Classical epidemiology has identified risk factors, including smoking, alcohol, obesity, and diabetes, among others. Yet these factors do not fully explain why people with similar exposures have different outcomes. Our challenge is to connect what people experience throughout life – their environmental and lifestyle exposures, or the exposome – with inherited susceptibility and the molecular changes that precede disease.

Risk has a history

At the Spanish National Cancer Research Centre, our Genetic and Molecular Epidemiology Group approaches PDAC as a multifactorial disease. PanGenEU, a multicentre European case-control study involving 28 centres across six countries, brings together epidemiological data and biological samples to investigate that complexity.

In our published PanGenEU analyses, current smokers had approximately 1.7 times the odds of pancreatic cancer compared with never smokers. Clearly the links between tobacco smoke and cancer in general are well known, but this does strengthen the case for this approach to examine the links between exposure and biological damage.

But it is by looking at diabetes that we see why timing is an essential component of accurately determining risk. Long-standing type 2 diabetes can be associated with increased pancreatic cancer risk, but diabetes developing shortly before diagnosis may also be a manifestation of a cancer of unknown primary. This situation can obscure understanding of distinct biological processes and therefore distinct opportunities to intervene.

Weight trajectories, glycemic control, and the interval between diabetes and cancer diagnoses can therefore be more informative than a single yes-or-no classification. And this idea – the collation of more nuanced information – applies across the exposome.

Measuring what changes

Capturing the pancreatic cancer exposome is exceptionally challenging. Diet, medications, pollutants, and microbial communities change over time, often in concert. Their effects may depend on age, exposure duration, or biological context.

Questionnaires can miss relevant exposure windows. Molecular measurements provide information but usually offer only snapshots. A metabolite measured at diagnosis may reflect an earlier exposure, altered metabolism, or the effects of the cancer itself. Distinguishing among these explanations is crucial for prevention.

Integrative molecular epidemiology provides a framework for investigating them. It integrates exposure histories and phenotypes with the genome, epigenome, metabolome, and microbiome. Each layer reveals aspects of susceptibility, biological responses, or potential disease markers.

Our published work on genetic susceptibility, including a multilayer post-genome wide association study analysis, shows how association signals can be examined in their functional context. A risk locus – a specific chromosomal region where genetic variation is linked to an increased susceptibility or probability of developing a particular disease – initiates an inquiry into the genes and regulatory processes through which it might act.

The microbiome offers another avenue. In collaboration with Peer Bork’s laboratory, we identified a fecal microbial signature associated with PDAC. This supports further investigation of stool-based markers, while leaving an essential question open: can a signature detected in patients with cancer identify risk before diagnosis?

Putting the pieces into place

Collecting several molecular layers does not automatically establish causation. A biological feature could mediate an exposure’s effect, modify that effect, confound an association, or emerge as a consequence of disease. Each role has implications for prevention.

DNA methylation, for example, may help link environmental exposures to biological responses. An association with PDAC does not establish that a methylation change causes cancer. Its interpretation depends on when it occurred, which cells were measured, and how it relates to other risk factors.

We need study designs that place these observations in time. Prospective cohorts with repeated exposure assessments and samples collected before diagnosis are crucial. Standardised collection procedures and compatible exposure definitions will help researchers compare findings across populations.

Artificial intelligence can support the analysis of complex, multimodal data. Its value depends on the questions we ask and on how we test its predictions. Independent validation and rigorous causal inference remain essential, especially when the intended application is to identify people who currently appear healthy.