Proteomic Signature of BMI and Risk of Type 2 Diabetes

Obesity is a major risk factor for type 2 diabetes (T2D), and substantial evidence has shown that managing obesity is an effective way in the treatment of T2D (1). Given the key role of obesity in the development of T2D, the U.S. Preventive Services Task Force recommends overweight and obesity as the primary criteria for screening adults for diabetes in primary care settings (2). However, overweight and obesity are highly metabolically heterogeneous phenotypes that are imperfectly measured by BMI (3,4). Increasing evidence has shown that a normal BMI does not necessarily indicate normal glucose metabolism or an absolute low risk of T2D; likewise, a high BMI does not necessarily indicate abnormal glucose metabolism (5–8). An estimated 15% of U.S. adults with a normal BMI have abnormal blood glucose status; thus, if BMI alone is used as the basis for blood glucose screening, a considerable proportion of high-risk individuals will be missed, delaying treatment or intervention (5). The American Medical Association recently adopted a policy criticizing BMI as an imperfect measure of weight and encouraging clinicians to use methods that combine BMI with other valid measures to diagnose obesity in the future (9).

Blood proteomics profiling has the potential to bridge the gaps between BMI and heterogeneity of obesity. Watanabe et al. (10) recently identified 74 proteins related to BMI using least absolute shrinkage and selection operator regression implemented in a machine learning method, which were trained in the Arivale cohort and externally tested in the TwinsUK cohort. They observed that the proteomic signature of BMI was significantly related to blood glucose, insulin, and insulin resistance levels, regardless of the BMI levels (10). Notably, they also found that individuals with a normal BMI but higher levels of proteomic signature of BMI had poor metabolic health (10). However, whether the proteomic signature of BMI can predict the incidence of T2D is unclear, and if so, whether integrating the proteomic signature of BMI with actual BMI can improve T2D risk stratification.

Therefore, based on 41,427 participants from UK Biobank who were free of diabetes at baseline and had complete data on the proteomic measure and actual BMI, we prospectively assessed the association between the proteomic signature of BMI and the risk of incident T2D. We also evaluated the utility of combining the proteomic signature of BMI and actual BMI in predicting T2D.

The UK Biobank is a large population-based cohort study recruiting >0.5 million participants, aged 40 to 69, at 22 assessment centers throughout England, Wales, and Scotland from 2006 to 2010. The details of the study design have been described elsewhere (11). A subset of representative sample of 52,701 participants was selected to examine the proteomic profiling on blood plasma samples (12). The main analysis included 41,427 participants after 3,079 participants with prevalent diabetes and 8,195 participants with missing values on targeted proteins were excluded. The study was approved by the North West Multi-Centre Research Ethics Committee (Manchester, U.K.) and the Tulane University (New Orleans, LA) Biomedical Committee Institutional Review Board, and written informed consent was obtained from all participants.

In the UK Biobank, the proteomic profiling on blood plasma samples was performed using the antibody-based Olink Explore 3072 platform. A total of 1,463 proteins analytes across four panels (Cardiometabolic, Inflammation, Neurology, Oncology) were measured for the current release version. Details of the participant selection, Olink proteomics assay, data processing, and quality control have been thoroughly documented elsewhere (13). In this study, to create a proteomic BMI signature, we used preidentified BMI-related proteins derived from a recent study by Watanabe et al. (10). A total of 74 proteins related to BMI were identified by training in 1,277 individuals from the Arivale cohort, using least absolute shrinkage and selection operator regression implemented in a machine learning method, and tested in an external cohort, TwinsUK (10). Of the 74 preidentified proteins, 67 were available in this study, and all the proteins passed the quality control. Seven proteins were missing, which included CCL3_2, FGF21_1, FGF21_2, FIGF, IL6_1, IL6_2, and VSIG2. We used a weighted method to calculate the proteomic BMI (pro-BMI) score based on the 67 plasma proteins. Each protein was standardized with a z-score using mean = 0 and SD = 1 before analysis, and each protein was weighted by its relative effect size (β-coefficient) associated with BMI obtained from the previous study (Supplementary Table 1) (10). We calculated the pro-BMI score by using the equation: pro-BMI score = (β1 × Protein 1 + β2 × Protein 2+…+β67 × Protein 67). Similar weight score calculation methods are widely used in the genetic risk score, metabolomic risk score, and others (14–16). We also calculated a BMI-specific proteomic score by excluding proteins related to both BMI and waist-to-hip ratio (WHR) from the pro-BMI score (Supplementary Table 1) (10). Missing values were imputed by using the limit of detection for any given protein.

Information on the diagnosis of T2D events was collected through medical history and linkage to data on hospital admissions and questionnaires. The outcome was T2D, which was defined according to the ICD-10 Revision codes E11. Follow-up time was calculated from the date of baseline to diagnosis of outcome, death, or the censoring date (December 2022), whichever occurred first. Previous studies have assessed the accuracy of using hospital inpatient records and ICD codes in the UK Biobank to diagnose T2D and have shown that these data are reliable enough for epidemiological studies (17).

ANCOVA (generalized linear models) and χ² were used for comparison of continuous and categorical variables, respectively, between participants with quartiles of pro-BMI score. Hazard ratios (HRs) and 95% CIs were estimated in Cox proportional hazards models to evaluate the associations of the pro-BMI score with the risk of incident T2D, and follow-up years were used as the underlying time metric. The proportional hazards assumption was tested by the Kaplan-Meier method and Schoenfeld residuals, and no violation was found. Several covariates were adjusted in the models, including age (continuous), sex (male, female), self-reported race (White, non-White), Townsend Deprivation Index (continuous), education levels (≤10 years, >10 to <20 years, ≥20 years), smoking status (current, former, never), regular physical activity (yes, no), moderate drinking (yes, no), healthy diet score (continuous), high cholesterol (yes, no), and hypertension (yes, no). Details of the assessment of covariates are described in Supplementary Methods. To test whether the pro-BMI score was significantly associated with T2D independent of the well-known obesity indicators, we conducted several sensitive analyses: 1) further adjustment for actual BMI; 2) further adjustment for WHR; 3) further adjustment for the polygenic risk score (PRS) for BMI; and 4) further adjustment for both actual BMI, WHR, and PRS for BMI. We also conducted two more sensitivity analyses by excluding proteins related to both BMI and WHR from the pro-BMI score (10) and excluding participants with high cholesterol, hypertension, or high WHR (WHR ≥0.85 for men or WHR ≥0.9 for women). For the associations between individual proteins and outcomes, each protein was included in the model separately, and false discovery rates were used to account for multiple comparisons. In addition, we conducted stratified analyses by actual BMI levels, tertile categories of WHR, and tertile categories of PRS for BMI, and combined categories of actual BMI and WHR levels. The multiplicative interaction between the pro-BMI score and these obesity-related indicators (categorical variable) were accessed by adding a cross-product term to the models. Poisson regression was used to calculate crude incidence rates according to combined categories of actual BMI and WHR levels, expressed as the number of events per 1,000 person-years. Exact Poisson CIs were used.

Moreover, to estimate the discrimination ability of models before and after inclusion of the pro-BMI score, we assessed the Harrell C index (18). To evaluate the improvement in prediction performance gained by adding the pro-BMI score, category-free net reclassification improvement (NRI^>0) (19) and the integrated discrimination index (IDI) (20) were assessed. Bootstrapping with 1,000 runs was used to estimate CIs. The Harrell C index is a natural extension of the receiver operating characteristic curve area to survival analysis and a parameter describing the performance of a given model applied to the population under consideration (18). NRI offers a simple intuitive way of quantifying improvement offered by new markers, and NRI^>0 is more objective and comparable across studies (19).

Statistical analyses were conducted with SAS 9.4 software (SAS Institute, Cary, NC) and R 3.6.1 software (R Foundation for Statistical Computing, Vienna, Austria). All P values were two-sided, and P < 0.05 was considered statistically significant.

This study was conducted using the UK Biobank Resource, approved project number 29256. The UK Biobank will make the source data available to all bona fide researchers for all types of health-related research that is in the public interest, without preferential or exclusive access for any persons. All researchers will be subject to the same application process and approval criteria as specified by the UK Biobank. For more details on the access procedure, see the UK Biobank website: https://www.ukbiobank.ac.uk/register-apply.

The baseline characteristics of participants according to quartiles of the pro-BMI score are presented in Table 1. Compared with participants with a lower level of pro-BMI score, participants with a higher level of pro-BMI score were more likely to be women, less likely to be White, had lower socioeconomic levels, less likely to have a regular physical activity, to be a moderate drinker, and to eat healthier. The participants with a higher level of pro-BMI score tended to have a higher BMI, a higher WHR, and a higher prevalence of high cholesterol and hypertension. In addition, the pro-BMI score was highly correlated with actual BMI, with a Pearson correlation of 0.66.

The association between the pro-BMI score and the risk of T2D is presented in Table 2. During a median follow-up of 13.7 years, 2,030 incident events of T2D were documented. After adjustment for age, sex, self-reported race, Townsend Deprivation Index, education levels, smoking status, physical activity, moderate drinking, healthy diet score, high cholesterol, and hypertension, per 1-SD increase in the pro-BMI score was significantly associated with an 80% higher risk of T2D (hazard ratio [HR], 1.80; 95% CI 1.74–1.87). Compared with the lowest quartile group, the HRs (95% CI) for the quartile 2 to quartile 4 groups of pro-BMI score were 1.84 (1.49–2.29), 3.15 (2.58–3.85), and 8.12 (6.72–9.82), respectively (P < 0.001 for trend). A linear association was observed between the pro-BMI score and T2D risk (Supplementary Fig. 1). After further adjustment for BMI, WHR, or PRS for BMI, the association was only slightly attenuated. The observed association also remained stable even after further simultaneous adjustment for BMI, WHR, and the PRS for BMI. Moreover, if we excluded those proteins that associated with both BMI and WHR, the association between the BMI-specific proteomic score and T2D risk remained significant, with an HR of 1.32 (95% CI 1.24–1.40) per 1-SD increase in the multivariable adjusted model. The association was not appreciably changed if we excluded participants with high cholesterol or hypertension or a high WHR (≥0.90 in men and ≥0.85 in women) (Supplementary Table 2).

The associations between individual proteins and the risk of T2D are shown in Supplementary Fig. 2. In sensitivity analysis, we excluded individual proteins that were significantly associated with T2D from the pro-BMI score and performed the analysis again (Supplementary Table 3). We found that the strength of the observed association weakened after excluding LEP, FABP4, or IGFBP1. The HR (95% CI) for per 1-SD increase in the pro-BMI score decreased from 1.80 (41.74–1.87) to 1.63 (1.58–1.69), 1.60 (1.55–1.65), or 1.64 (1.58–1.70), respectively.

We performed several stratified analyses to assess the association between the pro-BMI score with the risk of T2D by obesity indicators, including actual BMI levels (18.5 to <25 kg/m², 25 to <30 kg/m², and ≥30 kg/m²) and tertiles categories of WHR and PRS for BMI (Fig. 1). Within each BMI level, we observed that the pro-BMI score was consistently associated with a higher T2D risk (P = 0.307 for interaction) (Fig. 1,A). Compared the highest with the lowest quartiles, the adjusted HRs of T2D were 3.39 (95% CI 1.98–5.81), 5.50 (95% CI 4.00–7.57), and 2.02 (95% CI 1.14–3.59) for participants whose BMI ranged from 18.5 to <25 kg/m², 25 to <30 kg/m², and ≥30 kg/m², respectively (all P < 0.001 for trend). If we performed the analysis by replacing BMI levels with tertiles categories of WHR or PRS for BMI, similar results were observed (Fig. 1B and C). Moreover, we further classified participants according to the combined categories of actual BMI and WHR levels, including normal BMI (18.5 to <25 kg/m²) and normal WHR (<0.90 in men and <0.85 in women), normal BMI and high WHR (≥0.90 in men and ≥0.85 in women), high BMI (BMI ≥25 kg/m²) and normal WHR, and high BMI and high WHR. We observed that a higher pro-BMI score was still significantly associated with a higher risk of T2D in the four combined groups (all P < 0.001 for trend) (Fig. 2).

Next, we evaluated the discrimination and reclassification performance of the model by adding the pro-BMI score to a reference model (Table 3). Adding the pro-BMI score alone to the reference model with age, sex, and actual BMI significantly increased the C statistic by 0.023 (95% CI 0.018–0.027). Adding the pro-BMI score alone to the reference model also provided significant improvement in reclassification (NRI^>0 0.486; 95% CI 0.427–0.549), with event-specific components of 0.198 (95% CI 0.174–0.232) and no event-specific component of 0.288 (95% CI 0.269–0.305), and a significant IDI (0.007; 95% CI 0.006–0.009). Adding WHR alone to the reference model with age, sex, and actual BMI also led to a significant increase in the C index (difference 0.021; 95% CI 0.017–0.026), which was roughly equivalent to the increase for the pro-BMI score. The C index did not change PRS for BMI alone was added to the reference model. Moreover, if the pro-BMI score alone was added to a reference model that already included age, sex, BMI, WHR, and PRS for BMI, the pro-BMI score could still significantly increase the C index by 0.013 (95% CI 0.010–0.016) and provide an IDI of 0.005 (95% CI 0.004–0.007). Reclassification also made a significant improvement (NRI^>0 of 0.453; 95% CI, 0.392–0.512). In sensitivity analysis, if we extended the reference model by further including self-reported race, Townsend Deprivation Index, education levels, smoking status, moderate drinking, healthy diet score, high cholesterol, and hypertension, similar increases were observed in the C index as well as a significant NRI and IDI when the pro-BMI score was added to the reference model (Supplementary Table 4).

In this large study of 41,427 U.K. adults who were free of diabetes at baseline, we observed that the proteomic signature of BMI, as assessed by the pro-BMI score, was significantly associated with T2D risk independent of actual BMI, WHR, PRS for BMI, and other traditional risk factors. We found that when added to models that included age, sex, and actual BMI, the pro-BMI score alone significantly improved the risk prediction of T2D risk, with an effect size similar to that of adding WHR alone to the model. Moreover, we found that adding the pro-BMI score to a model that already included BMI and WHR still significantly but modestly improved the prediction of T2D risk.

Very few previous studies have evaluated the associations of proteomic signatures of BMI with T2D risk. Bao et al. (21) identified 10 BMI-specific proteins and 22 WHR-specific proteins in 4,203 participants and found that WHR-specific proteins, but not BMI-specific proteins, were significantly associated with a higher risk of diabetes. Their study had limited statistical power and only considered the association between individual proteins and diabetes risk, making their findings difficult to interpret. Compared with their study, our study had a larger sample size, included a larger number of proteins (67 proteins vs. 57 proteins), and analyzed BMI-related proteins jointly. We observed that the overall proteomic signature of BMI was significantly associated with T2D risk, even after excluding preidentified proteins that associated with both BMI and WHR. Moreover, the association between the overall proteomic signature of BMI and T2D risk remained stable after excluding any individual protein from the pro-BMI score, suggesting that the observed association represented cumulative effects from multiple BMI-related proteins.

In a recent cross-sectional study, Watanabe et al. (10) found that participants with higher levels of proteomic signature of BMI had significantly higher levels of glucose, insulin, and insulin resistance than those with lower levels of proteomic signature of BMI when their actual BMI levels were similar. Compared with participants included in the study by Watanabe et al. (10), our population was older, less likely to be women, and more likely to be White. Our study extended their findings and found that the proteomic signature of BMI was consistently associated with a higher risk of T2D within each actual BMI level, each WHR level, and each PRS for BMI level in a large prospective study. Moreover, our results indicate that several individual proteins, including LEP, FABP4, and IGFBP1, may have important roles in the association between the pro-BMI score and T2D risk. LEP is a pleiotropic adipokine that plays a key role in the regulation of energy homeostasis, and increasing evidence suggests that LEP is also critical for glycemic control (22). Welsh et al. (23) observed that higher levels of leptin were significantly associated with higher diabetes risk but not cardiovascular disease risk in an elderly population. In the Jackson Heart Study, Bidulescu et al. (24) found that leptin is directly related to T2D events, but this relationship appears to be mediated by insulin resistance. Consistent with the above epidemiology studies, our results showed that higher LEP levels were associated with a higher T2D risk in a large U.K. cohort (23,24). Moreover, animal studies have also reported that leptin replacement therapy could improve insulin resistance in muscle and liver in individuals with lipodystrophy (25). FABP4 is an adipokine primarily expressed in adipocytes and macrophages, and previous evidence from animal studies has shown that FABP4 plays a key role in development of insulin resistance and diabetes (26). Several epidemiology studies have also shown that higher FABP4 levels are associated with higher T2D risk and other obesity-related diseases (27,28). IGFBP-1 is a part of the IGF system, which is produced in the liver. Growing evidence has shown that IGFBP-1 is involved in the regulation of glucose metabolism (29,30). A recent study implicated IGFBP1 in glycemic physiology during pregnancy and suggested a role for placental IGFBP1 deficiency in the pathogenesis of gestational diabetes mellitus (31). These results provide insights into the molecular mechanisms underlying the link between adiposity and T2D, facilitating the discovery of new therapeutic targets.

Moreover, we found that the proteomic signature of BMI held promise as an adjunctive diagnostic indicator with actual BMI to better distinguish an individual’s T2D risk. Almost all current T2D screening guidelines used BMI as the primary criterion (32), but this has been shown to miss a large proportion of people at high risk for T2D (33). Notably, despite BMI having some limitations, there is currently no recognized indicator that can replace it (34,35). Genomics had high hopes, but several previous studies have shown that the PRS does not improve or only slightly improved BMI prediction of T2D risk (36,37). Fat distribution is a major contributor to the heterogeneity in obesity defined by BMI, and several guidelines for obesity risk assessment have recommended considering measures of fat distribution (waist or WHR) in addition to BMI (38,39). An intriguing finding in the current study was that adding the proteomic signature of BMI alone to a model with age, sex, and BMI resulted in a significant increase in the C index, roughly equivalent to the increase of adding an indicator of fat distribution (WHR) alone. It should be noted that when the proteomic signature of BMI is added to a model that included BMI, WHR, and PRS for BMI, the proteomic signature of BMI can still lead to a significant increase in the C index. The stratified analyses by combined categories of actual BMI levels and WHR levels further supported these findings. We found that the proteomic signature of BMI was consistently associated with higher T2D risk regardless of the combination types of BMI and WHR. More research is needed to verify our findings and determine whether the proteomic signature of BMI is clinically useful for screening purposes.

Strengths of our study include the well-validated plasma proteomic measurements in a well-established large cohort, the long follow-up period, and the wealth of covariates. More importantly, this is the first study to evaluate the association between the overall proteomic signature of BMI and T2D risk.

The current study has several potential limitations. First, it is important to note that although our results suggest that adding a pro-BMI score to a predictive model that includes BMI significantly improves the risk prediction of T2D, the magnitude of the improvement is modest. Owing to the high cost of current proteomic measurement, the significance of using the pro-BMI score in predicting disease risk is limited at this stage. Second, although most BMI-related proteins were covered in our study (67 of 74), several proteins were not covered. Third, the plasma proteomic measurements were only available at baseline, and we were unable to analyze the association of changes in proteomic signature of BMI with T2D risk. Fourth, most of the participants in this study were White, and whether our findings could be generalized to other racial groups is unclear. Fifth, this is an observational study, and the associations between proteomic signature of BMI and T2D risk cannot be interpreted as causal relations.

In summary, our study shows that the overall proteomic signature of BMI was significantly associated with the risk of T2D independent of actual BMI and other conventional obesity indicators. The proteomics of BMI can modestly help the actual BMI improve the accuracy of individual T2D risk prediction.

Funding. The study was supported by grants from the National Institutes of Health, National Heart, Lung, and Blood Institute (HL071981, HL034594, HL126024) and the National Institute of Diabetes and Digestive and Kidney Diseases (DK115679, DK091718, DK100383, DK078616), the Fogarty International Center (TW010790), and Tulane Research Centers of Excellence Awards. L.Q. is also supported by National Institutes of Health, National Institute of Diabetes and Digestive and Kidney Diseases (P30DK072476) and National Institute of General Medical Sciences (P20GM109036).

The funding sources and sponsor had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

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

Author Contributions. X.W. and H.M. contributed to acquisition, analysis, or interpretation of the data. X.W. and H.M. conducted the statistical analysis. X.W., H.M., and L.Q. conceived and designed the study. X.W., H.M., and L.Q. drafted the manuscript. All authors contributed to critical revision of the manuscript for important intellectual content. L.Q. 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.