Methylation scores for smoking, alcohol consumption, and body mass index and risk of seven types of cancer

Pierre‐Antoine Dugué, Chenglong Yu, Allison Hodge, Ee Ming Wong, Jihoon E. Joo, Chol‐Hee Jung, Daniel F. Schmidt, Enes Makalic, Daniel D. Buchanan, Gianluca Severi, Dallas R. English, John L. Hopper, Roger L. Milne, Graham G. Giles, Melissa C. SoutheymedRxiv 2021 · DOI

Introduction

In epidemiological research, understanding how lifestyle factors like smoking, alcohol consumption, and obesity contribute to cancer risk has traditionally relied on self-reported questionnaires. However, this method is inherently limited by "noise": participants may forget details, underestimate their consumption, or be unaware of passive exposures like second-hand smoke. Furthermore, questionnaires cannot capture the unique biological response an individual has to these exposures.

This study explores a more objective alternative: DNA methylation. By analyzing chemical tags on the DNA molecule that regulate gene expression, researchers can read a "biological ledger" of past and current exposures. This paper evaluates whether Methylation Scores (MS)—weighted averages of methylation levels at specific genomic sites—can more accurately predict the risk of seven different cancer types compared to traditional methods.

Methodology and Data

The researchers utilized data from the Melbourne Collaborative Cohort Study (MCCS), a prospective study involving over 41,000 participants. The analysis focused on 3,123 incident cases of seven cancers: colorectal, gastric, kidney, lung, prostate, urothelial, and B-cell lymphoma. These cases were matched with healthy controls based on age, sex, and country of birth.

To calculate the Methylation Scores for smoking, BMI, and alcohol consumption, the team employed two distinct weighting strategies:

  1. Predictive Weights: Derived from machine learning models (Lasso regression) designed to accurately predict current lifestyle factors.
  2. Association Weights: Derived from Epigenome-Wide Association Studies (EWAS) that identify sites most strongly linked to the exposure, regardless of their predictive accuracy for the behavior itself.

The associations between these scores and cancer risk were estimated using conditional logistic regression, adjusting for cell composition and self-reported data to determine if the biological markers provided independent predictive value.

Key Insights

The Independence of Biological Markers

A primary finding of the study was that methylation scores remained significantly associated with cancer risk even after adjusting for self-reported lifestyle data. This suggests that DNA methylation captures "hidden" risk factors—such as cumulative damage over decades or individual susceptibility—that surveys fail to record.

Cumulative Damage vs. Current Status

Interestingly, scores based on simple associations (EWAS) occasionally outperformed machine-learning-optimized predictive scores in identifying cancer risk. This implies that for oncological outcomes, the total biological impact on the methylome is a more relevant metric than how accurately blood markers can guess a person's current weight or current smoking status.

Tissue Specificity and Systemic Signals

While the study utilized peripheral blood samples, the smoking-related MS was most predictive for cancers of the respiratory and urinary tracts (lung and urothelial). This highlights how systemic epigenetic changes in the blood can serve as effective proxies for risk in distant, specific organs.

Results and Predictive Power

The study observed several significant associations between MS and cancer incidence:

  • Smoking & Lung Cancer: Smoking MS showed the strongest link, with Rate Ratios (RR) between 1.5 and 1.7 per standard deviation increase.
  • B-cell Neoplasms: BMI and alcohol MS were particularly relevant for mature B-cell neoplasms, showing moderate associations (RR ~1.3–1.4).
  • Discriminatory Accuracy: The inclusion of MS in traditional risk models (which include age, sex, and smoking history) improved the Area Under the Curve (AUC):
  • Lung Cancer: +7% improvement.
  • B-cell Neoplasms: +8% improvement.
  • Other Cancers: Approximately +3% improvement.

While smoking scores were highly reflective of actual smoking status (explaining up to 56% of variance), BMI and alcohol scores were more modest, explaining 11-15% of the variance in those behaviors.

Limitations and Future Directions

Despite the promise of these biological signatures, several challenges remain. There is no "one-size-fits-all" score; different versions of the MS performed better for different cancers, suggesting that a panel of multiple scores may be necessary for comprehensive screening. Additionally, because the study used peripheral blood, the methylation patterns observed may not perfectly mirror the changes occurring in the specific tissues where the tumors originate (e.g., the colon or lungs).

Finally, while the improvements in predictive power (AUC) are statistically significant, they are currently modest. The authors suggest that while these scores could eventually supplement or replace questionnaire data in research settings, their immediate clinical utility for population-wide screening requires further refinement. Future research will likely focus on integrating these scores into complex synthetic biomarkers and exploring genomic regions specific to cancer subtypes.

Summary updated

Abstract

ABSTRACT Methylation marks of exposure to health risk factors may be useful markers of cancer risk as they might better capture current and past exposures than questionnaires, and reflect different individual responses to exposure. We used data from seven case-control studies nested within the Melbourne Collaborative Cohort Study of blood DNA methylation and risk of colorectal, gastric, kidney, lung, prostate and urothelial cancer, and B-cell lymphoma (N cases=3,123). Methylation scores (MS) for smoking, body mass index (BMI), and alcohol consumption were calculated based on published data as weighted averages of methylation values. Rate ratios (RR) and 95% confidence intervals for association with cancer risk were estimated using conditional logistic regression and expressed per standard deviation increase of the MS, with and without adjustment for health-related confounders. The contribution of MS to discriminate cases from controls was evaluated using the area under the curve (AUC). After confounder adjustment, we observed: large associations (RR∼1.5-1.7) with lung cancer risk for smoking MS; moderate associations (RR∼1.2-1.3) with urothelial cancer risk for smoking MS and with mature B-cell neoplasm risk for BMI and alcohol MS; moderate to small associations (RR∼1.1-1.2) for BMI and alcohol MS with several cancer types and cancer overall. Generally small AUC increases were observed after inclusion of several MS in the same model (colorectal, gastric, kidney, urothelial cancers: +3%; lung cancer: +7%; B-cell neoplasms: +8%). Methylation scores for smoking, BMI, and alcohol consumption show independent associations with cancer risk, and may provide some improvements in risk prediction.
BibTeX
@misc{dugue2021methylation,
  author = {Pierre‐Antoine Dugué and Chenglong Yu and Allison Hodge and Ee Ming Wong and Jihoon E. Joo and Chol‐Hee Jung and Daniel F. Schmidt and Enes Makalic and Daniel D. Buchanan and Gianluca Severi and Dallas R. English and John L. Hopper and Roger L. Milne and Graham G. Giles and Melissa C. Southey},
  title = {{Methylation scores for smoking, alcohol consumption, and body mass index and risk of seven types of cancer}},
  year = {2021},
  doi = {10.1101/2021.02.08.21251370},
  howpublished = {medRxiv},
}
Figure