Pathways Targeted by Antidiabetes Drugs Are Enriched for Multiple Genes Associated With Type 2 Diabetes Risk

Abstract

Genome-wide association studies (GWAS) have uncovered >65 common variants associated with type 2 diabetes (T2D); however, their relevance for drug development is not yet clear. Of note, the first two T2D-associated loci (PPARG and KCNJ11/ABCC8) encode known targets of antidiabetes medications. We therefore tested whether other genes/pathways targeted by antidiabetes drugs are associated with T2D. We compiled a list of 102 genes in pathways targeted by marketed antidiabetic medications and applied Gene Set Enrichment Analysis (MAGENTA [Meta-Analysis Gene-set Enrichment of variaNT Associations]) to this gene set, using available GWAS meta-analyses for T2D and seven quantitative glycemic traits. We detected a strong enrichment of drug target genes associated with T2D (P = 2 × 10⁻⁵; 14 potential new associations), primarily driven by insulin and thiazolidinedione (TZD) targets, which was replicated in an independent meta-analysis (Metabochip). The glycemic traits yielded no enrichment. The T2D enrichment signal was largely due to multiple genes of modest effects (P = 4 × 10⁻⁴, after removing known loci), highlighting new associations for follow-up (ACSL1, NFKB1, SLC2A2, incretin targets). Furthermore, we found that TZD targets were enriched for LDL cholesterol associations, illustrating the utility of this approach in identifying potential side effects. These results highlight the potential biomedical relevance of genes revealed by GWAS and may provide new avenues for tailored therapy and T2D treatment design.

Introduction

Genome-wide association studies (GWAS) have uncovered >65 common DNA variants associated with type 2 diabetes (T2D) that collectively explain ∼10% of the genetic contribution to T2D susceptibility, mostly in populations of European ancestry (1,2). Hundreds of additional common variants (minor allele frequency >1%) are predicted to be associated with T2D, with modest odds ratios based on mixed-model and polygenic analyses of large GWAS meta-analyses (3–5). In addition, dozens of other single nucleotide polymorphisms (SNPs) have been found to be associated with various quantitative glucose and insulin-related traits (6–11), which are intermediate phenotypes of T2D and, hence, may help in better understanding the pathophysiology of diabetes. These GWAS have been extended by the use of large, custom-made genotyping arrays, which include the top SNP associations for a variety of metabolic traits that approached, but did not quite achieve, genome-wide significance (12); when deployed across large numbers of additional samples, genome-wide significance has been reached for 10 additional association signals with T2D (3) and 41 for a number of quantitative glycemic traits (13).

Several genes that lie near established SNPs associated with T2D or a related glycemic trait encode direct or indirect targets of antidiabetes medications (Table 1 and Supplementary Table 1). Specifically, the first two reproducible associations with T2D were reported for missense polymorphisms in two candidate gene regions (encoding the peroxisome proliferator–activated receptor γ [PPARG] and the islet ATP-sensitive potassium channel Kir6.2/sulfonylurea receptor SUR1 complex [KCNJ11/ABCC8]), which were selected because they happen to be targets of antidiabetic medications (thiazolidinediones [TZDs] and sulfonylureas, respectively) (14,15). This observation led us to ask whether additional genes that function in biological pathways or processes affected by various antidiabetes medications may also contain common variants associated with T2D or related glycemic traits, albeit less strongly (i.e., weaker effects, lower penetrance in the population), but when analyzed together, their combined effect may surpass statistical significance. In other words, we reasoned that proof of their successful modulation by pharmacological means might indicate their involvement in T2D pathophysiology.

Table 1.

Target genes and pathways of nine classes of FDA-approved antidiabetes medications

Medication class	Mechanism of action	Physiological effect	Tissue	Number of target genes*	Target genes of antidiabetes medication class*	Target genes in LD to known T2D or glycemic trait–associated SNPs†	References
Insulin	Downstream signaling from insulin/IGF-I receptor	Signals glucose uptake	Liver, muscle, fat	17	ACACA, AKT1, FOXO1, GAB1, INSR, IRS1, IRS2, IRS4, PIK3CA, PIK3R1, SLC2A1, SLC2A2, SLC2A3, SLC2A4, SLC2A5, SHC1, TRIB3	IRS1: T2D, fasting insulin, insulin resistance; SLC2A2 (GLUT2): fasting glucose	11,19,42
Metformin (biguanide)	AMP-activated protein kinase pathway; complex I inhibition	Improves insulin sensitivity; reduces gluconeogenesis	Liver, muscle	6	PRKAA2, SLC22A1, SLC22A2, SLC2A1, SLC2A4, STK11	—	20,30
TZDs	PPARG receptor pathway		Fat, muscle, liver	42	ABCA1, ACSL1, ADIPOQ, ADIPOR1, ADIPOR2, APOA1, APOC3, CCL2, CD36, CPT1A, CPT1B, CPT2, CRP, EDN1, F3, FACL2, FGB, GK, HSD11B1, ICAM1, IL6, IRS1, IRS2, MMP9, NFKB1, NFKB2, NOS2A, PDK4, PIK3CA, PIK3R1, PPARA, PPARD, PPARG, PTGS1, RETN, RXRA, SCARB1, SLC2A4, SLC27A1, SORBS1, TNF, VCAM1	PPARG: T2D; IRS1: T2D, fasting insulin	11,21,22,42
Sulfonylureas	ATP-sensitive K channel inhibition	Increases insulin production and secretion from β-cells in pancreas and/or decreases glucagon secretion from α-cells in pancreas	Pancreas, liver, fat	5	ABCC8, ABCC9, KCNJ11, SLC2A4, TNF	KCNJ11: T2D ABCC8: T2D	23–25
GLP-1 receptor agonists	GLP-1 receptor pathway		Pancreas, GI tract	18	ATF4, BCL2, CASP12, CPA1, DDIT3, EIF2S1, GCG, GLP1R, HSPA5, JUNB, NGFB, NGFR, PDX1, PPP1R15A, PPYR1, PRKACA, XBP1, XIAP	PDX1: fasting glucose‡	16–18
DPP4 inhibitors	Inhibits DPP4 degradation (i.e., GLP-1, GIP)		Diffuse	20	ADCYAP1, CCL11, CCL5, CCL22, CXCL9, CXCL10, CXCL11, CXCL12, DPP4, DPP8, DPP9, FAP, GHRH, GIP, GIPR, GRP, GLP2R, NPY, PYY, TAC1	GIPR: T2D‡, fasting glucose‡, 2-h glucose, 2-h insulin	18,26
Amylin mimetics	Amylin receptor pathway		Pancreas, GI tract	2	IAPP, IDE	IDE: T2D	27
Meglitinides	ATP-sensitive K channel inhibition		Pancreas, liver, fat	2	ABCC8, KCNJ11	KCNJ11/ABCC8: T2D	28
α-Glucosidase inhibitors	Inhibits α-glucosidase enzymes, α-amylase inhibition	Affects sugar absorption in gut by preventing digestion of carbohydrates	Small intestine, pancreas	2	AMY2A, GAA	—	29

GI, gastrointestinal.

*Genes involved in pathways targeted by antidiabetes medications were compiled from the literature.

†Validated SNP associations are based on GWAS of European descent individuals; details of associations are in Supplementary Table 1.

‡These genes were added to our drug target gene list before their association with T2D or a related glycemic trait was reported (3). Genes in boldface refer to drug targets in LD to SNPs associated with T2D or glycemic traits.

Currently, multiple classes of antidiabetic medications are approved for clinical use, including insulin, biguanides, sulfonylureas, TZDs, meglitinides, α-glucosidase inhibitors, GLP-1 receptor agonists, dipeptidyl peptidase 4 (DPP4) inhibitors, and amylin mimetics (16–30). All medications help to decrease glucose blood levels through different mechanisms of action. These vary from increasing insulin secretion by pancreatic β cells to increasing insulin sensitization in target tissues (e.g., muscle, fat) and inhibiting glucose absorption in the gastrointestinal tract. Some of the more recently approved medication classes are the result of rational drug design (e.g., GLP-1 receptor agonists, DPP4 inhibitors), some from empirical experimentation (e.g., metformin), and others from a combination of the two (e.g., TZDs [ciglitazone discovered through in vivo compounds screening (31) and successive analogs designed against the subsequently identified drug target PPARG (32)]).

The targeted pathways and downstream effects of these drug classes have been investigated to a greater or lesser extent through human, animal, and cell culture studies (16–30). We attempted to leverage existing GWAS data sets to comprehensively test whether these pathways contain multiple genes, in addition to those already known, that harbor natural genetic perturbations that may influence risk of T2D. We further evaluated whether this approach could be used to predict unintended phenotypic effects of drug treatment by examining the genetic basis of a known nonglycemic effect of TZD drugs.

Research Design and Methods

Construction of the Antidiabetes Drug Target Gene Set

The source of genes to include in the drug target gene set were culled from PubMed searches (initially performed in 2007 [before the wave of GWAS publications] and repeated in 2010) of original reports and reviews on the mechanism of action of nine Food and Drug Administration (FDA)–approved antidiabetes medication classes at the time: insulin, biguanides (metformin), sulfonylureas, TZDs, meglitinides, α-glucosidase inhibitors, GLP-1 receptor agonists, DPP4 inhibitors, and amylin mimetics (16–30). More recent classes of antidiabetes drugs (cholesterol-binding resins, dopamine agonists, or sodium-glucose co-transporter-2 inhibitors) have not yet reached widespread use and, thus, were not studied here. Drug class targets and their downstream effectors were considered broadly for each drug class. Proteins were considered a potential target of the drug if they were directly affected by the drug or a direct downstream mediator of the known drug pathway, considering human, animal, and cell culture studies. There was significant overlap of target genes among certain drug classes because their mechanisms of action are largely the same (e.g., sulfonylureas, meglitinides). Proteins in the drug pathway whose genes did not have a validated human genome location were excluded from the analysis.

GWAS Meta-analyses Analyzed

The present analysis was based on the SNP association P values from the following GWAS meta-analyses: 1) DIAGRAMv3 (DIAbetes Genetics Replication And Meta-analysis version 3) T2D meta-analysis: 12,171 T2D cases and 56,862 controls across 12 GWAS of individuals of European descent and ∼2.5 × 10⁶ genotyped and imputed SNPs (data can be downloaded at http://diagram-consortium.org/downloads.html) (3); 2) Metabochip T2D meta-analysis: 21,491 T2D cases and 55,647 controls across 25 studies of individuals of European descent and 1,178 T2D cases and 2,472 controls from one study of individuals of Pakistani descent (PROMIS [Pakistan Risk of Myocardial Infarction Study]), with a total of 22,669 cases and 58,119 controls (3) (the Metabochip custom array comprises 196,726 SNPs, 5,057 of which are T2D “replication” SNPs that capture the strongest independent autosomal association signals from the DIAGRAMv3 GWAS meta-analysis; individuals in the DIAGRAMv3 meta-analysis are independent from those in the Metabochip T2D meta-analysis); 3) seven MAGIC (Meta-Analyses of Glucose and Insulin-related traits Consortium) meta-analyses of glucose and insulin-related traits: 9–23 GWAS of 15,000–46,000 participants without diabetes and 2.3–2.7 × 10⁶ genotyped and imputed SNPs, depending on the trait (data can be downloaded at www.magicinvestigators.org/downloads) (7–9); and 4) three GWAS meta-analyses of plasma LDL cholesterol (LDL-C), HDL cholesterol (HDL-C), and triglyceride levels: 46 GWAS in ∼95,000–100,000 individuals total (data can be downloaded at www.sph.umich.edu/csg/abecasis/public/lipids2010) (33). The global lipids blood measurements were taken after >8 h of fasting. The lower-bound minor allele frequency of SNPs in these meta-analyses is 1%.

Discovery Step 1: Gene Set Enrichment Analysis of GWAS SNP Data Using MAGENTA

To test whether a set of antidiabetes drug target genes contains multiple genes associated with T2D more than would be expected by chance, we applied a Gene Set Enrichment Analysis (GSEA) method (an approach originally developed for gene expression) that we previously adapted for GWAS data called Meta-Analysis Gene-set Enrichment of variaNT Associations (MAGENTA) (34). Intuitively, MAGENTA tests whether multiple genes associated with a disease or trait cluster in a given biological pathway or set of functionally related genes. It does so by testing whether the distribution of gene association P values of all genes in a gene set of interest is skewed toward low P values compared with the (close to uniform) P value distributions of randomly sampled gene sets of equivalent size (Fig. 1). A significant skewness below a given P value cutoff (enrichment cutoff) would suggest that the gene set is enriched for multiple genes associated with the tested complex disease or trait. To gain statistical and explanatory power, MAGENTA tests for enrichment not only of genes with strong SNP associations (e.g., ones that pass multiple hypothesis correction, P < 5 × 10⁻⁸) but also of genes with modest SNP associations (associations that have not yet reached genome-wide significance [e.g., P = 10⁻⁴–10⁻³] due to insufficient power afforded by finite GWAS sample sizes). This is supported by polygenic analyses of GWAS that suggest that various complex phenotypes, including T2D risk, are influenced by hundreds of modest associations that have not yet been detected due to insufficient statistical power (3–5).

Figure 1.

An overview of the study design, analytical steps, and questions addressed. The strategy addressed a number of key questions about the relationship between human genetic associations with T2D or related glycemic traits and antidiabetes drug targets. A similar strategy can be applied to other diseases and traits. 2-h glucose and 2-h insulin, glucose or insulin plasma levels measured 2 h after an oral glucose tolerance test; HbA_1c, a measure for long-term glycemia; HOMA-B, a measure for β-cell function; HOMA-IR, a measure for insulin resistance.

We applied MAGENTA to all genotyped and imputed SNP associations from the aforementioned GWAS meta-analyses DIAGRAMv3, MAGIC, and global lipids GWAS meta-analyses, as described quantitatively and in detail by Segrè et al. (34). Briefly, first, MAGENTA scores all genes in the genome by assigning each gene the most significant local SNP association P value within −110 kilobases (kb) upstream and +40 kb downstream the transcript start and end sites, respectively (boundaries chosen to capture potential regulatory causal variants in addition to coding variants within the gene itself). The gene association scores are subsequently corrected for confounding effects, such as gene size, local SNP density, and linkage disequilibrium (LD)–related properties, using stepwise multivariate linear regression analysis (because larger genes are more likely to carry a SNP with a more significant P value than smaller genes by chance, as larger genes contain more SNPs) (34). Second, the adjusted gene association P values, used to rank genes in the genome with respect to their likelihood of association with the given trait, are used to estimate gene set enrichment P values for each gene set of interest. The gene set enrichment P value calculated by MAGENTA assesses the overrepresentation of highly ranked gene association P values above an enrichment cutoff, compared with multiple randomly sampled gene sets from the genome, with equal gene set size. The 75th percentile of the association P values of all genes in the genome (which corresponded, e.g., in DIAGRAMv3 to an adjusted gene P < 0.3) was used as the enrichment cutoff. Physical proximity along the chromosome between two or more genes in a given gene set was corrected for by collapsing them to one effective gene, retaining the gene with the most significant adjusted association P value. Only genes on autosomal chromosomes were analyzed, which led to the exclusion of three drug target genes on chromosome X. The HLA region was removed due to high LD and gene density in the region, making it difficult to disentangle the putative causal gene if an association signal exists in the region. One of the antidiabetes drug target genes, TNF lies in this region. The MAGENTA software package can be downloaded at www.broadinstitute.org/mpg/magenta.

The set of validated T2D SNPs used in this work (Supplementary Table 3) included 55 associated loci identified in GWAS meta-analyses of populations of European descent, including DIAGRAMv3 and Metabochip. Genes near the validated T2D SNPs were defined using the larger of two boundaries around each SNP: ±100 kb or an LD-based boundary defined by proceeding to r² > 0.5 on either side of the SNP, then to the nearest recombination hotspot, and finally adding an additional 50 kb on either side. Genome build 36 (hg18) was used for chromosome positions.

Replication Step 2: Modified GSEA of Metabochip Meta-analysis

To test if the gene set enrichment results obtained by MAGENTA replicated in an independent study, we applied a modified GSEA method for genetic association data that we developed in Morris et al. (3) to a separate, larger meta-analysis of multiple association studies genotyped on the Metabochip array. A modified GSEA approach was needed to account for the bias in the Metabochip SNP design, which contains a subset of SNPs that unevenly cover the genome (196,725 replication and fine-mapping SNPs) compared with the less biased genome-wide SNP arrays (∼2 × 10⁶ SNPs) used in DIAGRAMv3, MAGIC, and the lipid GWAS (analyzed in the discovery step). The Metabochip is a custom array designed to follow-up nominal associations for T2D and 22 other metabolic and cardiovascular traits in a more cost-effective manner than genome-wide SNP arrays and contains 5,057 T2D replication SNPs chosen based on the top independent association signals in DIAGRAMv3 (12). This has enabled the genotyping of top-ranked metabolic SNPs in an additional ∼21,500 T2D cases and ∼55,600 controls. Specifically, the modified GSEA method tests for enrichment of antidiabetes drug target genes among all genes near a set of top T2D-associated SNPs based on the Metabochip meta-analysis. First, for the enrichment cutoff, we used a set of high-confidence T2D SNPs/loci: 137 T2D loci, which included 53 established (P < 5 × 10⁻⁸); 6 highly probable (P < 5 × 10⁻⁷); and 78 probable (more modest) T2D loci with a posterior probability (confidence score) >75%, as chosen in Morris et al., based on a mixture model fitted to the Metabochip T2D meta-analysis z scores (for a list of the 137 T2D SNPs, see Supplementary Table 15 in Morris et al. [3], excluding the monogenic genes). Second, we tested two SNP-to-gene mapping definitions: 1) nearest gene and 2) an LD-based boundary defined by proceeding to r² > 0.5 on either side of the SNP, then to the nearest recombination hotspot, and finally adding an additional 50 kb on either side. SNPs with no genes in LD were assigned the nearest gene. SNPs with the same nearest gene were collapsed to one locus. SNPs with more than one proximal gene in the drug target gene set were counted as one instance to reduce inflation of the gene set enrichment P value due to physical clustering of subsets of drug target genes along the genome (3). For each gene set gs and a set of T2D SNPs/loci l with m SNPs, we calculated the probability of observing at least k T2D loci of a total m loci with one or more proximal genes that belong to gene set gs, given that n of N Metabochip SNPs analyzed have one or more proximal genes in gene set gs (using a hypergeometric probability distribution) (Eq. 1):

N refers to the number of Metabochip SNPs in the null set (defined next) plus the m top T2D SNPs/loci. To account for differences in coverage of the Metabochip replication SNPs across all genes in the genome, we generated an empirical null distribution of for randomly sampled sets of SNPs, matched for SNP number and local gene density with the T2D SNPs/loci set l. The adjusted gene set enrichment P value is the fraction of randomly sampled SNP sets of equal size to the T2D SNP set l with the same or more significant hypergeometric probability (Eq. 1) than that of the actual T2D SNP set (Eqs. 2–4):

For the null SNP set, we used the full set of Metabochip replication SNPs after LD pruning (r² < 0.05, using CEU [Utah residents with Northern and Western European ancestry] HapMap samples as the reference population), excluding the lead and proxy (CEU r² ≥ 0.1) SNPs of previously established T2D SNPs, 5,057 T2D replication SNPs, SNPs near monogenic diabetes genes, and QT-interval replication SNPs (used as a negative control). This resulted in a set of 16,408 null SNPs. Similar GSEA results to those in Table 5 were obtained when using as the null SNP set 1,600 SNPs with the lowest posterior probability of being associated with T2D of 3,408 LD-pruned Metabochip T2D replication SNPs (posterior probability < 5% of belonging to an alternative distribution in the mixture model of Metabochip z scores) (data not shown).

Table 5.

Replication of T2D association enrichment signal in antidiabetes drug target set in an independent T2D meta-analysis (Metabochip)

Genes mapped to established and high-confidence set of T2D SNPs*	Number of Metabochip replication SNPs (null set) near one or more drug target genes†	Number of top-ranked T2D SNPs near one or more drug target genes	Gene set enrichment P value	Genes near established or high-confidence T2D SNPs
Nearest gene	48	5	0.003	PPARG1, KCNJ111, IRS11, GIPR2, ACSL13
Genes in LD‡	83	7	0.04	PPARG1, KCNJ111, ABCC81, IDE1 IRS1 GIPR2, NFKB13, ACSL13

*T2D SNP set tested includes 137 loci: 59 established or highly probable SNPs and 78 high-confidence T2D SNPs based on Metabochip analysis (described in research design and methods).

†This set includes a null set of 16,408 LD-pruned Metabochip replication SNPs that does not contain SNPs in LD to previously established T2D SNPs, ∼5,000 T2D replication SNPs, monogenic diabetes genes, or QT-interval replication SNPs (see research design and methods ).

‡LD boundaries are defined in research design and methods.

¹Genes near previously established T2D SNPs.

²Gene near T2D SNP found in the joint T2D meta-analysis of DIAGRAMv3 and Metabochip, which did not reach genome-wide significance in DIAGRAMv3 alone.

³Genes near T2D SNPs that have not yet reached genome-wide significance but have a high posterior probability of being associated with T2D based on Metabochip analysis (3).

Results

We tested the hypothesis that biological pathways targeted by antidiabetes medications may be enriched for multiple genes modestly associated with T2D, more than would be expected by chance, by applying the GSEA approach implemented in MAGENTA (34) to a compiled list of 102 direct or indirect target genes of one or more of the nine classes of antidiabetes medications (described in research design and methods). The antidiabetes drug target genes ranged from 1 to 41 per drug class (Table 1). The study design and analyses performed on the drug target gene set and its subclasses are described in Fig. 1.

Using the largest available T2D GWAS meta-analysis of ∼12,000 cases and ∼57,000 controls (DIAGRAMv3), we found that the full set of antidiabetes drug target genes was significantly enriched for multiple genes that carry SNPs modestly associated with T2D risk (P = 1.7 × 10⁻⁵, 1.8-fold enrichment [i.e., we predict that about one-half the target genes with association scores above the 75th percentile enrichment cutoff are modestly associated with T2D]) (Table 2 and Fig. 2). MAGENTA suggests that 18 of the 41 loci (44 genes) above the 75th percentile enrichment cutoff (expected number of genes above cutoff, 23) (see columns 3–5 in Table 2) are true associations with T2D risk, 4 of which map onto known T2D loci (see column 8 of Table 2) and 14 of which may contain new SNP associations with T2D. Additional follow-up analyses and experiments are needed to identify the 14 true-positive novel gene associations. The top-ranked 44 drug target genes and their most significant local T2D SNP P values are listed in Supplementary Table 2.

Table 2.

GSEA of T2D and glycemic trait associations in the antidiabetes drug target gene set

GWAS meta-analysis	Nominal MAGENTA enrichment P value*	Number of OBS genes/loci above enrichment cutoff	Number of EXP genes/loci above enrichment cutoff	Excess number of genes/loci above enrichment cutoff (OBS − EXP)†	Enrichment fold (OBS/EXP)	Number of genes near validated GWAS SNPs‡	Genes near validated GWAS SNPs‡
T2D	1.7 × 10−5	41**	23	18	1.78	6***	PPARG, IRS1§, KCNJ11/ABCC8‖, IDE, GIPR¶
Fasting glucose	0.078	31	24	7	1.29	1	SLC2A2
HOMA-IR	0.11	29	24	5	1.21	0	—
2-h insulin	0.24	27	24	3	1.13	1	IRS1
HOMA-B	0.27	26	23	3	1.13	0	—
Fasting insulin	0.29	26	24	2	1.08	0	—
2-h glucose	0.74	20	23	0	0.87	0	—
HbA1c	0.75	20	23	0	0.87	0	—

The 2-h glucose and 2-h insulin concentrations were measured after an oral glucose tolerance test. EXP, expected; OBS, observed.

*The gene set enrichment P value was calculated by MAGENTA using a 75th percentile enrichment cutoff.

**44 genes had scores above the enrichment cutoff, but 3 genes were removed from GSEA to correct for physical clustering along the genome (see Table 4).

***Only 4 loci contributed to enrichment signal. See next two footnotes for explanation.

†Estimated number of antidiabetes drug targets that may be true associations with T2D, 14 of which have not yet reached genome-wide significance.

‡Genes were mapped onto 55 established T2D SNPs using the larger of the two boundaries around each SNP: ±100 kb or LD r² > 0.5 (see research design and methods and Supplementary Table 3).

§The gene association P value of IR1S did not surpass the enrichment cutoff because the established T2D GWAS SNP near IRS1 lies farther away than the gene boundaries used in MAGENTA (+110 kb/−40 kb).

‖KCNJ11/ABCC8 were collapsed to one effective gene in the GSEA due to their physical proximity.

¶GIPR was added to our drug target gene list before its association with T2D reached genome-wide significance in a joint meta-analysis of DIAGRAMv3 and Metabochip (3).

Figure 2.

Distribution of T2D gene association P values of antidiabetes drug targets. To visualize the enrichment of multiple modest associations with T2D among antidiabetes drug target genes, we plotted the noncumulative distribution of adjusted gene association P values (calculated with MAGENTA) for all the antidiabetes drug targets (99 autosomal genes), as shown in the first track (red line). The following two tracks display from top to bottom the individual gene P values (represented by vertical lines) for the insulin targets subset and the TZD targets subset. Common insulin and TZD targets are shown in blue. The dashed line marks the 75th percentile enrichment cutoff.

Given the positive results, we next asked whether the antidiabetes drug target gene set might also be enriched for multiple genes associated with glucose or insulin-related traits, which are intermediate phenotypes or risk factors of T2D. We repeated the aforementioned analysis by applying MAGENTA to seven GWAS meta-analyses of 15,000–46,000 nondiabetic individuals (MAGIC) for the following glycemic traits: fasting glucose levels, fasting insulin levels, 2-h glucose or 2-h insulin plasma levels following an oral glucose tolerance test, a measure for β-cell function (HOMA-B), a measure for insulin resistance (HOMA-IR), and a measure for long-term glycemia (glycated hemoglobin [HbA_1c]). None of the glycemic traits showed a significant overrepresentation of multiple modest gene associations in the antidiabetes drug target gene set (Table 2).

Given the strong enrichment of drug target gene associations with T2D, we asked whether the observed enrichment signal was primarily driven by genes targeted by one or a subset of the nine classes of antidiabetes medication classes. To address this, we tested for enrichment of T2D associations in individual drug class target subsets for four of the nine classes of drugs that contained at least 10 target genes (chosen as the lower bound for statistical power considerations). We found that insulin targets and TZD targets were the primary drivers of the collective enrichment signal (P = 0.001, 2.5-fold enrichment, and P = 0.02, 1.6-fold enrichment, respectively) (Table 3 and Fig. 2). However, although not enriched, incretin targets were also among the top-ranked drug target genes based on their T2D association P values (e.g., the DPP4 inhibitors GIP, GLP2R, GRP, GIPR) (Table 4 and Supplementary Table 2).

Table 3.

GSEA of T2D associations in individual drug class target sets before and after removing genes in established T2D loci

	All genes analyzed				Excluding genes in LD to validated T2D SNPs
Antidiabetes drug target gene set	Number of genes analyzed*	Nominal MAGENTA enrichment P value	Enrichment fold (OBS/EXP)	Excess number of genes above enrichment cutoff (OBS − EXP)	Number of genes analyzed*	Nominal MAGENTA enrichment P value	Enrichment fold (OBS/EXP)	Excess number of genes above enrichment cutoff (OBS − EXP)	Genes in LD to validated T2D SNPs†
All nine classes of drugs	92	1.7 × 10−5	1.8	18	87	4 × 10−4	1.7	15	PPARG, IRS1‡, KCNJ11/ABCC8, IDE, GIPR
Insulin	16	0.001	2.5	6	15	8 × 10−4	2.5	6	IRS1
TZDs	38	0.02	1.6	6	36	0.01	1.7	6	PPARG, IRS1
DPP4 inhibitors	18	0.15	1.4	2	17	0.22	1.5	2	—
GLP-1 receptor agonists	17	0.38	1.3	1	17	0.49	1.3	1	—

EXP, expected; OBS, observed.

*Number of genes analyzed after excluding genes in HLA region (one gene) and genes on sex chromosomes (three genes) and correcting for physical clustering of subsets of genes within a given gene set along the genome (six genes).

†Genes were mapped onto 55 established T2D SNPs using the larger of two boundaries around each SNP: ±100 kb or LD r² > 0.5 (see research design and methods and Supplementary Table 3).

‡The gene P value of IRS1 did not pass the enrichment cutoff because the validated T2D GWAS SNP near IRS1 lies farther away than the gene boundaries used in the MAGENTA analysis.

Table 4.

Top-ranked antidiabetes target genes above enrichment cutoff based on their DIAGRAMv3 T2D gene association P values

Gene	Description	T2D medication class	MAGENTA T2D gene association P value	Best local SNP rs number	Best local SNP DIAGRAMv3 P value	Gene in LD to established T2D SNPs
IDE	Insulin-degrading enzyme	Amylin mimetics	1.55 × 10−15	rs7911264	4.50 × 10−13	+
PPARG	Peroxisome proliferator–activated receptor γ	TZDs	6.32 × 10−9	rs11709077	1.12 × 10−9	+
KCNJ11*	Potassium inwardly rectifying channel, subfamily J, member 11	Sulfonylureas, meglitinides	6.88 × 10−4	rs5215	4.36 × 10−6	+
ABCC8*	ATP-binding cassette, subfamily C (CFTR/MRP), member 8	Sulfonylureas	1.22 × 10−3	rs5215	4.36 × 10−6	+
GIP	Gastric inhibitory polypeptide	DPP inhibitors	7.62 × 10−3	rs3809770	1.03 × 10−4	−
ACSL1	Acyl-CoA synthetase long-chain family member 1	TZDs	1.83 × 10−2	rs735949	7.76 × 10−5	−
IRS2	Insulin receptor substrate 2	Insulin, TZDs	1.89 × 10−2	rs1330545	1.02 × 10−4	−
GLP1R	Glucagon-like peptide 1 receptor	GLP-1 receptor agonists	2.39 × 10−2	rs1929902	5.35 × 10−5	−
GLP2R	Glucagon-like peptide 2 receptor	DPP inhibitors	2.87 × 10−2	rs17743194	4.28 × 10−5	−
NFKB1	Nuclear factor of κ light polypeptide gene enhancer in B-cells 1	TZDs	3.01 × 10−2	rs4648055	2.04 × 10−4	−
ADIPOQ	Adiponectin, C1Q, and collagen domain containing	TZDs	3.22 × 10−2	rs7649121	2.12 × 10−4	−
CXCL9†	Chemokine (C-X-C motif) ligand 9	DPP inhibitors	3.83 × 10−2	rs13131187	3.13 × 10−4	−
CXCL11†	Chemokine (C-X-C motif) ligand 11	DPP inhibitors	3.86 × 10−2	rs13131187	3.13 × 10−4	−
CXCL10†	Chemokine (C-X-C motif) ligand 10	DPP inhibitors	3.89 × 10−2	rs13131187	3.13 × 10−4	−
SLC2A2	Solute carrier family 2 (facilitated glucose transporter), member 2	Insulin	5.33 × 10−2	rs8192675	6.99 × 10−4	−
TRIB3	Tribbles pseudokinase 3	Insulin	5.85 × 10−2	rs1555318	4.42 × 10−4	−
GAA	Glucosidase, α; acid	α-Glucosidase inhibitors	6.14 × 10−2	rs2361710	9.28 × 10−4	−
GAB1	GRB2-associated binding protein 1	Insulin	7.07 × 10−2	rs300938	6.97 × 10−4	−
VCAM1	Vascular cell adhesion molecule 1	TZDs	8.35 × 10−2	rs1932351	5.58 × 10−4	−
NGF	Nerve growth factor (β-polypeptide)	GLP-1 receptor agonists	8.94 × 10−2	rs11466094	5.63 × 10−4	−
CPT2	Carnitine palmitoyltransferase 2	TZDs	9.48 × 10−2	rs1288351	7.58 × 10−4	−
NFKB2	Nuclear factor of κ light polypeptide gene enhancer in B-cells 2 (pT9/p100)	TZDs	9.64 × 10−2	rs7897947	2.44 × 10−3	−
DDIT3	DNA damage–inducible transcript 3	GLP-1 receptor agonists	1.13 × 10−1	rs813516	3.20 × 10−3	−
GRP	Gastrin-releasing peptide	DPP inhibitors	1.16 × 10−1	rs9951619	5.96 × 10−4	−
SLC2A4	Solute carrier family 2 (facilitated glucose transporter), member T	Insulin, sulfonylureas, metformin, TZDs	1.21 × 10−1	rs4562	2.23 × 10−3	−
TAC1	Tachykinin, precursor 1	DPP inhibitors	1.28 × 10−1	rs17168923	1.72 × 10−3	−
SLC22A1	Solute carrier family 22 (organic cation transporter), member 1	Metformin	1.34 × 10−1	rs8191811	1.15 × 10−3	−
ADIPOR1	Adiponectin receptor 1	TZDs	1.45 × 10−1	rs782810	1.62 × 10−3	−
SLC2A3	Solute carrier family 2 (facilitated glucose transporter), member 3	Insulin	1.45 × 10−1	rs3782681	2.13 × 10−3	−
GIPR	Gastric inhibitory polypeptide receptor	DPP inhibitors	1.68 × 10−1	rs8108269	3.12 × 10−3	+
AMY2A	Amylase, α2A (pancreatic)	α-Glucosidase inhibitors	1.70 × 10−1	rs1058607	6.75 × 10−3	−
CPT1A	Carnitine palmitoyltransferase 1A (liver)	TZDs	1.82 × 10−1	rs2507833	2.77 × 10−3	−
FOXO1	Forkhead box O1	Insulin	2.02 × 10−1	rs12874490	3.00 × 10−3	−
HSD11B1	Hydroxysteroid (11-β) dehydrogenase 1	TZDs	2.03 × 10−1	rs3737913	1.78 × 10−3	−
ACACA	Acetyl-CoA carboxylase α	Insulin	2.04 × 10−1	rs3744589	2.17 × 10−3	−
CCL22	Chemokine (C-C motif) ligand 22	DPP inhibitors	2.05 × 10−1	rs8102	3.70 × 10−3	−
PTGS1	Prostaglandin-endoperoxide synthase 1 (prostaglandin G/H synthase and cyclooxygenase)	TZDs	2.06 × 10−1	rs12005866	2.54 × 10−3	−
PIK3CA	Phosphatidylinositol-T,5-bisphosphate 3-kinase, catalytic subunit α	Insulin, TZDs	2.07 × 10−1	rs6443629	3.79 × 10−3	−
SLC2A5	Solute carrier family 2 (facilitated glucose/fructose transporter), member 5	Insulin	2.13 × 10−1	rs4908803	2.44 × 10−3	−
PDX1	Pancreatic and duodenal homeobox 1	GLP-1 receptor agonists	2.19 × 10−1	rs9581940	2.37 × 10−3	−
SCARB1	Scavenger receptor class B, member 1	TZDs	2.46 × 10−1	rs1031605	1.95 × 10−3	−
PPARD	Peroxisome proliferator–activated receptor δ	TZDs	2.55 × 10−1	rs10484578	3.55 × 10−3	−
CPA1	Carboxypeptidase A1 (pancreatic)	GLP-1 receptor agonists	2.61 × 10−1	rs10263705	4.01 × 10−3	−
IAPP	Islet amyloid polypeptide	Amylin mimetics	2.82 × 10−1	rs11045995	2.65 × 10−3	−

The listed genes are the top-ranked antidiabetes drug targets with an adjusted T2D gene association P value (based on DIAGRAMv3) above the 75th percentile enrichment cutoff. One-half of these genes are predicted by MAGENTA to be true associations with T2D that have not yet reached genome-wide significance.

*,†These are each a cluster of two or three genes physically adjacent to each other along the chromosome that share the same most significant local T2D SNP and hence were collapsed to one effective gene in the GSEA. TNF was excluded from the GSEA because it lies in the HLA region. IRS1 did not pass the enrichment cutoff because it lies >110 kb from the T2D GWAS SNP (beyond the SNP-to-gene mapping boundaries used), and hence its gene association score was weak.

Because 6 of the 102 drug target genes lie in five validated loci associated with T2D (Table 1 and Supplementary Table 1), we asked whether the T2D enrichment signal observed in the antidiabetes drug target set was mainly due to genes in known association regions or whether it was also driven by additional new associations of modest effect sizes that have not yet reached genome-wide significance (due to insufficient GWAS sample size). To test this, we excluded all genes near established T2D SNPs from the analysis (listed in Supplementary Table 3 for T2D SNP list and research design and methods for boundary definition) and reran MAGENTA on the full list of drug target genes and on the drug-specific target subsets. The enrichment signal still remained significant, although it decreased by an order of magnitude (P = 4 × 10⁻⁴, Bonferroni-corrected cutoff P < 0.003 accounting for 17 hypotheses tested in Tables 2 and 3), as may be expected when removing a portion of the signal. These results suggest that the enrichment is due not only to genes near known associations but also to ∼14 additional new genes of modest effects (odds ratio of modest associations above enrichment cutoff are in the range [0.85−1.24], with the exception of one value of 1.73; P values of best local SNPs range between 5.4 × 10⁻⁶ to 7 × 10⁻³).

To test the reproducibility of the T2D enrichment signal in the antidiabetes drug target set, we tested whether the results replicated in an independent association study of T2D. For this, we used the large-scale T2D association meta-analysis of ∼24,000 T2D cases and ∼58,000 controls genotyped on Metabochip and applied a modified GSEA method that accounts for the SNP bias of this custom array (see research design and methods for details). Of note, we observed a nominal enrichment of drug target genes among the nearest gene or the genes in LD to a set of 137 established or high-confidence T2D SNPs determined based on a mixture model of the Metabochip meta-analysis z scores (see research design and methods) (P = 0.003 and 0.04, respectively) (Table 5). Some of the new gene associations that replicated were ACSL1 and NFKB1 (TZD targets) and GIPR (which encodes the receptor for the incretin hormone GIP and reached genome-wide significance in the joint analysis of DIAGRAMv3 and Metabochip meta-analyses [3]).

Finally, we asked whether GSEA of human genetic association data could help to predict unintended, secondary phenotypic effects of drug treatment by testing for enrichment of associations in a drug target gene set, with a phenotype that is not directly targeted by the specific drug. We tested the hypothesis that targets of TZDs may be enriched for genetic associations with cardiovascular risk factors, such as circulating lipids. This is based on a potential effect of rosiglitazone on increased risk of myocardial infarction (35–37) (see FDA Advisory Committee Minutes at http://www.fda.gov/downloads/AdvisoryCommittees/CommitteesMeetingMaterials/Drugs/EndocrinologicandMetabolicDrugsAdvisoryCommittee/UCM369180.pdf) and on the observation that both rosiglitazone and pioglitazone affect lipid levels (22). Thus, we tested whether TZD target genes were enriched for multiple modest associations with circulating LDL-C, HDL-C, or triglyceride levels. We applied MAGENTA to three GWAS meta-analyses of LDL-C, HDL-C, and triglyceride blood levels across ∼100,000 individuals (33) and found that the 38 TZD targets were significantly enriched for genes associated with LDL-C levels with modest effect sizes (P = 0.0007, twofold enrichment) (Table 6). Of note, this result was specific to the TZD target set (see other antidiabetes drug class target sets in Table 6). We observed nominal enrichment of triglyceride levels among the TZD targets (P = 0.06) and the full set of antidiabetes drug targets (P = 0.03). No significant enrichment was found for HDL-C associations. In addition to the known LDL-C gene locus APOC3/APOA1, this analysis proposes that 8 of the 20 top-ranked genes based on their LDL-C association scores may be novel TZD targets associated with LDL-C plasma levels (listed in Supplementary Table 4).

Table 6.

Testing for potential nonglycemic effects of TZDs or other antidiabetes drug classes on global lipid plasma levels through GSEA of genetic associations

		LDL-C GWAS meta-analysis		Triglyceride GWAS meta-analysis		HDL-C GWAS meta-analysis
Antidiabetes drug target gene set	Number of genes analyzed*	Nominal MAGENTA enrichment P value	Excess number of genes above enrichment cutoff	Nominal MAGENTA enrichment P value	Excess number of genes above enrichment cutoff	Nominal MAGENTA enrichmentP value	Excess number of genes above enrichment cutoff	Antidiabetes drug target genes in LD to validated lipid SNPs‡
TZDs	38	0.0007†	9	0.06	4	0.35	1	APOA1 (TG, HDL-C, LDL-C), IRS1 (HDL-C, TG), SCARB1 (HDL-C)
All nine classes of drugs	94–95	0.01	10	0.03	8	0.22	3	APOA1 (TG, HDL-C, LDL-C), IRS1 (HDL-C, TG), SCARB1 (HDL-C)
Insulin	16	0.19	2	0.35	1	0.35	1	IRS1 (HDL-C, TG)
GLP-1 receptor agonists	17	0.65	0	0.10	3	0.24	0	—
DPP4 inhibitors	18–20	0.91	0	0.67	0	0.88	0	—

TG, triglyceride.

*Number of genes analyzed after excluding genes in HLA region and genes on sex chromosomes and correcting for physical clustering along the genome of genes within a given gene set. The number varies a bit between GWAS meta-analyses due to slight differences in SNP coverage between traits. Because of the physical proximity of APOC3 and APOA1, these genes were collapsed into one effective gene in the enrichment analysis (choosing the more significant gene P value) to prevent inflation of the gene set enrichment P value.

†Passes Bonferroni correction, P < 0.003.

‡Based on 95 loci associated with global lipid traits in (33).

Discussion

Meta-analyses of GWAS have yielded dozens of genetic variants that are overrepresented in cases of T2D compared with nondiabetic controls (3,6,38–44). The robustness of the evidence for their association with T2D is based on a stringent threshold for genome-wide significance that accounts for the number of independent tests that are possible among the ∼10⁶ common variants in the human genome (45). Because most common variants have modest effect sizes on common disease (e.g., disease risk odds ratio 1.05–1.10), although the adoption of this strict standard minimizes type I error, it leads to a high number of false-negative associations that remain undetected. The design of custom arrays that facilitate large-scale replication genotyping in many samples (12) can rescue some of these signals through increased power, but a large fraction (numbering in the hundreds to low thousands by some estimates) are yet to be discovered (3,4,39).

There is, therefore, a need to integrate additional tools to mine GWAS data sets in a hypothesis-driven but systematic manner, which can raise the prior probability of association while maintaining quantitative statistical standards. Other domains of biology can be brought to bear on GWAS data under the reasonable assumption that not every variant in the genome carries the same low prior probability of association with a given phenotype. In addition, queries that set each gene as the functional biological unit (gene-based tests) can increase statistical and explanatory power by considering association statistics of all variants that span a given gene and by collectively analyzing sets of genes that function in common pathways. The adaptation of GSEA to GWAS data sets as embedded in MAGENTA accomplishes all these goals, accounting for differences in genetic and physical properties between genes (34).

In this study, we applied MAGENTA to gene sets constructed under the reasonable assumption that genes whose protein products are targeted by drugs used to treat T2D are likely to influence glycemia when modified by naturally occurring variation. The present analysis was predicated on the initial observation that the first two reported and confirmed genetic associations with T2D implicated genes that encode T2D drug targets (14,15). Based on the empirical evidence of these existing T2D associations, we postulated that variants in other genes that encode drug targets have a higher likelihood of association with T2D than the genomic average, if they or variants in LD to them affect function or expression of the gene. Indeed, we found significant enrichment of modest common variant associations with T2D in pathways targeted by antidiabetes medications. The results were replicated in an independent study genotyped on the Metabochip, persisted after removing validated T2D loci, and suggest that 15–20 of the top-41 modest associations with T2D prioritized by the present GSEA are worthy of further investigation (about one-half of the 41 top-ranked loci listed in Table 4), mostly driven by genes involved in insulin and TZD signaling. These results are also consistent with the nominal enrichment observed in both the DIAGRAMv3 and the Metabochip meta-analyses for T2D associations in the PPAR signaling pathway, known to be targeted by TZDs (P < 0.04) (3). These findings highlight insulin sensitivity networks as a common nidus of potential T2D associations.

In addition, some of the top-ranked T2D associations lie near genes that encode hormones and their receptors, such as adiponectin and its receptor and incretins and their receptors (e.g., GIP, GIPR). Of note, some of the antidiabetes target genes are monogenic diabetes genes, including PDX1, INSR, KCNJ11, ABCC8, and PPARG (the latter three are also associated with the common form of T2D).

Although the present GSEA helped to hone in on a shortened list of candidate T2D-associated genes, additional functional analyses and experiments will be required to decipher which 15–20 of the ∼40 top-ranked drug target genes represent true T2D associations. Possible approaches include examining their expression levels in relevant T2D tissues or elucidating the phenotypic consequences of perturbing these genes in model systems. Further genetic studies with larger sample sizes may provide additional statistical support. Genes such as ACSL1, NFKB1, and GIPR that replicated in two independent genetic studies (DIAGRAMv3 and Metabochip) are top candidates for follow-up.

Because a large fraction of drugs that enter clinical trials today fail due to toxicity (46), we examined a test case to gauge the utility of this approach for detecting secondary, undesired phenotypic effects of drugs. Specifically, we provided human genetic support for a potential causal role of LDL-C blood level alterations in the potential increased incidence of myocardial infarction in people with T2D following treatment with TZDs (35). Although LDL-C blood levels have been shown to increase in response to TZDs (22), the clinical significance of this process is not yet clear. Further investigation of the top-ranked TZD target genes that are most likely to be associated with LDL-C based on MAGENTA analysis and that drove the observed enrichment signal may help to shed light on the LDL-dependent mechanism through which TZDs may affect risk of myocardial infarction in people with T2D. In the future, unbiased mining of genetic associations with a range of common diseases and traits may help to propose putative side effects of drugs for testing during drug development.

The finding that common DNA variants in genes that encode known drug targets are enriched for T2D associations supports the reciprocal notion that existing genetic associations from GWAS could guide us to novel relevant drug target genes or pathways. Furthermore, the present work may have useful applications for future genetic, pharmacogenetic, or drug development studies: 1) We expect that ongoing deep sequencing studies and/or larger GWAS focusing on functional variation might uncover novel genetic variation in our prioritized T2D-associated loci, 2) we highlight drugs and targets worthy of dedicated pharmacogenetic studies that might help to stratify the population into likely responders and nonresponders, 3) we suggest potential alternative drug targets for established drug classes, and 4) we provide additional evidence that might help to prioritize some of these genes in future GWAS for drug response if a suggestive signal of association with T2D is detected at one of these loci. One such example is SLC22A1, which encodes a liver-specific metformin transporter, because it is nominally associated with metformin response (47) and is among our top-predicted T2D-associated drug target genes.

The approach can be refined with more granular types of drug target definitions, such as genes whose expression varies in response to drug perturbations in relevant tissues or cell types, and can be applied to any complex disease or quantitative trait with available GWAS data and knowledge of drug targets. Of note, in concordance with the present findings, a recent study found that genes associated with rheumatoid arthritis or genes that interact with the disease genes through protein-protein interactions are enriched for targets of approved drugs for rheumatoid arthritis (48). An extension of this approach to examining low frequency or rare variation in drug target genes and pathways may be instrumental for personalized treatment design.