide homozygosity, relatedness, minor allele frequencies (MAFs), SNP call rate, sample call rate and Hardy–Weinberg equilibrium6,7,35,36. Genotype data of GPC-1 were then imputed using HapMap phase II CEU (Utah residents with Northern and Western European ancestry from the CEPH collection) as a reference panel including ∼2.5 million SNPs6 and, alternatively, a reference panel from 1000 Genomes Project phase 1 version 3 was used to impute the genotype data of GPC-2 (refs. 7,35,36). Poorly imputed SNPs (r2 < 0.3 or imputation quality (proper_info) < 0.3 (ref. 6) or 0.4 (refs. 7,35) and low MAF (<0.01 (ref. 6) or (refs. 7,35)) were excluded in the meta-analyses, resulting in a total number of 1.1 million–6.6 million SNPs7,35 across cohorts of GPC.

In the UK Biobank first release genetic data of 152,729 participants (June 2015), about two-thirds of the sample was genotyped using Affymetrix UK Biobank Axiom array, and the remaining were genotyped using the Affymetrix UK BiLEVE Axiom array3. Outlier, multiallelic and low-MAF (<1%) SNPs were excluded from phasing and imputation procedures. The reference panel of imputation was based on the 1000 Genomes Phase 3 and UK10K haplotype panels3. Further quality control procedures were applied after imputation, yielding a total of 8,268,322 SNPs for further analyses3.

Genotyping, imputation methods and the association analysis method used in the deCODE sample were previously described49. A total of 676,913 autosomal SNPs were typed using Illumina SNP chips49. SNPs with low MAF (<0.1%) and low imputation information (<0.8) were excluded and 99.5% of SNPs remained after imputation.

Genome-wide association analysis.

Association tests were performed by regressing personality traits on imputed dosages of SNPs in the 23andMe sample. Age, sex and the top five PCs45 for population structure correction were included as covariates, and P values were computed using likelihood ratio tests. For all five personality traits, the correlation structure of SNPs was determined by an LD matrix of 9,270,523 autosomal SNPs generated from European reference sample in 1000 Genomes Project phase 1 version 3 within 1,000,000 bp (1 Mb)50,51 using PLINK 1.07 (ref. 52). The original 13,341,935 SNPs were reduced into 9,270,523 SNPs in our subsequent analyses (e.g., LD correlation structure is used to determine LD-independent SNPs). All SNP positions were mapped to Genome Reference Consortium Human Build 37 (GRCh37) and UCSC Genome Browser human hg19 assembly. We made QQ plots with GWAS summary statistics of the 23andMe sample. The QQ plots lie along the expected null line for large P values (P > 10−3), indicating that the GWAS results are not inflated by population stratification or cryptic relatedness. This pattern is consistent with the genomic inflation factors (λ)53 close to 1, as shown in Supplementary Figure 6.

In each cohort of GPC-1 (ref. 6) and GPC-2 (refs. 7,35), linear regressions with covariates of sex, age and PCs were conducted for association tests using dosage data. The meta-analyses of GWAS results of cohorts for GPC-1 and GPC-2 were performed by the inverse-variance method using METAL54. SNPs available in one cohort only were excluded. The totals of 2,305,461, 2,305,682 and 2,305,640 SNPs were available for traits of agreeableness, conscientiousness and openness (respectively) in GPC-1, as well as 6,941,603 SNPs for extraversion and 6,949,614 SNPs for neuroticism in GPC-2. Genomic inflation factors (λ) are 1.01, 1.01, 1.03, 1.02 and 1.02 for agreeableness, conscientiousness, extraversion, neuroticism and openness, respectively.

Meta-analysis of 23andMe and GPC samples.

Given improved power for detection of genetic effects with larger sample sizes in GWAS, we performed a combined meta-analysis of 23andMe and GPC samples using METAL54 on the basis of the sample-size based method. To assess the quality of meta-analysis, SNPs with heterogeneity P < 0.05 were excluded. Eight significant LD-independent SNPs were identified after removing correlated SNPs at LD r2 > 0.05 that are within 1 Mb of the top SNP. In Table 1, the percentage of variance explained by each SNP is calculated using equation: (z2/(n-k-1+z2)) × 100, where z is the z value for each SNP controlling for covariates, n is the sample size for each SNP and k is the number of covariates in the regression model (k = 7 for age, sex, and top five PCs)55,56.

Conditional analysis within 1-Mb region of significant SNPs.

We performed a conditional analysis57 within the 1-Mb genomic region of each of the six LD-independent SNPs. In our study, we used 1000 Genomes Project reference panel of European ancestry to estimate LD correlations (r2) and excluded SNPs correlated at LD r2 > 0.9 with the top associated SNP within a 1-Mb window. We did not detect additional significant SNPs conditional on the top SNPs under the stringent GWAS threshold. However, for the significant loci in 8p, several SNPs still showed substantial association signals (P ∼ 10−7) conditioning on the top SNPs, rs6981523 or rs2164273.

Genetic correlation analysis.

We used the LD Score regression method to examine the pattern of genetic correlations (r g )9,58 across personality traits within and between 23andMe and GPC samples (Fig. 3a, Supplementary Fig. 1 and Supplementary Table 4) on the basis of GWAS summary statistics. The LD Score for each SNP measures the amount of pairwise LD r2 with other SNPs within 1-cM windows from 1000 Genomes Project reference panel of European ancestry. All SNPs were filtered by LD Score regression built-in procedures, including imputation quality (INFO) > 0.9 and MAF > 0.1, and merged to SNPs in HapMap 3 reference panel. Approximately 0.8 million–1.1 million SNPs (Supplementary Table 2) were retained to estimate genetic correlations.

We also examined genetic correlations among the five traits, which have been estimated previously using a twin design59,60, and unrelated individuals' SNP data from a relatively smaller sample, in which many estimates did not converge19. Our LD Score regression analysis based on a large sample provided additional contribution to this effort.

We further quantified genetic correlations between personality traits and psychiatric disorders, including schizophrenia61, bipolar disorder62, major depressive disorder63, ADHD61, autism spectrum disorder61 and anorexia nervosa64.

Query for eQTL database.

We queried eQTL evidence for our significant SNPs from the Brain eQTL Almanac (Braineac)65,66. The results are listed in Supplementary Table 1. We display the brain region with the lowest P value for each SNP among all 10 brain regions. To check the rank of eQTL P values of six LD-independent SNPs in the Braineac database, we randomly selected 50,000 SNPs and queried the database to extract the lowest P value for each SNP, resulting in a total of 36,190 SNPs with eQTL results. To match allele frequencies and distances to transcription start site (TSS) with the significant SNPs, the randomly selected SNPs were stratified into four groups: (i) within transcript, (ii) downstream 0–200 kb, (iii) upstream 0–200 kb and (iv) upstream 200–400 kb. SNPs that fell outside these ranges were removed. The SNPs in the 'within transcript' group were further stratified into three subgroups according to allele frequencies. This procedure resulted in six distributions of eQTL P values that matched the significant SNPs in terms of allele frequencies and TSS, and these were used to determine the ranking of eQTL associations (Supplementary Tables 1 and 5). Two SNPs were ranked highly for their significance as eQTL compared to randomly sampled eQTL markers with matched allele frequencies and distances to TTS from the Braineac database (top 10–20% ranking, rs6981523; top 20–30% ranking, rs9611519; Supplementary Table 5).

Colocalization analysis between GWAS and eQTL.

To investigate whether GWAS-significant SNPs and their eQTLs were colocalized with a shared candidate causal variant, we performed a colocalization analysis, COLOC, that uses Bayesian posterior probability to assess colocalization18. The SNP-associated locus was defined as within a 1-Mb window18 for each of the six SNPs (Table 1). The prior probabilities that the locus is associated with only trait 1 (i.e., personality traits), only trait 2 (i.e., eQTL) and both are 10−5, 10−4 and 10−6, respectively. The posterior probabilities (PP0, PP1, PP2, PP3 and PP4) for five hypotheses (H 0 , no association with either trait; H 1 , association with trait 1, not with trait 2; H 2 , association with trait 2, not with trait 1; H 3 , independent association with two traits, two independent SNPs; H 4 , association with both traits, one shared SNP)18 were calculated to determine which hypothesis is supported by the data. A limitation of this analysis is the potentially low power in the small eQTL sample (N = 134).

SNP-concordant test for the top GWAS signals.

To investigate concordance of SNP effects between personality traits and psychiatric disorders, we followed a procedure similar to one described previously67,68 by counting the number of same-direction effect sizes for the LD-independent top SNPs (P < 10−4) in the pairwise phenotype data and calculated the proportion of the same-direction effects in the total number of LD-independent top SNPs. The one-sided P value for the proportion of pairwise phenotypes was computed using a binomial test to examine the deviation from 0.5 for the proportion. In Supplementary Figure 2, a heat map of the proportions of the same-direction effect for pairwise phenotypes shows a similar pattern with a heat map of genetic correlations in Figure 3a.

Hierarchical clustering analysis.

We performed hierarchical clustering analysis using dissimilarity measures (1-genetic correlation) implemented in hclust function of R to investigate and display relationships between personality traits and psychiatric disorders. On the basis of genetic correlations, the more highly correlated phenotypes were grouped in the same clusters and displayed by a dendrogram (Supplementary Fig. 3), showing an agreement with classifications of the loading plot (Fig. 3b).

Data availability.

GPC-1 and GPC-2 summary statistics are available at http://www.tweelingenregister.org/GPC/; Psychiatric Genomics Consortium (PGC) summary statistics (schizophrenia, bipolar disorder, major depressive disorder, ADHD, autism spectrum disorder and anorexia nervosa) are available at https://www.med.unc.edu/pgc/results-and-downloads. The top 10,000 SNPs for five personality traits from the 23andMe discovery data set are available in Supplementary Data Sets 1,2,3,4,5. The full GWAS summary statistics for the 23andMe discovery data set will be made available through 23andMe to qualified researchers under an agreement with 23andMe that protects the privacy of the 23andMe participants. Please contact D.A.H. (dhinds@23andme.com) for more information and to apply for data access. All other data reported in the paper are included in the paper and Supplementary Materials.

URLs.

LDlink, http://analysistools.nci.nih.gov/LDlink/?tab=ldpair; US National Human Genome Research Institute GWAS catalog, https://www.ebi.ac.uk/gwas/; LocusZoom, http://locuszoom.sph.umich.edu/locuszoom/; Braineac (UK Brain Expression Consortium), http://www.braineac.org/; LD Score regression, https://github.com/bulik/ldsc; GCTA-COJO (conditional and joint genome-wide association analysis), http://cnsgenomics.com/software/gcta/cojo.html; METAL, http://csg.sph.umich.edu//abecasis/metal/; PLINK 1.07, http://pngu.mgh.harvard.edu/~purcell/plink/; Ethical and Independent Review Services, http://www.eandireview.com.