Type 2 Diabetes (T2D) is one of the fastest growing diseases in the United States and other developed nations (Nugent, 2008; Hu, 2011). In the last three decades, the number of Americans diagnosed with diabetes has tripled (from 5.6 to 20.9 million), making this a public health concern (CDC (Center for Disease Control), 2013). T2D is a chronic metabolic disease, characterized by high levels of glucose in the blood, and frequently caused by a deficiency of insulin secretion and/or the development of insulin resistance (the inability of cells to respond to the insulin). If not treated properly, it can produce kidney failure, blindness, and circulatory problems. (Manzella, 2007; Buijsse et al., 2011; Hu, 2011; Sanghera and Blackett, 2012). The interplay of environmental (i.e., sedentary life, obesity, lack of exercise, poor diet) and genetic factors (i.e., familial contribution), contribute to the etiology and epidemy of T2D, in addition to an estimated heritability of 26% (Poulsen et al., 1999). Since 2007, Genome Wide Association Studies known as GWAS, have identified and confirmed more than 50 loci associated with the development of T2D (Steinthorsdottir et al., 2007; Lindgren et al., 2009; Shu et al., 2010; Voight et al., 2010; Morris et al., 2012). Several genes identified so far are involved in encoding proteins necessary for insulin secretion, glucose metabolism, and beta-cell function, which are components that enable insulin production and insulin receptor activation in the body (Sladek et al., 2007; Steinthorsdottir et al., 2007; Yasuda et al., 2008).

Previous studies that have included genetic profiling and scores in T2D preventive models, have shown only a slight increase in predictive ability. Generally, the use of genetic variants provides a small contribution in terms of prediction accuracy due to their small effects, especially if compared to the use of age and clinically measured variables, such as BMI, and triglyceride levels and known risk factors for this disease (Saxena et al., 2007; Lyssenko et al., 2008; Voight et al., 2010; Vazquez et al., 2012). As of today, while there is excitement with the possibility of a more personalized medicine, medical professionals do not consider genotypic information as a variable in assessing patients' risk of developing T2D (Katsios, 2010; Lyssenko and Laakso, 2013). In several studies where risk assessment models have been built and tested, a few deficiencies have been noticed that could possibly have influenced their models' predictive ability. Such deficiency may arise due to the use of a model that so far does not capture the complexity of polygenic signals and their interaction with covariates. In addition, an ideal risk assessment model would incorporate the interplay of a substantial number of small-effect genes and several phenotypic variables (e.g., BMI) related to the development of T2D in order to get a more realistic and precise prediction (Lindstrom and Tuomilehto, 2003). However, by incorporating other phenotypes (also heritable) into the risk assessment models, pleiotropic genetic effects shared by both traits could be explained. BMI is an easy to measure phenotype, highly associated to diabetes and obesity and shown to be a strong predictor of diabetes (Lyssenko et al., 2008; Meigs et al., 2008). Nevertheless, it is possible that after accounting for BMI, the inclusion of SNP variants associated to T2D, may not improve prediction accuracy any further. However, this is an unanswered question.

To address these problems, we applied two statistical models (logistic regression, and a neural network) to data from the Framingham Heart Study, and incorporated 65 SNPs that are confirmed to be associated with T2D (Morris et al., 2012) to estimate genetic and non-genetic effects in the prediction of T2D. Since non-genetic factors play a predominant role in whether genetically predisposed individuals progress on to T2D (Poulsen et al., 1999), we considered including BMI information at the onset of T2D, and importantly including genetic by BMI interactions in the predictions of T2D.

In this paper we investigated the effects of including genetic information in preventive risk assessments for T2D, while using different modeling approaches (LR and NN). The effect of including genetic information was examined by adding 65 candidate SNPs for T2D and computing a genetic score based on these SNPs.

Of the 65 SNPs analyzed, 7 SNPs that are located in 4 genes (GLIS3, TCF7L2, LGR5, and PTPRD), showed a strong association with Type 2 Diabetes. In addition, IGF2BP2 and GCKR have been identified by several meta-analyses (Dupuis et al., 2010; Heid et al., 2010; Speliotes et al., 2010; Morris et al., 2012) as risk genetic variants for Type 2 Diabetes with effects in WHR. The SNPs: rs780094, rs7756992, rs7955901 are in the GCKR, CDKAL1, and LGR5 gene regions; with annotated functions of insulin production, pancreatic cell growth, and glucose homeostasis, respectively. GLIS3 has been listed as a diabetes susceptibility gene due to its role in the generation of pancreatic beta cells; an alteration in the expression of this gene could repress the generation of beta cells, and may be involved in pancreatic dysfunction (Dupuis et al., 2010; Nogueira et al., 2013). TCF7L2 was observed to have a relationship with BMI in both the DIAGRAM and GIANT consortiums (Lindgren et al., 2009; Morris et al., 2012). It has demonstrated to lower insulin secretion by affecting β-cell responsiveness to insulin; it is also found in chromatin regions in islets (Kiessling and Ehrhart-Bornstein, 2006; Sladek et al., 2007; Lyssenko et al., 2008; Mccarthy and Zeggini, 2009). The gene PTPRD (protein tyrosine phosphatase receptor type D) provides a component needed to trigger the reactions for the linkage of the insulin receptor to tissue. However, it was excluded as a risk gene for Type 2 Diabetes by Bektas et al. (2001) since none of the mutations did segregate with diabetes. IRS1 showed an association with BMI through SNP-by-BMI interaction. This genetic variant, with an increased interaction with multiple proteins, has been associated with T2D and obesity, and could lead to the development of insulin resistance (Rung et al., 2009; Caruso et al., 2014).

When analyzing the effects of the inclusion of genetic variants in the prediction of this disease, our results suggest that a vast number of SNPs provide a modest enhancement in the predictive ability of the models. Improvement of these discriminative values, show that the added SNPs capture genetic risk. However, when the interaction of the SNPs by environment (BMI) was included in the model, no further increase was seen. The consistency of AUC throughout the models, with the use of both Neural Network and Logistic Regression, suggests that the use of different statistical approaches neither aided nor reduced the predictive ability of the models. The limitation in predictive accuracy seems to be associated to factors other than the statistical model, such as: the size of the training sample, the number of SNPs included in the model, missing heritability issues and low heritability of the trait. A few concerns about SNPs information, were observed. The first pertains to the imputation uncertainty of the SNPs, since it was not fully taken into account in our analyses. Nevertheless, an alternative methods that consider imputation uncertainty are proposed by Marchini and Howie (2010). Secondly, biases could have been produced in the SNPs estimates due to family structure; nevertheless, since the number of families within our sample is large, it is considered to be of minor importance. In our sample of 5245 subjects, 2073 subjects were aggregated from 495 families, (these families contained subjects with at least one relative in the sample), moreover, the size of these families was 4.19 ± 6.40 (mean ± s.d) members per family.

The most commonly identified covariates used in assessment analyses that provide a high AUC (0.60–0.80) as a clinical baseline model have been: age, high blood pressure, and glucose levels between other covariates (Hu et al., 2001; Lyssenko et al., 2008; Meigs et al., 2008; Cooke et al., 2012). Due to the small effects and marginal change that genotyped data provides in risk prediction, they have been used in only a few models to quantify individual disease risk and thus to facilitate personalized management of T2D risk. The ability and the effects of including genetic information into risk prediction, have been widely studied but are still limited. Previous risk assessments were SNPs associated to T2D were included, slightly improved their predictive ability when compared to baseline clinical covariates (Lyssenko et al., 2008; Meigs et al., 2008; Van Hoek et al., 2008; Katsios, 2010; Bao et al., 2013; Lyssenko and Laakso, 2013; Talmud et al., 2014). In her study, Van Hoek et al. (2008), incorporated 18 SNPs, together with age, sex, and BMI and achieved an AUC of 0.68, yielding only a approximately 2% increase when compared to the baseline model. Furthermore, Lyssenko et al. (2008), evaluated the inclusion of a genetic score built with 16 SNPs; in addition to, multiple clinical covariates and achieved a discriminative value of 0.74. The addition of a modest amount of SNPs into risk prediction was lately studied by Talmud et al. (2014), with the use of 65 SNPs found by the DIAGRAM consortium, which were the same used in this study. A genetic score and clinical covariates such as: BMI, triglyceride levels and fasting glucose, altogether with a large data set, resulted in an AUC of 0.75. This last result is consistent with our results in the model GENS_BMI. A limitation of our study is that we did not take into account other clinical variables that have shown some degree of association with diabetes, such as triglyceride levels, high blood pressure, LDL or HDL, which could have enhanced our results. The Framingham Heart Study provides these variables, but there are missing values in many exams and subjects. To avoid reducing sample size, we only included BMI longitudinally (i.e., account for BMI at the first diabetes record), and we found that genetic signal from the SNPs is captured beyond what could be explained by the BMI. BMI estimated effect on diabetes may result biased since we incorporated BMI as the BMI at first diabetes diagnosis for diabetic subjects and last BMI on record for healthy subjects. However, preliminary analysis (not included in the paper) show us that the effect and their significance, for BMI and other covariables in the models, are insensitive to alternative ways to account for BMI, such as, BMI at the first exam, or maximum BMI of the subjects observed period. Despite our limitations, our study can provide important remarks. The effect of genetic information in the improvement of the prediction accuracy, was evaluated in our models by incorporating 65 SNPs both directly and into a genetic score. In addition, we looked at the inclusion of gene-environment (BMI) and gene-gene interaction into risk prediction. Also, a classical logistic regression and a Neural Network (a non-parametric classification algorithm) were explored.

Prevalence of T2D is highest among individuals with a BMI ≥ 40 kg/m² (Bays et al., 2007). The increase in central adiposity and percent body fat is associated with an increased risk of T2D; however, not all obese or overweight patients develop T2D, and of those who do, just a proportion is genetically predisposed. Our results show, in agreement with the literature, that BMI serves as a prediction enhancer for T2D. Predictive accuracy yielded better estimates in the baseline model that included BMI; and this was further improved when the genetic effect was also incorporated, giving an AUC difference of a approximately 8% when compared to baseline. Interaction between BMI and the genes: CILP2, HNF1B, and HMGA2 in relation to T2D, was found and reported in Table 6. HNF1B is a homodimer in charge of the nephron and pancreas development. Mutations in this gene region could result in the development of diabetes. In addition, HMGA2 has transcriptional regulating factors which play a role in adipogenesis and fat storage, inducing obesity.

In summary, this study confirmed the association of 21 genetic variants with T2D. It was observed that individuals who have a high genetic score may have increased probabilities of developing Type 2 Diabetes. Also, accounting for genetic information, either by including SNPs or a Genetic Score in the regression, led to an improvement in prediction accuracy (AUC) of approximately 2%. However, modeling strategies such as Neural Network or Logistic Regression did not yield differences in terms of prediction. We also showed that the inclusion of BMI into the risk assessment models, improved the predictive accuracy by approximately 8%. Furthermore, the risk assessment model yielded a modest increment in prediction accuracy when including genetic risk score, even after accounting for BMI. This small improvement suggests that there is still genetic signal involved in the development of T2D, yet to be captured, that could produce effects beyond the increase in BMI. In summary, marker information in addition to commonly used baseline covariates such as BMI, could lead to an overall modest improvement of predictive performance.

All individuals that helped in the writing process of this manuscript are listed as authors and co-authors, and were part of: the formation of the research, recompilation and management of the data, data analysis and interpretation as well as the redaction and edition of this manuscript.