Development of a novel prognostic risk model for pancreatic adenocarcinoma exploiting multi-omics data

This paper proposes a novel approach for developing prognostic risk models for pancreatic adenocarcinoma (PDAC) using data from multiple omics layers, including clinical data, micro-RNA (miRNA), copy number variation (CNV), proteomics, phosphoproteomics, transcriptomics, and radiomics. Through a multi-step process, a highly reduced set of prognostically significant features is selected from an initial pool of over 135,000 features. Specifically, the first step involves the selection of omics layers that produce the highest C-index for prognosis prediction when training a Cox proportional hazards model within a leave-one-out (LOO) cross-validation (CV) framework. Dimensionality reduction is performed as a preliminary step using principal component analysis (PCA) to address the high dimensionality of the data before training the Cox model. In the second step, the SelectKBest algorithm is employed to identify the most significant features within the selected layers. This procedure results in the development of a phosphoproteomics-based risk model comprising only four features, achieving a C-index of 0.65 under LOO-CV. Additionally, a transcriptomics-based risk model is derived from the phosphoproteomics model. Its performance is compared against risk models in the literature, consistently demonstrating superior C-index values across all test datasets used in this study.