Sequence data and association statistics from 12,940 type 2 diabetes cases and controls

To investigate the genetic basis of type 2 diabetes (T2D) to high resolution, the GoT2D and T2D-GENES consortia catalogued variation from whole-genome sequencing of 2,657 European individuals and exome sequencing of 12,940 individuals of multiple ancestries. Over 27M SNPs, indels, and structural variants were identified, including 99% of low-frequency (minor allele frequency [MAF] 0.1–5%) non-coding variants in the whole-genome sequenced individuals and 99.7% of low-frequency coding variants in the whole-exome sequenced individuals. Each variant was tested for association with T2D in the sequenced individuals, and, to increase power, most were tested in larger numbers of individuals (>80% of low-frequency coding variants in ~82 K Europeans via the exome chip, and ~90% of low-frequency non-coding variants in ~44 K Europeans via genotype imputation). The variants, genotypes, and association statistics from these analyses provide the largest reference to date of human genetic information relevant to T2D, for use in activities such as T2D-focused genotype imputation, functional characterization of variants or genes, and other novel analyses to detect associations between sequence variation and T2D.


Background & Summary
Genome wide association studies (GWAS) have provided a valuable but incomplete window into the genetic basis of type 2 diabetes (T2D) 1 . Common (minor allele frequency [MAF]>5%) variants at over 100 loci have been robustly associated with disease risk, but most have not yet been translated to causal variants, effector transcripts, or disease mechanisms 2 . Because common variants from GWAS have modest effect sizes, and because those previously published explain in aggregate only 10-15% of the genetic basis of T2D 3 , it has been hypothesized that variants unexplored by GWAS might have a greater impact on efforts to understand or treat T2D 4,5 .
To produce a more complete catalogue of rare, low-frequency, and common variants, the GoT2D and T2D-GENES consortia analysed whole-exome and genome sequence data in up to 12,940 individuals (6,504 T2D cases and 6,436 controls; Fig. 1a, Tables 1 and 2) 3 . First, to interrogate lower-frequency (MAF>0.5%) variation genome-wide, 2,657 Northern and Central European individuals were selected by GoT2D (1,326 cases, 1,331 controls) and characterized with a combination of low-pass (~5x) wholegenome sequencing, deep (~82x) whole-exome sequencing, and high-density (2.5M) SNP genotyping. Genetic variants from these assays were incorporated into a phased integrated panel (WGS panel), capturing an estimated 99% of variants, genome-wide, present in more than 0.5% of individuals (Table 3). Second, additional individuals from 10 cohorts spanning five ethnicities (European, Hispanic, South Asian, East Asian, and African American) were characterized by deep (~82x) whole-exome sequencing by T2D-GENES. The resultant T2D-GENES exome sequence data were combined with the GoT2D exome sequence data to produce a second panel of variation (WES panel), capturing an estimated 99.7% of coding variants present in more than 0.5% of the combined 12,940 individuals (Tables 4 and 5).
Each variant was tested for association with T2D under an additive genetic model. To increase power, variants were then assessed in larger sample sizes via one of two means (Fig. 1b). Coding variants were analysed in 79,854 additional individuals (28,305 T2D cases, 51,549 controls) via the Illumina Exome Array, which captures 81.6% of European MAF>0.5% coding variants in the WES panel. Non-coding variants (and coding variants absent from the Exome Array) were analysed in up to 44,414 additional individuals (11,645 cases, 32,769 controls) via statistical genotype imputation; after quality control, this analysis included 89% of variants observed in three or more individuals in the WGS panel. Each variant was tested for association with T2D in the additional individuals, under an additive genetic model, and association statistics were then combined with those from the sequence data via meta-analysis.
Collectively, these analyses suggest a limited role for low-frequency variation in the genetic basis of T2D 3 . However, they also demonstrate an ability to identify novel hypotheses about the effects of gene inactivation 6 , a resource of coding variants for calibrating cellular assays 7 , and a catalogue of noncoding variants for use in statistical or functional fine mapping of GWAS signals 3 . The WGS panel also provides a novel resource for genotype imputation 8 , with increased resolution for T2D-specific variants relative to the 1000 Genomes (1000G) reference panel, as well as a means to calibrate simulation-based models of population history 9 or disease genetic architecture 10 .

Methods
These methods are a modified version of the descriptions contained in Fuchsberger et al. 3 .

Ethics statement
All human research was approved by the relevant institutional review boards and conducted according to the Declaration of Helsinki. All participants provided written informed consent.

WGS (GoT2D integrated) panel generation
Ascertainment of individuals. Individuals were sampled from four studies: the Finland-United States Investigation of NIDDM Genetics (FUSION) Study (493 cases, 486 controls), KORA (101 cases, 104 controls), the UKT2D Consortium (322 cases, 322 controls), and the Malmö-Botnia Study (410 cases, 419 controls). All individuals were of Northern or Central European ancestry. Cases were preferentially lean, had (relatively) early onset T2D, or had a familial history of T2D; controls by comparison were preferentially overweight or had low fasting glucose levels 11 . To decrease the likelihood of selecting T2D cases who in fact had type 1 diabetes (T1D) or monogenic forms of diabetes (such as Maturity Onset Diabetes of the Young), cases with an age of diagnosis below 35, testing positive for GAD antibodies, or with a first-degree relative known to have T1D were not included. Statistics of the 2,657 individuals ultimately included in the association analysis are provided in Table 1.
Many of these individuals were measured for cardiometabolic phenotypes other than T2D, including glucose and insulin, anthropometrics, lipids, and blood pressure (  (24-marker panel) concordance, or estimated gender discordance were excluded from further analysis. Finally, SNPs monomorphic across all individuals, failed by the 1000G Omni 2.5 QC filter, or with Hardy-Weinberg equilibrium P o10 − 6 were excluded from analysis.
Processing, quality control, and variant calling of sequence data. Sequence data were processed and aligned to the human genome (build hg19) using the Picard (http://broadinstitute.github.io/picard/), BWA 12 , and GATK 13,14 software packages, following best-practice pipelines. Sequencing coverage of each individual was computed based on the fraction of target bases with >20 reads aligned (exome sequencing) or average number of reads aligned across all bases genome-wide (genome sequencing). Based on these metrics, we excluded from further analysis exome sequence data (from 151 individuals) with coverage ≤20x in >20% of the target bases and genome sequence data (from 68 individuals) with average coverage ≤5x.
Possible DNA contamination of sequence data was assessed using verifyBamID 15 , either by direct comparison of sequence and HumanOmni2.5 array genotypes (where available) or by indirect estimates of contamination based on HumanOmni2.5 array allele frequencies. DNA samples with estimated contamination >2% using either method were excluded from further analysis (data from 7 individuals in the exome sequencing dataset and 59 individuals in the genome sequencing dataset). Uncontaminated DNA sample swaps were also detected via comparison of sequence and array data and corrected prior to variant calling.
To identify single nucleotide variants (SNVs) from the whole-genome sequence data, we used two independent SNV calling pipelines: GotCloud 16 and the GATK UnifiedGenotyper 14 . We merged unfiltered SNV calls across the two call sets and then processed the merged site list through the SVM and VQSR filtering algorithms implemented by those pipelines. SNVs that failed both filtering algorithms were excluded from further analysis. To identify SNVs from the whole-exome sequence data, we used the GATK UnifiedGenotyper best-practices pipeline 14 .
To identify short insertions and deletions (indels) from the whole-genome sequence data, we called variants using the GATK UnifiedGenotyper best-practices pipeline. Because indels are known to have high false positive rates 17 , we applied more stringent criteria for indel QC than for SNV QC, excluding indels that failed either the SVM or VQSR filtering algorithms. To identify indels from the whole-exome sequence data, we used the GATK UnifiedGenotyper best-practices pipeline 14 .
To identify structural variants (SVs, or >100-bp deletions) from the whole-genome sequence data, we used GenomeSTRiP 18 . To increase sensitivity after initial discovery of SVs, we merged the discovered sites with deletions identified in 1,092 sequenced individuals from the 1000G Project 17 and then genotyped the merged site lists across the whole-genome sequenced individuals. After applying the default filtering implemented in GenomeSTRiP, pass-filtered sites variable in any of the individuals were identified as candidate variant sites. Among these candidate sites, we excluded variants in known immunoglobin loci to reduce the impact of possible cell-line artifacts. We did not call SVs from the whole-exome sequence data.
Integrated panel generation. We merged variants discovered from the three experimental platforms into one site list. For individuals who had data from each of the three platforms, we then calculated genotype likelihoods across all sites separately by platform: for the whole-genome sequence data, we used GotCloud; for the exome sequence data, we used the GATK UnifiedGenotyper; and for the HumanOmni2.5 data, we converted hard genotype calls into genotype likelihoods assuming a genotype error rate of 10 − 6 . If a site was not assayed by one of the three platforms, it was ignored in likelihood calculation for that platform.
We then calculated combined genotype likelihoods as the product of the genome, exome, and HumanOmni2.5 likelihoods, assuming independence across platforms. Following a strategy originally developed for the 1000G Phase 1 project 17 , we then phased the integrated likelihoods using Beagle 19 (with 10,000 SNVs per chunk and 1,000 overlapping SNVs between consecutive chunks) and refined phased genotypes using Thunder 20 as implemented in GotCloud (with 400 states).
Using the genotypes from the integrated panel, we performed principal component analysis (PCA) separately for each of the three variant types (SNVs, indels, SVs), using EPACTS on an LD-pruned (r 2 o0.20) set of MAF>0.01 autosomal variants (with variants in large high-LD regions 21,22 or with Hardy-Weinberg P o10 − 6 removed). Inspecting the first ten PCs for each variant type, we identified 43 outlier individuals based on PCs from SNVs and indels and 136 additional outliers based on PCs from SVs; these 179 individuals were excluded from further analysis. Additionally, 38 individuals with close relationships with other study individuals (estimated genome-wide identity-by-descent proportion of alleles shared >0.20) were excluded from further analysis.

WES (GoT2D+T2D-GENES Multiethnic) panel generation
Ascertainment of individuals. In addition to the individuals within the WGS panel, additional individuals, 10,242 of which were included in the final analysis, were chosen for whole-exome sequencing from 10 studies: the Jackson Heart Study (500 African-American cases, 526 matched controls), the Wake Forest School of Medicine Study (518 African-American cases, 530 matched controls), the Korea  Potential T1D or MODY cases were excluded via similar approaches as for the whole-genome sequencing experiment. Statistics of the individuals ultimately included in the association analysis are provided in Table 2.
As for the WGS panel, many individuals were measured for additional cardiometabolic phenotypes ( Table 6).
Exome sequencing. DNA samples were obtained and sequenced in the same manner as described for the exome sequencing component of the WGS panel.
Processing, QC, and variant calling. As for the exome sequence data within the WGS panel, sequence data for the WES panel were processed and aligned to the human genome (build hg19) using the Picard, BWA 12 , and GATK 13,14 software packages and best-practice pipelines. Genotype likelihoods were computed controlling for contamination. Hard calls (the GATK-called genotypes but set as missing at a genotype quality [GQ] o20 threshold 14 ) and dosages (the expected value of the genotype, defined as Pr (RX|data)+2Pr(XX|data), where R is the reference and X the alternative allele) were computed for each individual at each variant site. Hard calls were used only for quality control, while dosages were used in downstream association analyses. Multi-allelic SNVs and indels were dichotomized by collapsing alternate alleles into one category.
Individuals were excluded from analysis if they were outliers on one of multiple metrics: poor array genotype concordance (where available), high number of variant alleles or singletons, high or low allele balance (average proportion of non-reference alleles at heterozygous sites), or excess mean heterozygosity or ratio of heterozygous to homozygous genotypes. Within this reduced set of individuals, we then further excluded variants based on hard call rate ( o90% in any cohort), deviation from Hardy-Weinberg equilibrium (P o10 − 6 in any ancestry group), or differential call rate between T2D cases and controls (P o10 − 4 in any ancestry group).
The final WES panel contains genotypes for 13,008 individuals at 2.93M SNVs and 111.9 K indels. The set ultimately included in coding variant association analysis (after removal of individuals with close relatives or of uncertain ancestry) contains genotypes for 12,940 individuals at 2.89M SNVs and 110.2 K indels (Tables 4 and 5).
Assaying variants in larger sample sizes Imputation from the WGS panel. We carried out genotype imputation, using existing SNP array data, from the WGS panel into 44,414 individuals (11,645 T2D cases and 32,769 controls) from 13 studies participating in the DIAGRAM consortium. Each study performed quality control independently. A more detailed description of the analyzed individuals is available elsewhere 3 .
Exome array genotyping from the WES panel. We considered 28,305 T2D cases and 51,549 controls from 13 studies of European ancestry, each genotyped with the Illumina exome array. Studies independently called genotypes using the Illumina GenCall algorithm (http://www.illumina.com/ Documents/products/technotes/technote_gencall_data_analysis_software.pdf ) or Birdseed 23 . Individuals were excluded if they had a low call rate ( o99%), excess heterozygosity, high singleton counts, evidence of non-European ancestry, discrepancy between recorded and genotyped sex, or discordance with prior SNP array or genotyping platform fingerprint data (where available). Variants were excluded if they had a low call rate ( o99%), deviation from Hardy-Weinberg equilibrium (P o10 − 6 ), GenTrain score o0.6, cluster separation score o0.4, or a suspect intensity plot based on manual inspection. After quality control, missing genotypes were re-called using zCall 24 , and additional quality control was performed to exclude poorly genotyped individuals (call rate o 99% or excess heterozygosity) or variants (call rate o99%). A more detailed description of the analyzed individuals is available elsewhere 3 .

Association analysis
WGS panel single variant analysis. For each variant in the WGS panel, we tested for association between genotype and T2D in the 2,657 sequenced individuals. We used a logistic regression framework (assuming an additive genetic model) with the Firth bias-corrected likelihood ratio test 25,26 to test for significance. Tests were adjusted for sex, the first two PCs computed based on genotypes from the HumanOmni2.5M array, and an indicator function for observed temporal stratification based on sequencing date and center.
Analysis of imputed datasets. In each of the thirteen studies within which variants from the WGS panel were imputed, SNVs with minor allele count (MAC) ≥ 1 were tested for T2D association under an additive genetic model. Association tests were adjusted for study-specific covariates and performed using either the Firth, likelihood ratio, or score tests as implemented in EPACTS (https://genome.sph.umich. edu/wiki/EPACTS) or SNPTEST 27 . Residual population stratification for each study was accounted for using genomic control 28 . Association statistics were then combined across studies, using a fixed-effects sample-size weighted meta-analysis as implemented in METAL 29 .
WES panel single variant analysis. For each variant in the WES panel, we tested for association between genotype and T2D in the 12,940 sequenced individuals. We computed separate association statistics for each ancestry group using EMMAX 30 . Additionally, we performed association tests using the Wald statistic, adjusting for ethnic-specific principal components after exclusion of related individuals. For each test, we calculated genomic control inflation factors and corrected association summary statistics (P-values and standard errors) to account for residual population structure. We subsequently performed a fixed-effects meta-analysis of ancestry-specific association summary statistics for each variant using (i) a sample-size weighting of P-values from the EMMAX analysis and (ii) an inverse-variance weighting of effect size estimates from the Wald analysis. For the final results, P-values were taken from the EMMAX analysis, and effect size estimates from the Wald analysis.  Analysis of exome array datasets. In each study within which exome array genotyping was applied, variants were tested for association with T2D via both the EMMAX and Wald tests. For the Wald test, related individuals were excluded and statistics were adjusted for study-specific principal components. For each study, P-values and standard errors were corrected based on the calculated genomic control inflation factor. Variants were then combined via a fixed-effects meta-analysis. EMMAX P-values were combined via a sample-size weighted analysis, and Wald effect sizes were combined via an inverse-variant weighted analysis. For the final results, P-values were taken from the EMMAX analysis, and effect sizes were estimated from the Wald analysis.
Gene-level analysis. We first generated four variant lists ('masks') based on functional annotations and observed allele frequencies. Annotations were computed based on transcripts in ENSEMBL 66 (GRCh37.66) using CHAoS v0.6.3, SnpEFF v3.1 31 , and VEP v2.7 32 . We then identified variants predicted by at least one of the three algorithms in at least one mapped transcript to be protein-truncating ('for example, nonsense, frameshift, essential splice site), denoted PTVs, or other protein-altering (for example, missense, in-frame indel, non-essential splice site), denoted missense. We additionally used a previously described procedure 33 to identify subsets of missense variants bioinformatically predicted to be deleterious: those annotated as damaging by each of Polyphen2-HumDiv, PolyPhen2-HumVar, LRT, Mutation Taster, and SIFT were considered to meet 'strict' criteria, while those annotated as damaging by one of these algorithms was considered to meet 'broad' criteria. We then calculated the MAF of each variant based on the highest frequency across each of the five ancestry groups. We finally combined these annotations to produce four masks: the PTV-only mask included PTVs, the PTV+NS strict mask included variants in the PTV-only mask as well as those meeting 'strict' criteria for deleteriousness, the PTV +NS broad mask included variants in the PTV-only mask as well as those with MAF o 1% meeting 'broad' criteria for deleteriousness, and the PTV+missense mask included variants in the PTV+NS broad mask as well as those with MAF o1% annotated as missense.
We performed gene-level analysis using the MetaSKAT software package (v0.32) 34 , employing the SKAT v0.93 library to perform a SKAT-O 35 analysis within each ancestry group as well as across all ancestry groups via meta-analysis. Within each ancestry group, we assumed homogenous allele frequencies and genetic affects and adjusted for ethnic-specific axes of genetic variation after exclusion of 96 related individuals. For the meta-analysis, we used the MetaSKAT option to analyze genotype-level data, allowing for heterogeneity of allele frequencies and genetic effects between ancestry groups. All analyses were completed using the recommended ρ vector for SKAT-O: (0, 0.12, 0.22, 0.32, 0.52, 0.5, 1).

Code availability
All analyses were performed using publically available software packages, using versions and parameters as described above.

Data Records
Genotypes and phenotypes from the WGS and WES panels are available at the European Genomephenome Archive (EGA, Data Citation 1 and Data Citation 2) and the database of Genotypes and Phenotypes (dbGAP, Data Citation 3 to Data Citation 12).
The data in EGA are covered under a single data use agreement, which complies with all of the cohortspecific data use restrictions. While this does limit data access according to the criteria of the most restrictive cohort, it is the only mechanism through which the entire WES and WGS panels are available  to investigators. Additionally, the EGA contains data from one cohort that could not be released to a USbased repository. To download either the WGS or WES panel from the EGA, investigators must obtain approval from a data access committee (DAC, t2dgenes-got2d-dac@broadinstitute.org) to analyze data from all cohorts included in the study. The requester will receive an application packet that includes a project proposal document and a Data Transfer Agreement (DTA). The requester must then provide to the DAC a short description of their study, the proposed use of the data, an approval from the Institution's IRB, and a signed DTA. Assuming IRB approval and an executed DTA, the process for obtaining final approval from the DAC takes 4-6 weeks. Once approved, investigators can download either a single VCF file with genotypes from all individuals in the WGS panel (Data Citation 1) or a single VCF file with genotypes from all individuals in the WES panel (Data Citation 2). If investigators cannot obtain approval to analyze all cohorts in the WES panel (e.g., commercial uses) they can download cohort-specific data from dbGAP. Each cohort in dbGAP is subject to distinct data use restrictions, and investigators can obtain separate VCF files, as well as the raw sequence reads, for each of the cohorts.
The WGS and WES panels are accompanied by exclusion lists of variants and individuals (Data Citation 1 and Data Citation 2). The WGS VCF file contains data from the full set of 2,874 individuals and 28.45M variants that passed QC, with additional lists provided containing the 2,657 individuals and 17.69M variants included in association analysis. The WES VCF file contains the full set of 13,008 individuals that passed QC, with additional lists provided containing the 3.04M variants that passed QC (the VCF includes a small number of variants that failed QC) and the 12,940 samples and 2.93M variants included in association analysis. Additionally, the WES dataset includes lists of variants included in genelevel analysis for each of the four analyzed masks.
Five datasets of association statistics are also available for download. Association statistics for variants in the WGS panel are available from the whole-genome sequenced individuals or from those with imputed genotypes (Data Citation 1). Association statistics in the WES panel are available from the whole-exome sequenced individuals or from those genotyped on the exome array (Data Citation 2). Additionally, gene-level association statistics from the whole-exome sequenced individuals are available for each of the four variant masks (Data Citation 2).
A description of the datasets is available in Table 7. All association statistics are also available for browsing and searching via the public Type 2 Diabetes Knowledge Portal at www.type2diabetesgenetics. org. Through the portal, users can construct queries to find variants satisfying specified annotations and association thresholds, both across the WES and WGS analyses as well as other GWAS datasets. Users can also dynamically construct a set of variants within a gene and obtain a P-value from aggregate association analysis within the WES individuals.

Evaluation of variants in the WGS panel
We evaluated the variant sensitivity (fraction of true variant sites detected) of the WGS panel, based on the 2,538 individuals with data from all three experimental platforms (low-pass whole-genome sequencing, whole-exome sequencing, and HumanOmni2.5M array genotyping). To assess the sensitivity  of low-pass whole-genome sequencing alone, we computed the fraction of variants detected from wholeexome sequencing that were also detected by low-pass whole-genome sequencing. Sensitivity estimates were 99.8, 99.0, and 48.2% for common (MAF>5%), low-frequency (0.5%o MAF o5%), and rare (MAFo0.5%) SNVs, respectively, and >99.9, 93.8, and 17.9% for common, low-frequency, and rare short indels, respectively. We also assessed the coding SNV sensitivity of low-pass whole-genome sequence data combined with exome sequence data, based on the proportion of HumanOmni2.5 SNVs detected by either sequencing platform. Because HumanOmni2.5 SNVs are enriched for common variants, we calculated an averaged sensitivity at each allele count, weighted by the number of exomedetected variants given the allele count. Sensitivity estimates were 99.9, 99.7, and 83.9% for common, lowfrequency, and rare variants, respectively. These sensitivity estimates provide lower bounds on the sensitivity of the full WGS panel, which combines HumanOmni2.5 SNP array data as well as the two types of sequence data (Fig. 2a). We further evaluated the genotype accuracy of the WGS panel for each of the three classes of variant (SNVs, indels, and SVs). Across chromosome 20, concordance of low-pass whole-genome-sequencebased SNV genotypes with exome-sequence-based genotypes was 99.86%, with homozygous reference, heterozygous, and homozygous non-reference concordances of 99.97, 98.34, and 99.72%, respectively. Concordance between exome-sequence-based SNV genotypes and HumanOmni2.5 genotypes was 99.4%, with homozygous reference, heterozygous, and homozygous non-reference concordances of 99.97, 99.69, and 99.88%, respectively. For indels genotyped with both low-pass whole-genome-sequence data and exome-sequence data, concordance was 99.4%, with homozygous reference, heterozygous, and homozygous non-reference concordances of 99.8, 95.8, and 98.6%, respectively.
To evaluate the genotype accuracy of SVs detected from the low-pass whole-genome sequence data, we took advantage of the 181 individuals in our study who were previously included in the WTCCC array-CGH based structural variant detection experiment 36 . Taking the WTCCC data as a gold standard, we estimated genotype accuracy across 1,047 overlapping SVs (with reciprocal overlap >0.8) genome-wide. The overall genotype concordance was 99.8%, with homozygous reference, heterozygous, and homozygous non-reference concordances of 99.9, 99.6, and 99.7%, respectively.

Evaluation of variants in the WES panel
We assessed the overall sequencing quality of individuals in the WES panel by computing distributions of global statistics, stratified by reported ancestry (Fig. 2b). After quality control, the number of nonreference variants, mean heterozygosity, and average allele balance (fraction of non-reference reads at heterozygous sites) per individual approximately matched a Gaussian distribution within each ancestry. Concordance between genotypes from exome-sequence data and those from independent SNP arrays was above 99% for the vast majority of individuals, with non-reference concordance above 99.5% for individuals genotyped on the (highest-quality) OMNI array.
We also assessed bulk properties of indels within the WES panel. The length distribution of indels showed an excess of variants with lengths a multiple of three, as expected. Additionally, principal components computed from indels alone closely matched those computed from SNVs and indels together (Fig. 2c).

Evaluation of imputation from the WGS panel
We computed the mean imputation quality as measured by the average squared correlation between imputed genotypes and actual genotypes from leave-one-out cross-validation analysis. For variants of allele count ≈100 or above in the WGS panel (corresponding to a frequency of 1.8%), average r 2 values were in excess of 0.8 for Finnish individuals and in excess of 0.6 for British individuals (Fig. 3a).

Evaluation of exome array sensitivity
We assessed the overlap of variants present on the exome array with those observed in the WES panel. As the exome array primarily contains SNVs that are predicted to be protein altering, we focused on nonsense, essential splice site, and missense variants; only variants passing QC in both sequence and array data were included in the assessment. The fraction of variants in the WES panel on the exome array was highest for Europeans, at 81.6%, and lowest in African-Americans, at 49.0% (Fig. 3b).

Evaluation of association tests
We used the genetic power calculator (http://zzz.bwh.harvard.edu/gpc/) to estimate power to detect T2D association for each of the single variant analyses. All calculations assumed a T2D prevalence of 8%. Figure 4 shows power estimates under optimistic scenarios, in which a variant is present in the WGS panel (Fig. 4a), observed at equal frequency in all ancestries in the WES panel (Fig. 4b), imputed with high (r 2 = 0.8) accuracy (Fig. 4c), or included on the exome array (Fig. 4d). We also computed power for less optimistic scenarios 3 ; if, for example, a MAF 1% variant is present in only one ancestry, it must have an odds ratio of~3.5 to achieve significance of P = 10 − 4 , rather than an odds ratio of~1.8 were it to be present in all ancestries.  For gene-based tests in the WES panel, we used a simulated haplotype data set (http://cran.r-project. org/web/packages/SKAT/vignettes/SKAT.pdf) and estimated power as a function of (i) the phenotypic variance, under a liability scale, explained by additive genetic effects and (ii) the percentage of variants that were causal (50% or 100%). As for single-variant power calculations, we considered variants of constant frequency across all five ancestry groups, as well as variants specific to one ancestry group. These calculations suggested 3 that, even under optimistic scenarios, genes must explain >1% of genetic variance in order to achieve a moderately significant (P o10 − 3 ) association in the WES panel.
To ensure that association statistics were well calibrated, we computed quantile-quantile (QQ) plots comparing observed statistics to those expected under the null distribution. The vast majority of statistics matched the expected distribution (suggesting good calibration of the association tests) with a deviation from the null for common variant associations from the WGS panel, imputed genotypes, and exome array genotypes (suggesting power to detect known positive control T2D-associated non-coding common variants).

Usage Notes
The WGS and WES panels may be useful for simulation-based approaches that require individual-level genotypes and phenotypes. In this case the full list of 'QC+' variants (not merely those included in the T2D analysis) should be used, as the association analysis omitted very rare variants that might be useful in other settings. The WGS panel can also serve as a reference panel for genotype imputation, particularly in cases where an excess of haplotypes from T2D cases are required. Although more recent and larger efforts such as the Haplotype Reference Consortium 8 will provide greater imputation power for most use cases, the WGS panel is not restricted based on minor allele count and includes indels and SVs.  The most valuable data from the T2D-GENES and GoT2D studies are likely the catalogues of T2D association statistics for low-frequency and common variation. These statistics may prove useful for fine mapping or functional studies of T2D GWAS signals, in which enumeration of potential causal variants is required, or for 'reverse genetic' approaches, in which estimates of the phenotypic effects of variants with strong molecular effects are desired. For this usage, we advise investigators to query the T2D Knowledge Portal at www.type2diabetesgenetics.org as a first step, as its goal is to provide a simple and continuously updated means to query these and other association statistics. The portal is designed specifically for queries about individual variants or those within a single gene or genomic locus, as well as variant-or gene-level analyses for which investigators desire to adjust included variants, covariates, or individuals. Users should note that data from the T2D-GENES and GoT2D studies are included in, and thus should not be combined with data from, the Exome Aggregation Consortium (exac.broadinstitute.org).
Should investigators desire access to all association statistics, genome-wide, the files at EGA should be used (as the T2D Knowledge Portal does not support bulk download of association statistics). For single variant associations, a statistic in the largest available sample size should be used. Coding variant association statistics present in both the WES panel analysis and the exome chip analysis can be safely combined via meta-analysis, as can non-coding variant association statistics present in both the WGS panel analysis and the imputation-based analysis; statistics should not, however, be combined across the non-coding and coding variant analyses. Association results from the sequence or exome chip data should not need to be filtered, but it is advisable to filter results from the imputation-based analysis according to a threshold on imputation quality (e.g., r 2 >0.3). For gene-level analyses, investigators should first use the aggregate association statistics and then dissect results by examining variant-level statistics for each variant in the mask.
The pre-computed statistics should be sufficient for most investigators. Cases where recalculation of associations may be appropriate include (a) conditional analyses, such as variant association controlling for additional individual phenotypes or genotypes at different variants, (b) association analyses with phenotypes other than T2D, and (c) novel statistical tests. In these cases, usage of the tests and inclusion of the covariates described in the methods section is recommended, and only variants and individuals present in the final 'QC+' analysis lists should be included.