Low-frequency variants in HMGA1 are not associated with type 2 diabetes risk.

It has recently been suggested that the low-frequency c.136-14_136-13insC variant in high-mobility group A1 (HMGA1) may strongly contribute to insulin resistance and type 2 diabetes risk. In our study, we attempted to confirm that HMGA1 is a novel type 2 diabetes locus in French Caucasians. The gene was sequenced in 368 type 2 diabetic case subjects with a family history of type 2 diabetes and 372 normoglycemic control subjects without a family history of type 2 diabetes. None of the 41 genetic variations identified were associated with type 2 diabetes. The lack of association between the c.136-14_136-13insC variant and type 2 diabetes was confirmed in an independent French group of 4,538 case subjects and 4,015 control subjects and in a large meta-analysis of 16,605 case subjects and 46,179 control subjects. Finally, this variant had no effects on metabolic traits and was not involved in variations of HMGA1 and insulin receptor (INSR) expressions. The c.136-14_136-13insC variant was not associated with type 2 diabetes in individuals of European descent. Our study emphasizes the need to analyze a large number of subjects to reliably assess the association of low-frequency variants with the disease.

Insulin resistance and pancreatic β-cell dysfunction are the two physiologic hallmarks of type 2 diabetes, and the expectation was that the majority of susceptibility loci identified in genome-wide association studies (GWASs) would map to either of these diabetes-related phenotypes. Yet, the majority of these loci can be directly mapped to the pancreatic β-cell or are of unknown biologic consequence (1). Few loci were associated with insulin resistance (2,3). Furthermore, the known common variants account for <10% of the overall estimated genetic contribution to type 2 diabetes predisposition (3). There is obviously a “dark matter” that remains to be discovered (4). A growing number of researchers are turning to rare genetic changes with strong effects as important contributors (5–8). Because whole-genome sequencing requires computational power and remains costly for large case-control studies, candidate gene sequencing may be an effective approach to identify rare and low-frequency variants associated with common diseases.

In 2005, Foti et al. (9) observed that lack of HMGA1 causes insulin resistance and diabetes in humans. This gene encodes a nonhistone protein that binds on AT-rich segments in the minor groove of DNA and regulates gene transcription (10,11). HMGA1 is a positive regulator of both insulin receptor (INSR) and insulin-like growth factor-I receptor (IGF-IR) genes and is induced by the Wnt/β-catenin pathway, which is known to play a key role in type 2 diabetes risk (12–16). In humans, two rare Mendelian mutations were identified only in diabetic patients and associated with impaired HMGA1 expression: a hemizygous deletion of HMGA1 (two carriers from the same family) and a heterozygous single-nucleotide deletion in the 3′ untranslated region of HMGA1 (c.*369del) (9). Chiefari et al. (17) recently observed that a low-frequency variant in HMGA1 (c.136–14_136–13insC) was strongly associated with type 2 diabetes risk and decreased expression of HMGA1 and INSR genes. All these elements suggest that rare variants located in this gene can be strongly associated with type 2 diabetes through their effects on insulin resistance. However, extensive analysis of the HMGA1 gene in additional populations is necessary to validate previous findings and identify novel variants potentially associated with type 2 diabetes risk (18).

RESEARCH DESIGN AND METHODS

Study populations.

For the sequencing of the HMGA1 gene, a screening group of 368 French type 2 diabetic case subjects and 372 French normoglycemic control subjects was selected. The case subjects were composed of those with at least two first-degree relatives with type 2 diabetes, collected by the UMR CNRS 8199 laboratory. French control subjects were composed of those from the general population study Data from an Epidemiological Prospective Study on the Insulin Resistance Syndrome (DESIR), who were selected because they remained normoglycemic after 9 years of follow-up and had no family history of diabetes (19).

The c.136–14_136–13insC variant was also analyzed in a replication group of 4,538 French type 2 diabetic case subjects and 4,015 French normoglycemic control subjects (Table 1). Type 2 diabetic case subjects were composed of those from the UMR CNRS 8199 laboratory (n = 377), the Non-Insulin-Dependent Diabetes, Hypertension, Microalbuminuria or Proteinuria, Cardiovascular Events, and Ramipril (DIABHYCAR) study (n = 2,095), the DESIR study (n = 295), the Biological Atlas of Severe Obesity (ABOS) study (n = 117), and the Corbeil-Essonnes Hospital (n = 1,654) (20). Control subjects were composed of subjects from the UMR CNRS 8199 laboratory (n = 1,306), the DESIR study (n = 2,291), the ABOS study (n = 207), and the Fleurbaix-Laventie Ville Santé study (n = 211). The DIABHYCAR study design and results have been reported (20,21). Participants were selected on the basis of type 2 diabetes, treatment with oral antidiabetic agents on enrollment, high urinary albumin concentration, age ≥50 years, and serum creatinine concentration ≤150 μmol/L. French patients were selected by their general practitioners, and ≥98% were Caucasian (20). The ABOS study was first designed to analyze the association between glycemic or ponderal status and tissue gene expression and has been extended to provide longitudinal follow-up and samplings of a cohort of 900 morbid obese patients subjected to bariatric (weight loss) surgery. The Fleurbaix-Laventie Ville Santé study is a longitudinal epidemiologic study on the determinants of weight gain (22).

TABLE 1

Association between the c.136–14_136–13insC variant and type 2 diabetes in French individuals

All type 2 diabetic case subjects and normoglycemic control subjects of this study were defined according to the 2003 American Diabetes Association criteria (type 2 diabetes: fasting plasma glucose ≥7.0 mmol/L or treated for diabetes; normoglycemic: fasting glucose <5.6 mmol/L). Only French Caucasians were analyzed. Informed consent was obtained from each subject, and our study protocol was approved by local ethics committees.

GWAS data were made available from the DIAGRAM consortium (3). All sample characteristics and statistics remain the same as previously described (3). However, any overlapping samples from the Diabetes Gene Discovery Group (DGDG) study (French samples) were excluded. Total sample size in the seven-study meta-analysis remained at 7,451 type 2 diabetic case subjects and 38,290 control subjects. Correlations between polymorphisms and the c.136–14_136–13insC variant were estimated using pilot release data from the CEU population of the 1000 Genomes project (23).

Sequencing.

Overall, 4.5 kb of the HMGA1 gene were sequenced, including the entire promoter region (−591 base pairs [bp] before the transcription start site 1 to +980 bp after), 8 exons, and 16 intron-exon junctions and untranslated regions. HMGA1 gene structure was based on RefSeq transcripts (National Center for Biotechnology Information, Bethesda, MD) (NM_002131, NM_145899, NM_145901, NM_145902, NM_145903, NM_145905), genomic sequence (GenBank NT_007592.15). Primers were designed using Primer-BLAST (National Center for Biotechnology Information) and were validated not to amplify the pseudogene regions using In-Silico PCR (University of California, Santa Cruz, CA) (Supplementary Table 1). PCR amplifications were purified using Agencourt AMPure (Agencourt Bioscience Corporation, Beverly, MA) on Biomex NX (Beckman Coulter, Fullerton, CA) after validation of a single band of excepted size on agarose gel. Purified PCR amplification products were then used for sequence reaction with Big Dye Terminator Cycle Sequencing Ready Reaction Kit 3.1 (Applied Biosystems, Foster City, CA). Sequence products were purified with Agencourt CleanSEQ (Agencourt Bioscience Corporation) on Biomek NX read on an ABI Prism 3730XL DNA Sequencer and assembled and analyzed with Variant Reporter, Seqscape (Applied Biosystems), and Sequencher software (Gene Codes Corporation, Ann Arbor, MI). All variants were validated by resequencing of forward and reverse strands.

Genotyping.

Genotyping was performed using the High Resolution Melting (HRM) method using MeltDoctor HRM Master Mix (Applied Biosystems) on a Lightcycler 480 (Roche, Indianapolis, IN). HRM curve analysis was performed on LightCycler 480 software, and positive signals were confirmed by sequencing. There was a concordance rate of >99% between genotyping and sequencing results for the 368 type 2 diabetic case subjects and 372 normoglycemic control subjects of the screening study.

Intermediate metabolic traits.

BMI is the weight in kilograms divided by the square of the height in meters. The homeostasis model assessment (HOMA)-2–insulin resistance and HOMA2-B indices were calculated using the HOMA calculator based on fasting glucose and fasting insulin levels in subjects in a steady-state situation (fasting plasma glucose: 3–25 mmol/L; fasting plasma insulin: 20–400 pmol/L; available at http://www.dtu.ox.ac.uk/homacalculator). HbA1c was only available in a subset of the DESIR and UMR CNRS 8199 samples. The insulinogenic index (IGI) was calculated on the basis of the oral glucose tolerance test (OGTT) measures as IGI = (30 min plasma insulin − fasting plasma insulin)/30 min plasma glucose, where plasma insulin is in pmol/L and plasma glucose is in mmol/L (24). The insulin sensitivity index (ISI) was calculated on the basis of OGTT measures as ISI = 10,000/√(fasting plasma glucose × fasting plasma insulin × mean OGTT glucose × mean OGTT insulin), where plasma insulin and OGTT insulin are in mU/L and plasma glucose and OGTT glucose are in mg/dL (25). The disposition index (DI) was calculated on the basis of OGTT measures in subjects with a fasting plasma glucose ≥70 mg/dL as DI = ISI × 100 × 30 min plasma insulin × (30 min plasma glucose × [30 min plasma glucose – 70 mg/dL]), where plasma insulin is in mU/L and plasma glucose is in mg/dL (26). Indices based on OGTT measures were only available for a subset of UMR CNRS 8199 samples. Waist-to-hip ratio was calculated as waist circumference in centimeters divided by hip circumference in centimeters and was only available in DESIR and UMR CNRS 8199 samples. Only normoglycemic subjects were analyzed to avoid possible modifying effects of type 2 diabetes and treatments on metabolic traits.

Expression study.

The expression study was conducted in 149 Swedish obesity-discordant siblings (included in the Swedish Obese Subjects [SOS] Sib Pair study) using Merlin under a dominant genetic model (27,28). Average family size was 4.34, and median age (first to third quartiles) was 44 years (36–62). Gene expression data were available from 347 siblings, and peripheral blood was available from all subjects. Gene expression was measured in subcutaneous adipose tissue using the Affymetrix Human U133PLUS2.0 microarrays for HMGA1 (probe-sets 206074_s_at and 210457_x_at) and INSR (probe-sets 213792_s_at, 207851_s_at, 226450_at, 226212_s_at, and 226216_at). A Genetic Power Calculator was used to estimate the power of our study in identifying association between the c.136–14_136–13insC variant and HMGA1 and INSR gene expression levels (29). To this aim, we assessed the heritability of the HMGA1 and INSR transcription levels on the Swedish sample and used observations from the study by Chiefari et al. (17) to estimate both the minor allele frequency (MAF) of the c.136–14_136–13insC variant and its effects on the gene expression levels. The association study was carried out by fitting a linear regression model to each probe-set within a variance component framework to account for correlation between gene expression levels within each family (30).

Statistical analyses.

Given the low allele frequency of the c.136–14_136–13insC variant, only the dominant genetic model of inheritance was considered. In our screening analysis, the statistical power of our study design was sufficient to detect low-frequency variants with strong effect sizes (type I error rate of 0.05) (Supplementary Fig. 1). MAFs between case subjects and control subjects were compared using the χ2 test and Monte Carlo simulations (n = 2,000). Associations of the c.136–14_136–13insC variant with type 2 diabetes and age at diagnosis were analyzed using logistic regression and linear regression models, respectively. No associations between the c.136–14_136–13insC variant and the confounding factors (i.e., age at examination, sex, and BMI) were detected using adjusted linear and logistic regression models (P = 0.39, P = 0.42, P = 0.21, respectively). Furthermore, these confounding factors had no influence on the association between the c.136–14_136–13insC variant and type 2 diabetes (P = 0.81, P = 0.87, P = 0.59, respectively). No between-cohort heterogeneity in the genotype distributions of the c.136–14_136–13insC variant and no deviation from Hardy-Weinberg equilibrium were detected (Supplementary Table 2). Linear regression models were used to compare quantitative intermediate traits between normoglycemic and type 2 diabetic subjects. All statistical models were adjusted for age, sex, and BMI (if appropriate). When necessary, metabolic traits were log-transformed to satisfy the assumption of normality. Heterogeneity between studies was assessed by the Woolf’s test and quantified by the Higgins statistic (I2) (31,32). Because of strong heterogeneity between previous and current findings, we applied a random-effects meta-analysis to estimate summary odds ratios (ORs) 1, 3, and 4 (33). However, a fixed-effects meta-analysis was performed to estimate the summary OR 2 given that no heterogeneity was observed between summary OR 1 and DIAGRAM data. Statistical power was assessed using Quanto (version 1.2.4). All P values were two-sided. The Bonferroni correction was calculated to address the problem of multiple comparisons. SPSS (v. 14.0.2; SPSS, Chicago, IL) and R statistics (v. 2.10.1) were used for general statistics.

RESULTS

Screening of the HMGA1 gene.

The HMGA1 gene was sequenced in 368 type 2 diabetic case subjects having at least two first-degree relatives with type 2 diabetes and 372 normoglycemic control subjects prospectively followed during 9 years and without family history of diabetes. Overall, 41 genetic variants were detected, including four variants (c.136–14_136–13insC, rs41269028, rs2780219, and rs17038619) previously identified by Chiefari et al. (17) (Fig. 1 and Supplementary Table 3). None of them were differentially distributed between normoglycemic control subjects and type 2 diabetic case subjects. However, the c.136–14_136–13insC C allele tended to be more frequent in control subjects compared with type 2 diabetic case subjects (MAF 2.82 vs. 1.49%, P = 0.09) (Table 1, Supplementary Table 3, and Supplementary Fig. 2). Given that this low-frequency variant was supposed to be strongly associated with type 2 diabetes, we verified this hypothesis by analyzing larger groups of case and control subjects.

FIG. 1.

Genetic variants identified in the HMGA1 gene by sequencing. A total of 41 genetic variants were identified when sequencing the HMGA1 gene in 368 type 2 diabetic subjects and 372 control subjects from our screening group of samples. Coding regions are represented by hatched boxes.

Association of the c.136–14_136–13insC variant and type 2 diabetes.

An additional 4,538 type 2 diabetic case subjects and 4,015 normoglycemic control subjects were genotyped for the c.136–14_136–13insC variant. Genotype distributions and ORs are reported in Table 1. No association between the c.136–14_136–13insC variant and type 2 diabetes risk was observed (screening + replication groups: OR 1.07, CI 0.84–1.37, P = 0.57). Furthermore, no association between the c.136–14_136.13insC variant and the age at diagnosis was observed (P = 0.61). Given that Chiefari et al. (17) analyzed two independent groups of Italian control subjects and one group of Italian type 2 diabetic patients, we performed two meta-analyses on the association of the c.136–14_136–13insC variant and type 2 diabetes risk (dominant model of inheritance) including either control group 1 (n = 2,544, personal interviews to determine the lack of family history of type 2 diabetes and related conditions) or control group 2 (n = 784, self-reported medical questionnaire).

By including the Italian control group 1, we detected a strong heterogeneity between studies (Woolf test: χ2 = 67.08, df = 2, P = 2.66 × 10−15; Higgins test: I2 = 97% [94%; 98.5%]). In a total of 9,154 case subjects and 7,889 control subjects of European descent (Italian, U.S., and French), no association between the c.136–14_136–13insC variant and type 2 diabetes was detected in our meta-analysis (summary OR 1 = 2.95, 95% CI 0.76–11.54, P = 0.12) (Fig. 2). By using CEU data generated from the 1000 Genomes pilot project, we also identified 10 polymorphisms that were perfectly correlated (r2 = 1) with the c.136–14_136–13insC variant (Supplementary Table 4). Pairwise correlations (r2) between these 10 polymorphisms were high in HapMap 3 (0.77–1.00). Disease association statistics were available for these polymorphisms from a recently published DIAGRAM consortium meta-analysis of 8,130 type 2 diabetic case subjects and 38,987 control subjects (3). After exclusion of samples already tested in this present study, the total sample size was 7,451 case subjects and 38,290 control subjects. Disease association test statistics ranged from P = 0.31 to P = 0.89 for these 10 polymorphisms. The rare T allele (HapMap CEU MAF = 2.7%) of the polymorphism closest to the gene (rs9394200: distance ∼5 kb), which tags the risk-increasing insertion (OR 0.94, 95% CI 0.82–1.07, P = 0.32), was selected to assess the significance of this signal in further meta-analysis. In a total of 16,605 case subjects and 46,179 control subjects of European descent (Italian, U.S., French, DIAGRAM consortium), no association with type 2 diabetes was detected (summary OR 2 = 0.95, 95% CI 0.83–1.08, P = 0.44) (Fig. 2).

FIG. 2.

Meta-analysis on the association between the c.136–14_136–13insC variant (or correlated polymorphism) and type 2 diabetes. na, not applicable; summary OR 1, Italian + U.S. + French data; summary OR 2, Italian + U.S. + French + DIAGRAM data.

By including the Italian control group 2 (MAF 1.66%) with another meta-analysis comprising 9,154 case subjects and 6,129 control subjects of European descent (Italian, U.S., and French), a between-study heterogeneity was also detected (Woolf test: χ2 = 8.29, df = 2, P = 0.02; Higgins test: I2 = 76% [20.6%; 92.7%]). Again, no association was observed between the c.136–14_136–13insC variant and type 2 diabetes (summary OR 3 = 1.49, 95% CI 0.98–2.26, P = 0.06). Similar conclusions were reached when this meta-analysis was extended to the DIAGRAM data (Italian, U.S., French, DIAGRAM consortium) in 16,605 case subjects and 44,419 control subjects of European origin (summary OR 4 = 1.13, 95% CI 0.73–1.77, P = 0.58).

Effects on metabolic traits in normoglycemic control subjects.

Normoglycemic control subjects were then analyzed for type 2 diabetes–related quantitative traits (Table 2). Carrying a c.136–14_136–13insC C allele was not found to influence fasting glucose, fasting insulin, HOMA2 indices, IGI, ISI, DI, HbA1c, waist-to-hip ratio, and BMI.

TABLE 2

Effects of the c.136–14_136–13insC variant on metabolic traits in normoglycemic control subjects

Effects on HMGA1 and INSR expressions.

The HMGA1 and INSR expressions were assessed in the subcutaneous adipose tissue of 347 siblings from the SOS Sib Pair study. The expression levels of both INSR and HMGA1 in the subcutaneous adipose tissue were above average (74 and 71% of the most expressed gene, respectively). Maximum likelihood estimate of the minor allele frequency for the c.136–14_136–13insC variant was 3.30% (47 heterozygous carriers). For the power calculation, we assumed MAF of 1% for c.136–14_136–13insC. Heritabilities, estimated through Merlin in the Swedish sample and averaged among all alternate transcripts from the same gene, were 70 and 27% for HMGA1 and INSR, respectively. Effect sizes (r2) were estimated as d2/(d2 + 4) from Cohen’s d (34), calculated using the gene transcript mean levels and CIs as observed between carriers and noncarriers of the c.136–14_136–13insC variant in the study by Chiefari et al. (17). Average r2 values between type 2 diabetic carriers and both type 2 diabetic and healthy noncarriers were 0.66 for HMGA1 and 0.74 for INSR. We observed that in the study by Chiefari et al. (17), the type 2 diabetic noncarriers showed slightly lower expression levels of HMGA1 and INSR than the healthy noncarriers control subjects (P < 0.05). By assuming a conservative lower QTL effect size of 30% in our sample (where no type 2 diabetic carriers were included), we observed a statistical power >80% at α = 0.001 to identify a dominant effect for the c.136–14_136–13insC on the HMGA1 and INSR gene expression levels in our sample. The association study showed that the c.136–14_136–13insC variant was not associated with changes in INSR and HMGA1 expressions in the Swedish sample (P > 0.12) (Supplementary Table 5).

DISCUSSION

It has been suggested that the candidate gene approach may be a practical and logical alternative to GWAS of complex genetic traits (35). In fact, the study of genes with functional polymorphisms and coding for proteins involved in disease-related pathways may increase statistical efficiency by minimizing multiple comparisons. Furthermore, the GWAS approach is still unable to reliably assess the association between rare variants and complex diseases such as type 2 diabetes (6,35–37). However, candidate gene studies have brought a majority of false-positive results in the past, and a stringent study design is necessary to reach a statistical power that allows valid conclusions.

Low-frequency variants with strong effects may contribute to type 2 diabetes risk through their impact on insulin resistance (9,17,38–41). Although the HMGA1 gene was not part of the GWAS signals, there was some evidence for its role in insulin resistance and type 2 diabetes development (3,9,13,17,42–44).

However, our data suggest that low-frequency variants, including c.136–14_136–13insC, are not associated with type 2 diabetes in the French population. Furthermore, the c.136–14_136–13insC variant had no significant effects on intermediate quantitative traits. Given the low statistical power of our screening step, we cannot exclude that other rare variants increasing type 2 diabetes susceptibility may exist in the HMGA1 gene. Therefore, we emphasize the need to sequence many more individuals to reliably identify rare variants associated with common diseases.

Our data are not in line with those recently reported by Chiefari et al. (17), who observed that the c.136–14_136–13insC variant was strongly associated with type 2 diabetes risk in populations of European descent (OR 1.64–15.77). In our study (4,906 type 2 diabetic subjects vs. 4,387 normoglycemic subjects), we had a statistical power of 80% to detect a dominant OR of 1.29 for a MAF of 2.61% (type I error = 0.05). Furthermore, large meta-analyses in Europeans (including Italian control group 1 or 2) detected no association between the c.136–14_136–13insC variant (or correlated polymorphism) and type 2 diabetes. Of note, strong between-study heterogeneity was observed, mainly driven by the low allele frequencies in both groups of Italian control subjects (Fig. 2). This heterogeneity may be due to different inclusion criteria in the Italian, U.S., and French studies (questionnaire, personal interview, biological measures). In the largest Italian control group (control group 1), Chiefari et al. (17) conducted personal interviews to select those without a family history of type 2 diabetes or related conditions to reduce the number of people with insulin resistance. Apart from the fact that interviews are not always reliable, it seems unlikely that the low allele frequency in the Italian control group (MAF 0.22%) is only due to an absence of family history of type 2 diabetes. In fact, in our screening group of samples, a 13-fold higher allele frequency was observed in French control subjects with similar features (MAF 2.82%). Furthermore, the allele frequency of the c.136–14_136–13insC variant tended to be even higher in normoglycemic control subjects without a family history of type 2 diabetes than in type 2 diabetic case subjects with a strong family history of type 2 diabetes (MAF 1.49%). Finally, the effect size was not different between the screening group (ascertained for their family history of type 2 diabetes) and the replication group (not ascertained for their family history of type 2 diabetes) of French individuals (Woolf test: χ2 = 0.1, df = 1, P = 0.76), contrary to what was observed in Italians by Chiefari et al. (17).

We observed nominal deviation from the Hardy-Weinberg equilibrium when case and control subjects were analyzed together (P = 0.02), reflecting an underrepresentation of C/C homozygotes. Chiefari et al. (17) reported that the absence of homozygous individuals for this variant in their study was consistent with previous observations indicating that mice homozygous for the Hmga1 null allele have reduced fertility and die prenatally. However, the only homozygous carrier of the c.136–14_136–13insC C allele was aged 47 years, normoglycemic, without a family history of type 2 diabetes, and the mother of three children. Therefore, we hypothesize that their underrepresentation is rather due to sampling effects (modest sample size given the low allele frequency).

Although Chiefari et al. (17) reported strong associations between the c.136–14_136–13insC variant and type 2 diabetes (1.64 < OR < 15.77), no strong effects on type 2 diabetes–related intermediate traits were detected. These results are consistent with those of Chiefari et al. (17), who did not observe any variation in fasting insulin levels. They also showed a decreased expression of HMGA1 and INSR in blood monocytes of type 2 diabetic subjects carrying the c.136–14_136–13insC variant compared with both type 2 diabetic and healthy noncarriers, whereas no association was identified in our study using subcutaneous adipose tissue of normoglycemic individuals. Chiefari et al. (17) compared a subset of case subjects carrying the c.136–14_136–13insC variant with subsets of wild-type case and control subjects. Their selection criteria were not specified, and no comparison between control subjects (carriers vs. noncarriers) was reported. Therefore, the expression variation detected in monocytes may have been biased. Furthermore, point mutations affecting INSR expression were previously described but not sequenced in these samples (45–47). Finally, the regulation of INSR and HMGA1 expressions may not be the same in monocytes and in key tissues involved in insulin resistance (e.g., adipose tissue, liver, and skeletal muscle). However, additional European individuals need to be analyzed to confirm the lack of functional effects.

In conclusion, our data suggest that the c.136–14_136–13insC variant in the HMGA1 gene is not associated with type 2 diabetes risk in Europeans, contrary to what was reported by Chiefari et al. (17). The rapid development of cost-efficient next-generation sequencing should allow reliable detection of associations between rare mutations and type 2 diabetes in large populations (48–50).