Identification of Novel Causal Blood Biomarkers Linking Metabolically Favorable Adiposity With Type 2 Diabetes Risk

A weighted polygenic risk score (PRS) underpinning metabolically favorable adiposity was designed and validated in two independent cohorts, including the UK Biobank and the Outcome Reduction With Initial Glargine Intervention (ORIGIN Trial). We next tested the PRS for association with a panel of 238 cardiometabolic biomarkers in the ORIGIN Trial to identify associated biomarker blood concentrations. We then investigated the identified biomarkers for causation using a Mendelian randomization approach applied to type 2 diabetes risk and validated through epidemiological association analyses with prevalent type 2 diabetes and incident major adverse cardiovascular events (MACE), as well as through predictive model performance for MACE incidence, in the ORIGIN Trial (Fig. 1).

The UK Biobank is a prospective cohort of >500,000 individuals (aged 40–69 years) recruited from multiple centers across the U.K. between 2006 and 2010 that collected extensive phenotypic and genotypic data (15). From the full data set release issued in July 2017, a total of 343,735 unrelated individuals of British origin had genotyping data suitable for analysis (Supplementary Data). Among those individuals, 1,863 had missing phenotypic data, leaving 341,872 for our analyses (Supplementary Table 1).

The design of the ORIGIN Trial has previously been described (16). Between 2003 and 2005, a total of 578 clinical sites in 40 countries enrolled 12,537 participants with type 2 diabetes, impaired glucose tolerance or impaired fasting glucose levels, and additional cardiovascular risk factors. Following random allocation to two therapies using a factorial design (insulin glargine vs. standard care, and n-3 fatty acid supplement vs. placebo), participants were followed for a median of 6.2 years to monitor cardiovascular events and other health outcomes. At baseline, 8,494 participants provided blood samples for proteomic analyses, among whom 5,078 consented to genomic analyses (Supplementary Data). Finally, 1,931 participants of European Caucasian ancestry and 2,216 participants of Native Latin ancestry were included in our genetic analyses (Supplementary Table 2).

For both cohorts, single nucleotide polymorphisms (SNPs) were excluded on the basis of low call rates (<99%), deviation from Hardy-Weinberg equilibrium (P < 10⁻⁶), low minor allele frequency (<0.01), and low imputation quality (INFOscore <0.6).

After completion of the ORIGIN Trial, 1 mL serum from each participant was transported to Myriad RBM, Inc. (Austin, TX), to quantify 284 biomarkers (Luminex) related to metabolic and cardiovascular diseases. As previously validated (17), 238 biomarkers from 8,401 participants were deemed suitable for analysis (Supplementary Table 3), with biomarkers that were not normally distributed being first log transformed. All biomarkers were subsequently standardized to a mean of 0 and an SD of 1, with the exception of biomarkers with a high proportion of low or undetectable concentrations, which were analyzed as ordinal variables. Individuals without specific biological ancestry information were removed, resulting in a final sample of 8,197 for our biomarker analyses.

We partitioned ∼1.9 million GWAS-identified SNPs for BMI into three mutually exclusive categories (i.e., protective, deleterious, or neutral effect on type 2 diabetes risk) using a polygenic regional correlation approach (18). First, we divided the genome into small regions (n = 2,554; median size 750 kbp). Second, while all BMI variants were retained without assumptions of the significance threshold and aligned to the BMI-increasing alleles, we tested each region for polygenic correlation between BMI and type 2 diabetes using a weighted maximum likelihood approach. Polygenic correlation was then used to test whether the BMI-increasing genetic effect of each region was linked to a corresponding increase, decrease, or neutral effect on type 2 diabetes risk. BMI genome-wide summary association statistics from the Genetic Investigation of ANthropometric Traits (GIANT) consortium (7) and GWAS meta-analysis for type 2 diabetes from the DIAbetes Genetics Replication And Meta-analysis (DIAGRAM) consortium (8) were used as inputs for the analysis. We identified 27 genomic regions with nominal evidence (P < 0.05 for correlation) of a negative genetic correlation between BMI and type 2 diabetes, which were assigned to the “metabolically favorable adiposity” category. A total of 251 genomic regions showed nominal evidence of a positive correlation and were assigned to the “metabolically deleterious adiposity” category. The remaining 2,239 regions had neutral effects on type 2 diabetes risk (P ≥ 0.05 for correlation). Regional correlation could not be estimated for 37 regions, as no summary statistics for BMI or type 2 diabetes were available for those regions (Supplementary Fig. 1 and Supplementary Table 4). Third, PRSs corresponding to the aforementioned categories were derived. The metabolically favorable adiposity PRS encompassed 20,532 SNPs from the 27 regions, while the metabolically deleterious adiposity PRS encompassed 190,807 SNPs from the 251 regions. PRSs were calculated by using a weighted approach where the weight of each SNP (i.e., regression coefficient extracted from the GIANT consortium for BMI [7]) was corrected for linkage disequilibrium (LD), such that all SNPs of a given region were retained in the PRSs irrespective of LD. Indeed, PRSs were calculated for each participant of the UK Biobank cohort and the ORIGIN Trial as follows: ⁠, where was the normalized genotype of each SNP, b_d was the regression coefficient for each SNP (expressed in kg/m² increase per risk allele), and η_d was the LD adjustment for each SNP, which was based on the sum of the pairwise disequilibrium of each SNP over the neighboring 200 SNPs upstream and downstream (“GraBLD” R package) (19).

Associations of the metabolically favorable adiposity PRS with baseline BMI and type 2 diabetes prevalence were first performed in the UK Biobank cohort using linear regression and logistic regression models, respectively. All individuals were unrelated and of British Caucasian origin. The models were adjusted for age, sex, and the first 10 principal components. Associations of the metabolically favorable adiposity PRS with baseline BMI and type 2 diabetes prevalence were next replicated in the ORIGIN Trial using linear and logistic regression models. Since the ORIGIN Trial is a multiethnic cohort, models were tested separately in Europeans and Latin Americans, with adjustment for age, sex, and the first five principal components, and then meta-analyzed using fixed-effects models (“meta” R package).

Associations of the metabolically favorable adiposity PRS with the baseline serum concentrations of the 238 biomarkers were performed using linear regression and adjusted for multiple hypothesis testing using a conservative Bonferroni correction (P < 0.05/238 = 2.01 × 10⁻⁴). Models were tested separately in Europeans and Latin Americans, with adjustment for age, sex, and the first five principal components, and then meta-analyzed using fixed-effects models.

The possibility of a causal relationship between the identified biomarkers and type 2 diabetes risk was evaluated using a two-sample Mendelian randomization analysis (Supplementary Data). As instrumental variables, we selected a set of independent SNPs (r² < 0.1) within 300 kbp of the gene encoding the biomarker of interest (“cis SNPs”) that were associated with their respective biomarker concentrations in the ORIGIN Trial (P < 0.01). We then extracted the effect estimates of those SNPs on type 2 diabetes risk from the DIAGRAM consortium, including 12,171 type 2 diabetes case subjects and 56,862 control subjects (8). Finally, odd ratios (ORs) were determined by regressing the effect estimates from the type 2 diabetes associations on the biomarker association using the inverse-variance weighting method.

To further explore the associations of the identified biomarkers with baseline anthropometric measurements (BMI, waist-to-hip ratio [WHR]) and glycemic variables (fasting glucose and HbA_1c) and type 2 diabetes prevalence, we used linear regression for continuous variables and logistic regression for dichotomous outcomes.

Associations of the baseline biomarker concentrations (as a continuous variable and by tertiles of distribution) with the incidence of MACE—prospectively defined as nonfatal myocardial infarction, nonfatal stroke, or cardiovascular death—were assessed using Cox proportional hazards models. Subgroup analyses were performed to test the interaction of BMI strata (i.e., normal weight [BMI <25 kg/m²], overweight [BMI 25–29.9 kg/m²], and obesity [BMI ≥30 kg/m²]) with identified biomarker on MACE incidence. All models were adjusted for age, sex, and ancestry. The Cox regression model was also adjusted based on whether participants had type 2 diabetes at baseline.

We assessed the performance of adding the identified biomarkers to clinical risk models predicting cardiovascular events. The National Cholesterol Education Program (NCEP) Adult Treatment Panel III (ATPIII) definition for metabolic health was used as the referent clinical assessment of cardiometabolic risk (20). According to the NCEP ATPIII criteria, the metabolic health of each ORIGIN Trial participant was assessed based on a positive status for dysglycemia (i.e., type 2 diabetes, impaired glucose tolerance, or impaired fasting glucose), hypertension (i.e., systolic blood pressure ≥130 mmHg, diastolic blood pressure ≥85 mmHg, or antihypertensive drugs), hypertriglyceridemia (i.e., triglycerides ≥1.7 mmol/L or lipid-lowering drugs), low HDL cholesterol (HDL cholesterol <1.04 mmol/L for men and <1.29 mmol/L for women), and large waist circumference (i.e., waist circumference >102 cm for men and >88 cm for women). We evaluated the value of adding biomarker concentrations to these components for the prediction of MACE incidence in the ORIGIN Trial. We used the area under the receiver operating characteristic curve, the net reclassification index (NRI), and the integrated discrimination improvement (21). The performances of MACE prediction models were compared across three models: model 1, a “clinical model” including the NCEP ATPIII criteria, with adjustment for age, sex, and ancestry; model 2, a “biomarker model” including the biomarker level, with adjustment for age, sex, and ancestry; and model 3, a “clinical + biomarker model” in which the biomarker was added to the “clinical model.” Improvements in the predictive performance (NRI and integrated discrimination improvement) of model 2 or 3, compared with model 1, were tested using logistic regression models. The calibration (goodness-of-fit test) was measured by the Hosmer-Lemeshow χ² test (22) (“PredictABEL” R package [23]).

To test whether the identified biomarkers were specifically associated with metabolically favorable adiposity, we investigated the association between biomarker levels and a PRS underpinning “metabolically deleterious adiposity.” Associations of the metabolically deleterious adiposity PRS and the identified biomarkers were performed using linear regression, adjusted for age, sex, and the first five principal components of ancestry, and meta-analyzed across ancestries. Then, the coefficient regression of each biomarker concentration–PRS association was compared between the two PRSs using a two-tailed Z test.

All statistical analyses were conducted using R (version 3.3.2). Two-tailed P values of <0.05 were considered statistically significant, with adjustment for multiple hypothesis testing applied as appropriate.

We found that a 1 unit increase in metabolically favorable adiposity PRS was associated with a 0.73 kg/m² increase in BMI (β 0.73 per 1 unit PRS [95% CI 0.67–0.80], P < 2.0 × 10⁻¹⁶) in the UK Biobank cohort (Fig. 2). A higher metabolically favorable adiposity PRS predicted a lower type 2 diabetes prevalence with an OR of 0.73 per 1 unit increase in PRS (95% CI 0.68–0.78, P < 2.0 × 10⁻¹⁶) (Fig. 2). The PRS distributions among the UK Biobank participants with and without type 2 diabetes are shown in Supplementary Fig. 2. Associations of the metabolically favorable adiposity PRS with measured BMI and type 2 diabetes prevalence were significantly replicated in the ORIGIN Trial (Fig. 2). Quantile regression of the PRS along the BMI percentiles is presented in Supplementary Fig. 3.

We next investigated the metabolically favorable adiposity PRS for associations with 238 cardiometabolic serum biomarkers assayed in the ORIGIN Trial. After correction for multiple hypothesis testing, insulin-like growth factor–binding protein (IGFBP)-3 was the only biomarker significantly associated (P < 0.05/238). Lower metabolically favorable adiposity PRS predicted higher IGFBP-3 serum concentrations (β −0.28 SD per 1 unit PRS [95% CI −0.42 to −0.13], P = 1.63 × 10⁻⁴) (Table 1). In investigation of the relationship between IGFBP-3 level and the metabolically deleterious adiposity PRS (Supplementary Fig. 4), no significant association was found (P = 0.31). This sensitivity analysis thus positioned IGFBP-3 as a biomarker specifically associated with metabolically favorable adiposity and distinguishing genetically determined favorable adiposity from genetically determined deleterious adiposity (P = 0.0002 for the difference between the two IGFBP-3 concentration–PRS associations). Additionally, the IGFBP3 genetic variants were located within the region 7:38,969–39,738 and had a polygenic correlation P value ≥0.05. Thus, they were not included in the PRSs, suggesting that IGFBP-3 concentration–PRS associations were not driven by the IGFBP3 variants themselves.

Mendelian randomization was used to test for causation between IGFBP-3 serum concentrations and type 2 diabetes risk. Six IGFBP3 SNPs (rs788749, rs2462688, rs2280497, rs1542818, rs1496495, and rs10263580) were independently associated with IGFBP-3 serum concentrations (P < 0.01) in the ORIGIN Trial (Supplementary Table 5) and were included in the Mendelian randomization analysis. The association between each of the SNPs and type 2 diabetes risk was assessed using the DIAGRAM database (8). With use of inverse-variance weighted Mendelian randomization, the causal effect of IGFBP-3 on type 2 diabetes risk was estimated at 26% per 1 SD increase in genetically determined IGFBP-3 levels (OR 1.26 [95% CI 1.11–1.43], P = 0.0004) (Fig. 3).

We then assessed the epidemiological relationships between baseline IGFBP-3 serum concentrations and anthropometric (BMI and WHR) and glycemic (fasting glucose, HbA_1c, and prevalent type 2 diabetes) variables in ORIGIN Trial participants (N = 8,197). Consistent with our genetic findings pointing to higher IGFBP-3 levels as a marker for the lack of metabolically favorable adiposity, higher IGFBP-3 serum concentrations were associated with lower BMI (β −0.28 kg/m² per 1 SD [95% CI −0.39 to −0.16], P = 1.5 × 10⁻⁶) (Fig. 3). No association was found between IGFBP-3 levels and WHR (P = 0.32), even after adjustment for BMI (P = 0.80). A higher IGFBP-3 level was associated with higher fasting plasma glucose (β 1.46 mg/dL per 1 SD [95% CI 0.67–2.26], P = 0.0003), higher HbA_1c (β 0.03% [0.3 mmol/mol] per 1 SD [95% CI 0.01–0.05% (0.1–0.5 mmol/mol)], P = 0.002), and higher type 2 diabetes prevalence (OR 1.11 per 1 SD increase [95% CI 1.03–1.19], P = 0.005) (Fig. 3). The relationship with type 2 diabetes was not affected by adjustments for BMI (OR 1.12 [95% CI 1.04–1.20], P = 0.003) or WHR (OR 1.11 [95% CI 1.03–1.19], P = 0.005). Those findings were consistent with Mendelian randomization results (P = 0.11 for difference).

A total of 1,378 individuals among the 8,197 ORIGIN Trial participants suffered from one or more MACE over a median follow-up of 6.2 years. Higher IGFBP-3 concentration at baseline was associated with an increase in incident MACE (adjusted hazard ratio [aHR] 1.13 per 1 SD increase [95% CI 1.07–1.20], P = 1.98 × 10⁻⁵) (Fig. 3 and Supplementary Fig. 5). Subjects with an IGFBP-3 concentration in the highest tertile of the distribution had a 31% higher risk of experiencing a MACE compared with those having an IGFBP-3 concentration in the lowest tertile (aHR 1.31 [95% CI 1.15–1.50], P = 5.34 × 10⁻⁵) (Supplementary Table 6). Subgroup analyses by BMI categories suggested a trend toward greater association of IGFBP-3 concentrations with MACE incidence among the 1,218 subjects of normal weight (aHR 1.84 [95% CI 1.33–2.52], P = 1.8 × 10⁻⁴, highest vs. lowest tertile) compared with the two other BMI categories (P for interaction 0.06) (Supplementary Table 6).

To further explore the value of adding IGFBP-3 to standard clinical assessments of metabolic health (i.e., NCEP ATPIII components) in normal weight individuals, we evaluated the performance of predictive models for MACE across three different models, including “NCEP ATPIII,” “IGFBP-3,” and “NCEP ATPIII + IGFBP-3.” The addition of IGFBP-3 concentration to the NCEP ATPIII model did not significantly improve the area under the receiver operating characteristic curve for MACE incidence (P = 0.57) (Supplementary Fig. 6) but correctly reclassified a substantial number of subjects with normal weight with an NRI of 11.5% (P = 0.004), including 4.8% of subjects with incident MACE and 6.7% of subjects without incident MACE correctly reclassified (Supplementary Table 7). Consistent results were also observed in models fully adjusted for LDL cholesterol levels and smoking (NRI of 8.0%, P = 0.046). Moreover, the Hosmer-Lemeshow goodness-of-fit test showed adequate calibration for the three models (all P > 0.05) (Supplementary Fig. 7).

Using an integrated genomic-proteomic approach, we identified IGFBP-3 as a novel blood marker of metabolically favorable adiposity. We demonstrated that increased IGFBP-3 concentrations were causally related to increased type 2 diabetes risk using Mendelian randomization analyses, thus positioning IGFBP-3 as a causal biomarker linking the lack of favorable adiposity and increased type 2 diabetes risk. Our genetic findings were supported by epidemiological analyses, which showed that higher IGFBP-3 serum concentrations were associated with lower BMI, higher type 2 diabetes prevalence, and higher incidence of MACE. Finally, adding IGFBP-3 concentration to the most widely used clinical assessment for metabolic health may improve the risk prediction of incident cardiovascular events in normal weight individuals.

Although previous epidemiological studies found an association between IGFBP-3 levels and type 2 diabetes risk (24,25), causality was not assessed. Our Mendelian randomization analysis is the first to demonstrate the causal role of IGFBP-3 in the pathways between lack of favorable adiposity and type 2 diabetes. Several biological reports are consistent with this Mendelian randomization result. First, IGFBP-3 is expressed in multiple cell types and tissues, including subcutaneous and visceral adipose tissues (26). Second, transgenic mice overexpressing human IGFBP3 have reduced body weight, fasting hyperglycemia, and impaired glucose tolerance (27,28), while Igfbp3 knockout mice exhibit increased fat mass and are protected against glucose intolerance when fed a high-fat diet (29). Third, increased circulating IGFBP-3 seems to be involved in the induction of insulin resistance in adipocytes (30,31). Thus, the overall evidence suggests that a higher level of IGFBP-3 is a marker for the lack of favorable adiposity, which is characterized by a reduction of energy storage capacity. Insufficient storage capacity in turn exposes lean organs to the detrimental metabolic effects of ectopic lipid accumulation, with an ensuing increased risk of type 2 diabetes and cardiovascular diseases (32).

Distinguishing metabolically unhealthy individuals within the normal weight population has important clinical and public health implications. It can help stratify normal weight individuals for preventive interventions aiming at reducing cardiometabolic diseases and consequently optimize the cost-effectiveness of nutritional and lifestyle interventions (3). Unfortunately, this distinction is limited by the absence of universally accepted criteria to define the lack of metabolically favorable adiposity and further complicated by the fact that both favorable and deleterious adiposity are present in each individual in varying proportions. The latter could also explain the apparent heterogeneous association between IGFBP-3 concentration change and MACE incidence along the BMI categories. Although BMI is the most commonly used anthropometric index to estimate adiposity-related comorbidity risk, it has important limitations (33). Waist circumference, WHR, and waist-to-height ratio, as measurements of fat distribution, present some advantages over BMI (34) but do not fully discriminate visceral from subcutaneous fat either (35). Our results support the fact that cardiometabolic risk stratification based on the conventional clinical criteria of metabolic health remains unsatisfactory (36) and could be improved by adding relevant biomarkers to the cardiometabolic risk assessment. IGFBP-3 is the major binding protein of insulin-like growth factors (IGFs) and its blood concentration is routinely used in combination with IGF-1 blood concentration for the diagnosis and therapeutic management of growth hormone–related diseases (37). As demonstrated in our study, adding IGFBP-3 concentrations to clinical assessments for metabolic health may represent a new, easily applicable test to improve cardiovascular risk stratification in normal weight individuals.

From a therapeutic perspective, IGFBP-3 blood concentrations may be modulated by lifestyle interventions (26), raising the possibility of a targeted therapeutic strategy. However, no dietary pattern has been reported to modulate IGFBP-3 concentrations (38), and lifestyle interventions, including diet and/or exercise counseling, did not show an effect on IGFBP-3 concentrations (39). Nonetheless, this study had modest sample size (n = 439), and statistical power was correspondingly limited. Further work is required to fully delineate the metabolic and nonmetabolic effects of modulating IGFBP-3 concentrations (40).

The strengths of our study include its large sample size, the design of an innovative PRS to discover biomarkers of metabolically favorable adiposity, and the use of Mendelian randomization analysis to estimate a causal relationship between IGFBP-3 levels and type 2 diabetes risk. However, our study also has several limitations. First, the analysis was restricted to the biomarkers included in the assay panel, and more comprehensive multiplex platforms may discover additional biomarkers. Second, even if associations with biomarkers other than IGFBP-3 were present in our panel, our analyses may have been underpowered to detect them. Third, our analysis was restricted to individuals of European and Latin American ancestries, and the association of the metabolically favorable adiposity PRS with biomarkers may differ across other ancestries. Fourth, although we did not observe a relationship between IGFBP-3 levels and WHR, further studies with more detailed assessments of visceral and subcutaneous fat are required to pinpoint the anatomical distribution of adiposity in individuals having higher IGFBP-3 levels.

Our study identified IGFBP-3 as a novel and specific marker of metabolically favorable adiposity. IGFBP-3 was also identified as a causal mediator linking a lack of favorable adiposity to an increased risk of type 2 diabetes. These findings highlight the causal role of IGFBP-3 in the physiopathology of type 2 diabetes and suggests that this pathway could be targeted pharmacologically to prevent type 2 diabetes. IGFBP-3 levels also enabled us to stratify individuals for incident cardiovascular events, thus supporting the use of IGFBP-3 concentrations in defining criteria for the metabolically unhealthy normal weight phenotype. Further studies are still required to evaluate the diagnostic performance of IGFBP-3 in the general population and its usefulness for personalized interventions against cardiometabolic diseases.

Acknowledgments. The authors are thankful for all of the participants for contributing to this project, the GIANT and DIAGRAM consortia, and the UK Biobank for making their data available. Data analyses of the UK Biobank cohort were conducted under UK Biobank application number 1525.

Funding. The ORIGIN Trial and biomarker project were partly supported by the Canadian Institutes of Health Research (CIHR). Genetic analysis of ORIGIN Trial participants was supported by CIHR (award 125794 to G.P.). M.P. is supported by an award from Société Francophone du Diabète and a CIHR fellowship award (MFE-158092). G.P. is supported by the Canada Research Chair in Genetic and Molecular Epidemiology.

Duality of Interest. The ORIGIN Trial and biomarker project were also supported by Sanofi. The biomarker project was led by ORIGIN Trial investigators at the Population Health Research Institute (Hamilton, Ontario, Canada) with the active collaboration of Sanofi scientists. Sanofi directly compensated Myriad RBM, Inc., for measurement of the biomarker panel and the Population Health Research Institute for scientific, methodological, and statistical work. H.G. has received consulting fees from Sanofi, Novo Nordisk, Lilly, AstraZeneca, Boehringer Ingelheim, and GlaxoSmithKline and support for research or continuing education through his institution from Sanofi, Lilly, Takeda, Novo Nordisk, Boehringer Ingelheim, and AstraZeneca. G.P. has received consulting fees from Sanofi, Bristol-Myers Squibb, Lexicomp, and Amgen and support for research through his institution from Sanofi. G.P. is supported by the Cisco Professorship in Integrated Health Biosystems. S.H. is an employee of Sanofi. S.Y. has received research support for the ORIGIN Trial from Sanofi through his institution. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. M.P., J.S., S.M., and M.C. performed the statistical and bioinformatics analyses. M.P., J.S., S.M., M.C., S.H., S.Y., H.G., and G.P. contributed to the critical reading and revision of the manuscript. M.P., J.S., S.M., M.C., S.H., S.Y., H.G., and G.P. have approved the submitted version of this manuscript. M.P., H.G., and G.P. designed the study, planned the analyses, interpreted the results, and wrote the manuscript. S.H. suggested including IGFBP-3 in the biomarker panel. G.P. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Data Availability. No additional data are available.