Association Rules for predictive purposes applied to omics data

Abstract

Pancreatic ductal adenocarcinoma (PDAC) is the most lethal cancer globally, with limited improvements in prognosis over the past few decades. Immune infiltration plays a critical role in the progression of this disease, while genetic susceptibility is known to account for approximately 40% of the variability in immune system function. The primary objective of this study is to test the hypothesis that genetic susceptibility may modulate immune infiltration in pancreatic tumors. Immune profiles were categorized into three distinct clusters based on levels of tumor immune infiltration while genetic data, consisted of Single Nucleotide Polymorphisms (SNPs). All the information was derived from a previous study using pancreatic cancer patients from the TCGA database, with a total of 107 Caucasian patients and 117,486 SNPs. This study aims to utilize advanced statistical and Machine Learning techniques to classify individuals into the three clusters based on their genotypic data. Kaplan-Meier survival curves and the Cox proportional hazards model were employed to measure the impact of immune clusters on patient survival. For classification, penalized multinomial logistic regression models (Ridge, Lasso, and Elastic Net), alongside Random Forest and Association Rule-based classifiers, were applied. Model performance was assessed using the pairwise AUC metric, with Random Forest and Association Rule-based models yielding the best results. Establishing a link between immune infiltration and genetic susceptibility could offer new insights into pancreatic cancer research.