The Pharmacogenetics of Type 2 Diabetes: A Systematic Review

In 2013, there existed multiple pharmacologic interventions for the prevention and treatment of type 2 diabetes (1). However, all evaluations of known efficacious interventions reveal that some patients respond to treatment while others do not. As recognized by the American Diabetes Association in its 2012 statement on the management of hyperglycemia, care in type 2 diabetes must become more patient-centered (2), and the individualization of diabetes prevention and treatment based on genetic variation has great potential.

Narrative reviews have commented on the promise of pharmacogenomics of type 2 diabetes (3–6), and prominent individual studies have found statistically significant pharmacogenetic interactions associated with diabetes risk and glycemic outcomes (7–11). However, prior reviews have not systematically evaluated this literature to inform future research questions, and these reviews do not address the quality issues that affect the existing literature on diabetes pharmacogenetics. The clinical utility of genetic variation for tailoring diabetes medications rests on the identification of substantial and statistically significant pharmacogenetic interactions from internally valid studies and confirmation of their findings in varied populations based on race/ethnicity.

We conducted a systematic review of observational and experimental studies to determine if the effect of diabetes medications on diabetes incidence, HbA_1c, fasting glucose (FG), and postprandial glucose (PPG) varies by independent genetic variation in patients with impaired FG, impaired glucose tolerance, or type 2 diabetes. We hypothesized that 1) genetic variation associated with drug transporters, metabolizers, targets, and mechanisms of action would modify the effect of specific drugs and 2) the existing evidence would be insufficient to recommend clinical use of pharmacogenetic interactions because of a lack of well-conducted studies across diverse populations.

Senior members of the study were diabetes and obesity researchers with training in clinical epidemiology, clinical trials, and systematic review methodology (E.B., W.L.B., J.M.C., S.B., W.H.L.K., N.M.M., and M.O.G.) and genetic epidemiology (P.B., W.H.L.K., N.M.M., and M.O.G.). The team also included an experienced project manager with expertise in the conduct of systematic reviews (L.M.W.).

We searched the PubMed, EMBASE, Cochrane electronic databases and also manually searched key review articles, key journals’ tables of contents, and the references of included articles. Key journals were selected based on content area and ones that commonly published the included articles. The PubMed search and list of key journals are provided in Supplementary Tables 1 and 2. The electronic search included dates of database inception through 13 March 2013, and the manual search of tables of contents included January to March 2013. The search was limited to studies published in English.

We included original articles on the effect of Food and Drug Administration (FDA)-approved diabetes medications (Supplementary Table 3) on diabetes incidence, HbA_1c, FG, and PPG in adults with either type 2 diabetes or increased diabetes risk because of impaired FG (FG 5.55–6.94 mmol/L) or impaired glucose tolerance (2-h postload [75 g] glucose 7.77–11.04 mmol/L) by common genetic variation. We considered any independent genetic variation (e.g., single nucleotide polymorphisms [SNPs], copy number variants) eligible and excluded variation such as haplotypes. Eligible study designs were 1) controlled studies evaluating the effect of a drug for one allele/genotype versus another over time and 2) uncontrolled studies evaluating the effect (change in outcome or incidence of outcome) of a drug comparing one genotype/allele to another. We excluded studies of less than 24-h duration and did not include results for HbA_1c in studies shorter than 3 months.

We excluded case reports, case series, and cross-sectional studies; studies not written in English (due to lack of availability of resources to interpret these articles); and studies that included participants on more than one diabetes medication. We did not contact authors to obtain additional results from included studies.

Two investigators reviewed each title, abstract, and full-text article independently. A citation was advanced to abstract review if a single investigator included it. Abstracts and full-text articles were reviewed using a standardized and piloted eligibility criteria form, and disagreements were resolved through consensus.

We developed data abstraction forms based on included abstracts and articles. Data abstraction forms were piloted extensively and included information on study design, study population characteristics, genetic variation under study, and study results on outcomes of interest. Abstraction forms were completed using DistillerSR online systematic review management software. Two investigators abstracted data sequentially using the finalized standardized forms.

We developed quality abstraction forms based on the Strengthening the Reporting of Genetic Association Studies guidelines for reporting of genetic association studies (12). In the absence of guidelines for pharmacogenetic studies, we also incorporated recommendations from the HuGENet (Human Genome Epidemiology Network) HuGE Review Handbook (13) and prior methodological papers (14). We considered a study to be randomized if it randomized participants for the pharmacogenetic study and was not simply based on a prior randomized study. Forms captured elements of quality control of genotyping, including method of genotyping and genotyping call rate, and we considered a call rate ≥95% to be acceptable. We calculated genotype call rates when possible. We also recorded genotyping concordance as a genotyping quality metric. We considered selective reporting of interactions based on positive results and selection bias related to availability of genotyping (Supplementary Data). Two investigators evaluated the quality of each study independently, and disagreements were resolved through consensus.

We performed a qualitative synthesis of included studies’ results. We were unable to perform quantitative syntheses with meta-analyses because too few studies contained the same SNP–drug interactions with common outcomes.

Funding sources had no role in the design, conduct, analysis, or interpretation of the study.

Of 7,279 citations, we included 34 articles from 21 studies (7,8,10,15–45) comprised of 10,407 subjects (Fig. 1). The included articles used one of three study designs: 1) subanalysis of prior randomized controlled trials (RCTs; n = 13); 2) analysis of observational data (n = 8); and 3) nonrandomized, experimental, pre–post design without a control group (n = 13) (Table 1). None of the studies were de novo RCTs specifically designed to evaluate pharmacogenetic interactions. With the exception of the Stop Non-Insulin Dependent Diabetes Mellitus (STOP-NIDDM) trial (20–23) and the Diabetes Prevention Program (DPP) (7,8,17–19,36,37,45), all studies evaluated pharmacogenetic interactions in patients with diabetes. In the DPP, a randomized trial of metformin, a lifestyle intervention, and placebo for diabetes prevention, a broad candidate gene approach (more than 1,590 candidate gene loci) was taken to evaluate associations of SNPs with diabetes and interactions between genetic variants and the trial’s interventions (1 to 3.2 years of follow-up) (7,8,17–19,36,37,45). Genetics of Diabetes Audit and Research Tayside (GoDARTS) investigators performed retrospective analyses of observational data from patients with diabetes who had 12 to 18 months of follow-up using both a genome-wide (10) and candidate gene approach (15,16). The ethnic composition of each study is provided in Table 1. No study evaluated interactions for the other included FDA-approved medications of interest.

Genetic interactions with metformin were reported in 14 articles (Table 2; Supplementary Table 4) (7,8,10,15–19,34–38,45).

Genes encoding the metformin transporters, SLC22A1, SLC22A2, and SLC47A1, were each studied in four articles that evaluated different outcomes. For the SLC22A1 locus, rs683369 was associated with response to metformin with respect to diabetes risk in the DPP over 3.2 years (7), and three SNPs were not associated with HbA_1c in people with diabetes in two other studies (16,38). For SLC22A2, rs662301 was associated with risk of diabetes at 3.2 years in the metformin arm versus placebo in the DPP (7), and two SNPs were not associated with response to metformin with FG and HbA_1c as outcomes in three other articles (34,36,38). For the SLC47A1 locus, rs8065082 was associated with response to metformin for diabetes risk (7), and rs2289669 was not associated with metformin response in patients with diabetes (HbA_1c as the outcome) (34,38).

rs11212617 at the ATM locus predicted response to metformin for diabetes risk at 1 year in the DPP (37) and attainment of HbA_1c <7% (53 mmol/mol) at 18 months in GoDARTS (10). However, results for HbA_1c and FG at 1 year were not significant in the DPP (37). Neither GoDARTS or the DPP found a significant interaction for TCF7L2 SNPs for attainment of HbA_1c <7% (53 mmol/mol) (15) or diabetes incidence (17).

Statistically significant interactions between metformin and genetic variants were also reported for genes encoding additional proteins associated with AMP-activated protein kinase–dependent inhibition of gluconeogenesis [PRKAB2, PRKAA2, PRKAA1, STK11 (46,47), PCK1, PPARA (48), and PPARGC1A (7,49)], insulin secretion [KCNJ11 (7,18), ABCC8, CDKN2A/B (45), HNF4A, and HNF1B (7)]; and insulin sensitivity [ADIPOR2, ENPP1 (8), CAPN10, and GCK (7)] (Table 2; Supplementary Table 4).

Four studies evaluated the interaction between sulfonylureas and SNPs (Table 3; Supplementary Table 5) (15,24–26). Of two studies evaluating SNPs in CYP2C9 (25,26), the gene encoding the primary hepatic cytochrome P450 enzyme, which metabolizes sulfonylureas, one small study found a greater mean change from baseline in HbA_1c at 6 months by diplotype of rs1057910 (25). Notably, the sample size for the variant diplotype was very small (n = 2) (25).

GoDARTS evaluated the interaction between two TCF7L2 SNPs and sulfonylureas and reported a significant association with response to medication (15). Another study evaluating the interaction between two KCNJ11 SNPs and sulfonylureas did not find any differences in the change in FG across genotypes at 12 months (24).

Eight articles reported on genetic interactions with repaglinide (Table 3; Supplementary Table 6) (27–30,39–41,44). Of two SNPs evaluated in the SLC30A8 gene (28,40), rs13266634 was associated with response to repaglinide using HbA_1c, FG, and PPG at 8 weeks as outcomes (28). Similar, but nonsignificant, results were observed with the other SNP evaluated (D′ = 0.928 for rs168889462 and rs13266634) (28). Notably, rs13266634 has been one of the most replicated genetic risk variants in type 2 diabetes (50).

A single study reported a significantly different change in HbA_1c and PPG at 6 months (and similar, nonsignificant results for FG) by E23K genotype of KCNJ11, which encodes the potassium channel inhibited by binding of repaglinide to its receptor on the β-cell (27). SNPs in NEUROD1/BETA2, PAX4 (39), and UPC2 (44) predicted response to repaglinide for some glycemic outcomes (Table 3).

Three studies reported on interactions between pioglitazone and genetic variation (32,33,42) (Table 3; Supplementary Table 7). The Pro12Ala variant was associated with pioglitazone response in one (42) of two studies evaluating this SNP (32,42). A single study reported a significant effect of PTPRD rs17584499 genotype on PPG at 12 weeks but not on HbA_1c or FG (42).

Four studies reported on response to rosiglitazone by genetic variation (Table 3; Supplementary Table 8) (30,31,40,43). Individual studies reported significant interactions between the KCNQ1 (30) and RBP4 (43) loci and rosiglitazone for some, but not all, glycemic measures.

Interactions between acarbose and the PPARA, HNF4A, LIPC, PPARG2, and PPARGC1A loci were evaluated in the STOP-NIDDM trial with 3.3 years of follow-up for diabetes risk (Table 3; Supplementary Table 9) (20–23). Two of 11 SNPs from the PPARA locus were associated with response to acarbose (20). Of six SNPs from the HNF4A locus, two were associated with response to acarbose (21). Single SNPs at the LIPC and the PPARGC1A loci were also associated with response to acarbose (22,23).

We provide detailed results on the quality of included studies in Supplementary Table 10. None of the included studies was a prospective RCT designed to evaluate a pharmacogenetic interaction, and only 13 of 34 (38%) had a control group. Twenty-six of 34 (76%) studies did not report on losses to follow-up. Pharmacogenetic analyses were prespecified in 24 of 34 (71%) studies and were either not reported or not prespecified in the remainder of the studies. Sixteen of 34 (47%) studies addressed the issue of multiple comparisons (or only looked at a single SNP). Thirty-one of 34 studies (91%) addressed population stratification by adjusting for admixture or self-reported race/ethnicity or only included one race/ethnicity. All studies provided some information on method of genotyping. Only 14 of 34 (41%) reported on genotyping or SNP-specific call rate. Most studies (24 of 34 [71%]) did not report on genotyping concordance. Twenty-seven of 34 (79%) reported on testing for Hardy–Weinberg proportions. Studies did not report on masking of genotyping personnel, and 41% declared some form of industry support.

In this systematic review, we identified 34 articles on the pharmacogenetics of diabetes medications, with several reporting statistically significant interactions between genetic variants and medications for glycemic outcomes. Most pharmacogenetic interactions were only evaluated in a single study, did not use a control group, and/or did not report enough information to judge internal validity. However, our results do suggest specific, biologically plausible, gene–medication interactions, and we recommend confirmation of the biologically plausible interactions as a priority, including those for drug transporters, metabolizers, and targets of action. In particular, we recommend follow-up of the 1) SLC22A1, SLC22A2, SLC47A1, PRKAB2, PRKAA2, PRKAA1, and STK11 loci for metformin; 2) CYP2C9 and TCF7L2 loci for sulfonylureas; 3) KCNJ11, SLC30A8, NEUROD1/BETA2, UCP2, and PAX4 loci for repaglinide; 4) PPARG2 and PTPRD for pioglitazone; 5) KCNQ1 and RBP4 loci for rosiglitazone; and 6) PPARA, HNF4A, LIPC, and PPARGC1A loci for acarbose.

Given the number of comparisons reported in the included studies and the lack of accounting for multiple comparisons in approximately 53% of studies, many of the reported findings may be false positives. However, we expect interactions between response to medications and genes encoding their transporters, metabolizers, targets, and components of their pathways for action as observed in the included studies. The DPP reported significant interactions between metformin and loci for its transporters (SLC22A1, SLC22A2, and SLC47A1) (7). It deserves mention that positive findings were not replicated in other studies evaluating these loci (16,34,38), but outcomes (mean change in quantitative traits, achievement of HbA_1c <7% (53 mmol/mol) versus diabetes risk in the DPP) and follow-up time (6 to 18 months vs. 3 years in the DPP) differed in the other studies as well as did study design. The DPP also reported on significant interactions for loci associated with metformin pharmacodynamics (PRKAB2, PRKAA2, PRKAA1, and STK11) (7,46). The primary action of metformin is the inhibition of hepatic glucose production through inhibition of gluconeogenesis, and interactions with loci associated within this pathway (PCK1, PPARA, and PPARGC1A) were reported (7,48,49). Sulfonylureas and repaglinide bind to the sulfonylurea receptor (encoded by ABCC8), which then inhibits the function of the potassium channel encoded by KCNJ11 and causes β-cell depolarization and eventual insulin secretion. While we did identify interactions between repaglinide and KCNJ11 (27), this locus was not associated with sulfonylurea action in a single study that evaluated FG (24). Variation in CYP2C9, which encodes an enzyme that metabolizes sulfonylureas, was associated with response to sulfonylureas in one (25) of two studies (25,26). Finally, the thiazolidinediones activate peroxisome proliferator-activated receptor γ receptors, which regulate expression of genes important for sensitivity to insulin. Thus, variation in PPARG would likely affect response to this class of medications, and this was suggested in one (42) of the two studies (32,42) evaluating this for pioglitazone and was not evaluated for rosiglitazone.

Many putative loci were not evaluated in the included studies. Variation in the hepatic cytochrome P450 enzymes, which metabolize diabetes medications, would be expected to impact their effects, including variation in CYP3A4 and CYP2C8 for repaglinide (51), CYP2C8 and CYP2C9 for rosiglitazone (52), and CYP2C8 and CYP3A4 for pioglitazone (53). The transporter encoded by SLCOB1 transports repaglinide into hepatocytes for metabolism, and variation in this gene could affect the response to this medication (51). Acarbose primarily decreases intestinal glucose absorption by inhibiting brush border enzymes that hydrolyze carbohydrates and is mainly excreted fecally and does not seem to have obvious pharmacokinetic or pharmacodynamic targets.

Generally, we would also expect genetic variants that impact β-cell function to affect the response to insulin secretagogues and genetic variants that impact insulin sensitivity to affect response to insulin sensitizing medications. Also, because of its primary effect on PPG, genetic variation impacting glucose-stimulated insulin secretion would likely impact the response to acarbose. This rationale may explain other observed significant pharmacogenetic interactions (e.g., rosiglitazone–KCNQ1, repaglinide–SLC30A8, sulfonylurea–TCF7L2, and acarbose–HNF4A).

Prior work in this area has consisted of mainly narrative reviews, many of which have included the studies that we identified (3–6). We add to this literature by using a thorough and systematic approach with double review at all levels to identify as many studies as possible that have reported some interaction between individual diabetes medications of interest and diabetes risk and glycemic outcomes. Thus we present the state of the literature on the pharmacogenetics of type 2 diabetes, which lays the groundwork for directing future research efforts. Another novel contribution of our systematic review is the collection of detailed quality information from included studies, which aids in the interpretation of prior studies and illuminates areas for improvement and standardization.

The major limitation of the literature on the pharmacogenomics of type 2 diabetes is the lack of high-quality studies to identify and confirm findings for specific drug–SNP–outcome combinations: 1) The rationale for selection of loci and interactions studied was often not clear, which raises the concern of selective reporting of results and publication bias; in particular, we would be less suspicious of false-positive results in the setting of prespecified analyses based on prior evidence and/or biologic plausibility with adjustment for multiple comparisons. Therefore, it is likely that positive results were reported and that null results were not. 2) The small size of many of the studies does not exclude the possibility that interactions exist but could not be identified because of lack of power; we reported study results as significant based on a P value less than 0.05 but have noted results when P values were <0.20 when possible. Meta-analysis could help to address this issue, but the heterogeneity of studies with specific SNP–drug interactions, outcomes, and follow-up times differing across studies precluded quantitative synthesis with meta-analyses. While our qualitative synthesis summarizes the literature and suggests the existence of specific gene–drug interactions, we could not complete meta-analyses to quantify these observations. 3) Most studies did not have a placebo or other control group. Therefore, our inferences often relied on the results of a single medication intervention on change in or incidence of outcomes by genotype; these types of studies do not exclude the possibility that we are simply observing the effect of genotype and not specifically modification of the response to the medication. 4) Studies did not generally provide information to determine the potential for selection bias based on availability of genotyping information, on losses to follow up, or on the amount and handling of missing data. Regarding selection bias due to availability of genotyping, participant behaviors (e.g., adherence to intervention, follow-up) and outcomes (diabetes, death) may have differed between those with and without genotyping information; these kinds of differentiating characteristics in participants included in genetic analyses could impact the observed gene–drug interactions. 5) While studies did provide information on methods for genotyping, information on SNP-specific call rate was often not reported, and studies did not report on masking of personnel performing genotyping. 6) While none of the included studies were actually de novo RCTs, which would limit selection bias and confounding most completely, several articles were based on data from prior, well-conducted randomized trials (7,8,17–23,36,37,45). In the case of these trials, we would not expect that participant characteristics correlated with genotype would be related to assigned intervention and thus can feel more confident about the robustness of these studies. To address these issues with quality, we did tailor our inclusion criteria and abstraction tools to limit the inclusion of poor-quality studies and to understand the important potential sources of bias.

One limitation of our systematic review methodology is the exclusion of studies of patients on more than one diabetes medication. We sought to identify pharmacogenetic interactions that were based on a single drug and single genetic variant and wanted to avoid drug–drug–SNP interactions. Future work could address these types of interactions. Because of this exclusion, we did not include additional articles on the pharmacogenetics of diabetes in this review (11,54–56). These studies did report positive findings regarding gene–drug interactions based on pharmacodynamics [e.g., PPARG–rosiglitazone (54), SLC22A1–metformin (11), IRS1–sulfonylurea (55), and ATM–metformin (56)]. These studies confirm our hypotheses regarding the possibility of gene–drug interactions based on pharmacodynamics but are still individual studies different enough from the existing literature to preclude meta-analysis. We also excluded studies of non-FDA-approved medications and therefore did not include the article by Feng et al. evaluating interactions with gliclazide; this study did demonstrate a significant interaction between gliclazide and ABCC8 and KCNJ11 SNPs consistent with the pharmacodynamics of sulfonylureas (57). We limited our analyses to diabetes and glycemic outcomes (HbA_1c, FG, PPG) because these are more commonly and consistently measured and are also strongly associated with improvements in long-term complications and mortality (58). Future studies should evaluate other important efficacy and safety outcomes. Finally, because of study resource limitations, we excluded non-English language studies from our initial search and cannot estimate the number of otherwise-eligible studies that we excluded based on this.

We recommend that guidelines for the design, analysis, and reporting of pharmacogenetic studies of diabetes medications be developed to improve study quality and enhance comparability among studies; we have provided a prioritized list of quality and reporting items in Supplementary Table 11. The incorporation of response to medications based on genetic variation into clinical practice cannot occur without well-designed studies confirming significant pharmacogenetic interactions. Based on the limitations of the current literature, we recommend the following for future studies: 1) a priori specification of the SNPs and medications to be studied, 2) the use of experimental designs, 3) inclusion of a concurrent comparison group when possible, 4) agreement in the diabetes pharmacogenetics community regarding standardized outcomes and follow-up (e.g., HbA_1c at 3 months), 5) sufficient power for the primary outcome, 6) adjustment for multiple comparisons if multiple SNPs are examined, and 7) controlling for population stratification and relatedness. In addition, independent replication is important. We recommend that diabetes pharmacogenetics studies use current guidelines for reporting of genetic association studies (12) and that these guidelines be extended to emphasize information relevant to pharmacogenetic studies, including prespecified reporting of analyses with rationale, estimates of type 2 error, standardized reporting of medication interventions, and reporting of differences between genotyped and nongenotyped subjects when possible.

In conclusion, for all known efficacious diabetes preventive and therapeutic pharmacologic agents, some patients benefit or experience harm while others do not. In this systematic review, we find evidence of biologically plausible pharmacogenetic interactions for metformin, sulfonylureas, repaglinide, pioglitazone, rosiglitazone, and acarbose, but these results require confirmation in future studies to determine if an individual’s genetic information can be used to individualize the choice of prediabetes and diabetes pharmacologic management. Importantly, our results should guide the development of guidelines for the design, conduct, and reporting of studies of the pharmacogenetics of type 2 diabetes and other chronic conditions. These promising results show the potential of using genetic variation to tailor therapy for type 2 diabetes prevention and management.

Acknowledgments. The authors thank the following for their contribution to this article: Elisabeth Haberl (Johns Hopkins University, Baltimore, MD), Padmini Ranasinghe (Johns Hopkins University, Baltimore, MD), and Leslie Jackson (University of Indiana, Indianapolis, IN).

Funding. N.M.M. was supported by the National Institutes of Health/National Center for Research Resources (1KL2RR025006-01). M.O.G. was supported by the National Heart, Lung, and Blood Institute (5T32HL007024) and the National Institute of Diabetes and Digestive and Kidney Diseases (5T32DK-062707). P.B. was supported by the National Heart, Lung, and Blood Institute (5T32HL007024). S.B. was supported by the National Center for Research Resources (KL2 RR024990) and the National Center for Advancing Translational Sciences (KL2TR000440), National Institutes of Health. The Baltimore Diabetes Research Center also provided support for this study (P60 DK-079637).

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. N.M.M. conceived of the study, refined the study question and design, obtained data, interpreted results, and drafted and revised the manuscript. M.O.G., W.L.B., S.B., L.M.W., W.H.L.K., and J.M.C. refined the study design, obtained data, interpreted results, and contributed to the revision of the manuscript. P.B. and A.S. obtained data and contributed to the revision of the manuscript. E.B. refined the study design and contributed to the revision of the manuscript.