Prospective Association of Energy Balance Scores Based on Metabolic Biomarkers with Colorectal Cancer Risk

An estimated 55% of colorectal cancers diagnosed in the United States in 2014 were attributed to modifiable lifestyle factors, underscoring the importance of lifestyle in the development of colorectal cancer (1). Most of the evidence linking lifestyle and colorectal cancer comes from investigations of individual lifestyle factors without consideration of their downstream metabolic effects. Moreover, lifestyle risk factors are highly correlated and have synergistic effects on health (2).

Energy balance–related lifestyle risk factors associated with colorectal cancer include excess body fat, poor diet, physical inactivity, and sedentary behavior, which are particularly susceptible to clustering (3) and share common pathways in colorectal carcinogenesis, including chronic systemic inflammation and insulin resistance (4). While risk estimates are available for the association between individual energy balance–related factors and colorectal cancer (5), their combined influence is unclear.

Associations of multiple highly correlated lifestyle factors with disease outcomes are examined using indices or scores. These scores are rarely based on mechanisms or clinical biomarkers of disease risk. Reduced rank regression (RRR) provides a robust method to derive weighted lifestyle scores with an a priori definition of hypothesized mechanisms via use of biomarkers reflective of certain pathways or processes. RRR is a method for reducing large amounts of correlated explanatory variables to a smaller set of latent variables that maximize the explained variation in a single or set of responses variables, often an intermediate marker for disease risk (6). RRR has been used to identify associations between dietary patterns with disease risk (7, 8), but to date, no published study has developed a score of lifestyle factors using RRR.

In this analysis, we used RRR to derive and validate three energy balance scores in a subset of Cancer Prevention Study II (CPS-II) Nutrition Cohort participants with prediagnostic measures of high sensitivity C-reactive protein (hsCRP), C-peptide, and hemoglobin A_1c (HbA_1c), which are, respectively, established clinical markers for general inflammation, hyperinsulinemia, and hyperglycemia; the latter two together represent glucose homeostasis. We examined associations of the derived energy balance scores with incident colorectal cancer risk and of tumor molecular phenotypes (where available), among all eligible men and women in the CPS-II Nutrition Cohort.

Participants were from the CPS-II Nutrition Cohort, a prospective study of cancer incidence and mortality (9). Briefly, at enrollment (1992/1993), 184,185 men and women completed a 10-page self-administered questionnaire on medical history and lifestyle factors. Follow-up questionnaires were sent biennially starting in 1997 to update exposure information and to ascertain new cancer diagnoses. From 1998 to 2001, a subset of 39,371 participants provided nonfasting blood samples. Blood samples were shipped chilled overnight to a central repository for long-term storage. The CPS-II Nutrition Cohort is approved by the Institutional Review Board of Emory University (Atlanta, GA).

In this analysis, all participants who returned the 1999 survey were eligible (n = 151,342). Exclusions included: prevalent cancer except for nonmelanoma skin cancer (n = 28,472), loss to follow up (n = 4,497), unverified diagnosis (n = 174), invalid end-of-study time (n = 7), invalid dietary data (n = 2,184), and missing lifestyle data from all applicable survey cycles (1992, 1997, 1999; n = 1,019). The final analytic sample comprised 114,989 men and women.

Diet, physical activity, sedentary behaviors, and body mass index (BMI) were used to develop the scores because they represent a comprehensive characterization of modifiable exposures pertaining to energy balance. Intakes of 33 food groups were assessed in 1999 using a modified Willett food frequency questionnaire (FFQ, refs. 10, 11) or in 1992 using a modified Block FFQ (10, 11) if the 1999 FFQ was incomplete. Other self-reported information in 1999 (or in 1997 and 1992 if 1999 was missing) was used to characterize BMI and moderate-to-vigorous intensity physical activity (MVPA) MET-hours/week. Screen time hours/week, a valid proxy for sedentary time (12), was assessed in 1999 (or in 1992 if missing in 1999; not assessed in 1997). To mitigate the potential for misclassification bias between information on exposures carried forward (due to missingness) from past surveys and baseline in 1999, we weighted individuals in Cox models based on the proportion of the lifestyle factors that were measured at the 1999 survey. Complete information on dietary was factors considered to be one of four energy balance–related factors for the purpose of calculating the weight. Thus, individuals with complete data in 1999 were given a full weight of 1 and those with fewer available data in 1999 received a lower weight (e.g., 3/4 lifestyle factors available received a weight of 0.75), thereby having a smaller influence on HR estimates. Complete data in 1999 was available for 81.7% of participants. Missingness of the exposures was not associated with incident colorectal cancer.

Using nonfasting blood samples, circulating concentrations of hsCRP, C-peptide, and HbA_1c were measured from serum in a case–cohort study consisting of a random subcohort of 3,000 participants and 2,962 diabetes-related cancers (including colorectal cancers). All three biomarkers have been shown to be reliable and clinically useful when measured from a nonfasting state (13–15). The biomarkers represent states of metabolic dysfunction that develop over a period of multiple years.

Lab personnel were blinded to case–control status and all plates included blinded quality control samples. Human CRP Immunoassay (R&D Systems, Inc.), a quantitative sandwich enzyme immunoassay technique, was used to measure hsCRP. The coefficient of variation (CV) for hsCRP was 7.4%, with an intraclass correlation coefficient (ICC) of 99.8% for CPS-II samples. The C-peptide ELISA (Ansh Labs), an enzymatically amplified one-step sandwich-type immunoassay, was used to measure C-peptide. The CV for C-peptide was 7.7%, with an ICC of 97.5% for CPS-II samples. The HbA_1c assay is an enzymatic measurement in which lysed whole blood samples are subjected to extensive protease digestion. The CV for HbA_1c was 8.9%, with an ICC of 74.7% for CPS-II samples.

The 3,000 subcohort participants from the CPS-II nested case–cohort were used to derive energy balance scores. Seventeen participants were excluded for lack of biomarker data, and an additional 299 participants were excluded because of incomplete lifestyle information in 1999. We additionally excluded 186 participants who self-reported a diabetes diagnosis, as their reported lifestyle information may reflect postdiagnosis lifestyle changes. Ten participants were excluded from models with hsCRP because levels were indicative of acute inflammation (>40,000 μg/L). Of the remaining 2,498 participants, 80% (n = 1,999) were used to derive energy balance scores and the remaining 20% (n = 499) were used for validation. Values for all three biomarkers were log-transformed. RRR was used to create a factor score, which is a single continuous value that represents a linear combination of the energy balance–related factors, as determined by eigenvectors of the biomarker covariance matrix. In all models, only the first factor score was retained as it represents the weighted combination of energy balance factors that explains the greatest amount of variation in the biomarkers. First, RRR was performed in a single model with men and women using all three biomarkers as dependent variables to identify which individual factors were most strongly associated with the biomarkers for the overall sample. The individual lifestyle factors that explained at least 1% of the variation in the biomarkers in the overall model were retained for future models to derive the scores. This approach aims to limit the set of explanatory variables to those that are most important in predicting biomarker values, which facilitates the interpretation and application of the scores. All remaining RRR models were then performed in men (n = 835) and women (n = 1,164), separately. For both men and women, three different RRR models were performed corresponding to three different combinations of biomarkers used in the model: (i) poor glucose homeostasis and inflammation (all three biomarkers); (ii) poor glucose homeostasis (C-peptide and HbA_1c); and (iii) inflammation separately (hsCRP alone). The RRR model produces a factor score, which represents the overall correlation of all input variables to the biomarker(s) as a single continuous number.

Validity of the scores was examined among the remaining 499 men and women with biomarker data who were not used to derive the scores. Age- and sex-adjusted multivariate linear regression models were used to examine if the energy balance scores were associated with the biomarker or combination of biomarkers used in their derivation.

Self-reported cancer diagnoses from follow-up surveys were verified through medical records, registry linkage, or death certificates. A total of 2,228 incident cases of colorectal cancer (1,778 colon, and 450 rectum) were verified between 1999 and June 30, 2015. Among the cases, 1,108 were men and 1,120 were women. We collected formalin-fixed, paraffin-embedded tumor tissue specimens to assess molecular characteristics of the tumors in a sample of 627 cases (16). PCR-based assessment was used to categorize microsatellite instability (MSI) status among 409 tumors, and was based on the Bethesda Consensus Panel (17). Tumors were classified as MSI-high if ≥30% of the markers showed instability, and microsatellite stable (MSS)/MSI-low if <30% of the markers showed instability. Classification was based on ≥5 interpretable markers (unless all four markers were unstable, in which case the tumor was classified as MSI-high). CpG island methylator phenotype (CIMP) status was determined for 498 tumors using the eight-panel MethyLight array (18); tumors were classified as CIMP-high if ≥6 of the 8 genes had a percent of methylated reference (PMR) value ≥10, and non-CIMP if tumors had <6 genes with a PMR ≥10. BRAF mutation status c.1799T>A (p.V600E) was determined for 426 tumors using a fluorescent allele–specific PCR assay (19). Sanger sequencing was used to classify mutations in KRAS codons 12 and 13 for 352 tumors, and was considered mutated if any mutation was found in either codon (20).

Energy balance scores for each CPS-II Nutrition Cohort study participant for whom we do not have biomarker data were computed using model-based parameters (i.e., RRR weights) in a way that mimics how the scores were calculated in the biomarker subsample. First, all lifestyle factors were centered and scaled, as is done at the onset of the RRR procedure. Then, each lifestyle factor was multiplied by the RRR model weights (similar to a regression coefficient) produced in the RRR models when deriving the scores. The weighted scores for all components were summed to calculate a final energy balance score for each participant.

HR and 95% confidence intervals (CI) for the associations of energy balance scores with colorectal cancer risk were computed using Cox proportional hazards regression among 114, 989 CPS-II Nutrition Cohort participants. Time-on-study was used for the time scale and was calculated from the date of completion of the 1999 survey until date of colorectal cancer diagnosis, date of death, date of last returned survey, or June 30, 2015, whichever came first. Each score was categorized into sex-specific quartiles; the reference group was the first quartile score representing the lifestyles correlated with the lowest levels of the energy balance biomarker(s). Trend tests were assessed by assigning the median value of the quartile as a continuous exposure variable. We also computed the HRs for a 1-SD increase in the scores. All models were adjusted for age, sex, race/ethnicity, smoking status, alcohol intake, nonsteroidal anti-inflammatory drug (NSAID) use, multivitamin use, and menopausal hormone therapy (women only). We examined associations stratified by age (<70, ≥70 years), sex, and anatomic subsite (colon, rectum). In sensitivity analyses, we included adjustment for diabetes that was excluded from the main models as it may lie on the causal pathway between lifestyle and colorectal cancer. No violations of the proportional hazards assumption were observed using the Likelihood Ratio Test.

Cox models using the duplication method were used to examine potential heterogeneity for the association between the combined energy balance score on molecular subtypes of colorectal cancer (21). Molecular phenotype data of tumors is dependent upon the availability of tumor tissue and selection bias may be introduced into the HR estimates if the availability of tumor tissue is dependent upon the colorectal cancer subtype (i.e., nonrandom missingness). Inverse probability weighting (IPW) was used to remediate this potential bias (22). Logistic regression was used among verified CPS-II colorectal cancer cases in this analysis to calculate the probability of selection into this study given their age at diagnosis, year of diagnosis, anatomic location of the tumor, sex, and stage. The inverse of that calculated probability served as the selection bias weight so colorectal cancer cases with phenotype data represents all colorectal cancer cases ascertained in the full analytic cohort of similar diagnoses.

Mean (SD) levels of hsCRP, C-peptide, and HbA_1c in the 2,498 participants with biomarker data were 3,838.3 μg/L (5,013.7), 5.5 nmol/L (3.0), and 5.3% (0.7), respectively. The following factors explained >1% of the variation in the biomarkers and were used to derive the energy balance scores: BMI, MVPA, screen time, and servings/day of fruit juice, dark green vegetables, cruciferous vegetables, red/orange vegetables, whole grains, red meat, cured meat, organ meat, other fish (i.e., non-fried fish), eggs, high-fat dairy, oily fats, solid fats, and sugar-sweetened beverages.

The results from RRR and scoring weights used to calculate each of the sex-specific energy balance scores with different combinations of biomarkers are shown in Table 1. The highest weights for each energy balance score was with BMI (range from 0.20–0.38). The amount of variation in the biomarkers explained by the energy balance factors varied across the derived scores (Table 1). The largest amount of variation in the biomarkers explained by the derived energy balance scores was for the hsCRP score in women (10.5%). The smallest amount of variation in the biomarkers explained by the scores was for the HbA_1c + C-peptide score in men (5.1%). All of the energy balance scores were strongly correlated (r > 0.83; Supplementary Table S1). Multivariate linear regression results indicated that all energy balance scores were associated with the biomarker(s) used in their derivation among the validation subset of participants (Supplementary Table S2).

Descriptive characteristics across quartiles of the combined energy balance score are shown in Table 2. Participants in the fourth quartile compared with the first quartile of the score were less likely to be college educated, never smokers or to report multivitamin use, and more likely to report NSAID use, current menopausal hormone use, and a personal history of diabetes. The mean (SD) and median time from baseline to colorectal cancer diagnosis was 6.5 (4.1) and 6.0 years, respectively.

The associations of the three energy balance scores with risk of incident colorectal cancer are shown in Table 3. The higher risk observed in the fourth quartiles, compared with the first quartile, ranged from 30% for the HbA_1c + C-peptide lifestyle score to 35% for the hsCRP-alone score and the combined score based on all three biomarkers. All HR estimates from continuous models indicated a 10% higher risk of incident colorectal cancer per 1-SD increase.

Results from IPW-weighted models for associations between the combined energy balance score and molecular subtypes are shown in Table 4. Statistically significant heterogeneity was observed for the association when stratified by CIMP status of the tumor, where the fourth quartile was associated with a 58% higher risk compared with the first quartile for non-CIMP tumors (HR = 1.59; 95% CI, 1.17–2.16) but not CIMP-positive tumors (HR = 0.72; 95% CI, 0.39–1.30). Other statistically significant estimates were observed for the fourth quartiles when examining MSS/MSI-L (HR = 1.55; 95% CI, 1.10–2.19), BRAF-wild type (WT; HR = 1.70; 95% CI, 1.21–2.38), and for KRAS-mutant tumors (HR = 2.00; 95% CI, 1.14–3.50).

Supplementary Tables S3–S6 show results for associations among strata of sex, age, and anatomic subsite, respectively. Stronger associations were observed in participants <70 years old, but otherwise little evidence of heterogeneity was observed. No substantive differences were observed after adjusting for self-reported diabetes.

We empirically derived three energy balance scores based on circulating levels of hsCRP, C-peptide, and HbA_1c. All three scores were associated with higher risk of incident colorectal cancer in a large study population of predominantly non-Hispanic White men and women. These results indicate that men and women whose lifestyles reflect high potential for systemic inflammation and poor glucose homeostasis are at a higher subsequent risk of developing colorectal cancer. The relative role of excess body fat in poor metabolic health and subsequent colorectal cancer risk was evident by consistently high scoring weights for BMI. This study further supports long held hypotheses that systemic biomarkers of metabolic health mediate part of the relationship between several modifiable behaviors and colorectal cancer risk (23, 24).

These biomarkers may reflect synergistic interactions in metabolic pathways that link unhealthy energy balance–related lifestyles to colorectal cancer risk. Prediagnostic levels of hsCRP levels, which has been used to evaluate a chronic inflammatory state, were positively associated with colorectal cancer risk in a meta-analysis of 18 studies (25). Proinflammatory conditions may promote tumor malignant progression, invasion, and metastasis (26). C-peptide is a marker of insulin production from the β-cells in the pancreas, uninfluenced by fasting status and with a longer half-life than insulin, and has been positively associated with colorectal cancer in multiple meta-analyses of prospective studies (27, 28). HbA_1c, a stable indicator of circulating glucose over the previous 2–3 months, also has been positively associated with colorectal cancer risk (28). Hyperglycemia may influence colorectal cancer etiology through multiple biologic mechanisms, such as through angiogenesis (29) or through mitogenic effects of insulin-like growth factor, among others (30).

Previous studies examining the relationship between lifestyle scores and colorectal cancer risk did not focus specifically on energy balance–related risk factors, nor were the scores derived based on empirical biomarker data; nonetheless, all reported statistically significant associations in the hypothesized direction (31–42). In the only other study that derived a score based on a biomarker (42), Tabung and colleagues developed a lifestyle score comprising BMI, physical activity, and 12 food groups based on circulating C-peptide concentrations (43). Similar to the combined energy balance score derived herein, positive weights were observed for BMI, solid fats, and fruit juice; a negative weight was observed for physical activity. In that study, the highest quintile of the score was associated with a 49% higher risk of colorectal cancer (CI, 1.10–2.01), with no heterogeneity observed by sex (42). Differences observed in the current scores, such as weighting and direction of some food groups, may be explained by our use of multiple biomarkers of downstream effects of energy balance, not solely a marker of hyperinsulinemia. For example, the association of high fat dairy was stronger for hsCRP than with C-peptide and HbA_1c, which would tend to weaken the associations with all three biomarkers combined into one score. Nevertheless, BMI, physical activity, and sugar-sweetened beverages were most strongly associated with the combined score compared with the other two scores, supporting energy balance as an important predictor of metabolic health. In contrast to the Tabung and colleagues score based on hyperinsulinemic potential, the present combined score was additionally based on inflammation and hyperglycemia, which may provide a more comprehensive characterization of poor energy balance and colorectal cancer etiology. As other biologic pathways may connect energy balance–related factors to colorectal cancer risk (44, 45), the scores in this analysis may explain only a portion of the total effect.

It is possible the consistently stronger association observed among individuals <70 years old at baseline is explained by the slightly attenuated correlation between BMI and adiposity in older individuals (46), which subsequently limits our ability to estimate adipose-related inflammatory, hyperinsulinemic, hyperglycemic status in the empirical scores. In addition, the relative contribution to these biomarkers from adipose tissue may be less in older individuals compared with age-related declines of metabolic function that are independent of adiposity (47, 48).

This is the first study to examine an aggregate measure of energy balance–related factors in relation to molecular phenotypes of colorectal cancer, although our limited power made it difficult to examine associations for rarer subtypes of colorectal cancer and our ability to test heterogeneity was limited to large differences. Even with limited power, we observed a differential association of the combined energy balance score on CIMP status, with the association limited to non-CIMP tumors. There is evidence that genes associated with epigenetic silencing via methylation, such as SIRT1, have decreased expression in obesity resulting in lower levels of methylation (49). Non-CIMP tumors are also usually MSS (50), which have shown more consistent associations with excess adiposity (51, 52). Although we did not observe a statistically significant association, we observed similar patterns of association for MSI status of tumors as we saw for CIMP status. Characterization of lifestyles that may promote the progression of certain mechanistic pathways indicative of molecular subtypes may be useful when monitoring patient risk profiles for personalized prevention, although more research in this area is needed.

Some limitations to our study should be noted. We did not have biomarker data on all participants, thus we can only hypothesize that the derived scores represent associations between energy balance–related factors and biomarkers in the entire cohort, as supported by the validation of the scores in a subset of participants with biomarker data. Furthermore, we did not have sufficient power to perform a traditional mediation analysis. Self-reported exposure data may introduce misclassification bias, and there are known limitations in using BMI to assess adiposity (53). The study population comprised predominately older, non-Hispanic white participants, thus our results may not be generalizable to other age or racial/ethnic groups. Limited data on molecular subtypes did not allow for adequate power to examine associations in rarer molecular subtypes; future studies with larger numbers should consider this research question. The use of IPW accounts for nonrandom missing subtype data to help mitigate the role of selection bias in our models of molecular tumor phenotypes. Furthermore, energy balance scores and general lifestyle exposures did not previously differ across colorectal cancer cases with and without available tumor tissue (16). The term “energy balance score” was used given the stronger weighting of BMI, physical activity, and the combination of fruit juice and sugar-sweetened beverages relative to other components in the scores; however, we recognize that non-energy balance–related pathways may also be involved. In addition, the biomarkers in this analysis are likely influenced by other external and inherited factors not included in our analysis, such as smoking with respect to inflammation. There are many strengths in our approach. We used data from a large, prospective study with detailed assessment on lifestyle factors and covariates. Use of RRR allowed for a priori identification of mechanistic pathways along with empirically based scoring. Furthermore, data on established clinical markers that provide a comprehensive characterization of energy balance–related metabolic function were used to derive the scores.

In conclusion, this analysis suggests that the clustering of energy balance-related lifestyle factors indicative of high levels of inflammation and poor glucose homeostasis are associated with higher risk of colorectal cancer. Focusing on energy balance-related factors that lower inflammation and ameliorate abnormal insulin/glucose levels may be effective methods for reducing risk of colorectal cancer, particularly for some molecular subtypes of colorectal cancer, and should be incorporated into public health recommendations. Future analyses should include other risk biomarkers and a more complete examination of associations of energy balance-related lifestyle factors with molecular subtypes of colorectal cancer in more populations with greater racial and ethnic diversity.

W.D. Flanders is an owner of Epidemiology Research and Methods. No potential conflicts of interest were disclosed by the other authors.

Conception and design: M.A. Guinter, S.M. Gapstur, M.N. Pollak, P.T. Campbell

Development of methodology: M.A. Guinter, P.T. Campbell

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): S.M. Gapstur, P.T. Campbell

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): M.A. Guinter, S.M. Gapstur, W.D. Flanders, K.I. Alcaraz, P.T. Campbell

Writing, review, and/or revision of the manuscript: M.A. Guinter, S.M. Gapstur, M.L. McCullough, W.D. Flanders, Y. Wang, E. Rees-Punia, K.I. Alcaraz, M.N. Pollak, P.T. Campbell

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): P.T. Campbell

Study supervision: S.M. Gapstur, P.T. Campbell

Other (assay lab work): M.N. Pollak

This work was supported by the American Cancer Society's Cancer Prevention Studies Postdoctoral Fellowship Program (to M.A. Guinter and E. Rees-Punia) and by American Cancer Society funds for the creation, maintenance, and updating of the Cancer Prevention Study-II cohort. The authors sincerely appreciate all Cancer Prevention Study-II participants and each member of the study and biospecimen management group. The authors acknowledge the contributions to this study from central cancer registries supported through the Centers for Disease Control and Prevention's National Program of Cancer Registries and cancer registries supported by the NCI's SEER Program.

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.