Interethnic analyses of blood pressure loci in populations of East Asian and European descent

ARTICLE Interethnic analyses of blood pressure loci in populations of East Asian and European descent Fumihiko Takeuchi1,2, Masato Akiyama3, Nana Matoba3, Tomohiro Katsuya4,5, Masahiro Nakatochi6, Yasuharu Tabara et al. # Blood pressure (BP) is a major risk factor for cardiovascular disease and more than 200 genetic loci associated with BP are known. Here, we perform a multi-stage genome-wide association study for BP (max N=289,038) principally in East Asians and meta-analysis in East Asians and Europeans. We report 19 new genetic loci and ancestry-specific BP variants, conforming to a common ancestry-specific variant association model. At 10 unique loci, distinct non-rare ancestry-specific variants colocalize within the same linkage disequilibrium block despite the significantly discordant effects for the proxy shared variants between the ethnic groups. The genome-wide transethnic correlation of causal-variant effect-sizes is 0.898 and 0.851 for systolic and diastolic BP, respectively. Some of the ancestry-specific association signals are also influenced by a selective sweep. Our results provide new evidence for the role of common ancestry-specific variants and natural selection in ethnic differences in complex traits such as BP. Fumihiko Takeuchi, Masato Akiyama, Nana Matoba, Tomohiro Katsuya, Masahiro Nakatochi, Yasuharu Tabara et al. # DOI: 10.1038/s41467-018-07345-0 OPEN Correspondence and requests for materials should be addressed to N.K. (email: [email protected]). # A full list of authors and their affiliations appears at the end of the paper. NATURE COMMUNICATIONS | (2018) 9:5052 | DOI: 10.1038/s41467-018-07345-0 | www.nature.com/naturecommunications 1 1234567890():,; High blood pressure is a major risk factor for cardiovascular disorders such as coronary heart disease and stroke. Approximately 10 million deaths each year can be attributed to high blood pressure globally1,2. An individual’s risk for high blood pressure is determined by genetic, environmental and demographic factors and their interaction. Genome-wide association studies (GWASs) and/or large-scale analyses by genecentric (or exome) variation arrays have identified over 200 genetic loci influencing blood pressure in predominantly European-descent populations (henceforth referred to as Europeans)3–8. The prevalence of high blood pressure is increased in people of East Asian ancestry, contributing to their increased risk of stroke9. The reasons for such ethnic differences remain to be clarified from the viewpoint of genetic susceptibility as well as lifestyle. Although the recent progression of GWAS in East Asians allows us to make a preliminary comparison of association signals between the populations10,11, the sample sizes of GWAS in East Asians have been generally much smaller than those in Europeans and under-powered for the comprehensive interethnic comparison at a genome-wide scale. Therefore, large-scale genome-wide association data in both ethnic groups are required for systematic, genome-wide interethnic comparison. Here, we perform a multi-stage GWAS with a discovery sample of 130,777 East Asian individuals and follow-up meta-analyses involving East Asians and Europeans (max N=289,038), to seek both transethnic and ancestry-specific genetic effects for five blood pressure phenotypes: systolic blood pressure (SBP), diastolic blood pressure (DBP), pulse pressure (PP), mean arterial pressure (MAP), and hypertension. We then seek interethnic genetic heterogeneity of GWAS results between East Asians and Europeans, followed by examination of natural selection as a potential mechanism underlying the ethnic differences in genetic susceptibility for blood pressure as well as other complex traits. We report ancestry-specific blood pressure variants and selection signals in this study. Results Genome-wide association analysis and lookup for replication. Adopting a joint analysis strategy12, we performed a GWAS, which consisted of stage 1 (discovery) and stage 2 (follow-up), and a replication study (Supplementary Fig. 1). In stage 1 of GWAS, we used genome-wide association data from 130,777 individuals of Japanese ancestry. Characteristics of participants, genotyping arrays, and imputation are summarized in Supplementary Tables 1, 2. Genomic control and intercepts from linkage disequilibrium (LD) score regression13 were calculated at each study level (λ GC =0.89–1.24 and LD Score regression intercept = 0.94–1.06), indicating no residual confounding biases such as population stratification (Supplementary Table 2). Since the LD Score regression intercept can account for polygenic effects and inflation due to large sample size13, we applied the LD Score regression intercept as a correction factor for cohorts with a sample size of >3000 individuals (BBJ in this study). Genomic control λ GC was used as a correction factor in the other studies. Quantile−quantile plots for each of the five blood pressure traits are presented in Supplementary Fig. 2. Phenotype-specific metaanalysis was carried out in the two-stage approach for both the East Asian-specific and transethnic meta-analyses (Supplementary Figs. 1, 3). Genome-wide association results in the stage-1 identified 13,003 SNPs with a Pvalue < 1.6×10–5against any blood pressure phenotype in East Asians. This set of 13,003 SNPs (sentinel SNPs listed in Supplementary Data 1) was followed up in 53,008 East Asian individuals (stage 2). Additionally, these 13,003 SNPs were examined in the transethnic stage with phenotype-specific results for Europeans (max N=105,253) from the International Consortium on Blood Pressure (ICBP) GWAS (N=69,909)3and the International Genomics of Blood Pressure (iGEN-BP) Consortium (N=35,344)10; there was no overlap in samples between the two data sets. Sentinel SNPs (smallest P value against any blood pressure phenotype) that (i) reached P< 5×10–8in combined meta-analysis of stages 1 and 2 and (ii) showed evidence of support (P< 0.05) in the stage 2 metaanalysis alone are reported as novel loci in this study. We identified 19 previously unreported loci; 15 loci in East Asian-specific analyses and 4 additional loci in the transethnic meta-analysis (Table 1and Supplementary Data 2). By lookup in an independent replication sample of Europeans from the UK Biobank (N= 422,771)14 plus East Asians from the China Kadoorie Biobank (N =94,201)15, we examined associations at our list of 19 sentinel SNPs. With the exception of four SNPs, 15 sentinel SNPs showed significant (P< 0.00263 =0.05/19) blood pressure association with the concordant direction of allelic effects (Supplementary Data 2), thus validating the loci. Regional association plots are shown for the 19 newly identified loci in Supplementary Fig. 4. Associations of the 19 sentinel SNPs with other blood pressure phenotypes are demonstrated in Supplementary Data 3. In the discovery stage, we also replicated blood pressure associations at previously reported loci, which included 36 loci at genome-wide significance and further 179 loci at nominal significance (P< 0.05) (Supplementary Data 4). Functional annotations for new loci. To identify candidate genes at the newly identified blood pressure loci, we examined whether any of the association signals (sentinel blood pressure SNP and SNPs in East Asian LD r2> 0.80) were coding or associated with gene expression and other traits. At three loci, the sentinel SNPs were nonsynonymous, and 4 of 19 novel loci contained SNPs (in LD of r2> 0.80 with the top eVariant) associated with expression quantitative trait loci (eQTLs) in at least one tissue in the Genotype-Tissue Expression (GTEx) database (Supplementary Tables 3–5). At two candidate gene loci, proxy SNPs (rs760077 at MTX1 and rs3825942 at LOXL1) were nonsynonymous and associated with eQTLs. Furthermore, seven sentinel SNPs and/or their proxy SNPs (r2≥0.95) were previously reported to be significantly associated with non-blood pressure traits (Supplementary Data 5), including a sentinel SNP (rs11642015) at the FTO locus on 16q22, whose proxies (r2=0.97–0.99) have been reported to associate with body mass index and type 2 diabetes16. In our study, rs11642015 was significantly associated with SBP, MAP, and PP (P=1.9×10–12–1.3×10–9) with consistent reproducibility in both stages of East Asian analyses (Supplementary Data 1, 3). In addition, rs11642015 was recently identified to be significantly associated with SBP in multi-ancestry GWAS metaanalysis incorporating gene−smoking interaction17. Interethnic heterogeneity of GWAS results. In the present study, the availability of genome-wide association data from >100,000 individuals for both East Asians and Europeans separately motivated us to perform additional analyses of systematic, genome-wide interethnic comparison. We used transethnic association summary statistics available for both East Asian (N= 158,645 from stage 1 and iGEN-BP) and European (max N= 105,253 from ICBP and iGEN-BP) GWAS results in the subsequent analysis of interethnic heterogeneity. We defined interethnic heterogeneity as heterogeneity of genetic (or allelic) impact on SBP between the ethnic groups. Using GWAS data sets, we compared the genetic impact at transethnic SNPs and detected a total of eight interethnic heterogeneity loci—two significant (P< 5×10–8) and six suggestive (5×10–8≤P< 1×10–6) loci (Fig. 1a and Supplementary Data 6). In this study we distinguished the allelic ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-07345-0 2NATURE COMMUNICATIONS | (2018) 9:5052 | DOI: 10.1038/s41467-018-07345-0 | www.nature.c om/naturecommunications impact from allelic effect-sizes as previously defined by Brown et al.18; allelic impact is the genotype−phenotype correlation coefficient, which is approximately a product of allelic effect and minor allele frequency (MAF). Seven of the eight loci with interethnic heterogeneity were annotated to the previously reported blood pressure loci; sentinel blood pressure SNPs at half of them (i.e., four loci near the CACNB2,C10orf107,SH2B3 and DPEP1 genes3,5) were found to be in LD (r2≥0.2) with the SNPs showing some evidence for interethnic heterogeneity. The two loci with significant interethnic heterogeneity were on 12q24 and 10q21 and both contained multiple association signals (Fig. 1b). For the region on 12q24 spanning 1.5 Mb, two independent association signals, each specific to Europeans (near rs3184504 at SH2B3) and East Asians (near rs671 at ALDH2), had been identified19. We found that both of the signals were responsible for the discordant direction of allelic effects on 12q24 (Fig. 1c and Supplementary Fig. 5a, b). Similarly, we observed two independent association signals near the C10orf107 transcript on 10q21.2 (Fig. 1c and Supplementary Fig. 5c, d). The derived alleles of ancestry-specific sentinel SNPs on 10q21 (rs4590817 and rs145193831 specific to Europeans3and East Asians respectively) arose from a haplotype shared between ethnic groups, containing multiple transethnic SNPs. The discordant direction of effects for the shared haplotypes could be explained by alternation of effects attributable to the derived alleles of rs4590817 (decreasing in Europeans) and rs145193831 (increasing in East Asians) (Supplementary Fig. 6 and Supplementary Data 7). Ancestry-specific SNP loci. A total of 750 previously reported SNPs (listed in Supplementary Data 4) plus 19 newly identified SNPs could be classified into 485 loci by regarding two SNPs at most 500 kb apart to belong to the same locus. After exclusion of 39 loci (MAF < 0.01 in both East Asians and Europeans, or no data available in GWAS data sets for both populations), 446 loci were retained and categorized into two groups—group 1 and group 2. Group 1 consisted of 382 loci with MAF ≥0.01 in both populations and group 2 consisted of 64 loci with potential ethnic specificity, i.e., MAF < 0.01 in either East Asians or Europeans. Group 2 was further classified into group 2a (46 loci with MAF < 0.01 in one population and MAF ≥0.05 in the other) and group 2b (18 loci with MAF < 0.01 in one population and 0.01 ≤MAF < 0.05 in the other) (Supplementary Fig. 7). With regards to interethnic heterogeneity of association signals, we assumed two distinct scenarios: whether the underlying causal variants are shared between the ethnic groups or not. However, due to substantial interethnic differences in LD structure, it is not always feasible to distinguish between the two. First, as an example of the potential nonshared causal variant (or ancestry specificity), we examined interethnic comparability of genetic impact on blood pressure at 48 loci (46 loci in group 2a plus 2 target loci with potential ancestry specificity—C10orf107 and CACNB2—included in group 1; Supplementary Fig. 7), where sentinel common (MAF ≥0.05) blood pressure SNPs originally reported in a given ethnic group were monomorphic or MAF < 0.01 in the second ethnic group3–8,19. Then, we investigated interethnic heterogeneity at non-rare (MAF ≥0.01 in both ethnic groups) blood pressure loci (group 1 in Supplementary Fig. 7) that might be shared between the ethnic groups as described later. Considering the observations on 12q24 and 10q21, we explored common proxy SNPs forming a haplotype shared between ethnic groups at the locus (denoted as haplo-SNPs), for which the most significant interethnic heterogeneity of genetic impact was detected (Supplementary Fig. 8a–c). At a total of 11 loci (or 10 unique loci when the ALDH2 and SH2B3 loci on 12q24 were combined) (Supplementary Figs. 6, 9 and Supplementary Data 7), haplo-SNPs showed significant (P< 1.5×10–4under region-wise correction) heterogeneity between two ethnic groups. At 8 of 11 loci, we found that distinct common ancestry-specific variants colocalized within the same LD block and that the direction of effects for the proxy shared SNPs was discordant between the ethnic groups, similar to 12q24 and 10q21. On 5q14, for instance, a genome-wide significant association of rs112862634 with SBP, DBP, and MAP was detected in East Asians of this study (Supplementary Data 1), while SBP association of rs10059921 was previously reported in its vicinity (456 kb apart from Table 1 Genetic loci newly identified to be associated with blood pressure Sentinel SNP Chr Position EA/NEA EAF Trait NEffect P Genome-wide significant and replicated rs2990220 1 155,190,254 A/T 0.83 MAP 183,654a−0.41 (0.06) 2.2×10−12 rs6772151 3 46,896,499 A/C 0.29 DBP 156,503a0.28 (0.05) 7.8×10−9 rs17622152 3 183,520,112 A/G 0.47 MAP 183,759a−0.25 (0.04) 2.0×10−8 rs12209106 6 1,621,042 T/G 0.68 DBP 160,436a0.28 (0.05) 6.4×10−9 rs78399431 7 1,141,470 A/G 0.24 MAP 179,411a0.30 (0.05) 9.6×10−9 rs2125067 10 48,434,420 C/G 0.12 SBP 179,003a0.60 (0.10) 4.8×10−9 rs2305013 11 120,340,060 A/T 0.85 SBP 180,894a−0.59 (0.09) 5.6×10−10 rs5006548 12 32,692,233 T/G 0.16 HT 71,847a0.09 (0.02) 2.2×10−8 rs1535464 14 100,793,431 A/G 0.10 SBP 183,690a−0.61 (0.10) 3.5×10−9 rs66978877 19 18,455,657 T/C 0.55 HT 68,850a0.07 (0.01) 4.5×10−9 rs6021247 20 50,108,980 A/G 0.58 SBP 183,785a0.37 (0.06) 5.0×10−9 rs3853476 5 141,817,754 A/G 0.58 MAP 244,831b−0.20 (0.03) 6.0×10−9 rs10821808 10 62,390,646 A/G 0.58 SBP 288,917b−0.29 (0.05) 3.4×10−9 rs4418728 10 94,839,724 T/G 0.62 DBP 256,118b−0.20 (0.03) 1.5×10−8 rs1078967 15 74,222,987 T/C 0.15 SBP 265,280b0.42 (0.07) 5.6×10−9 Genome-wide significant but not replicated rs2076460 1 27,972,058 C/G 0.30 SBP 174,846a−0.42 (0.07) 3.6×10−9 rs11642015 16 53,802,494 T/C 0.21 SBP 174,917a0.58 (0.08) 1.9×10−12 rs9303509 17 64,530,887 A/C 0.40 SBP 183,769a0.37 (0.06) 3.9×10−9 rs66658258 20 61,462,502 C/G 0.58 DBP 164,638a0.28 (0.05) 1.0×10−8 Position is Build 37; EA: effect allele; NEA: non-effect allele; EAF: effect allele frequency; N: sample size (aEast Asians only; bwith European follow-up samples); Effect: as unit change in blood pressure (SE) per effect allele copy (SBP, DBP, PP, MAP) or as log odds ratio per effect allele (HT) NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-07345-0 ARTICLE NATURE COMMUNICATIONS | (2018) 9:5052 | DOI: 10.1038/s41467-018-07345-0 | www.nature.c om/naturecommunications 3 a –log10(P) 0.00 0.01 0.02 0.03 Genetic impact in EUR 0.00 0.01 0.02 0.03 Genetic impact in EUR 0.1 0.2 0.3 0.4 0.5 0.6 r2 to r2 to rs4590817 10q21 EUR rs4590817 0.1 0.2 0.3 rs3184504 12q24 EUR rs3184504 b Genetic impact in EAS r2 to 0.2 0.4 0.6 rs145193831 rs145193831 Position on Chr 10 (Mb) 63.4 63.5 63.6 EAS –0.02 0.00 0.02 Genetic impact in EAS Position on Chr 12 (Mb) r2 to rs671 rs671 112.0 112.5 113.0 EAS –0.02 0.00 0.02 0.2 0.3 0.1 12q24 c H3 H4 H1 H2 H9 H6 H7 H8 H3 H4 H1 H2 H9 H6 H8H5 Ancestral haplotype Ancestral haplotype rs3184504_T EAS EUR rs671_A (ALDH2) H3 H2 H6 H7 H5 H4 H3 H2 H6 H8 H7 H4 Ancestral haplotype Ancestral haplotype rs4590817_C rs145193831_T EAS EUR 10q21 (SH2B3) 20 15 10 5 0 123456789 Chromosome 11 13 15 18 22 Fig. 1 Interethnic heterogeneity of genetic impact of SBP. aManhattan plot showing results for genome-wide scan of genetic impact heterogeneity. The genetic impact at transethnic SNPs were compared between two populations of different ancestries using GWAS data sets. bRegional plots on 12q24 and 10q21, where there were multiple SNPs with significant (P< 5×10–8) evidence for interethnic heterogeneity (see Supplementary Data 6). Bordered circles represent SNPs with significant interethnic heterogeneity. Transethnic SNPs were plotted in two panels at each locus; genetic impacts of each SNP are denoted separately for Europeans (EUR, top panel) and East Asians (EAS, bottom panel) on 12q24 (left) and 10q21 (right) such that genetic impacts in Europeans are positive. In the individual regional plots, the correlation of ancestry-specific sentinel SNP to other SNPs at the locus is shown on a scale from minimal (blue) to maximal (red); the sentinel SNPs thus benchmarked are rs3184504 (EUR specific) and rs671 (EAS specific) on 12q24 and rs4590817 (EUR specific) and rs145193831 (EAS specific) on 10q21. The position of ancestry-specific sentinel SNP is indicated by an arrow head. cPhylogenetic relationships of ancestry-specific sentinel SNPs with transethnic haplotypes detectable in Europeans (top) and East Asians (bottom) on 12q24 (left) and 10q21 (right). Each node corresponds to a haplotype and the SNPs appear on the edges. The edge width reflects the haplotype frequency in the corresponding ethnic groups. At each locus, blood pressure increasing and decreasing haplotypes and derived, ancestry-specific alleles are colored in red and blue, respectively ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-07345-0 4NATURE COMMUNICATIONS | (2018) 9:5052 | DOI: 10.1038/s41467-018-07345-0 | www.nature.c om/naturecommunications rs112862634) in Europeans8. It turned out that rs112862634 was in strong LD (East Asian LD r2=0.95) with a haplo-SNP (rs6882046) at this locus and distinct common ancestry-specific variants with mutually inverted genetic effects—Europeanspecific rs10059921 (MAF =0.09 in EUR) and East Asianspecific rs78245349 (MAF =0.46 in EAS)—did colocalize in this region (Supplementary Fig. 8a and Supplementary Data 7). At the remaining 3 (of 11) loci, alternate rare ancestry-specific SNPs were likely to exist in the second ethnic group, although they were not detectable in our search of public databases. We designated these as a common ancestry-specific variant association model as discussed below. We hypothesized that there were three major combinations of East Asian-/European-specific SNPs and their resultant direction of effects for haplo-SNPs forming a shared haplotype at the locus, as schematically shown in Supplementary Fig. 6b. In accordance with this notion, we detected three types in this study (Supplementary Fig. 9), among which the first and major type (32 of 48 loci in group 2a) consisted of the cases with mutually inverted genetic effects as explained above. The second type consisted of those with distinct ancestry-specific variants showing concordant directions of effect such as the FGR locus (Supplementary Figs. 8b, 9). The third type consisted of those with distinct ancestry-specific variants showing discordant genetic effects, one of which appeared to be almost neutral such as the GNAS/EDN3 locus (Supplementary Fig. 8c). However, without using larger sample sizes, it appeared to be difficult to show statistically significant interethnic heterogeneity in particular, for those in the second or third type. At one locus (near HSD17B1 on 17q21), a haplo-SNP could not be selected (Supplementary Fig. 8d and Supplementary Data 7), presumably because of the ancestry-specific LD structure and the modest strength of association in the index ethnic group (Europeans at the locus) of this study. For loci with potential ancestry specificity (i.e., MAF < 0.01 in one population and 0.01 ≤MAF < 0.05 in the other; 18 loci classified as group 2b in Supplementary Fig. 7), we did not investigate interethnic heterogeneity of association signals because of difficulties in the relevant test for rare (MAF < 0.01) and low-frequency (0.01 ≤MAF < 0.05) genetic variants by using imputed GWAS results20. Heterogeneity at variants polymorphic in both ancestries.In addition to the ancestry-specific loci, we investigated interethnic heterogeneity at non-rare (MAF ≥0.01 in both ethnic groups) blood pressure loci that might be shared between the ethnic groups; 382 tested loci were either previously reported or newly identified in the present study (denoted as group 1 in Supplementary Fig. 7a). Since ICBP and iGEN-BP (European) data were imputed with HapMap SNPs, approximately one-third of group-1 SNPs were unavailable in our European GWAS data sets. Thus, 242 (out of 382) loci in group 1 were subjected to interethnic comparison of genetic impact on a lead blood pressure trait (Supplementary Data 8). Although majority of them appeared to show concordant effects (correlation coefficient r=0.754), nine sentinel SNPs (3.7%) showed significant (P hetero < 2.1×10–4) interethnic heterogeneity (Supplementary Fig. 10). Genetic impacts were more prominent in Europeans than in East Asians at eight of nine loci apart from rs1451538 in SLC28A1, at which genetic impacts were prominent in East Asians but not in Europeans (Supplementary Data 8). There were no proxy SNPs near each of the eight loci in the same LD block (Supplementary Fig. 11), which could have shown stronger association signals in East Asians than the sentinel SNPs originally reported in Europeans due to potential interethnic differences in LD structure, if any. Of note is the finding on 10q23 near PLCE1, there was another SBP association signal at rs7080472 in East Asians (P= 3.9×10–8in the combined samples; Supplementary Data 1) despite the absence of prominent association at rs932764, whose association was previously reported4and prominent in Europeans (Supplementary Fig. 11). rs7080472 was located in the LD block next to the one for rs932764 (East Asian LD r2=0.003 between rs7080472 and rs932764). On 10q21 near C10orf107, a DBP association signal was previously reported at rs153044021, which we found to be in LD (European LD r2=0.48) with an ancestryspecific SNP at the locus, rs4590817, aforementioned (Supplementary Data 7). Also, on 10p12 near CACNB2, a DBP association signal was previously reported at rs18133533, which we found to be in LD (European LD r2=0.56) with an ancestryspecific SNP at the locus, rs12258967. These indicated that interethnic heterogeneities identified for non-rare transethnic variants on 10q21 and 10p12 were the cases for which common ancestry-specific variants were actually responsible. By calibrating the proportion in the group-1 subset, in which blood pressure GWAS results for interethnic comparison were available for 242 (of 382) loci, we estimated the proportion of loci showing significant interethnic heterogeneity within the total blood pressure loci tested (N=446). The estimated proportion was 2.5% each in group 1 and group 2a, where the C10orf107 and CACNB2 loci were counted in group 2a (Supplementary Fig. 7b). Genetic correlation and power of GWAS. As an approach to quantitatively evaluating the interethnic differences in blood pressure GWAS results, we estimated the genetic correlation using summary statistics of the entire spectrum of GWAS associations18.Wefirst estimated the SNP-based heritability (h2)of SBP and DBP (Fig. 2). For SBP, h2estimates in our study were 0.107 (SE 0.007) for East Asians and 0.086 (SE 0.009) for Europeans and lower than a previously reported UK Biobank estimate of 0.156 (SE 0.004)22 calculated by the moment-matching method in Europeans. This discrepancy was likely due to the methodological differences in SNP-based heritability analyses between the studies but does not appear to affect geneticcorrelation estimates themselves23. Also, the h2of DBP was almost comparable between the ethnic groups in this study. Then, we found that the genetic correlations in SBP and DBP were 0.898 (SE 0.040) and 0.851 (SE 0.046) respectively, and significantly different from 1 (P=0.005 for SBP and P=0.0007 for DBP). This indicated that the allele-substitution effect-sizes differed significantly between the two ethnic groups despite the reportedly substantial genetic overlap in blood pressure traits (Supplementary Data 4). To estimate the degree of interethnic overlap and nonoverlap of blood pressure loci, we further calculated the power of GWAS of different sample sizes (i.e., 100K, 200K, and 500K) based on heritability parameters (see details in Supplementary Methods) via modeling, computing and random sampling (Fig. 3and Supplementary Figs. 12, 13). Similar to Europeans, the recent progresses of GWAS in East Asians prompted us to investigate different sample sizes in preparation for much-larger transethnic meta-analysis. When GWASs of the same size were carried out for SBP and DBP, it was expected that an almost equivalent number of genome-wide significant loci could be identified in both East Asians and Europeans but the number of overlap was less than half. We extended the interethnic analyses to other complex traits such as plasma lipid level, anthropometric measurement, and type 2 diabetes using published GWAS summary statistics of relatively large number of samples (Supplementary Table 6). Although genetic correlation appeared to be varied among the NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-07345-0 ARTICLE NATURE COMMUNICATIONS | (2018) 9:5052 | DOI: 10.1038/s41467-018-07345-0 | www.nature.c om/naturecommunications 5 complex traits examined (Fig. 2), we found that the proportion of nonoverlap [(nonoverlap) / (overlap +nonoverlap)] was relatively consistent across the traits for the same sample size; 0.71–0.82 for 100K, 0.65–0.78 for 200K and 0.46–0.70 for 500K (Fig. 3and Supplementary Fig. 12). As the sample sizes in both ethnic groups become larger, we can expect a higher proportion of interethnic overlap; nevertheless, more than or nearly half of the genome-wide significant loci may not overlap between the ethnic groups for GWAS of the same sample size. Selective sweeps at ancestry-specific loci. Subsequently, we created a list of ancestry-specific loci for SBP, DBP and other complex traits in which an SNP-trait association was genomewide significant in one ethnic group (e.g., East Asians) but no significant association signal was detectable in another (e.g., Europeans) due to low allele frequency (MAF < 0.05) (Supplementary Data 9). For the loci with the same SNPs being monomorphic in the second ethnic group, our selection criteria for ancestry-specific loci could be regarded stringent in that the absence of locus-wide significant association signals in the vicinity (≤500 kb) of the tested SNPs was required. A larger number of significant loci had been reported in Europeans compared to East Asians, reflecting the differences in sample size of GWAS conducted to date (mean for five traits was 81,991 in East Asians vs. 202,390 in Europeans) (Fig. 4and Supplementary Fig. 14). Thus, the total number of ancestryspecific loci across the examined traits was smaller in East Asians (10 loci) than in Europeans (63 loci). While it was most prominent for height, the sentinel SNPs at the ancestry-specific loci tended to have both lower MAF (0.20 ± 0.04 in East Asians, 0.16 ± 0.01 in Europeans) and genetic impact (0.020 ± 0.002 in East Asians, 0.014 ± 0.0004 in Europeans) across the traits. Among a list of ancestry-specific loci for multiple traits, we identified evidence of a positive selection at five unique loci using a highly sensitive algorithm, haploPS24 (Fig. 5and Supplementary Data 9). For blood pressure, a sentinel SNP rs56174355 on 17q23 previously reported to be associated with DBP only in Europeans8 was localized to a region with evidence of positive selection in East Asians (Fig. 5d). In this region, we observed the long haplotypes at high frequencies (i.e., 70–80%) to be selected exclusively in East Asians, on which the present-day major allele (G of rs56174355) could reside, whereas the minor allele (T of rs56174355) was associated with lower DBP in Europeans. Thus, a selective sweep in the region is considered to have retained the major allele that was likely beneficial in the populations of East Asian ancestry; conversely, this has reduced MAF in East Asians (T allele: 0.03 in East Asians vs. 0.10 in Europeans). We found similar examples for the traits other than blood pressure in four regions: rs12748152 for LDL-C and triglycerides, rs17031005 for T2D, rs11862222 for height and rs4253772 for total cholesterol (Fig. 5a−c, e). There was a significant (P Binomial =4.2 × 10–5) increase in the incidence of recent selection signals at the ancestry-specific loci, given that a total of 405 distinct genomic regions were identified to show evidence of positive selection across 14 populations worldwide24. Discussion Our GWAS in 183,785 East Asian individuals identified 15 new genetic loci influencing blood pressure phenotypes and 4 additional loci when combined with European individuals (max N= 289,038) (Table 1). Of the 19 newly identified loci, 15 loci were replicated in an independent sample of Europeans (N=422,771) plus East Asians (N=94,201) (Supplementary Data 2). A notable feature of this study is the use of a relatively large discovery-stage sample size in populations of non-European descent, thereby enabling us to identify a number of genetic loci that have not been reported by GWAS meta-analysis in Europeans (Fig. 3). By combining the East Asian data with European data, we were also able to seek interethnic genetic heterogeneity of GWAS results for blood pressure between the two ancestries (Fig. 1) as well as other complex traits. In particular, the present study provides examples for interethnic genetic heterogeneity, although the incidence may not be high, discovering two remarkable phenomena: (1) the colocalization of distinct ancestry-specific variants that are not rare and can exert mutually inverted genetic effects between the ethnic groups and (2) the potential involvement of natural selection in the occurrence of ancestry-specific association signals. Among genetic loci identified in East Asians, of note is the finding that at two loci on 1p35 and 3p21, the latter of which resides near the association signal previously reported in Chinese25, sentinel SNPs (rs2076460 and rs3774447) appear to be specific to East Asians; i.e., in Europeans the corresponding SNPs were monomorphic and no significant association signals were detectable in the vicinity (Supplementary Data 7). These support the possible presence of multiple East Asian-specific associations as well as European-specific ones. 0.00 0.25 0.50 0.75 1.00 Trans-ancestry genetic correlation (95% CI) EAS vs. EUR SBP DBP HDL LDL TC TG T2D BMI Height Heritability (95% CI) 0.00 0.25 0.50 0.75 1.00 EAS SBP DBP HDL LDL TC TG T2D BMI Height Heritability (95% CI) 0.00 0.25 0.50 0.75 1.00 EUR SBP DBP HDL LDL TC TG T2D BMI Height Fig. 2 Transethnic genetic correlation and SNP-based heritability. SNP-based heritability of SBP, DBP and other complex disease and phenotype traits is shown separately for East Asians (EAS) and Europeans (EUR) by using the published GWAS summary statistics (Supplementary Table 6). The whiskers are 95% confidence intervals of each value ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-07345-0 6NATURE COMMUNICATIONS | (2018) 9:5052 | DOI: 10.1038/s41467-018-07345-0 | www.nature.c om/naturecommunications It has been suggested that GWAS signals are produced by causal variants that are common and shared between ancestry groups26, with evidence for alternative rare variant association models (e.g., synthetic association27) that are assumed to be restricted to a limited number of loci. Apart from these models, we have discovered a new model in which genetic effects for transethnic SNPs that form a shared haplotype at a locus are driven by causal variants that are ancestry-specific but are not rare, which can be called a common ancestry-specific variant association model. We previously reported on 12q24 the East Asian-specific association signal at ALDH2 with blood pressure, which was located near the association signal at SH2B3 identified –0.02 0.00 0.02 –0.02 0.00 0.02 r = 0.12 Significant in NEUR = 105,253 Significant in NEAS= 158,645 DBP a Standardized effect-size in EUR Standardized effect-size in EAS –0.2 –0.1 0.0 0.1 0.2 –0.15 –0.10 –0.05 0.00 0.05 0.10 0.15 r = 0.05 Significant in NEUR = 173,082 Significant in NEAS= 31,732 LDL-C Standardized effect-size in EUR Standardized effect-size in EAS –0.04 0.00 0.04 –0.04 –0.02 0.00 0.02 0.04 Significant in NEUR = 158,186 Significant in NEAS= 25,066 r = 0.07 T2D Standardized effect-size in EUR Standardized effect-size in EAS –0.050 –0.025 0.000 0.025 0.050 –0.050 –0.025 0.000 0.025 0.050 Significant in NEUR = 322,154 Significant in NEAS= 158,284 r = 0.26 BMI Standardized effect-size in EUR Standardized effect-size in EAS Number of SNPs in bin 110 3106 –0.050 –0.025 0.000 0.025 0.050 –0.06 –0.03 0.00 0.03 0.06 Significant in NEUR = 253,279 Significant in NEAS= 36,227 r = 0.20 Height Standardized effect-size in EUR Standardized effect-size in EAS 100K 200K 500K 21 59 224 22 8 13 18 62 13 25 45 234 18 43 110 100K200K 500K Number of loci detectable in a single EUR GWAS, NEUR= Number of loci detectable in a single EAS GWAS, NEAS = b 28 66 198 59 17 28 43 137 21 41 78 383 26 55 135 100K200K 500K Number of loci detectable in a single EUR GWAS, NEUR= 100K 200K 500K Number of loci detectable in a single EAS GWAS, NEAS = 17 56 233 9579 32 9 17 27 144143787 100K 200K 500K Number of loci detectable in a single EAS GWAS, NEAS = 100K200K 500K Number of loci detectable in a single EUR GWAS, NEUR= 26 94 403 13 6 9 12 51 13 26 43 249 21 61 155 100K 200K 500K 100K 200K 500K Number of loci detectable in a single EAS GWAS, NEAS = 100K200K 500K Number of loci detectable in a single EUR GWAS, NEUR= 93 239 656 153 56 95 136 370 76 157 286 836 90 220 525 Number of loci detectable in a single EAS GWAS, NEAS = 100K200K 500K Number of loci detectable in a single EUR GWAS, NEUR= NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-07345-0 ARTICLE NATURE COMMUNICATIONS | (2018) 9:5052 | DOI: 10.1038/s41467-018-07345-0 | www.nature.c om/naturecommunications 7 in Europeans19. We also reported that these two association signals were phylogenetically independent, although a distance between the sentinel SNPs (rs671 and rs3184504) was relatively close (357 kb apart). Moreover, in the present study we have detected a number of transethnic SNPs in the 12q24 region to show highly significant heterogeneity of genetic impact on SBP between the ethnic groups (e.g., β EAS =−0.73 and β EUR =0.37, P het =9.74×10–21 at rs4766566), where inverted genetic effects are attributable to each of the ancestry-specific sentinel SNPs; a similar situation was also observed on 10q21 to reproduce this phenomenon (Fig. 1and Supplementary Data 6). Using ancestryspecific SNPs that are reported to reach genome-wide significance in either of the ethnic groups, we have found further evidence supporting the common ancestry-specific variant association model at 11 of 48 loci (23%) examined (Supplementary Data 7). This corresponds to 2.5% of total non-rare genome-wide significant blood pressure loci reported to date in populations of European and/or East Asian descent3–8,19 (Supplementary Fig. 7). Although it is beyond the scope of this study, part of the lowfrequency variants at group-2b loci (which constitute 4.0% of the tested blood pressure loci) may also be ancestry-specific20. These findings are important and should be kept in mind in the two well-known applications of transethnic GWAS, i.e., meta-analyses to increase the power for detecting new susceptibility loci and fine mapping. With the increase in sample size used for GWAS meta-analysis, it is expected that a larger number of genetic loci will be detected, and the distribution of such loci in the genome will become denser. Association signals annotated to the same locus are empirically defined such that a set of SNPs are bounded by pairwise correlation with the index SNP of r2≥0.1−0.3 within ±250–500 kb of the index SNP26,28. This is usually discussed in the context of locus heterogeneity rather than allelic heterogeneity. Apart from extreme cases in which an index SNP is monomorphic in the second ethnic group as above-mentioned, the cases in which a common variant in question is less common or even rare in the second ethnic group necessitate a greater sample size to achieve comparable statistical power for detecting a significant association. We should be careful in setting appropriate significance thresholds to maintain a balance between generating spurious associations and missing true modest associations in the second ethnic group. Hence, we chose ancestryspecific loci based on the Pvalue thresholds adjusted for the number of SNPs located ≤500 kb from the sentinel SNP. This set of loci may not exclude some cases with insufficient statistical power but can include the cases in which genetic impact at the locus is largely regarded as specific to the original ethnic group. Evidence of positive selection was observed at five unique loci among the list of ancestry-specific loci (Fig. 5and Supplementary Data 9). In addition to ancestry-specific loci, although the proportion appears to be relatively modest (approximately 2.5%), we have found significant interethnic heterogeneity of genetic impact at a number of blood pressure loci that are non-rare in both ancestries, with most of them originally reported in Europeans to date. It is assumed that the potential presence of modifier genes and/or gene−environment interactions can contribute to such interethnic heterogeneity but the overall influences and underlying mechanisms remain to be investigated. When combined with ancestry-specific variant associations (at group-2a or group2b loci in Supplementary Fig. 7), >5% of blood pressure loci are likely to show significant interethnic heterogeneity between East Asians and Europeans. According to our SNP-based heritability analysis, the genomewide correlation of causal-variant effect-sizes at SNPs common in both ancestry groups is 0.898 and 0.851 for SBP and DBP, respectively (Fig. 2). Part of the reduced interethnic correlation is attributable to transethnic variants that are common across populations but show substantial interethnic heterogeneity, although the proportion of such variants may not be high (e.g., 9 loci with interethnic heterogeneity detected in group-1; Supplementary Fig. 7). Even though they are not included in the SNPbased heritability analysis, ancestry-specific variants (at group-2a loci in Supplementary Fig. 7) can influence the per-allele effectsizes for a number of transethnic SNPs at the corresponding loci via LD, e.g., at the C10orf107 and CACNB2 loci. In summary, we identify a total of 19 genetic loci that have not been reported previously by GWAS meta-analysis, using relatively large discovery-stage sample size in East Asian populations. By comparing GWAS data for two ethnic groups, we have newly defined, so to speak, a common ancestry-specific variant association model, which should be brought to attention in the applications of transethnic GWAS. Methods Populations and genotyping. Description of the study design and phenotype measurement for each East Asian study (or cohort) participating in GWAS metaanalysis is provided in the Supplementary Methods. Descriptive statistics of the individuals, genotyping arrays, quality control filters, and genotype imputation applied to the individual studies are provided in Supplementary Tables 1, 2; 1000 Genomes Phase 3 reference panel was used for imputation in all studies except BBJ (1000 Genomes Phase 1) and TMM CommCohort Study (ToMMo 2KJPN panel plus 1000 Genomes Phase 3). SNP alleles were oriented to the forward strand of the GRCh37/hg19 reference sequence of the human genome. Collection of data and samples by the cohorts participating in the study was approved by respective research ethics committees, and written consent for participation was provided by all research participants. Phenotype modeling and SNP association analysis. For individuals taking antihypertensive therapies, blood pressure was imputed by adding 15 mmHg and 10 mmHg to SBP and DBP values, respectively. MAP and PP were calculated as MAP =(2 DBP +SBP)/3 and PP =SBP –DBP. In each study, the association of blood pressure (SBP, DBP, MAP or PP) with SNP allele dose was tested using linear regression adjusted for age, sex, and any study-specific covariates. Hypertensive cases were defined as follows: (i) SBP ≥160 mmHg and/or DBP ≥100 mmHg and/or on antihypertensive treatment and (ii) age of onset ≤65 years. Normotensive controls were defined as follows: (i) SBP < 130 mmHg and DBP < 85 mmHg and not on antihypertensive treatment and (ii) age ≥50 years. In each study, the association of a dichotomous trait of hypertension status with SNP allele dose was tested using logistic regression adjusted for sex and any study-specific covariates. The effect-sizes and standard errors estimated in linear and logistic regressions were used in subsequent meta-analysis. Fig. 3 Distribution of SNP effect-size in GWAS and power of GWAS. They are compared between East Asians and Europeans for DBP, low-density lipoprotein cholesterol (LDL-C), type 2 diabetes (T2D), body mass index (BMI) and height. aDistribution of SNP effect-size in actual GWAS conducted in East Asians (x-axis) and Europeans (y-axis). The effect-size of an SNP was standardized such that each of the trait and allele has a unit variance. The standardized effect-size equals the genetic impact. A positive effect-size indicates a higher trait value for the ALT allele compared to the REF allele of the 1000 Genomes (1000G) phase-3 data set. The horizontal and vertical bars to the bottom and right of the plots indicate the range of effect-sizes, in which genome-wide significant SNPs are localized. bThe expected numbers of genome-wide significant loci detectable in a single GWAS and their interethnic overlap. The number of SNPs was scaled to 1000G SNPs even for GWAS in which HapMap-derived SNPs were assayed. SNPs located ≤500 kb were regarded to be at the same susceptibility locus. The numbers of loci were inferred from the heritability model shown in Supplementary Fig. 13, where true observable effect-sizes were computed based on 100 trials of random sampling under the assumed heritability parameters (see Methods) ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-07345-0 8NATURE COMMUNICATIONS | (2018) 9:5052 | DOI: 10.1038/s41467-018-07345-0 | www.nature.c om/naturecommunications DBP Sentinel SNPs of loci associated in EAS Specific to EAS (1) Not specific (39) 0.00 0.01 0.02 0.03 0.04 Standardized effect-size in EAS (absolute value) 0.0 0.1 0.2 0.3 0.4 0.5 Minor allele frequency in EAS LDL-C Specific to EAS (0) Not specific (11) 0.00 0.05 0.10 0.15 0.20 Standardized effect-size in EAS (absolute value) 0.0 0.1 0.2 0.3 0.4 0.5 Minor allele frequency in EAS T2D Specific to EAS (2) Not specific (22) 0.000 0.025 0.050 0.075 Standardized effect-size in EAS (absolute value) 0.0 0.1 0.2 0.3 0.4 0.5 Minor allele frequency in EAS BMI Specific to EAS (6) Not specific (61) 0.00 0.01 0.02 0.03 0.04 0.05 Standardized effect-size in EAS (absolute value) 0.0 0.1 0.2 0.3 0.4 0.5 Minor allele frequency in EAS Height Specific to EAS (0) Not specific (70) 0.00 0.02 0.04 0.06 Standardized effect-size in EAS (absolute value) 0.0 0.1 0.2 0.3 0.4 0.5 Minor allele frequency in EAS Sentinel SNPs of loci associated in EUR Specific to EUR (8) Not specific (47) 0.00 0.01 0.02 0.03 0.04 Standardized effect-size in EUR (absolute value) 0.0 0.1 0.2 0.3 0.4 0.5 Minor allele frequency in EUR Specific to EUR (7) Not specific (65) 0.00 0.05 0.10 0.15 0.20 Standardized effect-size in EUR (absolute value) 0.0 0.1 0.2 0.3 0.4 0.5 Minor allele frequency in EUR 0.000 0.025 0.050 0.075 Specific to EUR (3) Not specific (36) Standardized effect-size in EUR (absolute value) 0.0 0.1 0.2 0.3 0.4 0.5 Minor allele frequency in EUR Specific to EUR (6) Not specific (68) 0.00 0.01 0.02 0.03 0.04 0.05 Standardized effect-size in EUR (absolute value) 0.0 0.1 0.2 0.3 0.4 0.5 Minor allele frequency in EUR 0.00 0.02 0.04 0.06 Specific to EUR (21) Not specific (459) Standardized effect-size in EUR (absolute value) 0.0 0.1 0.2 0.3 0.4 0.5 Minor allele frequency in EUR Fig. 4 Interethnic compatibility of GWAS results for DBP, LDL-C, T2D, BMI, and height. Each point in the plots represents a sentinel SNP with genome-wide significance in the GWAS summary statistics (Supplementary Table 6), plotted with its standardized effect-size (in y-axis) against minor allele frequency (in x-axis) for East Asians (EAS in the left column) and Europeans (EUR in the right column). SNPs specific to either of the ethnic groups are colored in red; ancestry-specific association was defined such that the sentinel SNPs at the corresponding loci reached genome-wide significance (P< 5×10–8) in one ethnic group but were non-polymorphic or rare (MAF < 0.05) in another ethnic group NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-07345-0 ARTICLE NATURE COMMUNICATIONS | (2018) 9:5052 | DOI: 10.1038/s41467-018-07345-0 | www.nature.c om/naturecommunications 9 University Medical Center Groningen, University of Groningen, Hanzeplein 1, 9713 GZ Groningen, Netherlands. 113 Department of Obstetrics and Gynecology, University Hospital of Oulu, University of Oulu, Oulu FI-90014, Finland. 114 Department of Clinical Sciences/Obsterics and Gynecology, University of Oulu, Oulu FI-90014, Finland. 115 Center for Cardiovascular Diagnostics and Prevention, Cleveland Clinic, Cleveland, OH 44195, USA. 116 Department of Cellular and Molecular Medicine, Lerner Research Institute, Cleveland Clinic, Cleveland, OH 44195, USA. 117 Keck School of Medicine, University of Southern California, Los Angeles, CA 90089, USA. 118 Research Centre for Prevention and Health, Glostrup University Hospital, 2600 Glostrup, Denmark. 119 Department of Epidemiology and Public Health, UCL, London WC1E 6BT, UK. 120 Heart Centre, Department of Cardiology, Tampere University Hospital, and University of Tampere School of Medicine, FI-33521 Tampere, Finland. 121 Department of Public Health & Clinical Medicine, Section for Family Medicine, Umeå universitet, SE-901 85 Umeå, Sweden. 122 Department of Epidemiology, University Medical Center Groningen, University of Groningen, Hanzeplein 1, 9713 GZ Groningen, Netherlands. 123 Department of Vascular Medicine, University Medical Center Utrecht, 3508 GA Utrecht, Netherlands. 124 Alzheimer Scotland Dementia Research Centre, University of Edinburgh, 7 George Square, Edinburgh EH8 9JZ, UK. 125 Academic Section of Geriatric Medicine, Institute of Cardiovascular and Medical Sciences, Faculty of Medicine, University of Glasgow, Glasgow G4 0SF, UK. 126 Institute of Clinical Molecular Biology, Christian-Albrechts-University of Kiel, Kiel 24105, Germany. 127 Department of Cardiology, Leiden University Medical Center, 2300 RC Leiden, Netherlands. 128 Department of Clinical Physiology, Tampere University Hospital, FI-33521 Tampere, Finland. 129 Department of Clinical Physiology, University of Tampere School of Medicine, FI-33014 Tampere, Finland. 130 Department of Obstetrics and Gynecology, Oulu University Hospital, PO Box 23FI-90029 Oulu, Finland. 131 Department of Medicine, Turku University Hospital, FI-20521 Turku, Finland. 132 Department of Medicine, University of Turku, FI-20014 Turku, Finland. 133 Durrer Center for Cardiogenetic Research, ICIN-Netherlands Heart Institute, 3511 GC Utrecht, Netherlands. 134 Institute of Cardiovascular Science, Faculty of Population Health Sciences, University College London, London WC1E 6BT, UK. 135 Department of Nutrition, Harvard School of Public Health, Boston, MA 02115, USA. 136 Institute of Reproductive and Developmental Biology, Imperial College London, Hammersmith Hospital, London W12 0HS, UK. 137 School of Medicine, University of Minnesota, Minneapolis, MN 55455, USA. 138 Biocenter Oulu, University of Oulu, P.O. Box 5000Aapistie 5A, FI-90014 Oulu, Finland. 139 Unit of Primary Care, Oulu University Hospital, Kajaanintie 50P.O.Box 20FI-90220 Oulu, Finland. 140 Department of Children and Young People and Families, National Institute for Health and Welfare, Aapistie 1, Box 310, FI-90101 Oulu, Finland. 141 Interuniversity Cardiology Institute of the Netherlands, Utrecht 3511 EP, Netherlands. 142 Department of Cardiovascular Sciences, University of Leicester, Glenfield Hospital, Leicester LE3 9QP, UK. 143 National Institute for Health Research Leicester Cardiovascular Biomedical Research Unit, Glenfield Hospital, Leicester LE3 9QP, UK. 144 Institute of Molecular GeneticsCNR, 27100 Pavia, Italy. 145 Clinical Institute of Medical and Chemical Laboratory Diagnostics, Medical University of Graz, Auenbruggerplatz 15, 8036 Graz, Austria. 146 Synlab Academy, Synlab Services GmbH, GottliebDaimler-Straße 25, 68165 Mannheim, Germany. 147 Department of Endocrinology, University Medical Center Groningen, University of Groningen, Hanzeplein 1, 9713 GZ Groningen, Netherlands. 148 National Hellenic Research Foundation, Institute of Biological Research and Biotechnology, Athens 116 35, Greece. 149 Department of Toxicogenomics, Maastricht University, Universiteitssingel 50, 6229ER Maastricht, Netherlands. 150 Oxford Centre for Diabetes Endocrinology and Metabolism, University of Oxford, Oxford OX3 7LE, UK ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-07345-0 16 NATURE COMMUNICATIONS | (2018) 9:5052 | DOI: 10.1038/s41467-018-07345-0 | www.nature.c om/naturecommunications