KoVariome: Korean National Standard Reference Variome database of whole genomes with comprehensive SNV, indel, CNV, and SV analyses

High-coverage whole-genome sequencing data of a single ethnicity can provide a useful catalogue of population-specific genetic variations, and provides a critical resource that can be used to more accurately identify pathogenic genetic variants. We report a comprehensive analysis of the Korean population, and present the Korean National Standard Reference Variome (KoVariome). As a part of the Korean Personal Genome Project (KPGP), we constructed the KoVariome database using 5.5 terabases of whole genome sequence data from 50 healthy Korean individuals in order to characterize the benign ethnicity-relevant genetic variation present in the Korean population. In total, KoVariome includes 12.7M single-nucleotide variants (SNVs), 1.7M short insertions and deletions (indels), 4K structural variations (SVs), and 3.6K copy number variations (CNVs). Among them, 2.4M (19%) SNVs and 0.4M (24%) indels were identified as novel. We also discovered selective enrichment of 3.8M SNVs and 0.5M indels in Korean individuals, which were used to filter out 1,271 coding-SNVs not originally removed from the 1,000 Genomes Project when prioritizing disease-causing variants. KoVariome health records were used to identify novel disease-causing variants in the Korean population, demonstrating the value of high-quality ethnic variation databases for the accurate interpretation of individual genomes and the precise characterization of genetic variations.


Results and Discussion
Construction of the Korean standard Variome: KoVariome. Since 2010, the Korean variome data center (KOVAC), as a part of the KPGP, has been recruiting volunteers to generate WGS and whole exome sequencing (WES) data. The current KoVariome (version 20160815) has been constructed based on WGS data from 50 unrelated Korean individuals who responded to questionnaires detailing body characteristics, habits, allergies, family histories, and physical conditions related to 19 disease classes (Supplementary Table S1). A total of 5.5 TB of high-quality paired-end WGS data were generated, containing an average of 31× coverage per individual (Table 1  and Supplementary Table S2). WGS data from each individual covered 95% of the human reference genome (hg19) on average. From these data, we identified approximately 3.8M SNVs (ranged 3.7-3.9M) and 0.5M indels (0.4-0.7M) per Korean individual (Table 1 and Supplementary Fig. S1A). The hetero-to-homozygosity ratio of the autosomal SNVs was 1.49, which is consistent with previously reported data 38 . The length distributions of the indel loci were symmetric, with the majority of indel sizes shorter than six bases (94.8% for insertions, 97.8% for deletions) ( Supplementary Fig. S1B). We identified approximately 20,097 (0.53%) SNVs and 258 (0.05%) indels in the coding regions including 10,394 (0.22%) non-synonymous changes per individual (Table 1).
Scientific REPORtS | (2018) 8:5677 | DOI: 10.1038/s41598-018-23837-x Novel KoVariome SNVs were counted by adding individual samples one by one (Fig. 1A), and the number of novel SNVs decreased logarithmically and became depleted after the 9 th donor. In total, we observed 59K novel SNVs, including 1.2K (2.03%) coding-SNVs, per individual. To assess the relatedness of the KoVariome individuals, we compared the pairwise genetic distance of KoVariome with those of family data (Fig. 1B). WGS data from thirty families were downloaded from the KPGP database, which included two monozygotic twins, 14 parent-children pairs, seven siblings, five grandparents-grandchildren, six uncles-nephews, and three cousins. We analyzed familial SNVs using the same method as in KoVariome and also compared genetic distances between the two groups (see Methods). The genetic distance among KoVariome individuals was higher (pi = 8.8e-4) than those found in the familial data, such as monozygotic twins (4.8e-4), siblings (6.7e-4), parent-child (6.8e-4), uncle-nephew (7.7e-4), grandparents-grandchild (7.8e-4), and cousins (8.2e-4). This verifies that no genetic bias was present in the sample collection stage and current KoVariome. In accordance with previous reports, the multidimensional scaling (MDS) of variants among Korean, Chinese, and Japanese individuals showed a clear separation of the three populations (Fig. S2) despite the geographical and historical associations between these groups 35,37 . These analyses reinforce the need for distinct KOREF and KoVariome reference resources to parse disease variants from demographic variants in this population.

Accuracy test of SNVs and indels in KoVariome.
We evaluated the accuracy of KoVariome SNV and indel predictions by comparing genotype results from the Axiom TM Genome-ASI 1 Array with WGS data from 35 individuals. A total of 503,694 SNV positions were compared, from which we obtained an average of 0.9993 precision (ranged: 0.9984-0.9996) and 0.9980 recall (ranged: 0.9817-0.9994) (Supplementary Table S3). In addition, there was a 99.65% (ranged: 98.62-99.87%) concordance of the SNVs called by the WGS and Axiom array calls. Compared to similar variome studies, this genotype accuracy was slightly lower than the high-depth trio data in the Danish population study (99.8%) 16 but higher than that of the Dutch population SNVs (99.4-99.5%) analyzed with intermediate depths 39 . The accuracy of the SNV calls was analyzed across the genome, and a total of 499,889 (99.24%) SNVs showed a genotype concordance higher than 0.99, while 0.4% of SNVs showed the genotype accuracy less than 0.95 (Supplementary Table S4). Similar levels of genotype concordances were observed in repetitive regions of the genome (99.56% of SNVs with the genotype correspondence >0.95, Supplementary Table S5), suggesting that SNV calling accuracy is not reduced in repetitive regions of the genome.
We also compared the accuracy of indel variant calls with the 1,981 indel markers on the Axiom TM Genome-ASI 1 Array. A genotype comparison showed an average accuracy of 98.49% for indels, which was slightly lower than those observed in SNVs (Supplementary Table S3), and comparable to the false positive (FP) rate for indels that was reported in the Danish data 16 . In terms of genomic loci, 1,343 (91.11%) indels showed perfect genotype concordance with array data and 1,446 (98.10%) indels had an accuracy higher than 90% ( Supplementary Fig. S3).
Genome-wide features of KoVariome. By merging the variants of 50 unrelated Korean individuals, we identified 12.7M SNVs and 1.7M small indels shorter than 100bp (Table 1); approximately 1.5 times the number of SNVs previously reported from preliminary KPGP data (8.5M) 33 . Both types of variants were primarily distributed in the non-coding regions (about 98%), including intergenic and intron regions (Supplementary Table S6 in the 1000GP variome (Supplementary Table S6); and based on allele frequencies, 4.6M (51.03%) and 4.4M (48.82%) of these SNVs were classified into the categories '1000GP common' and '1000GP low frequency' , respectively ( Fig. 2A). Most notably, 13,584 (0.15%) KoVariome SNVs were rarely observed in the 1000GP continental groups with a MAF < 0.1%. A similar distribution was observed with the indels, where 64.2% and 35.8% of the KoVariome indels were classified into the '1000GP common' (0.5M) and '1000GP low frequency' (0.3M) classes, respectively. Only ten indels were classified into the '1000GP rare' category. Almost all of the variants in the '1000GP common' category were also frequently observed in KoVariome, representing 4.5M (98.33%) SNVs and 0.5M (93.37%) indels in this class ( Fig. 2A and Supplementary Table S6). Surprisingly, however, roughly half of the variants in '1000GP low frequency' were classified as 'frequent in KoVariome' . This indicates that there exist a significant population specific biases for common and uncommon variants. When we compared the allele frequencies in five the continental 1000GP groups to KoVariome. In total, we observed 3.4M (77.19%) SNVs and 0.2M (74.21%) indels that were statistically enriched in at least one of the continental groups or the Korean population ( Fig. 2B), suggesting a population stratification. To further explore the population stratification, we identified the variants uniquely enriched in each continental group, and the enriched variants that were in common between the continental groups. In total, nearly three million (2.7M) SNVs and 156K indels were frequently found in the Korean population. Among them, 2.5M (95.20%) SNVs and 143K (94.47%) indels showed Korean specific enrichments, while the other enriched variants were shared by other continents (Fig. 2B). Among the five continental groups, as expected, EAS shared the largest number of enriched variants (89.5K SNPs and 5.3K indels) with the Korean population 40 .
Interpretation of the KoVariome-specific variants. Characterizing ethnicity specific variants is necessary to understand the demographic differences between populations and can be used to filter out low frequency clustered variants in a specific group. In KoVariome, there were 3.8M SNVs and 0.9M indels not observed in the 1000GP variome (Supplementary Table S6  To identify the possible clinical relevance of these KoVariome-specific frequent variants, we compared the genomic loci of these SNVs against the ClinVar database and identified six likely pathogenic loci with associated disease information ( Table 2). Two of these likely pathogenic SNVs (rs386834119 and rs1136743) were associated with autosomal recessive (AR) diseases, and therefore, no phenotypes were expected since all of the KoVariome SNVs were heterozygotes at these sites. We also observed a KoVariome allele (three males and two females) of a possible cancer-associated SNV (rs200564819) in RAD51D, which was previously reported to increase the risk of developing ovarian, breast, colorectal, lung, pancreatic, and prostate cancers 41 . It has been suggested that this allele truncates the RAD51D gene by interrupting a canonical splice site, though additional genetic data is needed to conclusively classify this allele as "pathogenic" 42 . Since there are five heterozygous rs200564819 alleles in KoVariome without any cancer incidence ( Table 2, Table S1), these variants may not be as high-risk in the Korean population; though the database size will need to be increased to verify the effect of such population-specific disease associated markers. In addition, we observed two pathogenic missense SNVs (rs121912678 and rs20016664) in the KoVariome population that have been previously reported to be associated with fibrodysplasia ossificans progressive (FOP) and Van der Woude syndrome (VWS), respectively ( Table 2). The rs121912678 SNV (chr2:g-158630626C>G) was rarely observed in the ExAc database (MAF = 0.0002), but the C>T mutation at this position was predicted to cause the FOP disease by constitutively activating the activin receptor type I (ACVR1) 43 . While the pathogenicity of R206P in ACVR1 due to a C>G mutation is not yet known, we suggest that it is likely benign because of the high MAF (0.14) of this allele in KoVariome without any FOP phenotypes, skeletal malformation, or progressive extraskeletal ossification recorded in the KPGP survey. Finally, the 400 th amino acid of the interferon regulatory factor 6 (IRF6) gene is known to be a hot spot of VWS, orofacial clefting disorders. Two pathogenic variants, R400W 44 and R400Q 45 , have been reported for VWS; however, the pathogenicity of R400P arisen by chr1:209961970C>G, as frequently seen in KoVariome, is not yet confirmed. A total of 14 heterozygous SNVs had no phenotype for VWS symptom, despite the AD inheritance pattern of this disease; and consequently, the R400P substitution seems to be benign. Taken together, the KoVariome-specific frequent variants demonstrate the importance of using population-scale health data to identify pathogenic loci in specific diseases, and for the accurate identification of benign variants that are not annotated because of population stratification.
Genomic distribution of rare variants. We investigated the proportion of the SNVs in four SNV classes (1000GP Common, 1000GP Low Frequency, 1000GP Rare, KoVariome Specific; Supplementary Fig. S4). Our analyses showed that a high portion of the coding SNVs were enriched in the '1000GP rare' class, while the SNVs in the non-coding regions were similarly distributed in all other variant classes. The portion of non-synonymous SNVs in the '1000GP rare' class was more than twice what was observed in the other classes. It is possible that these patterns are associated with purifying selection to rapidly remove deleterious alleles in the population 46 , though it was not possible to identify this pattern in frame-shift indels because of the small number of variants (981) in this class. To analyze the tendencies of purifying selection in KoVariome, we defined rare variant ratios (RVRs) as the number of SNVs in the 'rare in KoVariome' class divided by the number of SNVs in the 'frequent in KoVariome' class. We then compared RVRs across genomic regions (Fig. 2C). In both SNVs and indels, RVRs in the intergenic region were lowest (0.66), while similar levels of RVRs were observed in other non-coding regions (0.66-0.87). Under the assumption that mutations occur randomly throughout the genome, lower rates of RVR in non-coding regions suggest neutral selection with no or weak selection pressures in the population. Conversely, the highest RVR in frame-shift indels (1.45) suggests there was some purifying selection against these variants in the Korean population. Furthermore, about twice as many RVRs were observed in the non-synonymous (1.16) and splice-site (1.33) SNVs compared to intergenic regions. Although SNVs in the coding region can be deleterious to protein function, selection pressure on the non-synonymous and splice-site SNVs seem to be slightly lower than that of the frame-shift indels, as expected.

Interpretation of disease-causing variants among Korean individuals. Rare SNVs in an individual
genome are more likely to be pathogenic than common variants. Because genetic variants are known to be geographically clustered, characterizing population stratification is a critical first step to identifying disease-causing variants 47 . With this concept, we examined rare SNVs in each individual after filtering out SNVs that were classified as '1000GP common' , '1000GP low frequency' , or 'frequent' in KoVariome. From an average of 3.8M SNVs per individual, 3.4M (88.70%) and 0.4M (9.39%) SNVs were filtered out using the 1000GP variome or KoVariome, respectively ( Fig. 3A and Table 3). Overall, KoVariome allowed 1,231 (12.25%, median value) non-synonymous SNVs and 40 (24.01%) splice-site SNVs to be filtered out as common variants in the Korean population, which significantly improves the ability to pin-point disease causative variants.
After filtering, Korean donors had a median of 47,957 (1.26%) rare SNVs, most of which (98.33%) were located in non-coding regions. Among these, we observed an average of 219 (67.17%) non-synonymous SNVs and seven (0.87 %) splice-site SNVs per individual ( Fig. 3B and Table 2). On average, 166 (73.45%) of these SNVs were present in dbSNP (ver. 146), but not in the 1000GP variome (Fig. 3C) Table S7). In total, only 38 (0.5%) pathogenic rare SNVs in KoVariome were homozygotes and the remaining (99.5%) were heterozygotes. In addition, 29 (58%) of the donors had no homozygous pathogenic rare SNVs. To obtain clinical information concerning these pathogenic rare-SNVs, we searched the genomic loci for these SNVs against the ClinVar database. A total of 127 of the rare SNVs were found in ClinVar, 53 of which showed clear clinical significance. Eight (6.39%) and thirteen (10.24%) were listed as benign and likely benign in ClinVar, respectively, and not fatal for a specific disease. Conversely, 29 (22.83%) and three (2.36%) were pathogenic and likely pathogenic, respectively (Table 4). These rare SNVs contribute to pathogenicity according to their inheritance patterns, and a manual investigation of the inheritance type using the Online Mendelian Inheritance Man (OMIM) database identified seven AD and 17 AR SNVs for specific loci; although we failed to identify the inheritance types for eight SNV loci (Table 4). All 17 of the AR SNVs were heterozygous in KoVariome, so it was not possible to assign phenotypes to these loci. Within the donor group with pathogenic rare AD SNVs, we searched for phenotypes or familial histories associated with target diseases in the questionnaire. We identified a familial history for type II diabetes mellitus associated with rs121918673 allele in KPGP participants; however, one donor with the rs121918673 allele was   nondiabetic and reported no family history of this disease. Additionally, one donor was heterozygous for the rs121912749 allele, which has an AD association with spherocytosis, and this donor reported associated symptoms but no anemia (Supplementary Tables S1 and S7). However, it is clinically known that spherocytosis has heterogenetic symptoms ranging from asymptomatic to hemolytic anemia. These examples highlight the utility of population specific variation databases as an important resource for assessing the disease-relevance of genetic variants as a routine component of precision healthcare.  (Fig. 4B). A high proportion of SVs were specific to an individual genome (Fig. 4C), consistent with findings from the 1KJPN 17 . The portion of individual-specific SVs was greatest for INSs (92.51%), followed by INVs (88.87%), ITXs (68.93%), and DELs (47.82%) (Table S8). A substantial proportion of SVs (98.5% INSs and 61% DELs) were novel and were not previously deposited in the Database of Genomic Variants (DGV). Overall, the non-redundant combined SVs ranged in size up to 10M and all classes were enriched in the 1-2Kb size range (Fig. 4D, Supplementary Fig. S6).  Finally, we analyzed the SVs to determine whether they were enriched for repetitive elements. Within the SVs, we cataloged repeat types and searched for Korean-specific enrichments compared to those present in other populations. Among the SVs, we found that 13% contained short interspersed elements (SINEs), 20% contained long interspersed elements (LINEs), 3.4% contained DNA transposons, and 8.6% contained long terminal repeats (LTRs). The majority of SINEs were observed in DELs of 200-300bp, which is consistent with de novo assembled SVs 16 and the predicted SVs 15 . These results suggest that SVs are enriched for SINEs in the 1-4Kb INVs, and LINEs in the 4-40Kb INVs (Supplementary Fig. S6A). Additionally, simple repeats were predominantly observed in INSs (Figs. 4D) and 3-5Kb ITXs (Supplementary Fig. S6B).  Table S9). After filtering out spurious CNVs ( Supplementary Fig. S7), 161.74 (81.46%) deletions and 296.72 (88.29%) duplications remained from the original calls. In total, we predicted 2,038 non-redundant deletions and 1,564 non-redundant duplications, and the unified CNVs were approximately 5Kb-100Kb in length (Fig. 5A). When compared to the DGV, we identified 3.6K known CNVs, including 1,169 (57.36%) deletions and 846 (54.09%) duplications. Repeat composition analyses of CNV regions revealed that deletions smaller than 5K and duplications smaller than 10K contained a 20-fold more simple repeats compared to their overall frequencies in the human genome. In addition, SINEs were 2-fold more frequent in the >600 Kb deletions. These associations differ from the repeat distributions in SVs. By examining the genes in the unified CNVs, 869 (46.47%) deletions and 1,105 (70.65%) duplications were found to contain at least one gene. In addition, only two deletions and three duplications were conserved in the 50 Korean individuals (Table 5). Interestingly, a long 2M genomic block on chromosome 10, containing seven genes, was found to be duplicated an average of 4.22 times in the KPGP donors. Included among these genes is G protein regulated inducer of neurite outgrowth 2 (GPRIN2), which is associated with brain development and neurite outgrowth 49 . Previous reports identified this duplication in Asian, European, and Yoruba populations (three-six copies), while no duplications were reported in the chimpanzee, orangutan, or gorilla 22 . We also identified 444 CNVs conserved in 1000GP (Supplementary Table S10), which are probably shared East Asian CNVs and are not specific to Koreans. Five deletions and nine duplications were found to be enriched in the Korean population using the following criteria; i) odds ratio >10 comparing with CNV ratio in any continents, ii) p-values < 0.01, and iii) more than five individuals in KoVariome. Phenotypic features were examined by searching genes against the OMIM database, resulting in the identification of three deletions and three duplications containing genes associated with known phenotypes (Fig. 5B). A high copy number deletion of UDP glucuronosyltransferase family 2 member B17 (UGT2B17), which is associated with bone mineral density and osteoporosis 50 , was observed by comparing our Korean individuals with EUR, AFR, and AMR populations. This finding is consistent with previous studies which reported that 66.7% of Korean males have a deletion of this gene, compared to only 9.3% of Swedish males 51 . We also observed frequent deletions of acyl-CoA thioesterase 1 (ACOT1), which functions to maintain the cellular levels of acyl-CoA and free fatty acids 52 . We identified the duplication of hydroxycarboxylic acid receptor 2 (HCAR2) in 12% of the Koreans, which is associated with lipid-lowering effects 53 . We excluded the gene duplications of NBPF15 and HERC2 because they were located at the CNV break points. These CNVs will be useful for detecting Korean-specific genetic associations with specific phenotypes in future studies, which is especially important since CNVs are analyzed less often than SNVs even though they likely contain important disease-relevant variations.

Conclusions
To discover disease-causing genetic variants, researchers rely on comprehensive, population-specific databases containing the benign genetic variation present within specific ethnic groups. The KoVariome database was created to fill this need for the Korean population, and includes 5.5 TB of WGS data from 50 healthy, unrelated Korean individuals with corresponding health metadata. Using this database, we characterized all four variation types and identified 12.7M SNVs, 1.7M indels, 4K SVs, and 3.6K CNVs, many of which were novel or selectively enriched in the Korean population. Despite their close geographic proximity, the Korean population was shown to be genetically distinct from the Chinese and Japanese populations, highlighting the need for a Korean-specific variome to accurately identify rare disease variants in this population. Accordingly, a comprehensive comparative analysis of the population-specific variants within KoVariome was used to predict candidate loci, inheritance patterns, and genetic risk for several diseases, including cancer, fibrodysplasia ossificans progressive, Van der Woude syndrome, type II diabetes mellitus, and spherocytosis. As genetic tests become increasingly routine components of precision healthcare, KoVariome will be an invaluable resource for biomedical researchers and health practitioners; and will directly benefit patients by ensuring they are presented with the most accurate genetic predictions of disease risks.

Methods
Sample collection and data distribution. Since 2010, the Korean variome data center (KOVAC) recruited volunteers for the Korea Personal Genome Project (KPGP: http://kpgp.kr). All methods used in this study were carried out in accordance with relevant guidelines and regulations and were approved by the Institutional Review Board (IRB) of the Genome Research Foundation (GRF). Informed consent for study participation was acquired from all participants in accordance with the Korean Life Ethics bill, and all experimental protocols were approved by the GRF IRB. In addition to providing a blood sample for WGS, each individual responded to a questionnaire regarding body characteristics, habits, response to 16 allergies, family histories, and physical condition related to 19 disease classes (Table S1). Genomic DNA was extracted using a QIAamp DNA Blood Mini Kit (Qiagen, CA, USA) and 69 WGS libraries were constructed using TruSeq DNA sample preparation kits (Illumina, CA, USA). Sequencing was performed using Illumina HiSeq sequencers following the manufacturer's instruction. The homepage of KoVariome is http://variome.net. WGS data from 50 healthy unrelated Korean individuals were analyzed to create the KoVariome database, which was released through the national FTP portal server of the KOBIC (ftp://ftp.kobic.re.kr/pub/KPGP/) and distributed through GRF (http:// pgi.re.kr), and Variome.net. All data analyzed in this study were deposited in NCBI SRA (PRJNA284338) and accessions for each sample were listed in Supplementary Table S2. Analysis of SNVs and indels. The WGS data were processed according to a protocol that was evaluated by the technical committee of the Korean Research Institute of Standards and Science (KRISS). Genomic resources were downloaded from UCSC Genome bioinformatics (http://hgdownload.cse.ucsc.edu/goldenPath/hg19/big-Zips/), including the reference human genome (GRCH37/hg19), reference genes, and repeat annotations. Raw DNA reads were cleaned by Sickle (https://github.com/najoshi/sickle) with a quality score >20 and read length >50 bp. Cleaned paired-end reads were mapped to the human reference genome using BWA 54 and indels were realigned and recalibrated after removing the PCR duplicates. Finally, we identified SNVs and indels for each individual using the GATK UnifiedGenotyper (ver. GATK-Lite-2.3-9) 55 . To improve the quality of identified SNVs, we applied SNV meeting criteria of: i) read depth (DP) is 20× or higher, ii) mapping rate is 90% or higher. Low-quality indels were removed from future analyses using the following criteria: i) quality score <27 and DP <6, ii) heterozygous indels with mapped allelic valance less than 0.3.

Protein modeling of the variants.
To infer the functional effects of variations, we implemented SnpEff-3.3 56 . The deleterious effects of the non-synonymous SNVs were obtained by searching dbNSFP (ver. 2.9.1), a portal database providing deleterious non-synonymous SNVs 57 . We then predicted the effects of each variant on protein function using SIFT, Polyphen2, PROVEAN, MetaSVM, and MetaLE, and further annotated variants using the Interpro_domain and COSMIC (Catalogue of Somatic Mutations in Cancer, ver. 71) databases. Previously reported SNVs and indels were identified using the dbSNP database (ver. 146). All variants shorter than 50 bp were then stored in this database 58 . The databases ClinVar (ver. 20161101) 42 and OMIM (generated 2016- [11][12][13][14][15][16][17][18][19][20][21][22] 59 were searched to identify known pathogenic variants. Genetic distance calculation. The genetic distance (pi) between two samples was calculated using the following formula: where D is the nucleotide difference between two samples and N is the number of compared positions. The sum of the nucleotide difference was calculated between two samples for each genomic position, which ranged from 0-1. A homozygous genotype composed of a reference allele was adopted as the genotype for uncalled sites.
Multidimensional Scaling (MDS) analysis. Genotype data for 84 Chinese and 86 Japanese individuals were obtained from Phase 3 of the HapMap project 3 . A total of 1,387,956 SNV loci were merged with KoVariome. The PLINK program was used to remove the genomic loci with MAF < 0.05, call rates < 0.05, and SNPs in linkage disequilibrium blocks 60 . In total, 117,521 SNPs remained after filtering and were used in the MDS analysis. Five dimensional components were calculated in R with the distance matrix method "canberra" and MDS plots were generated using the MASS package 61 .
Accuracy of the SNVs. To measure the accuracy of SNV predictions, 35 individuals were genotyped with the Axiom TM Genome-Wide East Asian (ASI) 1 Array (Affymetrix, Inc.). The accuracy and recalls were analyzed using a contingency table constructed with the presence and absence of the alternative alleles analyzed from our pipeline and the genotyping results from the Axiom TM Genome-ASI 1 Array. The precision of calls was calculated by analyzing the concordance and denoted as true positive predictions (TP) from all predicted SNVs. The recalls were defined as TPs divided by the number of genotypes represented on the Axiom TM Genome-ASI 1 Array. The genotype accuracies were measured by analyzing the concordance of the genotypes between the GATK prediction and the results from the Axiom TM Genome-ASI 1 Array. The accuracy of the indel predictions were calculated by comparing genotypes between GATK predictions and the Axiom TM Genome-ASI 1 Array.

Structural variants.
We applied two programs, BD 62 and pindel 63 , to predict genome-wide SVs based on the discordant mate-pair and split-read information, respectively. From the bam files for each individual, insertions and deletions of a length between 100 and 1 Kb were predicted by pindel (ver. 0.2.4t) and those longer than 1Kb were predicted by BD (ver. 1.4.5) 64 . We next constructed unassembled genomic blocks ('N') from the hg19 reference genome and examined the SVs that overlapped with these unassembled genomic regions. From this analysis, we discovered a high portion of spurious SVs in these regions ( Supplementary Fig. S5), with the majority of them >100M in size. The following criteria were used to filter out spurious SVs; i) reciprocally >10% overlaps between SVs and un-assembled genomic blocks, ii) 'N's more than 50% coverage of SVs, and iii) more than 2 un-assembled genomic blocks in the predicted SVs. After filtering, we clustered SVs that reciprocally overlapped >70% in any individual. Unified SVs were defined by the average start and end positions in each SV cluster. The novelty of each SV was defined by comparing unified SVs with those in the DGV 65 , with 70% reciprocal overlaps.

Copy number variations.
CNVs were predicted with FREEC (ver. 10.6) using window size = 100, step size = 50, and breakpoint = 0.6 48 . The spurious CNVs were enriched in >1M in length ( Figure S7), which were filtered using the same criteria described in the SV methods above. Unified CNVs were constructed by merging individual's CNVs that reciprocally overlapped by >=70%. The start and end positions of the unified CNVs were defined as average position of the original calls. Known CNVs were defined by comparing with CNVs in the DGV database 65 .