A GENOMICS-INFORMED ENSEMBLE LEARNING FRAMEWORK FOR EARLY RISK PREDICTION OF PANCREATIC DUCTAL ADENOCARCINOMA FROM STRUCTURED CLINICAL PHENOTYPES

Abstract

Pancreatic ductal adenocarcinoma (PDAC) is one of the most aggressive cancers that afflicts humans, with very low five year survival rate (<10%) and a nearly fatal prognosis upon diagnosis. By contrast, the molecular pathogenesis of PDAC is highly defined – as the disease progresses through pancreatic intraepithelial neoplasia (PanIN), KRAS is mutated to an activating form and CDKN2A, TP53 and SMAD4 are inactivated. These genomic drivers are seldom exploited for population-level early detection, because tissue genotyping is invasive and rarely available before symptoms appear. In this study we ask whether the systemic phenotypic consequences of that molecular cascade, recorded in routinely collected clinical and laboratory variables, can flag individuals at elevated risk. We first formalise a generative genotype–phenotype model that treats clinical features as noisy observations of a latent molecular state, and then develop a stacked ensemble that combines Random Forest and XGBoost base learners with a logistic-regression meta-learner. The framework is specified through a hierarchy of equations spanning feature representation, class-imbalance correction, the two base learners, stacked generalisation and evaluation, and is assessed across four independent datasets spanning balanced and severely imbalanced class distributions using AUC, F1-score, recall, specificity and the Brier score. The ensemble achieved perfect discrimination on small, clean datasets (AUC = 1.00) but degraded to chance-level ranking (AUC ≈ 0.50) on large, highly imbalanced cohorts, where it collapsed onto the majority class. These contrasting outcomes show that dataset composition, label leakage and class imbalance, rather than the choice of algorithm, govern apparent accuracy. We interpret this behaviour through the molecular biology of PDAC and set out, with an explicit fusion model, how germline variant panels, circulating tumour DNA and methylation signatures could convert a phenotype-based screen into a genuinely genomics-integrated risk tool. The framework provides a transparent and reproducible scaffold for translating the established genetics of pancreatic cancer into early-detection strategies.