Proteomic Profiles of Body Mass Index and Waist-to-Hip Ratio and Their Role in Incidence of Diabetes.

CONTEXT: It is unclear to what extent the plasma proteome of abdominal fat distribution differs from that of body mass index, and whether the differences have clinical implications.
OBJECTIVE: To evaluate the difference between the plasma proteomic profiles of body mass index (BMI) and waist-to-hip ratio (WHR), and then examine the identified BMI- or WHR-specific proteins in relation to incidence of diabetes.
METHODS: Data were obtained from the Malmö Diet and Cancer-Cardiovascular Cohort study in the general community. Participants (n = 4203) with no previous diabetes (aged 57.2 ± 6.0 years, 37.8% men) were included. Plasma proteins (n = 136) were measured by the Proseek proximity extension method. BMI- and WHR-specific proteins were identified at baseline using a 2-step iterative resampling approach to optimize internal replicability followed by β coefficient comparisons. The identified proteins were considered internally replicated and were then studied in relation to incident diabetes by Cox proportional hazards regression analysis. The main outcome measure was incident diabetes over a mean follow-up of 20.3 ± 5.9 years.
RESULTS: After excluding 21 overlapping proteins and proteins that did not show significantly different associations with BMI vs WHR, 10 internally replicated proteins were found to be specific to BMI, and 22 were found to be specific to WHR (false discovery rate-adjusted P < .05). Of the WHR-specific proteins, 18 remained associated with diabetes risk after multivariate adjustments, whereas none of the BMI-specific proteins showed associations with diabetes risk.
CONCLUSION: Abdominal fat distribution was associated with some unique characteristics of the plasma proteome that potentially could be related to its additional risk of diabetes beyond general obesity.

The prevalence of obesity is high and rapidly increasing worldwide (1). Obesity carries a remarkable risk of cardiometabolic diseases, such as type 2 diabetes, which subsequently leads to high health and economic burden (1). Body mass index (BMI) has been traditionally used to measure obesity. However, it mainly reflects overall adiposity and fails to distinguish different fat depots. It has been argued that BMI is a “very poor proxy of health” since metabolic health can vary significantly for the same BMI (2, 3).

Fatness located at different body sites may exert paradoxical impacts on health. The abdominal (upper-body) fat is known to be detrimental (3, 4). In contrast, the protective role of gluteofemoral (lower-body) fat has been supported by its associations with improved lipid and glucose homeostasis and a decreased tendency to develop metabolic abnormalities (3, 5). Whereas the exact mechanisms remain to be analyzed, the variations of upper- and lower-body fat in lipolysis, fatty acid uptake, and adipokine profiles have been proposed as possible explanations for their divergent effects (3-5). As a ratio of waist-to-hip circumference (waist-to-hip ratio, WHR), WHR considers both the detrimental effects of abdominal fat and the beneficial effects of gluteofemoral fat, and it is widely used as a proxy of body fat distribution (6, 7).

WHR significantly improves the prediction of diabetes risk on top of BMI (8, 9). The genetic basis of WHR only marginally overlaps with that of BMI (7, 10), and the genetic predisposition to WHR adjusted for BMI is predictive of diabetes (11). All these results indicate that WHR may differ from BMI in biology. Recently, a Swedish cohort study reported that changes in BMI and WHR over a 10-year period were related to changes in different plasma proteins (6). Nevertheless, it remains unclear to what extent the plasma proteome of WHR differs from that of BMI, and more importantly, whether these differences have clinical implications, for instance could confer differential risk of diabetes.

In the present study, we aimed to investigate differences in protein profiles between BMI and WHR based on a targeted proteomics approach, and, of particular interest, to see if the identified BMI- and WHR-specific proteins are prospectively related to risk of diabetes, a common consequence of obesity. Our results will provide some insight into plausible mechanisms underlying the different roles of general obesity and abdominal fat distribution in health.

Materials and Methods

Study Participants

The Malmö Diet and Cancer (MDC) study is a large prospective cohort study, with participants enrolled from the general population of Malmö, Sweden (12). From 1991 to 1994, a random sample (n = 6103) from the MDC population was recruited to a subcohort named the Malmö Diet and Cancer-Cardiovascular Cohort (MDC-CC). A total of 4913 participants had fasting blood samples obtained and were free of diabetes and cardiovascular diseases (CVDs) at baseline. Among them, 4348 had blood samples sent for plasma protein analysis, out of which 4214 passed the internal quality control. Of the remaining participants, 11 with missing data on smoking (n = 5) or low-density lipoprotein cholesterol (LDL, n = 6) were excluded. Therefore, 4203 participants were left for the final analyses (Fig. 1). All participants signed written informed consent. The study was approved by the Regional Ethical Review Board in Lund, Sweden (LU 51/90) (LU 2012/762) and was in accordance with the Declaration of Helsinki.

Figure 1.

Study population flow chart.

Proteomic Analyses

Fasting plasma samples were stored at –80°C immediately after collection until protein analysis in 2015. Fifty-seven plasma proteins were measured using Olink Proseek® Multiplex Oncology I reagent kit and 92 were measured using Olink Proseek® Multiplex CVD I kit, from the SciLifeLab analysis service (Uppsala, Sweden). There was no overlap in proteins between these 2 panels. The reagents were based on proximity extension assay (PEA) technology (13) where a pair of DNA oligonucleotides linked to a set of matched antibodies was used to detect the target protein in a homogeneous assay. Upon binding of the 2 oligonucleotides, a DNA duplex was formed, which then was extended by a DNA polymerase to generate a real-time PCR amplicon. Since only correctly matched DNA string pairs could generate detectable and quantifiable signals, the PEA technology exhibited high specificity and excellent sensitivity (13, 14). Values were presented as normalized protein expression (NPX) units, which were real-time qPCR cycle values on a log2 scale (15). Quality control over the technical performance of the assays and the samples was performed using the Olink® NPX Manager software, the basis of which depends on 4 internal controls that are spiked into all samples and external controls in every Olink analysis. Values below the lower limits of detection as determined from a negative control were imputed as limit of detection/2. Proteins with a call rate less than 75% were removed (n = 13), resulting in 136 proteins available for proteomic analyses (see Table S1 for a full list of the included proteins (16)). Due to different call rates for proteins, samples used for different proteins varied between 3478 and 4198 subjects. Also see the Olink webpage (http://www.olink.com) for detailed information regarding proteomic panels, PEA technology, assay performance, quality control and validation.

Anthropometric Measurements

Baseline anthropometric measurements were conducted by trained nurses following standard procedures. Standing height (m) was measured using a fixed calibrated stadiometer. Weight (kg) was measured using balance-beam scales while subjects wore light clothing without shoes, and the values were recorded to the nearest 0.1 kg. Waist circumference (cm) was measured midway between the lowest rib margin and iliac crest, and hip circumference (cm) was determined as the largest circumference between waist and thighs. BMI was calculated as weight in kilograms divided by the square of height in meters (kg/m2), and WHR was calculated as waist circumference divided by hip circumference.

Other Measurements

Blood pressure (BP, mmHg) was measured with a mercury column sphygmomanometers after a 10-minute rest in a supine position. Medical history, medications, alcohol consumption, and smoking habits were assessed in self-administered questionnaires. Smoking status was categorized as nonsmokers or current smokers. High alcohol consumption was defined as >40 g of alcohol per day for men and >30 g per day for women. After an overnight fast, fresh blood samples were drawn to measure fasting blood glucose (mmol/L) and plasma lipids according to standard procedures at the Department of Clinical Chemistry, University Hospital, Malmö. LDL (mmol/L) was determined based on the Friedewald’s formula.

Ascertainment of Endpoints

All participants were free of diabetes at baseline and were followed until incident diabetes, emigration from Sweden, death, or the end of follow-up on December 31, 2016, whichever came first. Information on cases of new-onset diabetes was retrieved from validated national and local registers (17), including the Swedish National Diabetes Register, the regional Diabetes 2000 Register of the Scania region, the Malmö HbA1c Register, the Swedish Inpatient register, the Swedish Outpatient register (for hospital-based policlinics), and the nationwide Swedish Drug Prescription Register. In the Swedish National Diabetes Register and the Diabetes 2000 Register, new cases of diabetes were diagnosed according to established criteria (fasting plasma glucose concentration ≥7.0 mmol/L with 2 repeated tests on separate occasions). In the Malmö HbA1c Register, if an individual had at least 2 HbA1c records ≥6.0% (42 mmol/mol) with the Swedish Mono-S standardization system (corresponding to 7.0% [53 mmol/mol] according to the US National Glycohemoglobin Standardization Program) during the follow-up, he or she would be considered to have incident diabetes. In the Swedish Inpatient and Outpatient registers, diabetes was diagnosed by a senior physician. In the nationwide Drug Prescription Register, a filled prescription of insulin or glucose lowering medications (anatomical–therapeutic–chemical classification code A10) was required for diagnosis of diabetes.

Statistical Analyses

Correlation between obesity measures was estimated by Pearson’s correlation analysis. Baseline characteristics were assessed across quartiles of BMI and WHR, respectively. Sex-specific quartile limits were used for BMI or WHR to ensure equal sex ratios in each quartile. Categorical variables are presented as numbers (percentages) while continuous variables are presented as mean ± SD or median (interquartile range). Differences in continuous or categorical variables across quartiles were examined by linear or logistic regression, respectively.

First, the association of BMI (both in quartiles and per 1 SD change) with incidence of diabetes were determined using Cox proportional hazards regression with time-on-study as the timescale. WHR was adjusted for as a covariate. Other potential confounders were taken into account: age, sex, smoking, high alcohol consumption, systolic BP, LDL, and antihypertensive medication. The association of WHR with incidence of diabetes was similarly assessed after adjusting for BMI and other potential confounders.

Second, proteins associated with BMI or WHR were discovered with a 2-step iterative resampling (TSIR) approach to optimize internal replicability. This method was primarily used for loci discovery in genome-wide association studies where the power and efficiency have been proven (18). A 70:30 split into discovery and replication sets and a cutoff of 20 out of 100 replicated pairs showed excellent performance to determine proteins to be internally replicated while providing conservative type I error control (18). It has also been used for protein discovery in a recent study from MDC-CC (19). Similarly, in the present study, all participants were randomly divided into the discovery (2/3 of the population) and replication (1/3 of the population) sets. Random splits were repeated 100 times. In the first step, multivariate linear regressions were separately conducted for every single protein (as the dependent variable) to explore its association with BMI and WHR (as the independent variables). BMI and WHR were adjusted for each other to examine the associations of proteins with BMI adjusted for WHR and with WHR adjusted for BMI, respectively. Covariates taken into account were age, sex, smoking, high alcohol consumption, systolic BP, LDL, and antihypertensive medication. Afterward, the proteins with significant associations discovered in the first step were similarly studied in relation to BMI and WHR in the replication set. Both BMI and WHR were used as continuous variables (per 1 SD change) in the TSIR approach. P-values adjusted for false discovery rate (FDR) of .05 by the Benjamini–Hochberg method were always used to identify proteins with significant associations with BMI or WHR. Proteins were considered internally replicated if they were significantly (FDR-adjusted at .05) associated with BMI (or WHR) in the same direction in both the discovery and replication sets in at least 20 out of 100 of the randomly generated pairs.

Third, for each of the single internally replicated protein that was associated with BMI, but not with WHR (or vice versa), multivariate linear regression was similarly performed using the whole study sample. BMI and WHR were included simultaneously in the same model, and β coefficients were generated for BMI and WHR, respectively, regarding their associations with a specific protein. These 2 β coefficients were compared using an F-test. A protein was considered to be specific to BMI if it was associated with BMI but not WHR in the TSIR procedure, and was significantly more closely associated with BMI than with WHR in β coefficient comparisons. Likewise, WHR-specific proteins were identified. These proteins were then filtered by an orthogonality criterion (20). Proteins with large Spearman correlations (|r|>0.8) with others were removed by taking the protein from each cluster with more associations with BMI or WHR in the TSIR approach as its representative. After orthogonal filtering, Spearman correlations coefficients between each pair of BMI- or WHR-specific proteins were then visualized using heat-map graphs. To facilitate the interpretation of the findings, protein–protein interaction (PPI) network analysis was performed using STRING database (http://www.string-db.org). The interactions among the BMI- or WHR-specific proteins after the filtering were illustrated using Cytoscape 3.5 (http://www.cytoscape.org/) with MCODE (Cytoscape plug-in) used to carry out the module analysis. Gene Ontology functional and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were further conducted using DAVID 6.8 (https://david.ncifcrf.gov/) to show the critical biological implications associated with the identified proteins. Data from the Genotype-Tissue Expression (GTEx) project (https://gtexportal.org/home/) were used to show expression levels of genes corresponding to BMI- and WHR-specific proteins in tissues that are highly relevant to obesity and insulin resistance, such as subcutaneous and visceral adipose tissue, liver, skeletal muscle, and whole blood.

Longitudinal associations of BMI- or WHR-specific proteins with incidence of diabetes were investigated using Cox proportional hazards regression with Z-score standardized proteins and with time-on-study as the timescale. Hazard ratios (HRs) with 95% CIs were estimated after adjusting for age, sex, smoking, high alcohol consumption, systolic BP, LDL, antihypertensive medication, and BMI (for WHR-specific proteins) or WHR (for BMI-specific proteins). In a sensitivity analysis, the P values were corrected using a more strict FDR correction by taking all the protein-diabetes associations (multiple testing = 136 times) into consideration. Potential interactions between protein levels and sex were examined using multiplicative interaction terms in the Cox models. Volcano plots were drawn with Graphpad Prism 9 to illustrate the associations of 136 plasma proteins with BMI and WHR, respectively. All the other analyses were performed using the Statistical Analysis System version 9.3 for Windows (SAS Institute, Cary, NC, USA).

Results

Study Population Characteristics

A total of 4203 participants were included, with a mean age of 57.2 ± 6.0 years, 37.8 % of whom were men. The mean BMI was 25.8 ± 3.2 kg/m2 for men and 25.2 ± 4.0 kg/m2 for women, while the mean WHR was 0.94 ± 0.06 for men and 0.78 ± 0.05 for women. A moderate correlation was found between BMI and WHR (r = 0.58, P < .0001 for men and r = 0.35, P < .0001 for women). Baseline characteristics of participants across quartiles of BMI or WHR are shown elsewhere (Tables S2 and S3 (16)). WHR, but not BMI, increased with high alcohol consumption, while BMI, but not WHR, increased with age and decreased with smoking. Other cardiometabolic risk factors showed a similar increasing trend across BMI or WHR quartiles (all P for trend < .0001).

Predictive Values of BMI and WHR for Incidence of Diabetes

After a mean follow-up of 20.3 ± 5.9 years, 619 participants developed incident diabetes. As presented in Table 1, the association between WHR and diabetes was relatively stronger than the association between BMI and diabetes. The adjusted HR was 1.31 (95% CI 1.22-1.42, P < .0001) and 1.52 (95% CI 1.33-1.72, P < .0001) for per 1 SD change in BMI and WHR, respectively.

Table 1.

Incidence of diabetes in relation to sex-specific quartiles of BMI or WHR (n = 4203)

	Q1	Q2	Q3	Q4	P for trenda	Per 1 SD	P
BMI
Incidence of diabetes	108	115	137	259	—	619	—
Incidence per 1000 person years	4.94	5.23	6.36	13.1	—	7.27	—
Model 1b	Reference	1.02 (0.78, 1.32)	1.19 (0.92, 1.53)	2.30 (1.82, 2.91)	<.0001	1.43 (1.33, 1.53)	<.0001
Model 2c	Reference	0.95 (0.73, 1.24)	1.03 (0.80, 1.34)	1.77 (1.39, 2.26)	<.0001	1.31 (1.22, 1.42)	<.0001
WHR
Incidence of diabetes	96	160	152	211	—	619	—
Incidence per 1000 person years	3.81	6.7	8.23	12	—	7.27	—
Model 1b	Reference	1.31 (0.99, 1.74)	1.69 (1.29, 2.22)	2.48 (1.92, 3.19)	<.0001	1.71 (1.54, 1.89)	<.0001
Model 2c	Reference	1.24 (0.93, 1.64)	1.47 (1.12, 1.93)	1.90 (1.45, 2.47)	<.0001	1.52 (1.33, 1.72)	<.0001

Abbreviations: BMI, body mass index; SD, standard deviation; WHR, waist-to-hip ratio.

Analysis by Cox proportional hazards model.

Hazard ratios (95% CI) adjusted for age, sex, smoking, high alcohol consumption, systolic blood pressure, low density lipoprotein, and antihypertensive medication.

Adjusted for covariates in Model 1 plus WHR or BMI.

Plasma Proteins Specific to BMI or WHR

The association between plasma proteins and BMI and WHR, before applying the TSIR approach and without mutual adjustment for BMI and WHR, is shown in Figure S1 (16).

The TSIR approach identified 59 and 45 plasma proteins to be internally replicated to associate with BMI or WHR, respectively, including 21 overlapping proteins (Tables S4 and S5 (16)). Therefore, 38 proteins were associated with BMI but not WHR, while 24 were associated with WHR but not BMI. As shown in Table 2, out of these proteins, F-tests confirmed that 10 were specific to BMI, including 3 with positive associations. In addition, 22 proteins were identified to be specific to WHR; all of which except for proto-oncogene tyrosine protein kinase Src showed positive associations.

Table 2.

Internally replicated proteins that are specific to BMI or WHR (n = 4203)

BMI					WHR
Proteins	N	β (SE)	FDR-adjusted P1b	FDR-adjusted P2c	Proteins	N	β (SE)	FDR-adjusted P1b	FDR-adjusted P2c
AM	4131	0.27 (0.02)	1.00 × 10–30	6.11 × 10–07	PRSS8	4198	0.10 (0.01)	2.75 × 10–12	.00017
IL6	4131	0.13 (0.02)	2.43 × 10–14	.47	IL18	4131	0.16 (0.03)	4.27 × 10–07	.0069
TNFR1	4131	0.17 (0.02)	3.70 × 10–22	.05	CCL20	4131	0.16 (0.03)	5.07 × 10–07	.0074
CSF1	4131	0.15 (0.02)	9.26 × 10–17	.1	CCL4	4131	0.16 (0.03)	5.07 × 10–07	.0029
IL16	4131	0.14 (0.02)	1.42 × 10–14	.068	IL8	4131	0.16 (0.03)	8.40 × 10–07	9.44 × 10–05
RAGE	4131	–0.16 (0.02)	2.94 × 10–18	.05	CDKN1A	4091	0.13 (0.03)	2.34 × 10–06	.0019
GAL	4131	–0.15 (0.02)	8.48 × 10–17	.0051	CXCL6	4131	0.15 (0.03)	2.95 × 10–06	.003
VEGFD	4131	–0.16 (0.02)	1.14 × 10–18	.083	CXCL11	4194	0.14 (0.03)	6.88 × 10–06	.0019
EMMPRIN	4198	0.04 (0.01)	6.03 × 10–15	2.22 × 10–06	GDF15	4131	0.13 (0.03)	2.95 × 10–06	.0024
ILT3	4179	0.05 (0.01)	1.41 × 10–09	.71	CHI3L1	4131	0.13 (0.03)	5.48 × 10–06	.0031
IL12	4198	0.08 (0.01)	3.17 × 10–11	.27	CA125	4131	0.14 (0.03)	1.09 × 10–05	.0031
TRANCE	4131	0.12 (0.02)	3.57 × 10–11	.21	LAPTGFbeta1	4198	0.08 (0.02)	5.27 × 10–06	.003
EZR	4121	0.04 (0.01)	9.87 × 10–11	.05	CTSL1	4131	0.13 (0.03)	1.59 × 10–05	.039
TNFR2	4131	0.11 (0.02)	8.66 × 10–10	.55	TNFSF14	4131	0.12 (0.03)	.00016	.075
ESM1	4131	–0.11 (0.02)	8.66 × 10–10	.18	CASP3	4131	0.12 (0.03)	7.42 × 10–05	.003
MB	4131	0.09 (0.02)	6.22 × 10–08	.068	CXCL5	4198	0.11 (0.03)	.00013	.0046
TRAIL	4131	0.10 (0.02)	6.54 × 10–08	.52	HBEGF	4131	0.12 (0.03)	.00013	.0079
FRalpha	4198	–0.05 (0.01)	5.19 × 10–08	.21	CXCL13	4128	0.08 (0.02)	8.34 × 10–05	.012
KLK6	4131	–0.10 (0.02)	5.19 × 10–08	.024	IL7	4149	0.07 (0.02)	.00024	.0074
Gal3	4131	0.10 (0.02)	4.11 × 10–08	.016	Dkk1	4131	0.11 (0.03)	.00045	.003
NTRK3	4198	–0.04 (0.01)	3.10 × 10–08	.67	MMP7	4131	0.11 (0.03)	.00032	.0029
hK11	4131	–0.10 (0.02)	5.33 × 10–08	.0021	MICA	3726	0.07 (0.02)	.00056	.22
TRAILR2	4131	0.09 (0.02)	5.19 × 10–08	.83	OPG	4131	0.10 (0.03)	.0006	.0023
MMP10	4131	–0.10 (0.02)	2.47 × 10–08	.0051	SRC	4131	-0.11 (0.03)	.00032	.003
HE4	4198	–0.04 (0.01)	6.16 × 10–08	.00047
REG4	4198	–0.04 (0.01)	4.88 × 10–07	.016
PTX3	4131	–0.08 (0.02)	6.73 × 10–06	.13
mAmP	4131	–0.09 (0.02)	2.41 × 10–06	.00075
ITGA1	4198	–0.03 (0.01)	6.99 × 10–05	.73
ECP	4131	0.07 (0.02)	5.15 × 10–05	.73
ErbB4HER4	4198	–0.03 (0.01)	4.61 × 10–05	.22
MIA	4198	–0.03 (0.01)	8.10 × 10–05	.068
TM	4131	0.07 (0.02)	0.0001	.55
AGRP	4131	–0.07 (0.02)	8.10 × 10–05	.18
IL6RA	4131	0.07 (0.02)	0.00023	.21
Flt3L	4198	0.03 (0.01)	0.00028	.55
SCF	4131	–0.06 (0.02)	0.00042	.68
MPO	4131	0.06 (0.02)	0.00093	.99

AGRP, Agouti-related protein; AM, Adrenomedullin; BMI, body mass index; CA125, Ovarian cancer–related tumor marker CA 125; CASP3, Caspase-3; CCL20, C-C motif chemokine 20; CCL4, C-C motif chemokine 4; CDKN1A, Cyclin-dependent kinase inhibitor 1; CHI3L1, Chitinase-3-like protein 1; CSF1, Macrophage colony-stimulating factor 1; CTSL1, Cathepsin L1; CXCL11, C-X-C motif chemokine 11; CXCL13, C-X-C motif chemokine 13; CXCL5, C-X-C motif chemokine 5; CXCL6, C-X-C motif chemokine 6; Dkk1, Dickkopf-related protein 1; ECP, Eosinophil cationic protein; EMMPRIN, Extracellular matrix metalloproteinase inducer; ErbB4HER4, Receptor tyrosine-protein kinase erbB-4; ESM1, Endothelial cell-specific molecule 1; EZR, Ezrin; FDR, false discovery rate; Flt3L, Fms-related tyrosine kinase 3 ligand; FRalpha, Folate receptor alpha; GAL, Galanin peptides; Gal3, Galectin-3; GDF15, Growth/differentiation factor 15; HBEGF, heparin binding epidermal growth factor like growth factor; HE4, Epididymal secretory protein E4; hK11, Kallikrein-11; IL12, Interleukin-12; IL16, Interleukin-16; IL18, Interleukin-18; IL6, Interleukin-6; IL6RA, Interleukin-6 receptor subunit alpha; IL7, Interleukin-7; IL8, Interleukin-8; ILT3, Immunoglobulin-like transcript 3; ITGA1, Integrin alpha-1; KLK6, Kallikrein-6; LAPTGFbeta1, Latency-associated peptide transforming growth factor beta-1; mAmP, Membrane-bound aminopeptidase P; MB, Myoglobin; MIA, Melanoma-derived growth regulatory protein; MICA, MHC class I polypeptide-related sequence A; MMP10, matrix metalloproteinase-10; MMP7, matrix metalloproteinase-7; MPO, Myeloperoxidase; NTRK3, NT-3 growth factor receptor; OPG, Osteoprotegerin; PRSS8, Prostasin; PTX3, Pentraxin-related protein PTX3; RAGE, Receptor for advanced glycosylation end products; REG4, Regenerating islet-derived protein 4; SCF, Stem cell factor; SE, standard error; SRC, Proto-oncogene tyrosine-protein kinase Src; TM, Thrombomodulin; TNFR1, Tumor necrosis factor receptor 1; TNFR2, Tumor necrosis factor receptor 2; TNFSF14, Tumor necrosis factor ligand superfamily member 14; TRAIL, TNF-related apoptosis-inducing ligand; TRAILR2, TNF-related apoptosis-inducing ligand receptor 2; TRANCE, TNF-related activation-induced cytokine; VEGFD, Vascular endothelial growth factor D; WHR, waist-to-hip ratio.

A protein was considered to be specific to BMI if (1) it was internally replicated to associate with BMI but not WHR and (2) the β coefficient for its association with BMI statistically differed from that for its association with WHR in multivariate linear regressions (and vice versa for proteins specific to WHR). Proteins specific to BMI or WHR were highlighted in bold font.

Linear regressions with BMI and WHR adjusted for each other in the same model and adjusted for age, sex, smoking, high alcohol consumption, systolic blood pressure, low density lipoprotein, and antihypertensive medication. FDR-adjusted P1 < .05 was used as the cutoff.

F-tests were conducted to compare if the absolute values for weighs of BMI and WHR, as reflected by β coefficients, differed significantly in multivariate linear regressions. FDR-adjusted P2 < .05 was used as the cutoff.

With the orthogonal screening, all BMI-specific proteins showed weak to moderate correlations with each other (largest Spearman correlation |r|=0.61). As for WHR-specific proteins, heparin-binding endothelial growth factor-like growth factor and cyclin-dependent kinase inhibitor 1 showed high correlation with dickkopf-related protein 1 (Dkk1, r = 0.82) and caspase-3 (r = 0.83), respectively, and had more associations with WHR in the TSIR approach (see Table S5 (16)). After excluding Dkk1 and caspase-3, correlations between pairs of the remaining proteins are shown in Figure S2 (16). According to the PPI analysis (Figure S3 (16)), the most important module for WHR-specific proteins includes 9 nodes and 68 edges (Module 1). All the genes included are those encoding chemokines, chemokine receptors, or interleukins. Relatively loose interaction was observed among genes corresponding to the BMI-specific proteins. Figure S4 (16) shows the top Gene Ontology terms. Genes corresponding to WHR-specific proteins are primarily enriched in extracellular space, chemokine activity, cell–cell signaling, chemokine-mediated signaling pathway, inflammatory response, and neutrophil chemotaxis. Genes corresponding to BMI-specific proteins are primarily enriched in extracellular space, serine-type endopeptidase activity, and proteolysis. The KEGG pathway enrichment analysis (Figure S5 (16)) indicates that the genes corresponding to WHR-specific proteins are mainly involved in cytokine–cytokine receptor interaction. However, no significantly enriched KEGG pathways were found for BMI-specific proteins. According to Genotype-Tissue Expression (Figure S6 (16)), among the examined tissues that are highly relevant to obesity and insulin resistance, adipose tissue is a major production site for most of the BMI- or WHR-specific proteins. The expression levels of different transcripts can vary widely between subcutaneous and visceral fat. Those encoding commonly known inflammatory proteins are more likely to be expressed in visceral fat than in subcutaneous fat (eg, CCL4, CCL20, CXCL8, IL18, and CXCL6), and the encoded proteins (eg, C-C motif chemokines 4 and 20 [CCL4, CCL20], interleukins 8 and 18 [IL8, IL18], and C-X-C motif chemokine 6 [CXCL6], respectively) tend to be WHR-specific proteins.

Associations of BMI- and WHR-specific Proteins With Incidence of Diabetes

As presented in Table 3, only 2 (ie, adrenomedullin and galectin-3) of the 10 BMI-specific proteins were associated with increased risk of diabetes after adjusting for potential confounders (Model 1). None of the associations remained significant after additionally adjusting for WHR. Notably, 19 of the 22 WHR-specific proteins were positively associated with diabetes in Model 1. Adding BMI as a covariate only had minor impact on the associations. Significant HRs were still observed for 18 proteins (Model 2). In a sensitivity analysis, after performing a more strict FDR correction including 136 times of multiple testing, the significance for the protein–diabetes associations was generally consistent with those in Model 2. Proteins that remained significant in Model 2 were marked in red in the volcano plots (Figure S7 (16)). No statistically significant gender differences were noted in the associations between proteins levels and risk of diabetes (all P for interaction > .05).

Table 3.

Associations with diabetes incidence for internally replicated proteins that are specific to BMI or WHR (n = 4203)

Proteins		Model 1a		Model 2b		Sensitivity analysisc
	N	HR (95% CI)	FDR-adjusted P	HR (95% CI)	FDR-adjusted P	Adjusted P
BMI
AM	4131	1.16 (1.06, 1.27)	.0092	1.11 (1.01, 1.21)	.13	.055
GAL	4131	0.93 (0.85, 1.01)	.27	0.95 (0.87, 1.03)	.36	.32
EMMPRIN	4198	1.07 (0.98, 1.16)	.27	1.07 (0.99, 1.16)	.31	.15
KLK6	4131	0.95 (0.88, 1.04)	.34	0.96 (0.89, 1.05)	.44	.45
Gal3	4131	1.12 (1.03, 1.22)	.039	1.12 (1.03, 1.22)	.073	.02
hK11	4131	0.94 (0.87, 1.02)	.27	0.94 (0.87, 1.02)	.31	.23
MMP10	4131	0.95 (0.88, 1.04)	.34	0.95 (0.88, 1.04)	.36	.35
HE4	4198	1.03 (0.94, 1.12)	.52	1.04 (0.95, 1.13)	.50	.55
REG4	4198	0.97 (0.89, 1.05)	.45	0.98 (0.90, 1.06)	.54	.62
mAmP	4131	1.06 (0.98, 1.15)	.28	1.06 (0.98, 1.15)	.31	.24
WHR
PRSS8	4198	1.29 (1.18, 1.42)	3.09 × 10 –07	1.25 (1.15, 1.37)	1.38 × 10 –05	1.22 × 10 –05
IL18	4131	1.18 (1.09, 1.28)	.00034	1.14 (1.05, 1.24)	.0038	.0073
CCL20	4131	1.23 (1.14, 1.32)	3.63 × 10 –07	1.20 (1.12, 1.30)	1.38 × 10 –05	1.71 × 10 –05
CCL4	4131	1.15 (1.07, 1.24)	.00053	1.13 (1.05, 1.22)	.0036	.0071
IL8	4131	1.15 (1.06, 1.24)	.00082	1.15 (1.06, 1.24)	.0016	.0027
CDKN1A	4091	1.14 (1.05, 1.23)	.0023	1.11 (1.03, 1.21)	.011	.023
CXCL6	4131	1.19 (1.10, 1.29)	.00012	1.17 (1.08, 1.26)	.00093	.0013
CXCL11	4194	1.01 (0.93, 1.10)	.78	1.00 (0.92, 1.08)	.91	.95
GDF15	4131	1.20 (1.09, 1.31)	.00037	1.18 (1.08, 1.30)	.0016	.0027
CHI3L1	4131	1.14 (1.05, 1.24)	.0037	1.11 (1.02, 1.21)	.016	.034
CA125	4131	1.07 (0.98, 1.16)	.12	1.05 (0.97, 1.14)	.26	.36
LAPTGFbeta1	4198	1.17 (1.08, 1.27)	.00028	1.15 (1.06, 1.24)	.0016	.0027
CTSL1	4131	1.18 (1.08, 1.28)	.00037	1.14 (1.05, 1.24)	.0042	.0090
CASP3	4131	1.11 (1.02, 1.20)	.017	1.10 (1.01, 1.19)	.035	.066
CXCL5	4198	1.13 (1.04, 1.23)	.0044	1.12 (1.03, 1.22)	.012	.025
HBEGF	4131	1.14 (1.06, 1.24)	.0016	1.13 (1.04, 1.22)	.0042	.0090
CXCL13	4128	1.10 (1.02, 1.19)	.022	1.08 (1.00, 1.17)	.064	.11
IL7	4149	1.15 (1.06, 1.24)	.00082	1.14 (1.05, 1.23)	.0031	.0054
Dkk1	4131	1.14 (1.05, 1.24)	.0021	1.14 (1.05, 1.24)	.0036	.0071
MMP7	4131	1.20 (1.10, 1.31)	.00016	1.21 (1.11, 1.32)	.00014	.00022
OPG	4131	1.12 (1.02, 1.22)	.018	1.12 (1.03, 1.22)	.015	.031
SRC	4131	0.94 (0.87, 1.02)	.13	0.94 (0.87, 1.02)	.13	.19

Plasma proteins are standardized. Significant P values were highlighted in bold font.

Abbreviations: AM, Adrenomedullin; BMI, body mass index; CA125, Ovarian cancer-related tumor marker CA 125; CASP3, Caspase-3; CCL20, C-C motif chemokine 20; CCL4, C-C motif chemokine 4; CDKN1A, Cyclin-dependent kinase inhibitor 1; CHI3L1, Chitinase-3-like protein 1; CI, confidence interval; CTSL1, Cathepsin L1; CXCL11, C-X-C motif chemokine 11; CXCL13, C-X-C motif chemokine 13; CXCL5, C-X-C motif chemokine 5; CXCL6, C-X-C motif chemokine 6; Dkk1, Dickkopf-related protein 1; EMMPRIN, Extracellular matrix metalloproteinase inducer; FDR, false discovery rate; GAL, Galanin peptides; Gal3, Galectin-3; GDF15, Growth/differentiation factor 15; HBEGF, heparin binding epidermal growth factor like growth factor; HE4, Epididymal secretory protein E4; hK11, Kallikrein-11; HR, hazard ratio; IL18, Interleukin-18; IL7, Interleukin-7; IL8, Interleukin-8; KLK6, Kallikrein-6; LAPTGFbeta1, Latency-associated peptide transforming growth factor beta-1; mAmP, Membrane-bound aminopeptidase P; MMP10, Matrix metalloproteinase-10; MMP7, Matrix metalloproteinase-7; OPG, Osteoprotegerin; PRSS8, Prostasin; REG4, Regenerating islet-derived protein 4; SRC, Proto-oncogene tyrosine-protein kinase Src; WHR, waist-to-hip ratio. FDR-adjusted P < 0.05 was used as the cutoff.

Cox proportional hazards regressions adjusted for age, sex, smoking, high alcohol consumption, systolic blood pressure, low density lipoprotein, and antihypertensive medication.

Adjusted for covariates in Model 1 plus WHR or BMI.

P values generated in multivariate analysis (Model 2) were corrected using a more strict FDR correction by taking all the protein–diabetes associations (multiple testing = 136 times) into consideration.

Discussion

The present observational study identified plasma proteins that were specifically associated with BMI or WHR, respectively. WHR-specific proteins, but not BMI-specific proteins, were positively associated with increased risk of diabetes. From a proteomic perspective, we demonstrated for the first time that compared with general obesity, abdominal fat distribution may have some unique pathological features, which could potentially explain its ability to confer additional risk of diabetes.

Previously, several other studies have linked Olink proteomics to obesity. Ferreira et al (21) identified discriminant protein biomarkers for obesity (BMI ≥ 30 kg/m2) using a biostatistical method. Folkersen et al (22) found causal evidence for the associations of several proteins with BMI or WHR using a Mendelian randomization analysis. In addition, Pang et al (23) reported that only some of the BMI-associated proteins were predictive of CVD risk. However, none of those studies focused on the unique proteomic features of different obesity measures and their role in diabetes. WHR was only moderately correlated with BMI in our study. Meanwhile, in line with previous observations (8, 9), WHR was predictive of diabetes independent of BMI, clearly indicating that WHR identified different and unique aspects of obesity-related risk from that of BMI. Previous genome-wide association studies have revealed different genetic bases for BMI and WHR (7, 10). In addition, several studies indicate that BMI and body fat distribution are differently linked to circulating proteins (6, 24, 25). Typically, only a few well-studied proteins have been investigated (24, 25), such as adiponectin, CRP, IL6, and tumor necrosis factor α. Unlike prior studies, we focused mainly on the proteins that were specific for BMI or WHR. We also included a broader spectrum of proteins and adopted a TSIR approach (18, 19) to ensure internal replicability for identified proteins. More importantly, we took a further step to investigate whether the identified proteins could drive differential metabolic risks, which has not been explored earlier.

In our TSIR approach, a significant overlap was found between proteins associated with BMI and WHR. Most of the overlapping proteins are among the top proteins that show positive associations with BMI (eg, leptin [LEP], fatty acid-binding protein, adipocyte [FABP4] and IL1 receptor antagonist protein [IL1ra]) or WHR (eg, cathepsin D [CTSD], CCL3, tissue-type plasminogen activator [tPA]). LEP and FABP4 are adipokines with well-established roles in adipocyte biology (6). IL1ra and CCL3 are known to be involved obesity-associated inflammation (26). CTSD is a lysosomal aspartyl protease that could be activated by weight gain and contributes to adipocyte death (27). TPA, a serine protease, is also increased in obesity, possibly due to a compensatory response to impaired fibrinolysis caused by obesity-associated metabolic stress (28). BMI and WHR still share some proteomic features, which is not surprising since they both measure obesity. However, obesity is also a heterogeneous disorder. In this study, both BMI and WHR have their unique proteins. Most of these proteins are predominantly expressed in adipose tissue but levels of expression vary between subcutaneous and visceral fat. This suggests that the different protein profiles of BMI and WHR may at least partly result from discrepancy between protein expression in visceral and subcutaneous fat depots.

BMI- and WHR-specific proteins appear to be involved in different biological processes. Our results indicate that proteolysis is the primary biological process associated with the BMI-specific proteins. Consistently, a major proportion of the BMI-specific proteins are proteases (eg, kallikrein-6, kallikrein-11, matrix metalloproteinase [MMP] 10, and membrane-bound aminopeptidase P) or could importantly regulate proteolytic activity (eg, extracellular matrix metalloproteinase inducer and epididymal secretory protein E4 (29)). In line with this finding, previous data showed that proteolytic systems (eg, the ubiquitin–proteasome system and autophagy) are altered in obesity (30) and increased proteolysis in obesity may be related to consequences such as muscle catabolism (31). In contrast, WHR-specific proteins are primarily involved in cell–cell signaling, chemokine activity, and inflammatory and immune responses. Unlike a loose connection observed among BMI-specific proteins in the PPI network, WHR-specific proteins seem tightly connected, with chemokines, chemokine receptors, and interleukins forming the most important module. Together, these results support that heterogeneity exists in obesity. Compared with BMI, high WHR is more likely to occur together with signs of inflammation.

Most of the WHR-specific proteins remained predictive of diabetes risk even after multivariate adjustment. It is speculated that WHR-related inflammatory signals may at least partly explain or mirror why WHR contributes to metabolic risk in addition to BMI. Inflammation is a key link between obesity and diabetes (32). Among WHR-specific proteins that are predictive of diabetes, several have already been shown to be involved in adipose inflammation and diabetes-related pathogenesis (IL18, IL8, CCL20, CCL4, CXCL6 (26, 32, 33)). In contrast, the increased expression of growth/differentiation factor 15 might be a compensatory mechanism against inflammation but inadequate to reduce inflammation-related diabetes risk (15, 34). Some proteins might be involved in pathways related to adipogenesis or inflammation, for instance the canonical Wnt-β–catenin pathway (eg, Dkk1 (35) and MMP7 (36)) or its downstream adipogenic transcription factor, peroxisome proliferator–activated receptor gamma (eg, cathepsin L1 (37) and osteoprotegerin (38)), which has a pivotal role in regulating inflammation, adipogenesis, insulin synthesis and secretion, and pancreatic beta cell proliferation (39). Nevertheless, the observed associations could be causal or the elevated protein may simply reflect a pre-existing inflammatory state and diabetes risk.

In contrast, hardly any of the BMI-specific proteins are predictive of diabetes risk. One possible explanation might be that some of the BMI-associated proteins that could confer diabetes risk are proteins shared by BMI and WHR (eg, E-selectin (22)). In addition, even though some diabetes risk–related proteins showed association with BMI but not WHR, their association with BMI vs WHR did not significantly differ. These proteins (eg, receptor for advanced glycosylation end products (22)) were still excluded because we used relatively strict filter out strategies aiming at focusing on the specific proteins.

Strengths of the present study included the prospective design, the long-term follow-up with a high follow-up rate, and a large number of proteins measured at baseline. In addition, the adoption of the TSIR approach significantly increased internal validity. However, the absence of an external validation cohort was a major limitation which made it inappropriate to directly translate any specific protein into the clinical setting. Our study has several other limitations that need to be considered. First, targeted proteomic analysis was conducted on proteins known or suspected to be involved in tumors or CVDs. Thus other proteins of significance could have been missed. Second, both BMI and WHR could be subject to measurement errors. Even though both usually change slowly over time, it is possible that these anthropometric measures changed over the long follow-up period. However, if anything, this should be similar for both measures and reduce the relationships with diabetes. Third, although we excluded all individuals with prevalent diabetes at baseline, and a large majority of incident diabetes can be assumed to be type 2 diabetes, it is still possible that a few were latent autoimmune diabetes in the adult patients. This number can be assumed to be small though. Fourth, even though the diabetes registers have been validated previously (17), different parameters were used in different registries and the definition of incident diabetes was not standardized. Fifth, sample sizes for analyzing different proteins varied between 3478 and 4198 subjects, which might potentially restrict the direct comparison between the proteins. Last, due to the nature of an epidemiological design, residual confounding cannot be ruled out even after extensive adjustment for potential confounders.

Taken together, the present study indicates that, compared with general obesity, abdominal fat distribution carries a unique proteomic signature that can contribute to diabetes risk. Once better understood, the findings may help to characterize specific metabolic phenotypes of obesity or risk mechanisms, and suggest better strategies to prevent metabolic disorders related to abdominal fat distribution.