GAN-Enhanced Machine Learning and Metabolic Modeling Identify Reprogramming in Pancreatic Cancer

Abstract

Pancreatic Ductal Adenocarcinoma (PDAC) is one of the deadliest forms of cancer and presents a significant clinical challenge due to poor prognosis and limited treatment options. In this study, we developed a novel framework integrating genome-scale metabolic modeling (GSM) with machine learning to identify metabolic biomarkers and vulnerabilities in PDAC. We addressed the inherent class imbalance in cancer datasets by generating synthetic healthy samples using a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP), implementing a three-step biological filtration process to ensure their validity. Our approach achieved 94.83% accuracy in distinguishing between healthy and cancerous metabolic states. Systems-level analysis revealed three key dysregulated pathways: heparan sulfate degradation, O-glycan metabolism, and heme degradation. We identified impaired lysosomal degradation of heparan sulfate proteoglycans as a potential contributor to PDAC pathogenesis, providing a mechanistic explanation for the previously observed association between lysosomal storage disorders and pancreatic cancer. Additionally, we found that nervonic acid transport (MAR00336) was the most discriminative reaction between healthy and cancerous states, with gene-level analysis highlighting FABPs, SLC27As, ACSLs, and ACSBGs as key molecular drivers of metabolic reprogramming in PDAC. Overall, our multi-level approach connected genetic drivers to functional metabolic consequences, revealing coordinated upregulation of fatty acid transport and activation processes. These findings enhance our understanding of PDAC metabolism and present potential therapeutic targets, demonstrating the value of integrated computational approaches in cancer research.