Multimodal Integrative Prediction of Pancreatic Cancer

Abstract

While there has been some progress on discovering clinically validated biomarkers for early detection in pancreatic ductal adenocarcinoma (PDAC), several challenges remain. The most well-known biomarker, carbohydrate antigen 19-9 (CA19-9), has limited sensitivity and specificity for early-stage disease. Using data from a multicenter study with age-matched cohort ($n=203$ of healthy controls $n = 46$, pancreatitis controls $n = 36$, and diagnosed cases $n = 121$), we develop a multimodal approach integrating 2083 micro-RNAs (miRNA), 125 metabolites, and 3 circular protein biomarkers, for early detection in PDAC. We pursued a machine learning (ML) approach that selects an optimal panel of biomarkers (that is a subset of the full set of biomarkers included for analysis) resulting in the greatest discriminatory power for early detection of PDAC. In particular, the selected panel of biomarkers relies heavily on metabolites and produces greater than $95%$ area under the curve (AUC) for all classification tasks, while exhibiting sensitivity $\sim90%$ at $90%$ specificity for almost all tasks considered within the training sample. The performance under the multimodal panel generalizes well to a test sample that was held out when training the machine learning model. We propose a data adaptive approach to automatically select tuning parameters for prediction under the ML approach which produces high sensitivity ($\ge 0.94$) and reasonable specificity that are much improved compared to miRNA-based panel proposed recently in literature. The ML approach identifies biomarkers with the highest importance in classification, singling out the role of aminobutyric acid in discriminating between healthy controls vs cancer and pancreatitis vs early-stage cancer. Additional biomarkers such as fumaric acid, and miR-4455, with high importance scores are also discovered and included in the multimodal biomarker panel. Our analysis clearly illustrates the superior ability of the multimodal biomarker panel to elicit strong sensitivity for early detection in PDAC and discovers novel metabolite biomarkers that have clear discriminatory power.