Is Genetic testing useful to predict type 2 diabetes?

Abstract

The early identification of individuals at risk for type 2 diabetes (T2D) enables prevention. Recent genome-wide association studies (GWAS) have added at least 40 genetic variants to the list of already well characterized T2D risk predictors, including family history, obesity, and elevated fasting plasma glucose levels. Although these variants can significantly predict T2D alone and as a part of genotype risk scores, they do not yet offer clinical discrimination beyond that achieved with common clinical measurements. Future progress on at least two research fronts may improve the predictive performance of genotype information. First, expanded GWAS efforts in non-European populations will allow targeted sequencing of risk loci and the identification of true causal variants. Second, studies with longer prediction time horizons may demonstrate that genotype information performs better than clinical risk predictors over a longer period of the life course. At present, however, genetic testing cannot be recommended for clinical T2D risk prediction in adults.

Introduction

The Human Genome Project listed among its benefits the “earlier detection of genetic predispositions to disease”.¹ The decade that followed the sequencing of the human genome has seen the generation of enough genetic data to occupy the scientific community for decades more. Before the current genomic era, the prevalent genetic paradigm concerned monogenic and highly penetrant diseases such as cystic fibrosis and Huntington disease. Nonetheless, a considerable body of evidence had also suggested a large heritable component to complex diseases such as coronary artery disease, obesity, and diabetes mellitus. Now, large genome-wide association studies (GWAS) have begun to uncover some of the genetic determinants of these complex diseases. Current genetic discoveries challenge us to find their appropriate place in disease risk prediction, particularly in the context of other already well characterized clinical disease predictors.

Type 2 diabetes mellitus (T2D) exemplifies the complex interactions between environmental and genetic determinants. The increasing prevalence² of T2D and its associated comorbidities and healthcare expenditures make it a public health priority. Fortunately, the incidence of T2D can be delayed or prevented by maintaining healthy lifestyle behaviors.^3–5 However, the increasingly obesogenic global environment^6,7 is at odds with people’s attempts at lifestyle modification. The identification of population subgroups at particularly high risk for T2D might facilitate the targeting of prevention efforts to those who might benefit from them the most. Risk prediction has this identification and prevention as its goals.

It is reasonable to ask whether recent genetic associations improve the T2D risk prediction already achieved with traditional risk predictors alone. This review will summarize the data that address this question. We will review the known genetic and non-genetic risk predictors of T2D and compare the prediction models that incorporate them. We will then identify current limitations and future directions for the field of T2D genetics. In this review, we will use the term risk predictor to indicate a characteristic significantly associated with a future outcome (here, incident T2D). In contrast, a risk factor may be associated with a current (prevalent) and/or incident outcome.

Genetic associations with type 2 diabetes

The pace of accumulation of known T2D variants has mirrored the pace of advances in genetic technology. In 2003, linkage analysis (see Glossary) in samples of European ancestry found a T2D-associated variant on chromosome 10q,⁸ later identified in the TCF7L2 gene, which encodes a transcription factor in the Wnt signaling pathway.⁹ To date, it remains the variant with the largest effect size; each copy of the T allele (see Glossary) at rs7903416 carries a 1.4–1.5 increased odds of type 2 diabetes.¹⁰ Also by 2003, candidate gene studies (see Glossary) had associated T2D with variants in PPARG and KCNJ11, both genes encoding targets of diabetes medications.^11,12 The arrival of GWAS, however, considerably lengthened the list of T2D-associated variants. The sequencing of the entire human genome had made possible the development of genotyping platforms consisting of only a few hundred-thousand single-nucleotide polymorphisms (SNPs) that could statistically capture the majority of genetic variation in the human genome. By genotyping large samples of cases and controls on these platforms, investigators could scan the entire genome in a non-biased way for novel genetic associations with T2D.¹³ The first T2D GWAS in 2007 confirmed the association with TCF7L2 and identified new T2D associations with variants at SLC30A8 and HHEX.¹⁴ The same year, other groups replicated these findings and discovered additional associations at CDKAL1, IGF2BP2, and CDKN2A/B.^15–17 When research groups merged to form the Diabetes Genetics Replication and Meta-Analysis (DIAGRAM) consortium, the increased statistical power enabled the discovery of variants with smaller effects sizes.¹⁸ The most recent publication by the DIAGRAM group has brought the number of independent replicated genetic T2D associations to at least 38,¹⁹ with most variants increasing an individual’s relative odds of T2D by a modest 5–15% (Table 1).¹³

Table 1.

Thirty-eight genetic variants associated with type 2 diabetes (T2D) at genome-wide significance.

Locus	Marker	Chr	Type of variant	Risk allele/ other allele	Risk allele frequency	OR for T2D (95% CI)
KCNQ1	rs223789274 a	11	Intronic	C/T	0.61b	1.40 (1.34–1.47)
TCF7L2	rs79031469	10	Intronic	T/C	0.25	1.37 (1.28–1.47)18
DUSP9	rs594532619	X	8 kb upstream	G/A	0.12	1.27 (1.18–1.37)
CDKN2A/B	rs1081166115–17	9	125 kb upstream	T/C	0.79	1.20 (1.14–1.25)
IRS1	rs294364175	2	502 kb upstream	C/T	0.61	1.19 (1.13–1.25)
IGF2BP2	rs440296015–17	3	Intronic	T/G	0.29	1.17 (1.10–1.25)18
FTO	rs805013615, 16	16	Intronic	A/C	0.45	1.15 (1.09–1.22)18
KCNJ11/ABCC8	rs5219c/rs75711012 c	11	Missense:Glu23Lys/Ala1369Ser	T/C	0.50	1.15 (1.09–1.21)17
				G/T	0.40
THADA	rs757859718	2	Missense: Thr1187Ala	T/C	0.92	1.15 (1.10–1.20)
CENTD2	rs155222419	11	5′ UTR	A/C	0.88	1.14 (1.11–1.17)
PPARG	rs180128211	3	Missense: Pro12Ala	C/G	0.92	1.14 (1.08–1.20)15–17
HHEX	rs111187514	10	7.7 kb downstream	C/T	0.56	1.13 (1.08–1.17)17
NOTCH2	rs1092393118	1	Intronic	T/G	0.11	1.13 (1.08–1.17)
WFS1	rs180121419	4	Intron-exon junction	G/A	0.27	1.13 (1.07–1.18)
ADCY5	rs1170806776	3	Intronic	A/G	0.78	1.12 (1.09–1.15)
CDKAL1	rs775484015–17	6	Intronic	C/G	0.31	1.12 (1.08–1.16)
HNF1B	rs75721077, 78	17	Intronic	A/G	0.43	1.12 (1.07–1.18)19
SLC30A8	rs1326663414	8	Missense: Arg325Trp	C/T	0.75	1.12 (1.07–1.16)19
CDC123/CAMK1D	rs1277979018	10	Intergenic region	G/A	0.23	1.11 (1.07–1.14)
RBMS1/ITGB6	rs759373079	2	Intronic	C/T	0.57	1.11 (1.08–1.16)
TLE4	rs1329213619	9	234 kb upstream	C/T	0.93	1.11 (1.07–1.15)
HMGA2	rs153134319	12	43 kb upstream	C/G	0.10	1.10 (1.07–1.14)
JAZF1	rs86474518	7	Intronic	T/C	0.52	1.10 (1.07–1.13)
ADAMTS9	rs460710318	3	38 kb upstream	C/T	0.81	1.09 (1.06–1.12)
MTNR1B	rs1083096376	11	Intronic	G/C	0.30	1.09 (1.06–1.12)
TSPAN8/LGR5	rs796158118	12	Intronic	C/T	0.23	1.09 (1.06–1.12)
BCL11A	rs24302119	2	99 kb downstream	A/G	0.46	1.08 (1.06–1.10)
KCNQ1	rs23136219 a	11	Intronic	G/A	0.52	1.08 (1.06–1.10)
ZBED3	rs445705319	5	41 kb upstream	G/A	0.26	1.08 (1.06–1.11)
GCK	rs460751776	7	36 kb upstream	A/G	0.20	1.07 (1.05–1.10)
KLF14	rs97228319	7	47 kb upstream	G/A	0.55	1.07 (1.05–1.10)
OASL/HNF1A	rs795719719	12	20 kb downstream	T/A	0.85	1.07 (1.05–1.10)
PRC1	rs804268019	15	Intronic	A/C	0.22	1.07 (1.05–1.09)
PROX1	rs34087476	1	2 kb upstream	C/T	0.50	1.07 (1.05–1.09)
DGKB/TMEM195	rs219134976	7	Intergenic region	T/G	0.47	1.06 (1.04–1.08)
GCKR	rs78009476	2	Intronic	C/T	0.62	1.06 (1.04–1.08)
TP53INP1	rs89685419	8	Intronic	T/C	0.48	1.06 (1.04–1.09)
ZFAND6	rs1163439719	15	1.5 kb downstream	G/A	0.56	1.06 (1.04–1.08)

SNPs are ordered by odds ratio (OR) for T2D. Reference for each SNP is from the original report, and OR for T2D is from that reference cited unless otherwise noted. Risk allele frequencies are from HapMap CEU unless otherwise noted. Chr, chromosome. CI, confidence interval. kb, kilobase. Adapted from Billings LK, Florez JC. Ann NY Acad Sci. 2010 Nov; 1212:59–77.

^aThese SNPs are in low linkage disequilibrium (LD) in Europeans (r² = 0.01) and likely represent independent association signals.

^bRisk allele frequency from HapMap-JPT.

^cThese two SNPs are in strong LD in adjacent genes in Europeans (r² = 0.87).

Traditional type 2 diabetes risk predictors

As an “exposure” in the epidemiologic sense, a T2D-associated genetic variant is fixed from conception, and this permanence should give it some predictive advantage over other risk predictors that vary over the life course, such as body-mass index (BMI) and even family history. Nonetheless, routine clinical T2D risk prediction has incorporated inexpensive and easily measured predictors, which, as a result, are the gold standard against which any new predictor has to be judged.

T2D risk increases with age throughout adulthood,²⁰ and members of non-white race/ethnic groups have increased risk compared with whites.^21,22 Lifestyle factors associated with an increased T2D risk include cigarette smoking,²³ physical inactivity²⁴ and diets with high glycemic load²⁵ or trans-fat content.²⁶ Elements of the medical history also predict T2D: a family history of diabetes^20,27 or a personal history of gestational diabetes²⁸ markedly increase an individual’s personal T2D risk. Obesity strongly increases an individual’s risk for T2D: a BMI ≥ 30 kg/m² is associated with an odds ratio of 6.4 for T2D over 7 years in the Framingham Offspring Study, after adjustment for age and family history, compared with a BMI <25 kg/m².²⁰ This direct relationship between T2D risk and adiposity persists even in the normal range of BMI values. In the Nurses Health Study, the age-adjusted relative risk of T2D over 14 years was 5.0 in BMI range 24.0–24.9 ≥ 30 kg/m² compared with BMI values <22 kg/m².²⁹ Elevated values of blood pressure, another common clinical measurement, are also associated with risk of future T2D.^20,30

Not surprisingly, the strongest laboratory predictors of T2D measure glycemia. Higher fasting plasma glucose levels, even in the normal range, predict future T2D,³¹ while individuals with impaired fasting glucose (IFG, 100–125 mg/dL or 5.6–7.0 mmol/L) have an annualized relative risk of T2D of 4.7 compared to normoglycemic individuals. Impaired glucose tolerance (IGT) after an oral glucose tolerance test performs similarly as a predictor, although the combination of IFG and IGT carries much greater risk than either alone.³² In one Dutch study, 65% of people with IFG and IGT at baseline developed T2D over a mean follow-up of 6.4 years.³³ Elevated values of hemoglobin A1c also predict future T2D, and a recent meta-analysis demonstrated that values of 6.0–6.5% were associated with an annual T2D incidence of 25–50% over 5 years.³⁴ Measures of dyslipidemia, particularly elevated triglyceride and low high-density lipoprotein (HDL) levels, also predict future T2D independently of glycemic measures.^20,31

Clinical type 2 diabetes prediction models

While some of these T2D risk predictors, particularly lifestyle and dietary factors, are important for public health policy, they are less easily measured in busy clinical settings. Groups from several large prospective cohort studies, such as the Framingham Offspring Study and the Atherosclerosis Risk in Communities Study (ARIC), have developed prediction models for T2D in middle-aged adults using common clinical measurements.^20,35–37 These models have more similarities than differences and usually include measures of glycemia, adiposity, dyslipidemia, and blood pressure, in addition to age, family history of diabetes and, in some cases, race. The ability of these models to discriminate future cases from non-cases is usually very good, with C statistics (also called areas-under-the-receiver operating characteristic curve, see Glossary) ranging from 0.7 to 0.9 (Fig. 1). When these models are applied to racial groups other than those in the original validation studies, recalibration can preserve their discriminatory ability.³⁸ Other clinical models have additionally included lifestyle predictors such as fruit and vegetable consumption and physical activity, with modest improvement in prediction.^39,40

Fig. 1.

ROC curves for five type 2 diabetes prediction models from Framingham Offspring Study. The four clinical models have higher C statistics than the personal model. Reprinted with permission from Wilson, et al. Arch Intern Med. 2007;167(10):1068–1074. Copyright © 2007 American Medical Association. All rights reserved.

Genetic prediction of type 2 diabetes

Individual genetic variants, too, can predict incident T2D. In the Swedish Malmö Preventive Project, 11 SNPs individually predicted incident T2D in more than 15,000 participants over a median 24.8 years of follow-up, after adjustment for sex and baseline age. The odds ratio per risk allele ranged from 1.07 for the variant near HHEX to 1.30 for that in TCF7L2, compared with 1.67 for having a first-degree family member with diabetes and 1.84 for each standard deviation increase in BMI.⁴¹

These relatively modest effect sizes of individual SNPs have prompted efforts to combine SNPs into composite genotype risk scores. To construct these scores, an individual’s risk alleles at each T2D-associated locus (ranging from 0 to 2 copies at each locus) are summed. Commonly, each locus is weighted according to its T2D effect size from the original GWAS. For instance, with this approach, each TCF7L2 risk allele carries a greater weight than each HHEX risk allele. A simulation study by Janssens et al., however, suggests that such weighting might not substantially improve the C statistic for alleles with small effects sizes (odds ratio 1.10–1.25).⁴² To date, several T2D genotype risk scores have been published in prospective studies of middle-aged adults (reviewed in⁴³ and⁴⁴), and the results have been similar. For population-level T2D prediction, prospective cohort studies (Table 2) are preferable to crosssectional case–control designs, since the latter do not reflect the true disease incidence in the population. Genetic risk scores alone, consisting of between 2 and 40 SNPs, have C statistics for T2D prediction ranging from 0.54 to 0.68. In the Framingham Offspring Study, cumulative T2D incidence over 28 years of follow-up increased significantly with genotype risk score (Fig. 2), and each 1-point increase in the score increased the odds of T2D over 8 years by 12%. The group with the highest genotype scores had an odds ratio for T2D of 2.6 compared to those with the lowest scores.⁴⁵

Table 2.

Eight prospective studies comparing clinical and genetic type 2 diabetes prediction models.

Study	n	Clinical model and AUC		Genetic model (no. of SNPs) and AUC		Clinical + genetic model AUC	p
Rotterdam46	5297	Age, sex, BMI	0.66	18	0.60	0.68	<0.0001
Malmö41	16,061	Age, sex, BMI, FG, BP, TG, FH	0.74	11	0.63	0.75	0.0001
Botnia41	2770	Age, sex, BMI, FG, BP, TG, FH	0.79	11	0.68	0.80	0.23
Framingham Offspring45	2377	Age, sex, BMI, FG, BP, TG, FH	0.900	18a	0.581	0.901	0.49
Framingham Offspring47	3471	Age, sex, BMI, FG, BP, TG, FH	0.903	40a	0.606	0.906	0.04
Whitehall II80	5135	Age, sex, BMI, FG, BP, TG, FH	0.78	20	0.54	0.78	0.10
DESIR81	3442	Age, sex, average BMI over 9 years	0.82	3	0.56	0.83	NR
DESIR37	3817	WC, HTN, FG, GGT, FHD, smoking (men), BMI (women)	0.850 (men)	2	NR	0.851 (men)	NR
			0.917 (women)			0.912 (women)

P value compares areas under the receiver operating characteristic curves (AUCs) from clinical + genetic model to clinical model alone. BMI, body-mass index. BP, blood pressure. FG, fasting plasma glucose. FH, family history of diabetes. GGT, gamma-glutamyl transferase. HTN, hypertension. NR, not reported. TG, triglycerides. WC, waist circumference.

^aGenetic model also adjusted for sex.

Fig. 2.

The 28-year cumulative incidence of type 2 diabetes by genotype risk score in 2377 participants in the Framingham Offspring Study.⁴⁵ Reprinted with permission from the Massachusetts Medical Society, Copyright^© 2008.

However, the clinical T2D prediction models that consist of basic demographic, clinical, and laboratory predictors have C statistics ranging from 0.66 in the Rotterdam Study⁴⁶ to 0.90 in the Framingham Offspring Study (Table 2),^45,47 values superior to what genotype scores alone have yet achieved. Moreover, the addition of genotype risk scores to clinical prediction models only modestly improves the C statistic. For example, the C statistic improves from 0.903 to 0.906 with the addition of a 40-SNP score to the clinical model in the Framingham Offspring Study,⁴⁷ and from 0.74 to 0.75 in the larger Malmö Preventive Project.⁴¹

Although the C statistic has relevance for population-level prediction, it is not the best measure of clinical utility. A higher C statistic corresponds to a higher likelihood that a test can distinguish an individual who will develop incident T2D from one who will not. But for T2D prevention, the clinician is asked to risk-stratify an individual patient, not to choose which one of a pair of patients is likelier to develop T2D. For this reason, Pencina, et al., introduced the net reclassification improvement (NRI) index,⁴⁸ which compares the abilities of two prediction models to put individuals correctly into predicted risk categories (see Glossary). Using T2D risk categories of 0 to <2%, 2 to 9003C;8%, and ≥8% over 8 years in the Framingham Offspring Study, the addition of an 18-SNP genotype risk score to a model consisting only of sex and family history correctly reclassified only 2.6% of individuals (p = 0.22).⁴⁵ Genetic risk prediction may work better in younger individuals. In Framingham, the addition of a 40-SNP score to a full clinical model, achieved an NRI of 10.2% (p = 0.001) among those younger than 50 years at baseline.⁴⁷ It is important to note, however, that these categories of risk (0 to <2%, 2 to <8%, and ≥8% over 8 years) do not inform widely-accepted clinical prevention targets. This is in contrast to the Framingham Risk Score for 10-year risk of coronary disease, whose categories help tailor individual cholesterol level targets.⁴⁹ Genetic testing will likely not have a role in clinical T2D prediction unless its addition to prediction models correctly reclassifies individuals as having lower or higher risk than previously thought based on patient phenotype, and unless the prevention strategies targeted for the individual change as a result of the marginal information value afforded by the genetic test. Recently, a large multi-disciplinary European consortium recommended against the use of genetic information in targeting T2D preventive efforts.⁵⁰

Family history and heritability of type 2 diabetes

If genetic information does not improve T2D prediction compared with clinical prediction models, a next reasonable question is how genetic information compares with family history of T2D alone, itself a strong risk predictor. In the pre-genomic era, twin and family studies played an important role in separating the putative environmental and genetic components of T2D. Twin studies have estimated the genetic proportion of variance in T2D (or heritability, see Glossary) to be between 25% and 40%,^51,52 which suggests that both genetic and non-genetic factors contribute substantially to an individual’s T2D risk. Having one parent with T2D doubles an individual’s risk, and having two affected parents can increase an individual’s risk up to six-fold.^20,27 Family history does correlate with T2D genetic risk: data from the PPP-Botnia and Framingham Offspring Studies show that T2D genetic risk scores increase slightly with the degree of T2D family history (i.e. 0, 1, or 2 affected parents).⁵³ However, the known T2D genetic variants do not account for the strong relationship between family history and T2D risk. In the Malmö Preventive Project, a self-reported first-degree family history of diabetes carried an odds ratio for incident T2D of 1.62 after adjustment for clinical predictors. When added to this multivariate model, the genetic risk score was an independent predictor of T2D (odds ratio 1.12 per 1-point increase in score), but the effect of family history was unchanged (odds ratio 1.65).⁴¹ In the Framingham Offspring Study, the addition of a genetic risk score to the clinical T2D prediction model modestly decreased the adjusted odds ratio for first-degree family history from 1.72 to 1.64.⁴⁵ Indeed, the estimate of the degree of T2D heritability explained by known genetics variants is only 10%.⁵⁴ Measurement error in both the numerator and denominator of this heritability proportion may account for the discrepancy between family history and genetic variants as T2D predictors. First, a greater percentage of heritability (numerator) may be explained with the future discovery of rare and private variants and the identification of the true causal variants for which the discovered T2D-associated variants are imperfect proxies (see below). Second, the total T2D heritability (denominator) may be overestimated, since its calculation ignores the potential contributions of intrauterine effects, gene–gene and gene-environment interactions, and the insufficient accounting for shared environmental, behavior and normative factors in family studies.^55,56 However, the evidence above demonstrates that, compared with currently identified genetic variants, family history remains a more powerful T2D predictor that likely captures the genetic and environmental determinants of T2D risk.

Limitations of the current evidence base

The evidence to date does not justify the routine use of genetic testing for T2D prediction in clinical care. At present, certain limitations of the evidence base contribute to the modest performance of genotype information as a risk predictor. GWAS, including those that discovered the known T2D variants, rely on genotyping array platforms that may miss up to 20% of common variation in the human genome.⁵⁷ These platforms fail to capture an even greater percentage of common variation in non-European ancestral populations, due to lower linkage disequilibrium (LD, see Glossary) rates in those groups.⁵⁸ It is also worth noting that the majority of known T2D-associated variants (Table 1) are in non-coding intronic or even intergenic regions of the genome.¹³ This observation suggests that these variants may only be proxies in LD with the true causal variants. Fine-mapping and resequencing these regions should identify those causal variants, which should have stronger associations with T2D than their proxies and thus improve genetic prediction. This anticipated progress in defining the genetic architecture of T2D highlights the preliminary nature of the current state of genetic testing for T2D prediction. In an analysis of 5297 participants of the Rotterdam study, updating a T2D genetic risk test with the inclusion of more recently discovered T2D-associated variants changed the risk category (above or below average risk) for 34% of people.⁵⁹ Clinicians may engender mistrust from their patients if the genetic risk predictions they deliver are overturned by the next wave of genetic discovery.

Future directions

Although GWAS have likely identified most common (risk allele frequency ≥5%) T2D-associated loci in populations of European ancestry, T2D genetics is still an active area of research. The nascent search for interactions between genes and between genes and environmental factors (including diet⁶⁰ and physical activity⁶¹) has necessitated new analytic methods to preserve statistical power when looking for potentially small interaction effect sizes among trillions of combinations.⁶² It is not known how these future discoveries might be incorporated into risk prediction models.

Extending T2D genetic knowledge to non-European ancestral populations

The largest gap in the current evidence base for T2D genetic prediction is the racial homogeneity of studies to date. The largest T2D GWAS meta-analyses have taken advantage of the lesser degree of LD in European populations to identify genetic associations with T2D. Thus it is appropriate that T2D genetic prediction studies (Table 2) first applied these genetic discoveries to T2D prediction in populations of European ancestry similar to the ones in which they were discovered. Genetic prediction modeling in other racial groups, however, lags behind. It is well known that the risk allele frequencies for many cardiometabolic traits vary by race and ethnicity.^63–65 However, genetic association studies are generally finding that T2D-associated variants from European GWAS have similar effects in non-European populations.⁶⁶ Still, the European T2D-associated SNPs are likely much poorer proxies for the true causal variants in older ancestral populations such as Africans, which have more fragmented genomic LD. The Candidate gene Association Research Project (CARe), the largest T2D GWAS reported to date in African ancestral populations consisted of 8090 people,⁶⁷ compared to more than 140,000 in the most recent DIAGRAM meta-analysis.¹⁹ Of all the known T2D association from European populations, this consortium replicated only that at TCF7L2. Larger GWAS in more diverse populations will bring at least two advantages. First, it will extend the knowledge of T2D genetic architecture to those racial and ethnic groups disproportionately affected by the disease.²² Second, any T2D-associated variant replicated or discovered in older populations such as Africans will likely be in an LD region smaller than the corresponding European region, narrowing the genome necessary to map and resequence for identification of the causal variant(s).⁶⁸ The inclusion of these causal variants, instead of their originally identified proxy SNPs, have the potential improve T2D genetic prediction models for individual patients.

Specifying the time horizon of T2D genetic risk prediction

To address the limitations of the evidence base in genetic prediction, the U.S. Centers for Disease Control and Prevention sponsored the Genetic Risk Prediction Studies (GRIPS) Group. In 2011, GRIPS made several recommendations for standardizing the presentation of results from genetic prediction studies, including specifying the time horizon over which each risk prediction model applies.⁶⁹ Applying this recommendation to T2D genetic risk prediction highlights another research gap in the field: the limited range of ages and durations of follow-up over which T2D genetic prediction has been modeled. Fig. 3 shows that the prospective studies of T2D genetic prediction include participants with baseline ages generally in middle adulthood and, with the exception of the Malmö study, have follow-up durations of around ten years. This time horizon is familiar to clinicians who use the Framingham Risk Score for 10-year coronary risk stratification.⁴⁹ However, because obesity is increasingly affecting younger age groups⁷⁰ and T2D can be prevented with lifestyle modification,^3–5 a prediction model validated in younger age groups with a longer prediction time horizon may help target earlier prevention. A longer time horizon is also likely to improve the predictive value of genetic testing relative to other T2D risk predictors that vary with time, such as BMI and fasting glucose. Indeed, the Framingham Offspring Study showed that a 40-SNP genotype risk score significantly improved the T2D net reclassification improvement in participants younger than, but not older than, 50 years at baseline.⁴⁷ T2D genetic risk testing may prove to have clinical utility only in populations younger than those in whom prediction models have been published to date.

Fig. 3.

Baseline ages and time horizons of six prospective studies reporting genetic prediction of type 2 diabetes.^{41,46,47,80,81} Arrow origins and bars correspond to mean age +/−1 standard deviation and arrow length corresponds to mean follow-up time, unless otherwise noted. ^aMedian baseline age and interquartile range are shown. ^bMedian follow-up time is shown.

Conclusions

Genetic testing has many potential future roles in the clinical management and prevention of T2D. It may encourage patients to engage with their health and risk factors and motivate them for lifestyle modification.^71,72 Clinicians may be able to tailor T2D therapy to individual genotypes.⁷³ Genotype information does significantly predict T2D, but it is outperformed in adults, at least, by measurements already routinely incorporated into clinical medicine. At present, this lack of utility does not support the clinical use of genetic testing for T2D prediction in adults.

Practice points.

• About 40 genetic variants can significantly predict T2D, although, individually, most variants increase the odds of T2D only modestly (5–15%). “High” versus “low” genetic risk scores comprised of sums of variants are associated with a ~2-fold increased relative risk of incident T2D.

• Prediction models consisting of routine clinical risk factors such as family history, BMI, and elevated fasting plasma glucose outperform genetic prediction models in middle-aged adults of European ancestry.

• The addition of genotype information to clinical T2D prediction models does not meaningfully improve prediction for middle-aged adults of European ancestry.

Research agenda.

• Expanded GWAS efforts in non-European populations will allow targeted sequencing of risk loci and the identification of true causal variants that are likely to improve predictive performance for individuals.

• Studies with longer prediction time horizons may demonstrate that genotype information performs better than clinical risk predictors over a longer period of the life course.

Glossary

Allele

A possible form that a particular region of DNA may take in an individual’s genome. For example, an individual may have a T allele at rs7903146 in his maternal copy of TCF7L2 and a C allele in his paternal copy. In addition to single-nucleotide polymorphisms (SNPs) such as this example, allele may also refer to other forms of genetic variation, including deletions and insertions.

C statistic

Also known as the area under the receiver operating characteristic (ROC) curve (Fig. 1). For a given model that aims to predict an outcome (for example, incident type 2 diabetes), the ROC curve plots the sensitivity against 1 minus the specificity at all possible dichotomous cutpoints of the model. The area under that curve corresponds to the model’s discrimination, or the probability that it will distinguish a case from a non-case. The C statistic ranges from 0.5 (no discrimination, similar to a coin toss) to 1 (perfect discrimination).

Candidate gene study

A hypothesis-driven method to identify genotype-phenotype associations by examining genetic variations only within genes which, because of their known function or protein product, are thought likely to be implicated in the phenotype of interest.

Heritability

The proportion of the population variance in a phenotype (e.g. type 2 diabetes) due to genetic effects, often calculated from twin, family, and adoption studies. Underlying these studies is the assumption that siblings reared together share the same environment, at least in childhood, and thus any phenotypic variance between them is due to unshared genetic factors.

Genome-wide association study

A hypothesis-free method of identifying genotype-phenotype associations. In a large population of cases and controls, the phenotype of interest is tested for statistical association with each of several hundred-thousands of single nucleotide polymorphisms (SNPs), which because of linkage disequilibrium (LD), capture a majority of human genetic variation. To minimize type I error, a genome-wide association is considered statistically significant at p<5×10⁻⁸.

Linkage analysis study

A method to identify genotype-phenotype associations by looking at the inheritance patterns of certain genetic markers in families with high prevalence of the phenotype of interest.

Linkage disequilibrium (LD)

The state when two alleles are inherited together more frequently than by chance because of their close proximity on a chromosome. LD is often measured by r², the square of the correlation coefficient for the two allele frequencies, which ranges from 0 (linkage equilibrium) to 1 (perfect linkage disequilibrium). If LD is present between two loci, an allele at one genotyped locus can be a reasonable proxy for an allele at another, non-genotyped locus. Because recombination events degrade LD over generations, older ancestral groups such as Africans have lower LD values for neighboring alleles in a given chromosomal region than European ancestral populations.

Net reclassification improvement (NRI) index

A measure of whether a new prediction model meaningfully improves upon a previous model by correctly reclassifying individuals into risk categories (e.g. low-, medium-,and high-risk). The NRI is higher as it correctly upgrades the risk category of future cases (for example, incident type 2 diabetes) and correctly downgrades the risk category for non-cases. The NRI is lower as it incorrectly downgrades the risk category of future cases and incorrectly upgrades the risk category of non-cases.

Phenotype

An observed trait resulting from gene expression. Phenotypes can be discrete diseases like type 2 diabetes and coronary artery disease or quantitative traits such as fasting glucose or low-density lipoprotein (LDL) levels.