Background
Pancreatic cancer (PC) represents the fourth largest cause of death by cancer in Europe. Without progress in the prevention and management of this disease, it will become the third cause of cancer-related deaths by 2030. The main cause of its lethality is its usual late diagnosis, which could be addressed by discovering new biomarkers to define high-risk populations and screening. The worldwide non-homogeneous PC incidence suggests that multiple interrelated (non)modifiable factors may be involved. This complex interplay can be reflected as changes in DNA methylation profiles (DNAm). However, DNAm in peripheral blood has been only briefly explored as a PC-associated marker at the individual CpG level,  not for capturing methylome variability. To address this limitation, high-dimensional models integrating whole methylation profiles are needed.
Objective
To identify the smoking-based DNAm signatures associated with the PC risk.
Methods
We included 338 PC cases and 285 controls from the Spanish subset of the PanGenEU study, with epidemiological and DNAm data for >850K CpGs, obtained from leukocyte DNA using the Illumina Infinium HumanMethylationEPIC v1.0 array. PC cases were selected from those with early-stage tumors to prevent reverse-causation bias. Controls were matched to cases by age and sex. After preprocessing DNAm data using Maksimovic’s pipeline, we applied Bayesian kernel-based regressions using the BGLR package. Models were adjusted for age, gender, recruitment region, and immune cell composition. They included the DNAm-based similarity matrix among pairs of individuals, computed as the (co)variance matrix of the CpG beta values, and different combinations of established PC risk factors, including diabetes, smoking, nasal allergies and asthma status, and family history of PDAC. DNAm-based risk scores were then associated with PC risk factors to investigate whether these profiles capture their effects. We then estimated the proportion of the PC risk variability explained by the DNAm component from each model. Effect estimates of individual CpGs were obtained after applying X'M^-1u, where M and X are the pairwise individuals’ similarity and CpGs’ incidence matrices, and u contains the methylation risk scores.
Results
We estimated that 54-59% of the PC risk variance is explained by DNAm. The largest estimate was obtained with the baseline model, including adjustment variables and the kernel. Adding PC risk factors to the model decreased the estimated percentage of phenotypic variance explained by the epigenomic component, suggesting that the epigenomic component captures variability due to those risk factors. To prove this, we ranked the CpGs based on the absolute differences between the CpGs effect estimates obtained with the full model (including all PC risk factors) and those without smoking and found that the signature was significantly enriched with CpGs/genes previously found to be associated with smoking behaviour in EWAS/GWAS studies.
Conclusions/Implications
Our study highlights that the epigenomic component not only captures biological processes but also encodes signatures of known risk factors in PC contributing to its exposome definition.