Circulating Biomarker Score for Visceral Fat and Risks of Incident Colorectal and Postmenopausal Breast Cancer: The Multiethnic Cohort Adiposity Phenotype Study

This article is featured in Highlights of This Issue, p. 887

Excess body weight is a major risk factor for diabetes, cardiovascular disease, and cancer (1). It is estimated that overweight and obesity are responsible for 8% of cancers in high-income countries and this attributable burden is expected to increase further (2). An expert review recently concluded that a high body mass index (BMI) increases risk for 13 cancers, including colorectal and postmenopausal breast cancer (3). Excess adiposity alters physiologic functions of adipose tissue, leading to insulin resistance, chronic inflammation, and increased secretion of adipokines and sex steroids (4, 5). Markers of these conditions (e.g., insulin, leptin, sex steroids, adiponectin, C-reactive protein) have individually been associated with cancer promotion (4, 5). This abnormal metabolic profile is particularly dependent on the amount of visceral adipose tissue (VAT) in the intra-abdominal region because of its greater metabolic activity (6) and direct access to portal circulation (7, 8). Thus, VAT is hypothesized to be more relevant to risk of metabolic diseases and cancer (9–11). Most past epidemiologic studies investigating the relationship of body fat and cancer used BMI and/or waist circumference or waist-to-hip ratio (WHR), which are inadequate proxies for VAT, especially in older adults due the redistribution of adipose tissue toward the visceral compartment (12, 13). Thus, previous associations of excess body weight with cancer risk may be inaccurate and, in particular, underestimated for individuals or populations who have a propensity to deposit fat intra-abdominally (14, 15). At present, accurate measurement of VAT can only be achieved by imaging methods, for example, MRI and CT, that are not amenable to epidemiologic investigations. Thus, only a few small studies have directly examined the association of VAT with cancer risk (16, 17). Alternatively, we and others have explored using circulating biomarkers to aid in the prediction of VAT volume (18, 19).

We report here, following the TRIPOD guidelines (20), on the development of a blood-biomarker score representing VAT, as measured by MRI, to better estimate disease risk, in the large Adiposity Phenotype Study (APS) using cross-sectional and prospective data in a uniquely diverse population. We then tested whether this score predicts, independently of BMI, incident postmenopausal breast and colorectal cancer risks in a nonoverlapping sample of the Multiethnic Cohort (MEC).

The MEC is an ongoing prospective study in Hawaii and Los Angeles of lifestyle and genetic risk factors for cancer and other chronic diseases. In 1993–1996, 96,810 men and 118,441 women aged 45–75 years, and mainly of Japanese American, Native Hawaiian, European American, African American, and Latino ancestry, were recruited (21). The mailed baseline questionnaire collected information on demographics, weight and height, reproductive history, smoking, physical activity, and diet, using a validated quantitative food frequency questionnaire (QFFQ). Follow-up questionnaires were mailed every 5 years to update exposures; the 10-year follow-up survey included waist and hip circumferences self-measured with a provided tape measure. A fasting blood sample was collected on approximately 70,000 cancer-free cohort participants mostly in 2001–2005, forming the MEC Biospecimen Sub-Cohort. Institutional Review Boards at University of Hawaii (UH) and University of Southern California (USC) approved the protocol, and a written informed consent was provided by all participants. Blood components were separated within 4 hours of collection and stored in vapor-phase liquid nitrogen. Supplementary Figure S1 provides a flowchart of the study design.

The MEC APS was conducted in 2013–2016 among a subset of MEC members (15). Briefly, 1,861 healthy, nonsmoking, men and women aged 60–77 years, with BMI between 17.1–46.2 kg/m², were enrolled using stratified sampling according to sex, race/ethnicity, and six BMI categories. Exclusion criteria included reported BMI outside the range of 18.5–40 kg/m²; smoking in the past 2 years; soft or metal body implants or amputation; insulin or thyroid medications; and serious medical conditions (e.g., dialysis, chronic hepatitis, previous cancer diagnosis). During the clinic visit at the University of Hawaii (UH) or University of Southern California (USC), participants provided a blood sample after an overnight fast, completed questionnaires, and underwent anthropometric measurements, an abdominal MRI and a whole-body dual-energy X-ray absorptiometry (DXA) scan for total/regional fat measurement (15). The imaging protocol has been published (15) and is summarized in Supplementary Materials. Sixty participants with invalid MRIs were excluded, yielding a final sample size of 1,801. Institutional Review Boards at UH and USC approved the protocol, and a written informed consent was provided by all participants. The studies reported here were conducted in accordance to the ethical guidelines included in the Belmont report and U.S. Common rule.

Cancer diagnoses were identified by linkage of the MEC to the Surveillance, Epidemiology and End Results (SEER) registries for Hawaii and California through December 2013. In addition, deaths were ascertained by linkage to state vital statistics and the National Death Index. During a mean follow-up of 6 years after blood collection, 950 incident cases of invasive postmenopausal breast cancer and 831 incident cases of invasive adenocarcinoma of the colon or rectum with prediagnostic blood samples were identified in the Biospecimen subcohort. One control was matched to each case on area, birth year (±5 years), sex, race/ethnicity, date of blood collection (±3 years), and hours of fasting, and selected so that the control was alive and free of colorectal or breast cancer at the age of the case diagnosis. There was no overlap among participants of the APS and nested case–control studies.

Biomarkers selected for their reported associations to obesity-caused metabolic, hormonal, and inflammation dysfunctions were measured in the APS blood samples at the UH Cancer Center's Analytical Biochemistry Shared Resource (ABSR) in plasma or serum. Choice of the biomarkers is described in the Supplementary Materials and the list of all biomarkers considered is presented in Supplementary Table S1, along with assay and reproducibility information. These analytes were additionally assessed in samples from the MEC Biospecimen repository (established an average of 10.9 years earlier than APS) for 500 APS participants, randomly selected within APS recruitment strata, to evaluate the prospective association of the biomarker score in these stored samples with future VAT (from APS). The samples for the nested breast and colorectal cancer studies were also retrieved from the MEC Biorepository and analyzed using the same methods.

A least absolute shrinkage and selection operator (LASSO) model (22) regressed log VAT on anthropometric measures and the final APS biomarker panel to create an equation for the VAT prediction score. LASSO is a regularized regression technique where coefficients of unimportant variables are constrained to zero; lambda, the constraint parameter, was estimated using 10-fold cross-validation (23). The regression coefficients were the average of the parameters from models run on 100 bootstrap samples, to avoid overfitting. Separate models were run for men and women. While the list of important analytes were almost identical with matching directionality for men and women, the magnitudes of the coefficients differed considerably by sex. Ethnic-specific models were run by sex but they were not more predictive than the sex-specific model in any ethnic/racial group; in particular, African American sex-specific models did not provide better R² values than the general sex-specific models for this group. Therefore, the final models are sex-specific with the same list of predictors. Potential independent variables are listed in Supplementary Table S2, including log-transformed analytes and the anthropometry measures assessed at the APS clinic visit: BMI (weight in kg/height in m²), BMI squared, height (m), and height squared. The final model was determined through the following steps. Complete case analysis was performed with exclusion of the small number of subjects with any missing biomarker data. Sex-specific (n = 2) and sex/race-specific (n = 10) models were run allowing any of the independent variables to enter. A second modeling step was performed including only the independent variables with a standardized coefficient (with variances of predictors set to 1) > 0.3 in absolute value in one of the 12 models. The final model included only variables with a standardized coefficient > 0.3 in absolute value in one of the sex-specific models in the second modeling step. Coefficients are presented in standardized units, as well as natural units. R² values were computed from the final sex-specific models, by sex and race/ethnicity. Also the area under the ROC (AUROC) was computed to determine the ability of the predicted VAT score, to detect by logistic regression, visceral obesity, defined as ≥150 cm² (the median in APS), by sex and by race/ethnicity.

Each case and control in the breast and colorectal cancer studies was assigned a VAT score by applying the final sex-specific VAT prediction model using the analytes measured in the MEC Biospecimen subcohort and the anthropometry data from the baseline questionnaire. Conditional logistic regression of cancer incidence was used to determine the association with the VAT score, with strata defined by matching criteria rather than matched pair indicators to minimize sample loss due to missing data. Adjustment variables are specified in footnote to Table 3. To jointly consider the highly correlated variables of BMI and the VAT score and to distinguish their associations with cancer risk, BMI-adjusted VAT scores were created using the method of residuals (24), where the adjusted scores were defined as the residuals from the sex-specific regression of log VAT on log BMI, translated to the original units by adding a sex-specific median log VAT value. We report ORs and 95% confidence intervals (CI) by tertile of the BMI-adjusted VAT scores, where the tertile cutoff points were based on the sex-specific distribution among cases and controls, as well as a dose–response P value for the log VAT score. The heterogeneity of the associations with cancer risk by sex, race/ethnicity, BMI category or menopausal hormone therapy (MHT) use was evaluated by inclusion of interaction terms for indicators of subgroup (e.g., ethnicity) and the log VAT score. Polytomous models were run to compare the associations by tumor estrogen receptor status for breast cancer and by the number of years between blood draw and diagnosis of cases for both sites. Models were minimally adjusted for the matching factors and then for cancer-specific risk factors. The risk factors adjusted for in the breast cancer models are matching factors and age at blood draw, and the following variables at baseline: menopausal hormone therapy (ever/never), pack-years of smoking, hours/week of moderate-to-vigorous activity, family history of breast cancer in first-degree relatives (yes/no), type (natural/surgical), and age of menopause (<45, 45–49, 50–54, 55+), age at first live birth (<20, 20–29, 30+), number of children (0, 1, 2–3, 4+), ethanol (g/day), and log energy (kcal/day). The risk factors adjusted for in the colorectal cancer models are matching factors and age at blood draw, menopausal hormone therapy (yes/no/male), pack-years of smoking, vigorous activity, multivitamin use (use at least once a week in past year: yes/no), history of polyps, NSAID use (at least twice a week for 1 month or longer: yes/no), family history of colorectal cancer in first-degree relatives (yes/no), and log intakes of alcohol, dietary fiber, dietary folate equivalents from food or supplements, calcium from food or supplements, and energy. Models were additionally adjusted for BMI from the baseline or blood draw or waist circumference from the 10-year follow-up questionnaire. To maximize number of cases and controls available for analysis, we used complete case analysis with missing categories for grouped variables. However, the results are similar to those using complete case analysis with no missing categories and using multiple imputations.

The APS participants' characteristics are presented in Table 1 (the corresponding distributions by sex and ethnicity can be found in ref. 15). Each of the racial/ethnic and sex groups was approximately equally represented. The mean and SD for age was 69.3 ± 2.8 and 69.1 ± 2.7 years, and for BMI was 27.7 ± 4.3 and 28.0 ± 8.7 kg/m², in men and women, respectively. Women had a greater total fat mass and men, a greater VAT amount.

Supplementary Table S2 presents the regression coefficients from the preliminary LASSO model for each blood biomarker measured in APS, in addition to those for BMI, (BMI)², height and (height)². Table 2 gives the anthropometric variables and biomarkers selected for inclusion in the final VAT score, as well as their regression coefficients and overall contributions to the total VAT variance. The mean VAT score was 5.19 ± 0.40 in men and 4.79 ± 0.42 in women. The VAT score included total, HDL and LDL cholesterol, triglycerides, insulin, leptin, adiponectin, sex-hormone binding globulin (SHBG), and total carotene, in addition to BMI and height, and explained 64% of the variance in VAT among men and 67% among women. In ethnic-specific analyses, these figures varied from 46% in African American men to 74% in Japanese American women (Supplementary Table S3). For comparison, anthropometry alone [i.e., BMI, (BMI)², height, and (height)²] explained only 58% and 56% of the variance in VAT in men and women, respectively; the corresponding R²s for BMI were 0.53 and 0.52. The AUROC for visceral obesity (VAT ≥150 cm²) was 0.90 in men and 0.86 in women (Table 2), and varied from 0.84 in African American women to 0.93 in Japanese American men (Supplementary Table S3). The AUROC for the VAT score was higher (0.94 for men, 0.91 for women) when a cutoff point of 100 cm² was used, instead of 150 cm², for visceral obesity. The VAT score was found to predict future VAT reasonably well, evidenced by the R²s among the 500 APS participants based on MEC samples collected approximately 10 years earlier of 0.64 for men and 0.60 for women, compared with R²s of 0.71 and 0.69 for concurrent VAT prediction using their APS samples. The correlation between the log VAT scores using the APS and MEC biomarker values for the 500 participants was 0.94.

Table 3 shows the baseline characteristics of the MEC breast cancer cases and controls. Cases were heavier, had a greater VAT score, had fewer children, and were more likely to have ever smoked cigarettes than controls. Risk was increased by 63% and 46% for women in the overweight and obese category, respectively (Supplementary Table S4). The P_trend for the association of waist circumference with breast cancer was 0.03 in the 714 cases and 703 controls with this variable but was no longer significant (P = 0.18) after further adjustment for BMI. No association was found for WHR (P = 0.45).

The crude VAT score (unadjusted for BMI) was positively associated with breast cancer risk, with the following ORs across increasing tertiles: 1.00, 1.50 (95% CI: 1.19–1.88), 1.43 (1.13–1.82), after minimal adjustment for matching factors (P_trend < 0.0001). Table 4 presents the breast cancer OR by tertile of the BMI-adjusted VAT score. Risk of breast cancer was significantly increased for women in the upper versus lower tertile of the VAT score [multivariate-adjusted OR = 1.48 (1.16–1.89)], with a significant trend (P = 0.002). This association was not affected by including BMI in the model, while BMI remained independently associated with an increased breast cancer risk [the ORs for BMI in the multivariate adjusted model were 1.00, 1.65 (1.30–2.88) and 1.46 (1.10–1.92); P_trend < 0.01]. Risk estimates for breast cancer remained virtually unchanged after adjustment for waist circumference (Supplementary Table S5). The VAT score–breast cancer association was observed similarly in women with and without history of ever having used MHT (Supplementary Table S6; P_{heterogeneity} = 0.35), for ER⁺ and ER⁻ breast cancer (Supplementary Table S7; P_{heterogeneity} = 0.75), and by tertile of years between blood draw and diagnosis (P_{heterogeneity} = 0.88). In an analysis stratified by BMI category, the association appeared stronger for BMI < 25 kg/m² (Supplementary Table S8; P_{heterogeneity} = 0.51). Finally, although the association with the VAT score was suggested in three of the five racial/ethnic groups, it appeared stronger for Japanese Americans (P = 0.001, P_{heterogeneity} = 0.08; Supplementary Table S9).

Table 3 presents the baseline characteristics of the MEC colorectal cancer cases and controls. Cases were heavier, consumed more alcohol and less dietary fiber and calcium, were less likely to have ever used MHT among women, and had a greater VAT score than controls. Colorectal cancer risk was increased by 30% and 60% for individuals in the overweight and obese category, respectively (Supplementary Table S9). The P_trend for the association of waist circumference was 0.03 in the 559 cases and 608 controls with this variable but was no longer significant (P = 0.73) after further adjustment for BMI. No association was found with WHR (P = 0.72).

The crude VAT score (unadjusted for BMI) was positively associated with colorectal cancer risk, with the following ORs across increasing tertiles: 1.00, 1.20 (0.94–1.54), 1.31 (1.02–1.70), after minimal adjustment for matching factors (P_trend = 0.007). However, there was no association between the BMI-adjusted VAT score and colorectal cancer (P = 0.82; Table 4) before or after further adjustment for BMI overall, or by sex (Supplementary Table S10) or with further adjustment for waist circumference (Supplementary Table S5). However, a modifying effect was suggested for number of years between blood collection and the case diagnosis by tertiles (<4; 4–6; and ≥7 years; Table 5). The ORs adjusted for established risk factors rose across tertiles of VAT score [1.0, 0.98 (95% CI: 0.68–1.39), 1.24 (95% CI: 0.88–1.76); P_trend = 0.08] only when cases were diagnosed seven or more years after blood draw (P_{heterogeneity} = 0.06). This association was not affected by including BMI in the model and was observed similarly for colon and rectal cancer (P_{heterogeneity} = 0.94), but appeared strongest for Japanese Americans (Supplementary Table S11).

Sensitivity analyses performed using BMI at blood draw instead of BMI at cohort baseline showed very similar BMI-adjusted VAT score ORs for breast (810 cases/801 controls) and colorectal cancer (615 cases/645 controls; 1.00, 1.04 (95% CI: 0.81–1.33), 1.54 (95% CI: 1.19–2.98), P_trend = 0.001; 1.00, 0.88 (95% CI: 0.66–1.16), 0.94 (95% CI: 0.70–1.26), P_trend = 0.055, respectively).

In the large diverse sample of older adults who underwent an abdominal MRI and other measurements in the APS, we utilized circulating biomarkers to develop a score predicting VAT amount for estimating disease risk associated with VAT. Candidate biomarkers were known to be related to obesity-induced metabolic, hormonal, and inflammation dysfunctions. The final score, which included nine biomarkers, BMI, and height, explained 11% and 15% more of the variance in VAT than BMI alone in men and women, respectively. The VAT score was found in MEC to be significantly associated with risk of postmenopausal breast cancer, independently of BMI, waist circumference, and other risk factors, but not with risk of colorectal cancer. However, a colorectal cancer association was suggested in a latency analysis when tumors that occurred within 7 years of blood draw were excluded. There was no evidence that the VAT score–breast cancer association differed by MHT use or ER status for breast cancer. The score performed better, both in regard to predicting VAT and predicting cancer risk, in Japanese Americans, the population with the greatest VAT amount and worse in African Americans, the population with the smallest VAT amount (15).

Determinants of visceral adiposity include age, sex, race, physical activity, alcohol intake, and diet (25–27). VAT, more so than other fat depots, is thought to result in a metabolic, hormonal, and inflammatory milieu that is tumor promoting (4–6). For example, VAT is inversely associated with SHBG (28) and is more strongly associated than subcutaneous fat with insulin resistance and proinflammatory cytokines (6)—factors that may promote both breast and colorectal cancers.

Neamat-Allah and colleagues (19) developed a VAT score using blood biomarkers among Germans aged 48–80 years. They found that a model for MRI-measured VAT that included adiponectin, C-reactive protein (CRP), leukocytes, aspartate aminotransferase, gamma-glutamyl transferase, uric acid, and LDL explained 4% more of the VAT variance than anthropometry in men. In women, the final model included adiponectin, hemoglobin A1c, and triglycerides, and explained 5% more of the variance in VAT than anthropometry. The authors concluded against the use of these biomarkers as proxies for VAT in epidemiologic studies and recommended the use of imaging. In contrast, our VAT score, with a common set of circulating biomarkers for men and women and that overlapped with the Neamat-Allah score, performed better in explaining additional variance in VAT and was associated with breast and, possibly, colorectal cancer.

Prediction scores developed by Lee and colleagues (29) for DXA-determined total fat mass, based on anthropometric and demographic factors using NHANES 1999–2006 data, were found to be associated with diabetes in two cohort studies (30). Two of the biomarkers most highly correlated with predicted fat mass in NHANES, namely triglycerides and insulin, were also important predictors of VAT in our study. Conversely, CRP was highly correlated with predicted total fat mass in NHANES but was not included in our VAT score, implying that systemic inflammation may not be specific to VAT but indicative of total adiposity.

To our knowledge, no past study has directly investigated the relationship between VAT and colorectal and breast cancer risks; this is the first study to do so using a proxy variable. Our results are supported by previous reports of an association between VAT and cancer at all sites-combined (31, 32). Our data are also agree with studies that have consistently associated waist circumference or WHR, as estimates of visceral obesity, with breast cancer (33–35). As few of these studies have adjusted for BMI, whether these factors have an independent effect is unclear. Waist circumference did not have an independent effect in our data. However, central obesity, measured by a waist circumference >88 cm, was recently associated prospectively in the Women's Health Initiative with increased overall cancer mortality in postmenopausal women, even among women with normal BMI (36).

It is possible that our overall results for colorectal cancer were null because of the study's relatively short follow-up. The observed suggested association among participants with ≥7 years of follow-up would be consistent with studies showing an association between visceral fat, measured by CT, and increased risk of adenoma, a precursor for most colorectal cancer, even after adjustment for BMI, waist circumference and subcutaneous fat (37). If replicated, these results would suggest that VAT exerts its promotional effects only on early stages of colorectal cancer development. This is plausible given the long latency period demonstrated for other colorectal cancer risk factors, such as smoking and aspirin (38, 39).

The relationships that we observed between the VAT score and breast and colorectal cancers appeared to show a magnitude gradient from strongest in Japanese Americans, a population with a relatively low mean BMI, to nonexistent in African Americans. It is interesting to note that we documented a parallel variation in VAT levels across these ethnic groups in APS, with an almost 2-fold greater total fat mass-adjusted VAT in Japanese Americans compared with African Americans (15). It may be that the association of VAT with disease risk is only or more easily detectable in groups or individuals with a substantial amount of visceral fat at relatively low BMI levels than in individuals in the obese and very obese range.

Major strengths of the APS are its large size and multiethnic composition and the inclusion of a variety of metabolic, hormonal, and inflammation biomarkers. We cannot exclude that other candidate biomarkers, not assessed here, may be more predictive of VAT. Another limitation is the restricted age range represented (60–72 years); our results may not directly apply to younger individuals. Nevertheless, the prospective analysis on 500 APS participants using samples collected in MEC indicated that our VAT score was robust with a prediction ability extending to an earlier period of ten years or longer. Strengths of our breast and colorectal cancer studies include the prospective design, the large multiethnic samples independent of the APS study participants, the complete outcome data obtained through registry linkage, and the availability of comprehensive questionnaire data. The baseline risk factors, especially BMI, used in our nested case–control studies may have changed during follow-up. However, we were able to use BMI at time of blood draw in 86% of participants for breast and 77% for colorectal cancer, and found very similar associations with the VAT score. Because of the diversity of our population and the fact that biomarkers were mechanistically related to cancer, we believe that the findings obtained with our two-step approach should be highly generalizable to other settings.

In conclusion, this is the first study to develop a risk score for VAT that is robust and associated with postmenopausal breast cancer, independently of BMI and other risk factors. These findings provide more specific evidence than past studies for the importance of VAT in cancer risk. The VAT score will be useful for further exploring the association of visceral fat with the risk of cancers and other chronic diseases.

J.A. Shepherd has research grants from GE and Hologic unrelated to this project. No potential conflicts of interest were disclosed by the other authors.

Conception and design: L. Le Marchand, L.R. Wilkens, B.S. Kristal

Development of methodology: L. Le Marchand, L.R. Wilkens, K.R. Monroe, B.S. Kristal, T. Ernst

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): L. Le Marchand, K.R. Monroe, J.A. Shepherd, A. Franke, T. Ernst, U. Lim

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): L. Le Marchand, L.R. Wilkens, A.M. Castelfranco, G. Maskarinec, T. Ernst

Writing, review, and/or revision of the manuscript: L. Le Marchand, L.R. Wilkens, B.S. Kristal, I. Cheng, G. Maskarinec, M.A. Hullar, J.W. Lampe, J.A. Shepherd, A. Franke, T. Ernst, U. Lim

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): L. Le Marchand, L.R. Wilkens, K.R. Monroe, U. Lim

Study supervision: L. Le Marchand, L.R. Wilkens, K.R. Monroe, U. Lim

This work was supported by NCI grants P01 CA168530 (to L. Le Marchand, L.R. Wilkens, A.M. Castelfranco, K.R. Monroe, B.S. Kristal, I. Cheng, G. Maskarinec, M.A. Hullar, J.W. Lampe, J.A. Shepherd, A. Franke, T. Ernst, and U. Lim), U01 CA164973 (to L. Le Marchand and L.R. Wilkens), and P30 CA71789 (to L.R. Wilkens and A. Franke).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.