Main

Human genetics offers a powerful approach for better understanding and treating disease by identifying molecular alterations that are causally associated with physiological traits1. Common-variant array-based genome-wide association studies (GWAS) have associated thousands of genomic loci with hundreds of human traits2, and further analyses indicate that heritability of most complex traits is attributable to modest-effect common regulatory variants3. However, non-coding GWAS associations are challenging to assign to causal variants or genes4.

Protein-coding variants with strong effects on protein function or disease can offer molecular ‘probes’ into the pathological relevance of a gene5 and potentially establish a direct causal link6 between gene gain- or loss-of-function and disease risk7—especially when there is evidence of multiple independent variant associations (an ‘allelic series’) within a gene8. Several lines of evidence9 predict that strong-effect variants (allelic odds ratios > 2) will usually be rare (minor allele frequency (MAF) < 0.5%) and, in many cases, difficult to accurately study through current array-based GWAS and imputation strategies5. Whole-genome or whole-exome sequencing, by contrast, allows interrogation of the full spectrum of genetic variation.

Previous exome-sequencing studies have identified relatively few exome-wide significant rare-variant associations for complex diseases such as T2D10. This paucity of findings is in part due to the limited sample sizes of previous studies, the largest of which included less than 10,000 disease cases and fall short of the sample sizes that analytic9 and simulation-based calculations11 predict are needed to identify rare disease-associated variants under plausible disease models. To increase rare coding variant analysis power, we collected and analysed exome-sequencing data from 20,791 T2D cases and 24,440 controls—one of the largest analyses of exome-sequenced cases for T2D, specifically, and for any disease, more generally.

Genetic discovery from association analysis

Study participants (Supplementary Table 1) were drawn from five self-reported ancestries: (Hispanic/Latino (effective size (neff) = 14,442; 33.8%), European (neff = 10,517; 24.6%), African-American (neff = 5,959; 13.9%), East-Asian (neff = 6,010; 14.1%) and South-Asian (neff = 5,833; 13.6%)) and yielded equivalent statistical power to detect associations as a balanced study of around 42,800 individuals or a population-based study (assuming T2D prevalence of 8% and no ascertainment bias) of around 152,000 individuals. Power was improved compared to the previous largest T2D exome-sequencing study10 of 6,504 cases and 6,436 controls, increasing, for example, from 5% to 90% for a variant with MAF = 0.2% and odds ratio = 2.5 (Extended Data Fig. 1).

Exome sequencing to 40x mean depth, variant calling and quality control (Extended Data Fig. 2, Supplementary Methods, Supplementary Figs. 13 and Supplementary Table 2) produced a dataset with 6.33 million variants: 2.3% common (MAF > 5%), 4.2% low-frequency (0.5% < MAF < 5%) and 93.5% rare (MAF < 0.5%) (Supplementary Table 3). These include 2.26 million nonsynonymous variants and 871,000 insertions and deletions (indels), more than twice the number of variants that were analysed in a previous T2D exome-sequencing study10.

We first tested each variant, regardless of allele frequency, for T2D association (‘single-variant’ test; Methods and Extended Data Figs. 3, 4). Fifteen variants (in seven loci) exceeded exome-wide significance (P < 4.3 × 10−7 for coding variants12, P < 5 × 10−8 for synonymous or non-coding variants), including ten nonsynonymous variants (Fig. 1a and Extended Data Table 1). These 15 associations are a substantial increase over the single association that was reported in a previous T2D-exome sequencing study10 and illustrate again the value of multi-ancestry association analyses13—as only 9 out of 15 variants achieved P < 0.05 in European samples. However, only two variants were not previously reported by GWAS: a variant in SFI1 (rs145181683, Arg724Trp; Supplementary Fig. 4) that failed to replicate in an independent cohort (n = 4,522, P = 0.90; Methods) and a low-frequency (in Hispanic/Latino individuals; MAF = 0.89%) moderate-effect (odds ratio = 2.17, 95% confidence interval = 1.63–2.89) MC4R variant (rs79783591, Ile269Asn) that has previously been shown to decrease MC4R activity and to be associated with obesity and T2D in smaller studies14. Conditioning on body-mass index reduced but did not eliminate the MC4R Ile269Asn T2D association (P = 1.0 × 10−5).

Fig. 1: Exome-wide association analysis.
figure 1

a, Single-variant associations were calculated using the two-sided EMMAX test. Red line, P = 4.3 × 10−7. b, Gene-level P values, corrected for four analyses performed using the two-sided minimum P-value test. The most-significant genes are labelled. Red line, P = 6.5 × 10−7. c, SLC30A8 gene-level P values (left y axis, black line), calculated using a two-sided burden test after removing variants (in order of increasing single-variant P value) from the 1/5 1% mask (the strongest signal for SLC30A8). Dashed line, P = 0.05. Right y axis (blue line), estimated effect size (log10(odds ratio)). Blue shading, 95% confidence interval. Dotted line, effect size = 0. Single-variant n = 45,231 individuals. Gene-level n = 43,074 unrelated individuals.

Because single-variant analyses have limited power to detect rare-variant associations9, we next performed association tests for aggregations of variants within genes. Because numerous variant aggregation approaches (that is, ‘masks’) and gene-level tests are available, we developed a method (Methods, Extended Data Figs. 5, 6 and Supplementary Figs. 5, 6) to consolidate information across 14 analyses into four results per gene: burden9 and SKAT15 analyses, each of which were either summarized as the ‘minimum P value’ across masks or ‘weighted’ to estimate the effect of gene haploinsufficiency. We used an exome-wide gene-level significance threshold of P = 6.57 × 10−7 (Methods).

Using this strategy, gene-level associations were exome-wide significant for MC4R, SLC30A8 and PAM (Fig. 1b, Extended Data Table 2 and Supplementary Table 4), with variants from multiple ancestries contributing to each signal (Methods). All three genes lie within reported T2D GWAS loci and contain previously identified coding variant signals: the common variant Arg325Trp and 12 rare protective protein-truncating variants (PTVs) for SLC30A87,16, the low-frequency variants Asp563Gly and Ser539Trp for PAM10,17 and the low-frequency variant Ile269Asn for MC4R.

The associations in MC4R (combined MAF = 0.79%, minimum P = 2.7 × 10−10, odds ratio = 2.07, 95% confidence interval = 1.65–2.59) and PAM (combined MAF = 4.9%, weighted P = 2.2 × 109, odds ratio = 1.44, 95% confidence interval = 1.28–1.62) result largely from effects of the previously identified coding variants in these genes, although the MC4R signal remained nominally significant after removing Ile269Asn (P = 8.6 × 10−3; Supplementary Fig. 7) and the PAM signal remained nominally significant (P < 0.05) after removing the 35 strongest individually associated PAM variants (Supplementary Fig. 8). As illustrated by a recent study that identified a novel T2D risk mechanism through cellular characterization of PAM Asp563Gly and Ser539Trp18, variants identified in our study (uniquely from sequencing)6 could yield further insights into the T2D risk mechanism mediated by PAM.

In contrast to MC4R and PAM, the SLC30A8 signal (103 variants, combined MAF = 1.4%, weighted P = 1.3 × 10−8, odds ratio = 0.40, 95% confidence interval = 0.28–0.55) was not primarily driven by an individual variant (Arg325Trp (MAF > 1%) was not included in the gene-level analysis). The association was instead driven by 90 missense variants (weighted P = 3.9 × 10−7) and remained nominally significant (P < 0.05) even when we removed the 32 strongest individually associated SLC30A8 variants (Fig. 1c and Supplementary Fig. 9). Although SLC30A8 was first implicated in T2D over a decade ago16, the disease-associated molecular mechanism(s) through which SLC30A8 acts remain poorly understood19—in part because the common risk-increasing allele Arg325Trp and the rare risk-decreasing PTVs were both initially thought to decrease protein activity7,19. The protective allelic series from our analysis argues that decreased T2D risk is the typical effect of SLC30A8 missense variation—that is, it is not unique to haploinsufficiency—and provides many additional alleles that can be characterized to gain mechanistic insights.

To evaluate association evidence for genes other than MC4R, PAM and SLC30A8, we assessed the 50 most-significant gene-level associations from our study in two independent exome-sequencing datasets: 12,467 European or African-American individuals (3,062 T2D cases) from the CHARGE discovery sequencing project20 (Supplementary Table 5; 50 genes available) and 49,199 European individuals (12,973 T2D cases) from the Geisinger Health System (Supplementary Table 6; 44 genes available). In a meta-analysis of the three studies (Methods and Supplementary Table 7), MC4R (P = 6.9 × 10−14), PAM (P = 3.0 × 10−9) and SLC30A8 (P = 3.3 × 10−8) each became more significant. In addition, one gene, UBE2NL (P = 5.6 × 10−7)—which has few prior links to T2D or other complex traits—newly achieved exome-wide significance (http://www.type2diabetesgenetics.org/). All aspects of this association passed quality control (Methods and Supplementary Table 8), although further replication will be important to establish UBE2NL as a novel T2D-relevant gene.

More broadly, we observed an excess of directionally consistent associations (both odds ratio > 1 or both odds ratio < 1) between the original and replication analyses (31 out of 46 in CHARGE, one-sided binomial P = 0.013; 23 out of 40 in the Geisinger Health System, P = 0.21; overall P = 0.011; Supplementary Table 7), suggesting that several more of our top gene-level signals will reach exome-wide significance in future studies.

Further insights from gene-level analyses

Even if a gene-level association does not achieve exome-wide significance, it might still be of use to prioritize a gene as relevant to T2D8 or predict whether loss or gain of protein function increases disease risk7. To investigate potential insights that could be obtained by sub-exome-wide significant gene-level associations, we analysed 16 gene sets that were connected to T2D based on a variety of sources of evidence (for example, genes that contained diabetes-associated Mendelian variants, T2D drug targets21 or genes that have been implicated in diabetes-related phenotypes in mouse models22; Methods and Supplementary Table 9).

First, for each gene set, we investigated whether the genes within the set had more significant gene-level associations than expected by chance (Methods). In total, 12 out of 16 gene sets achieved P < 0.05 set-level associations (Fig. 2a–e and Supplementary Fig. 10), including T2D drug targets (P = 2.1 × 10−3), genes previously reported in mouse models of non-insulin-dependent diabetes (NIDD; P = 5.2 × 10−3) or impaired glucose tolerance (P = 7.2 × 10−6) and genes that contained common likely causal coding-variant T2D associations6 (P = 8.8 × 10−3 after conditioning on the common variants nearby these genes). Additionally, as previously described10, we observed a significant set-level association (P = 1.2 × 10−3) for genes implicated in maturity onset diabetes of the young (MODY; Fig. 2a, Supplementary Table 10), with nominal associations in four genes including PDX1 (weighted P = 1.7 × 10−4, odds ratio = 3.45, 95% confidence interval = 1.78–6.71, 65 variants). Rare variants in genes associated with MODY also demonstrated aggregate association with lower body-mass index (minimum P = 5.7 × 10−3) and lower fasting insulin (minimum P = 0.028), consistent with the known predominant variant risk mechanism of reduced insulin secretion in MODY23. Most gene set signals were driven by multiple genes in the set (Supplementary Table 11) and—compared with previous studies focused on PTVs24—consisted of substantial contributions from missense variants. Indeed, set-level P values from PTVs alone were >0.05 for almost all gene sets (Supplementary Fig. 11).

Fig. 2: Gene set analysis.
figure 2

ae, Rank percentiles (1 = highest) for gene-level associations (compared to matched genes) within 11 monogenic diabetes genes (548 matched genes) (a), 8 T2D drug targets (400 matched genes) (b), 31 genes linked to NIDD in mice (1,499 matched genes) (c), 323 genes linked to impaired glucose tolerance in mice (10,043 matched genes) (d) and 11 genes with common likely causal coding variants (537 matched genes) (e). P values are from a one-sided Wilcoxon rank-sum test between each gene set and comparison set. Labels indicate minimum, 25th percentile, median, 75th percentile and maximum. f, Estimated odds ratios, from the weighted burden test of the 5/5 mask, for 8 T2D drug targets (red, agonists; blue, inhibitors). Data are mean ± s.e. of log(odds ratio) from the burden test. n = 43,071 unrelated individuals.

Collectively, these results suggest that association strength at the gene level can be used as a potential metric to prioritize candidate genes relevant to T2D. For example, the set of 40 genes within T2D GWAS loci with gene-level P < 0.05 had a significant excess of protein–protein interactions among them (Methods and Supplementary Table 12), suggesting that this set may be enriched for ‘effector genes’ that mediate T2D GWAS associations6. Fully evaluating the relevance to T2D of these and other candidate genes will require further experimental work4.

In addition to prioritizing genes that are potentially relevant to T2D, we assessed whether gene-level analysis could help to predict whether gene inactivation increases or decreases T2D risk, as this is of high interest for the development of therapeutics8. We compared the odds ratios that were estimated from a gene-level weighted burden analysis to directional relationships that have been previously reported (Methods). Seven out of eight T2D drug targets showed concordance between genetic and therapeutic directions of effect (three out of four inhibitor targets had an odds ratio < 1, four out of four agonist targets had an odds ratio > 1; one-sided binomial P = 0.035; Fig. 2f). The only exception was KCNJ11 (odds ratio = 1.59, inhibited by sulfonylureas), for which the gene-level signal was driven by a known25 activating missense mutation (His172Arg); an analysis without this variant predicted the correct (odds ratio < 1) directional relationship. This finding is consistent with the known reciprocal roles of KCNJ11 in both diabetes and persistent hyperinsulinaemic hypoglycaemia of infancy.

Concordances between gene-level estimates of odds ratios and knockout effects in mice were more equivocal (for example, 7 out of 11 diabetes-associated genes had an odds ratio > 1, binomial P = 0.27; 137 out of 240 genes associated with increased circulating glucose had an odds ratio > 1, P = 0.016; Supplementary Fig. 12). The lower concordances for these gene sets, despite a trend towards lower-than-expected gene-level P values within them (Supplementary Fig. 10), highlight the known limitations of animal models26, which can be highly dependent on model conditions27, to predict human physiology. Candidate genes with significant but directionally unexpected gene-level associations may provide valuable insights into seemingly promising preclinical results: for example, the protective gene-level signal for ATM in our analysis (burden test of PTVs odds ratio = 0.50, P = 0.003) contradicts previous expectations—based on insulin resistance and impaired glucose tolerance in Atm knockout mice28—that ATM loss-of-function should increase T2D risk. Evidence is even less favourable that ATM haploinsufficiency strongly increases T2D risk, rejecting an odds ratio > 2 at P = 1.3 × 10−8. These observations could be relevant in the ongoing study of whether ATM has a role in metformin response29 or whether ATM activators are considered able to treat cardiovascular disease30.

Comparison of rare and common variant associations

Despite early arguments that rare-variant studies would considerably advance our understanding of complex diseases5, most genetic discoveries continue to be provided by studies of common variants, which can be studied in much larger sample sizes through array-based genotyping and imputation31. Previous quantitative analyses have similarly emphasized the main contribution of common variants to T2D heritability6,10, but they have lacked the sequencing data that are needed to fully evaluate the value added by rare variants (that is, direct sequencing in addition to array-based genotyping and imputation) to discover disease-associated loci, explain disease heritability and elucidate allelic series.

To compare discoveries that were possible from sequencing and array-based studies, we collected genome-wide array data within the same individuals that we sequenced (available for 34,529 (76.3%) individuals; 18,233 cases), imputed variants using best-practice reference panels32,33 and conducted a single-variant association analysis (‘imputed GWAS’; Methods and Supplementary Table 13). Out of 10 exome-wide significant nonsynonymous single-variant associations from the sequence analysis, 8 were detected in the imputed GWAS analysis (PAX4 Arg192His and MC4R Ile269Asn were not imputable), together with genome-wide significant non-coding variant associations in 14 additional loci (Fig. 3a and Supplementary Table 14). All 10 variants with significant single-variant sequence associations were also present on the Illumina Exome Array6. These results demonstrate the limited power of sequencing to detect single-variant associations beyond array-based genotyping and imputation, even before considering the much larger sample sizes enabled by the substantially lower cost of array-based genotyping.

Fig. 3: Comparison of exome-sequencing to array-based GWAS.
figure 3

a, Single-variant associations, calculated by two-sided Firth logistic regression, from an array-based imputed GWAS of the subset (n = 34,529) of samples in the exome-sequencing analysis for which array data were available. Labels and axes as described in Fig. 1a. b, The observed liability variance explained (LVE) by the top 19 exome-sequencing gene-level associations (red) and the top 19 imputed GWAS single-variant associations (maximum of 1 per 250 kb; blue) and their ratio (black). Signals ranked by liability variance explained. c, Gene rank percentiles from exome-sequencing gene-level analysis (x axis) and a previous multi-ancestry T2D GWAS13 (y axis). Genes are from the mouse NIDD gene set (Fig. 2c).

We next compared the contributions to T2D heritability of the strongest (common) single-variant associations from the imputed GWAS to those of the strongest (mostly rare-variant) gene-level associations from the sequencing analysis (Methods). The three exome-wide significant gene-level signals explain an estimated 0.11% (MC4R), 0.092% (PAM) and 0.072% (SLC30A8) of T2D genetic variance, only 10–20% of the variance explained by the three strongest independent common-variant associations in the imputed GWAS (TCF7L2, 0.89%; KCNQ1, 0.81%; CDC123, 0.35%; Fig. 3b). More broadly, fitting a previous exponential model of heritability34 to our data (Methods) estimated that the top 100 gene-level signals associated with T2D explained only 1.96% of genetic variance within our study. These results argue against a large contribution to T2D heritability from even the strongest gene-level signals, even after accounting for potential sources of downward bias in our calculations (see Methods).

We finally assessed whether an array-based GWAS could have detected the many potential allelic series that we observed from direct sequencing. Among the variants that contributed to the exome-wide significant gene-level associations in SLC30A8, MC4R and PAM, we estimate that 95.3% of variants are not imputable (r2 > 0.4; Methods) in the 1000 Genomes multi-ancestry reference panel32, 74.6% of those in Europeans are not imputable in the larger European-focused Haplotype Reference Consortium panel10 and 90.2% (79.7% of European variants) are absent from the Illumina Exome Array. Additionally, gene set associations (using gene ‘scores’; see Methods) from the imputed GWAS showed suggestive associations (four gene sets achieved P < 0.05, nine achieved P < 0.1; Supplementary Fig. 13) but were weaker than gene set associations from the sequencing analysis. Some of these gene set associations are detectable in larger array-based studies: analysis of a 110,000-sample multi-ancestry GWAS13 produced P < 0.05 for 12 out of 16 gene sets that we studied (Supplementary Fig. 14); however, the genes (and corresponding variants) that are responsible for the array-based gene set associations were mostly different from those responsible for the sequence-based associations, as the two methods often produced uncorrelated rank orderings of genes within gene sets (for example, r = −0.11, P = 0.57 for the mouse NIDD gene set; Fig. 3c).

Collectively, these results demonstrate the complementarity of array-based GWAS and exome sequencing, with the former favouring locus discovery and the latter enabling full enumeration of potentially informative alleles.

Inferences from nominally significant associations

The T2D drug targets analysed here illustrate the opportunities and challenges of using current exome-sequencing datasets in translational research. Rare-variant gene-level associations are significant across these targets as a set (Fig. 2b) and predict the correct T2D directional relationship for all but one gene (Fig. 2f). However, to detect—at exome-wide significance—the effect sizes estimated from our study with 80% power would require 75,000–185,000 sequenced cases (150,000–370,000 exomes in a balanced study, or 600,000–1,275,000 exomes from a population with a prevalence of T2D of 8%; Fig. 4a and Methods).

Fig. 4: Decision support from exome-sequencing data.
figure 4

a, Estimated power, as a function of future sample size (x axis), to detect T2D gene-level associations (two-sided test at P = 6.25 × 10−7) with aggregate frequency and odds ratios equal to those estimated from our analysis of 8 T2D drug targets (Fig. 2f). b, A proposed workflow for using exome-sequencing data in gene characterization. Depending on the prior belief in the disease relevance of a gene, the cost of experimental characterization and the benefit of gene validation, a decision to conduct the experiment could be informed by the posterior probability of the disease relevance of the gene, as estimated from exome-sequencing association statistics (available through http://www.type2diabetesgenetics.org/).

As a consequence, many of the modest associations (for example, P = 0.05) in current samples may point to clinically or therapeutically relevant variants or genes (Supplementary Fig. 15). The false-positive rate for these associations is expected to be greater than the false-positive rate for exome-wide significant associations35 and be further influenced by imperfect calibration of association test statistics. If this false-positive rate can be quantified using independent ‘truth’ data36, however, then a modest association signal could help to justify further experimentation on a gene based on the likelihood that it is a true association, the cost of the experiment and the benefit of success37 (Fig. 4b).

We developed and evaluated a method to quantify the false-positive association rate for nonsynonymous variants in our dataset by using independent data, modelling assumptions and prior data to map single-variant P values to estimated posterior probabilities of true, causal associations (PPAs) (Methods and Extended Data Fig. 7). Model parameters in the middle of the range that we explored (Methods and Extended Data Fig. 8) predict that 1.5% (95% confidence interval = 0.74–2.2%) of nonsynonymous variants that achieve P < 0.05 in our study are truly, causally associated with T2D, increasing to 3.6% (95% confidence interval = 1.4–5.9%) for P < 0.005 and 9.7% (95% confidence interval = 3.9–15.0%) for P < 5 × 10−4 (Supplementary Fig. 16). In this model, 541 (95% confidence interval = 270–810) of the 36,604 nonsynonymous variants with P < 0.05 in our study represent true, causal associations.

We next applied this method to variants within a curated set of 94 T2D GWAS loci (Methods), which might be expected to show further enrichment of true associations. Our model predicted that nonsynonymous variants within these loci had even higher PPAs: 2.0% (95% confidence interval = 0.048–4.0%) of such variants overall, 8.1% (3.6–12.4%) with P < 0.05 in our study and 17.2% (7.7–24.1%) with P < 0.005 were estimated to represent true, causal T2D associations. Of particular note are variants in these loci that not only achieve nominal significance (P < 0.05) in our analysis but also have moderate-to-large estimated effects on T2D risk (Supplementary Tables 15, 16), as we predict that a substantial number of these variants (for example, 76 (95% confidence interval = 29–117) out of 746 with estimated odds ratio > 2 and 50 (95% confidence interval = 19–77) out of 503 with estimated odds ratio > 3) show true, causal associations.

Outside of GWAS loci, many genes are suspected to be involved in T2D because of prior evidence from non-genetic sources (for example, animal studies22 or because of implication in related disorders23). To evaluate variants in such genes, we extended our PPA estimation approach to incorporate gene prior probabilities (or ‘priors’)38 (Methods and Extended Data Fig. 7d) and applied it to two sets of genes.

First, using a prior of 100% for genes associated with MODY—thus assuming that all genes implicated in MODY are relevant to T2D—our model predicts 24 variants (combined MAF = 1.1%) to have PPA ≥ 40% (Supplementary Table 17). Nine have estimated odds ratio > 3 in our study; as none of these were previously reported to be pathogenic MODY variants, they are therefore novel rare-variant candidates for use in the prediction of T2D risk. On the other hand, these results show that, once false-positive rates are empirically estimated rather than assumed, nominally significant variants (P = 0.05) in genes associated with MODY are still, in absolute terms, more likely to be false-positive rather than true associations39.

Second, as an example of a gene prior that was derived objectively (rather than subjectively), we used a mixture model approach40 to estimate the proportion of non-null associations across the mouse NIDD gene set (Methods), leading to a prior of approximately 23% for genes of which knockout causes NIDD in mice. Our model with this prior (Supplementary Table 18) predicts nonsynonymous variants that achieved P < 0.05 to have PPAs of 9.9% (PPAs of 24.6% for P < 0.005). In particular, we predict several nonsynonymous variants in MADD and NOS3 to have PPA ≥ 14% (Supplementary Table 19), suggesting links between variation in these genes and T2D based on combined evidence from human genetic studies and mouse models41,42.

Although these PPA calculations have limitations (Methods), they present a framework to use suggestive genetic signals to support cost–benefit estimates of ‘go/no-go’ decisions43 in the language of decision theory37 (Fig. 4b). To enable this strategy, we have made our exome-sequencing association results publically available through the AMP T2D Knowledge Portal (http://www.type2diabetesgenetics.org/), which supports queries of precomputed associations and further enables dynamic recomputations of associations with custom covariates and sample- and/or variant-filtering criteria.

Discussion

Our results provide a nuanced description of rare variation and its association with T2D, which might also apply to other complex diseases. Rare-variant gene-level signals are likely to be distributed across numerous genes; however, the vast majority of signals individually explain vanishingly small amounts of T2D heritability: more than one million samples may be required for rare-variant signals in validated therapeutic targets to become significant exome-wide. Even among the four genes that reached exome-wide significance in our analysis, two (MC4R and PAM) do not include unusually strong rare-variant associations but rather typically modest rare-variant associations that are boosted from nominal to exome-wide significance by low-frequency variants.

Thus, for biological discovery in many complex traits, such as T2D, exome sequencing and array-based GWAS seem complementary: locus discovery and fine mapping are achieved most efficiently using larger array-based GWAS, whereas rare coding variant allelic series—that could aid experimental gene characterization44 or provide confidence in disease-gene identification—are best discoverable through sequencing. For personalized medicine, exome sequencing may produce some rare variants with sufficient effect sizes (Supplementary Tables 12, 17) to provide viable contributions to the prediction of genetic risk; however, these are sufficiently rare to be best viewed as complements to rather than replacements for GWAS-derived polygenic risk scores45. Whole-genome sequencing might soon become sufficiently cost-effective to subsume both array-based GWAS and exome sequencing; even now, it is essential to expand imputation reference panels to power higher-resolution GWAS across all major ethnicities.

Our results suggest that, for now, maximizing the utility of exome sequencing will require drawing insights from associations that do not (yet) reach exome-wide significance. To help to interpret these suggestive associations, we present a principled and empirically calibrated Bayesian approach (Fig. 4, Extended Data Fig. 7 and Supplementary Table 18) to estimate the association probability for any variant in our dataset, highlighting its use to interpret variants in known disease genes and prioritize genes from animal model studies for further investigation. Results and customized analyses from our study can be accessed through a public web portal (http://www.type2diabetesgenetics.org/), advancing the use of exome-sequencing data across many branches of biomedical research.

Methods

A full description of the methods used in this study is available as Supplementary Methods.

Data reporting

The experiments were not randomized and the investigators were not blinded to allocation during experiments and outcome assessment.

Sample selection

We drew samples for exome sequencing from six consortia, most of which consisted of multiple studies and are described fully in Supplementary Table 1. T2D case status was determined according to study-specific criteria described in full in in Supplementary Table 1 and the Supplementary Methods. All individuals provided informed consent and all samples were approved for use by their institution’s institutional review board or ethics committee, as previously reported10,46,47,48. Samples that were newly sequenced at The Broad Institute as part of T2D-GENES, SIGMA and ProDiGY are covered under Partners Human Research Committee protocol 2017P000445/PHS ‘Diabetes Genetics and Related Traits’.

Data generation

The details of data generation, variant calling, quality control and variant annotation are described in full in the Supplementary Methods. In brief, for each consortium, sequencing data were aggregated (if previously available) or newly generated (if not) and then processed through a standard variant calling pipeline. We then measured samples and variants according to several metrics indicative of sequencing quality, excluding those that were outliers relative to the global distribution (Supplementary Fig. 1, Supplementary Table 2). These exclusions produced a ‘clean’ dataset of 49,484 samples and 7.02 million variants.

Following initial sample and variant quality control, we performed additional rounds of sample exclusion from association analysis (Extended Data Fig. 2). We also excluded the 3,510 childhood diabetes cases from the SEARCH and TODAY studies based on an analysis that suggested their lack of matched controls would induce artefacts in gene-level association analyses (Supplementary Fig. 17). These exclusions produced an ‘analysis’ dataset of 45,231 individuals and 6.33 million variants. A power analysis of this dataset is presented in the Supplementary Methods.

After these three rounds of sample exclusions, we estimated—within each ancestry—pairwise identity-by-descent values, genetic relatedness matrices and principal components for use in downstream association analyses. We used the identity-by-descent values to generate lists of unrelated individuals within each ancestry, excluding 2,157 individuals to produce an ‘unrelated analysis’ set of 43,090 individuals (19,828 cases and 23,262 controls) and 6.29 million non-monomorphic variants. We used this set of individuals and variants for single-variant and gene-level tests (described below) that required an unrelated set of individuals.

We annotated variants with the ENSEMBL Variant Effect Predictor49 (VEP, version 87). We produced both transcript-level annotations for each variant as well as a ‘best guess’ gene-level annotation using the –flag-pick-allele option (with ranked criteria described in the Supplementary Methods). We used the VEP LofTee (https://github.com/konradjk/loftee) and dbNSFP (version 3.2)50 plugins to generate additional bioinformatics predictions of variant deleteriousness; from the dbNSFP plugin, we took annotations from 15 different bioinformatics algorithms (listed in Extended Data Fig. 5) and then added annotations from the mCAP51 algorithm. As these annotations were not transcript-specific, we assigned them to all transcripts for the purpose of downstream analysis.

Although we incorporated both transcript-level and gene-level annotations into gene-level analyses (see below), all single-variant analyses reported in the manuscript or figures are annotated using the ‘best guess’ annotation for each variant.

Single-variant association analysis in sequencing data

To perform single-variant association analyses, we first stratified samples by cohort of origin and sequencing technology (with some exceptions described in the Supplementary Methods), yielding 25 distinct sample subgroups (Extended Data Fig. 3). For each subgroup, we performed additional variant quality control beyond that used for the ‘clean’ dataset, excluding variants according to subgroup-specific criteria described in Extended Data Fig. 3; in general, these criteria were strict—particularly for multiallelic variants and X-chromosome variants. We verified that these filters led to a well-calibrated final analysis through inspection of quantile–quantile plots within and across ancestries (Extended Data Fig. 4).

For each of the 25 sample subgroups, we then conducted two single-variant association analyses: one of all (including related) samples using the (two-sided) EMMAX test52 and one of unrelated samples using the (two-sided) Firth logistic regression test53. Both analyses included covariates for sequencing technology, and the Firth analysis included covariates for principal components of genetic ancestry (those among the first 10 that showed P < 0.05 association with T2D).

We then conducted a 25-group fixed-effect inverse-variance weighted meta-analysis for each of the Firth and EMMAX tests, using METAL54. We used EMMAX results for association P values and Firth results for effect size estimates.

Additional analysis of rs145181683

To assess whether the rs145181683 variant in SFI1 (P = 3.2 × 10−8 in the exome-sequencing analysis) represented a true novel association, we obtained association statistics from 4,522 Latinos55) who did not overlap with the current study. On the basis of the odds ratio (1.19) estimated in our analysis and the MAF (12.7%) in the replication sample, the power was 91% to achieve P < 0.05 under a one-sided association test. The observed evidence (P = 0.90, odds ratio = 1.00) did not support rs145181683 as a true T2D association. Further investigation of this lack of replication evidence suggested that, although the association from our sequence analysis is unlikely to be a technical artefact (genotyping quality was high), it could possibly be a proxy for a different (Native American-specific) non-coding causal variant (full details are available in the Supplementary Methods). Further fine-mapping and replication efforts will be necessary to test this hypothesis.

Gene-level analysis

For each gene, following previous studies10,56,57, we separately tested seven different ‘masks’ of variants grouped by similar predicted severity (defined in Extended Data Fig. 5). For each gene and each mask, we created up to three groupings of alleles, corresponding to different transcript sets of the gene; for many genes, two or more of these allele groupings were identical.

Before running gene-level tests, we performed additional quality control on sample genotypes. For each of the 25 sample subgroups (the same as used for single-variant analyses), we identified variants that failed subgroup-specific quality control criteria (shown in Extended Data Fig. 5) and set genotypes for these variants in all individuals in the subgroup as ‘missing’.

We conducted two gene-level association tests: a burden test, which assumes all analysed variants within a gene are of the same effect, and SKAT15, which allows variability in variant effect size (and direction); each of these tests is two-sided. We performed each test across all unrelated individuals with 10 principal components of genetic ancestry, sample subgroup and sequencing technology as covariates. As this ‘mega-analysis’ strategy was different from the meta-analysis strategy that we used for single-variant analyses, as a quality control exercise we conducted a single-variant mega-analysis and found that its results showed broad correlation with those from the original meta-analysis (Supplementary Fig. 18).

We then developed two methods to consolidate the 2 × 7 = 14 P values produced for each gene (described in full in Extended Data Fig. 5, Supplementary Methods and Supplementary Figs. 5, 6). First, we corrected the smallest P value for each gene according to the effective number of independent masks tested for the gene (variable, but on average 3.6), based on the gene-specific correlation of variants across masks58 (referred to as the minimum P-value test; Supplementary Fig. 19). Second, we tested all nonsynonymous variants (that is, missense, splice site and protein-truncating mutations), but weighted each variant according to its estimated probability of causing gene inactivation9 (referred to as the weighted test, which essentially assessed the effect of gene haploinsufficiency from combined analysis of protein-truncating and missense variants; Supplementary Fig. 6). We verified that these two consolidation methods were well-calibrated (Extended Data Fig. 6) and broadly consistent yet distinct: across the 10 most significantly associated genes, P values were nominally significant using both methods for 8 genes but varied by 1–3 orders of magnitude (Extended Data Table 2).

Because each gene mask could in fact represent up to three sets of alleles (owing to the transcript-specific annotation strategy that we used), for each of the four analyses multiple P values were possible for some genes. To produce a single gene-level P value for each of the four analyses, we thus collapsed (for each gene) the set of P values across transcript sets into a single gene-level P value using the minimum P-value test.

We used a conservative Bonferroni-corrected gene-level exome-wide significance threshold of P = 0.05/(2 tests × 2 consolidation methods × 19,020 genes) = 6.57 × 10−7. For each gene referenced in the manuscript, we report the P value and odds ratio from the analysis that achieved the lowest P value for the gene.

Gene-level analysis near T2D GWAS signals

In principle, a nearby common-variant association could lead to over- or underestimation of the strength of a gene-level association59. To assess whether differential patterns of rare variation across common-variant haplotypes could significantly affect our gene-level results, we conducted two analyses (described in the Supplementary Methods) and found no evidence that confounding from common-variant haplotypes was primarily responsible for the associations that were observed in our gene-level analyses.

Further exploration of significant gene-level associations

For our exome-wide significant gene-level associations (MC4R, PAM and SLC30A8), we conducted additional gene-level analyses to dissect the aggregate signals that were observed. First, we performed tests by progressively removing alleles in order of lowest single-variant analysis P value, in order to understand the (minimum) number of alleles that contributed statistically to the aggregate signal. Second, we performed tests conditional on each allele in the sequence (that is, calculating separate models with each individual allele as a covariate), and we then compared the resulting P values to the full gene-level P value, in order to assess the contribution of each allele individually to the signal. Finally, for MC4R, we conducted an analysis with an added sample covariate for body-mass index and found that it, as shown previously60,61, reduces the significance of both the Ile269Asn single-variant signal (P = 1.0 × 10−5) and the gene-level signal not attributable to Ile269Asn (P = 0.035).

To evaluate which ancestries contributed variants to MC4R, SLC30A8, and PAM, we calculated the proportion of variants in each signal unique to an ancestry and also compared the significance and direction of effect of each signal across ancestries. Across the three signals, 68.4% (287 out of 419) of variants in total were unique to one ancestry (63.9% for MC4R, 67.0% for SLC30A8 and 71.6% for PAM). Each signal had a direction of effect that was consistent across all five ancestries and each signal achieved P < 0.05 in at least two ancestries (MC4R in East-Asians and Hispanics; SLC30A8 in all ancestries other than African-Americans; and PAM in Europeans, South-Asians and Hispanics).

Analysis of exomes from the Geisinger Health System

We obtained gene-level association results that were previously computed from an analysis of 49,199 individuals (12,973 T2D cases and 36,226 controls) from the Geisinger Health System (GHS). Association statistics were available for 44 out of the 50 genes with the strongest gene-level associations from our study. A power analysis of the GHS replication analysis is available in the Supplementary Methods.

GHS sequencing data were processed and analysed as previously described24, and variants were grouped into four (nested) masks (roughly corresponding to the LofTee, 5/5, 1/5 1% and 0/5 1% masks; more details are available in the Supplementary Methods). For each mask, association results were computed using two-sided logistic regression under an additive burden model (with phenotype regressed on the number of variants carried by each individual) with age, age2 and sex as covariates. To produce a single GHS P value for each gene, we applied the minimum P-value procedure across the four mask-level results.

Analysis of exomes from the CHARGE consortium

We collaborated with the CHARGE consortium to analyse the 50 genes with the strongest gene-level associations from our study in 12,467 individuals (3,062 T2D cases and 9,405 controls) from their previously described study62,63. A power analysis of the CHARGE replication analysis is available in the Supplementary Methods.

Variants in the CHARGE exomes were annotated and grouped into seven masks using the same procedure as for the original exome-sequencing analysis. Burden and SKAT association tests were then performed in the Analysis Commons64 using a two-sided logistic mixed model65 assuming an additive genetic model and adjusted for age, sex, study, race and kinship. To produce a single CHARGE P value for each gene, we applied the minimum P-value procedure across the seven mask-level results, as for the GHS analysis.

Meta-analysis with CHARGE and GHS

We conducted a meta-analysis among our original burden analysis and those of CHARGE and GHS. For each gene, we selected the mask that achieved the lowest P value in our original analysis and conducted a two-sided sample-size weighted meta-analysis with the results from CHARGE and GHS for the same mask (or an analogous mask as defined in the Supplementary Methods).

Investigation of the UBE2NL association

We investigated the novel association that was found in the gene-level meta-analysis (UBE2NL, meta-analysis P = 5.6 × 10−7) in more detail. The UBE2NL burden signal was due to five PTVs in the original analysis (observed in 29 cases and 1 control; all of which had high (>45×) sequencing coverage; Supplementary Table 8) and was replicated at P = 0.02 in CHARGE; UBE2NL results were not available in GHS. As UBE2NL lies on the X chromosome, we conducted a sex-stratified analysis of the original samples and observed independent associations in both men (P = 5.7 × 10−4) and women (P = 1.6 × 10−3). UBE2NL does not lie near any known GWAS associations (http://www.type2diabetesgenetics.org/) and has few available references66,67,68, suggesting that it may be a novel T2D-relevant gene, although further replication will be important to establish its association.

Evaluation of directional consistency between exome-sequencing, CHARGE and GHS analyses

We examined the concordance of direction of effect size estimates (that is, both odds ratios of >1 or <1) between burden tests from our original exome-sequencing analysis and those from CHARGE and GHS. For the 46 genes advanced for replication with burden P < 0.05 for at least one mask (that is, ignoring those with evidence for association only under the SKAT model), we compared the direction of effect estimated for the mask with lowest P-value mask to that estimated for the same (or analogous) mask in the GHS or CHARGE analysis. We then conducted a one-sided exact binomial test to assess whether the fraction of results with consistent direction of effects was significantly greater than expected by chance.

Gene set analysis in sequencing data

We curated 16 sets of candidate T2D-relevant genes, defined in Supplementary Table 9 with criteria as specified in the Supplementary Methods. For each gene set, we constructed sets of matched genes with similar numbers and frequencies of variants within them (details are provided in the Supplementary Methods). A sensitivity analysis of this matching strategy is presented in the Supplementary Methods.

To conduct a gene set analysis, we then combined the genes in the gene set with the matched genes. Within the combined list of genes, we ranked genes using the P values observed for the minimum P-value burden test. We then used a one-side Wilcoxon rank-sum test to assess whether genes in the gene set had significantly higher ranks than the comparison genes.

Use of gene-level associations to predict effector genes

To assess whether gene-level associations from exome sequencing—which are composed mostly of rare variants independent of any GWAS associations—could prioritize potential effector genes within known T2D GWAS loci, we first assessed whether predicted effector genes (based on common-variant associations) were also enriched for rare coding variant associations. Our analysis (described in full in the Supplementary Methods) indicated that effector genes predicted from common coding variant associations do show significant enrichments (P = 8.8 × 10−3), but effector genes predicted from transcript-level associations do not (P = 0.72).

We then curated a list of 94 T2D GWAS loci, and 595 genes that were within 250 kb of any T2D GWAS index variant, from a 2016 T2D genetics review69 and observed 40 with a P < 0.05 gene-level signal (Supplementary Table 12), greater than the 595 × 0.05 = 29.75 expected by chance (P = 0.038). Only three (SLC30A8, PAM and HNF1A) were from the list that we curated of 11 genes with causal common coding variants6. We found that these 40 genes were significantly more enriched for protein interactions (P = 0.03; observed mean = 11.4, expected mean = 4.5) than the 184 genes implicated based on proximity to the index SNP (P = 0.64; observed mean = 21.1, expected mean = 21.9), although evaluation of the biological candidacy of these genes will ultimately require in-depth functional studies70. Rare coding variants could therefore, in principle, complement common-variant fine-mapping71,72 and experimental data4,70 to help to interpret T2D GWAS associations; however, our results indicate that much larger sample sizes and/or orthogonal experimental data will be required to clearly implicate specific effector genes. A full description of this analysis is included in the Supplementary Methods.

Use of gene-level associations to predict direction of effect

To assess whether gene-level association analyses of predicted deleterious variants could be used to predict therapeutic direction of effect, we compared odds ratios estimated from a modified weighted burden test procedure (described in the Supplementary Methods) to those expected for T2D drug targets (assuming agonist targets to have true odds ratios > 1 and inhibitors to have true odds ratios < 1). For a similar comparison to expectations for mouse gene knockouts, we used the relationship between mouse phenotype and human phenotype specified in the Supplementary Methods. Genes present in two gene sets with opposite expected direction of effects were excluded from this analysis.

Collection and analysis of SNP array data

To compare discoveries from our exome-sequencing analyses to discoveries possible from common-variant GWAS of the same samples, we aggregated all available SNP array data for the exome-sequenced samples (18,233 cases and 17,679 controls; Supplementary Table 13). After sample and variant quality control (described in the Supplementary Methods), we imputed variants from the 1000 Genomes Phase 332 (1000G) and Haplotype Reference Consortium33 (HRC) reference panels using the Michigan Imputation Server73. We used 1000G-based imputation for all association analyses and HRC-based imputation to assess the number of exome-sequence variants imputable from the largest available European reference panel (details available in the Supplementary Methods).

After imputation, we performed sample and variant quality control, as well as two-sided association tests, analogous to the exome-sequence single-variant analyses. In contrast to the exome-sequencing analyses, a quantile–quantile plot suggested that the associations from the EMMAX test were not well calibrated, and we therefore used only the Firth test (that is, for both P values and odds ratios) in the imputed GWAS analysis.

To conduct gene set analysis with the imputed GWAS data, we first used the method implemented in MAGENTA74 to calculate gene scores from the imputed GWAS single-variant association results. Following the same protocol as for gene set analysis from the exome-sequencing results, we then conducted a one-sided Wilcoxon rank-sum test to compare the gene scores to those of matched comparison genes. We followed the same approach for the gene set analysis that we conducted in a larger, previously published13 GWAS.

LVE calculations

To calculate LVEs, we used a previously presented formula75 (equations are available in the Supplementary Methods) to calculate the LVE of a variant with three genotypes (AA, Aa and aa) and corresponding relative risks (1, RR1 and RR2). When presenting the strongest LVE values for the imputed GWAS analysis, we only considered variants that were genotyped in at least 10,000 individuals to avoid potential artefacts that result from a spurious association in a small-sample subgroup. For gene-level LVE calculations, we used the variant mask with lowest P value to calculate LVEs. We also conducted a sensitivity analysis to bound the extent to which our gene-level LVE estimates might be biased downwards due to their inclusion of benign alleles; this analysis (described in full in the Supplementary Methods) produced upper bounds of gene-level LVEs that were at most twofold higher than the point estimates.

Prediction of LVE explained by the top 100 and top 1,000 gene-level associations

To forecast the LVE that will be explained once 100 (or 1,000) significant T2D gene-level associations are detected, we applied a previously suggested model34 in which the LVE of a gene is related to its rank in the overall gene-level P-value distribution. Specifically, the model is LVEn = ean + b where LVEn is the LVE of the gene with nth lowest gene-level P value. We fitted this model using linear regression to the top 50 genes in our analysis (Supplementary Fig. 20), yielding estimates of a = −0.044 and b = −7.07. We then calculated the LVE of the top 100 (or 1,000) genes by summing the actual LVE of the top three signals (which achieved exome-wide significance in our analysis) with the LVE predicted by the model for genes ranked 4–100 (or 4–1,000).

Estimated power to detect gene-level associations with T2D drug targets

To estimate the power of future studies to detect gene-level associations in genes with effect sizes similar to those for established T2D drug targets, we used aggregate allele frequencies and odds ratios estimated from our gene-level analysis and an assumed prevalence of K = 0.08 to calculate a proxy for true population frequencies and relative risks. For each gene, we used odds ratios and frequencies from the variant mask that yielded the strongest gene-level association. Because, on average, these drug targets had five effective tests per mask, we used an exome-wide significance threshold of α = 1.25 × 10−7 for power calculations. We calculated power as previously described76.

The ranges given in the main text (75,000–185,000 disease cases) represent the numbers from the power calculations for INSR (the drug target with the highest observed effect size) and IGF1R (the drug target with the lowest observed effect size other than KCNJ11 and ABCC8). We excluded KCNJ11 and ABCC8 from this reported range, given that a mixture of risk-increasing and risk-decreasing variants in these genes probably diluted their burden signals. We did not account for uncertainty in estimated odds ratios or aggregate variant frequency in these calculations, as no genes had 95% confidence intervals that that did not overlap odds ratio = 1.

Interpretation of suggestive associations

We quantified the PPA for nonsynonymous variants observed in our dataset as a function of association strength measured by single-variant P values. We define a true association as a variant that, when studied in larger sample sizes, will eventually achieve statistical significance owing to a true odds ratio ≠ 1. We distinguish true associations from causal associations: causally associated variants are the subset of truly associated variants in which the variant itself is causal for the increase in disease risk, as opposed to being truly associated due to linkage disequilibrium (LD) with a different causally associated variant (that is, an ‘LD proxy’). An overview of the method that we developed for PPA calculations is provided in Extended Data Fig. 7, and a full description of the method is included in the Supplementary Methods. Here, we outline the steps in the approach.

First, for various single-variant P-value thresholds in the exome-sequencing analysis, we calculated the fraction of variants that reached this threshold with directions of effect concordant with those of an independent exome array study10. For example, 61.3% of nonsynonymous variants within T2D GWAS loci that reached P < 0.05 in the exome-sequencing analysis had concordant directions of effect with the independent study, a fraction that decreased (as expected) for higher P-value thresholds (for example, 49.4% at P > 0.5) or when only variants outside of T2D GWAS loci were analysed (51.9% at P < 0.05).

Second, we derived an equation to convert the fraction of concordant associations to an estimated proportion of true associations. This value provides a PPA estimate, as a function of P value, for an arbitrary variant in the set initially used to calculate direction of effect concordances. We computed separate mappings for arbitrary nonsynonymous variants (using all exome-wide nonsynonymous variants) and one for nonsynonymous variants within GWAS loci (using only nonsynonymous variants within the 94 T2D GWAS loci). We note that the mapping produced from our analysis applies only to the results from the current study: because other studies have different sample sizes and may apply different statistical tests, the mapping would need to be recomputed to interpret the associations of other studies using the same method.

Third, we converted PPA estimates to estimates of the posterior probability of causal associations (PPAc). This conversion requires estimates of the fraction of coding variant associations that are causal (as opposed to LD proxies). We explored several values for this parameter, as described in the Supplementary Methods and shown in Extended Data Fig. 8.

Fourth, we extended PPA estimates to incorporate gene-specific priors by mapping posterior odds of causal association (POc) to a Bayes factor for causal association (BFc). This calculation requires a set of training variants with a known prior. For this training set, we use nonsynonymous variants within GWAS loci and modelling assumptions for their prior. Details of this model are described in the Supplementary Methods and a sensitivity analysis of its assumptions is shown in Extended Data Fig. 8.

Finally, as a preliminary estimate of a principled prior likelihood for genes in the mouse NIDD gene set, we estimated the proportion of non-null associations across all genes in the set. To use true prior data (rather than associations from the current study), we calculated gene-level P values for each gene in the set using the MAGENTA74 algorithm applied to a recent transethnic T2D GWAS13. We then used a previously developed approach40,77 that models the distribution of observed P values as a mixture of uniform (representing the null distribution) and beta (representing the non-null distribution) distributions, yielding a prior value of 23.2%.

Our PPAc calculations currently have several limitations. They apply only to single-variant associations and not (yet) to gene-level associations; extending them to apply to gene-level associations would avoid the possibility of conflicting results among variants within a gene but would require larger-scale gene-level replication data than that we had available in the current analysis. Additional work will also be needed to generate data and develop methods to estimate objective rather than subjective gene priors (researchers can often overestimate evidence of disease relevance for genes in which they have invested considerable effort), to reduce dependence of our conclusions on modelling assumptions (Extended Data Fig. 8) and to explore the extent to which the large number of variant associations that we predict from our data localize to specific gene or variant functional annotations78.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this paper.