Data-driven approach to predict survival of cancer patients

Abstract

We develop a novel method to identify patients 'with different disease risk level. Our method estimates the optimal partition (cutoff) of a single gene's expression level by maximizing the separation of the survival curves related to the high- and low risk of the disease behavior. We extend our approach to construct two-gene signatures, which can exhibit synergetic influence on patient survival. Using bootstrapping and statistical modeling, we evaluate the performance of our method by analyzing Affymetrix U133 data sets of two large breast cancer patient cohorts. Using 232-grade signature genes associated with different aggressiveness of breast tumor, we reveal a large number of gene pairs, which provides pronounced synergetic effect on patient's survival time and identifies patients with low- and high-risk disease subtypes. The selected survival significant genes are strongly supported by gene ontology (GO) analysis and literature data. Specifically, for the first time, we demonstrate that cyclin A2 or cyclin A and protein tyrosine phosphatase T (CCNA2- PTPRT) and megalin (LRP2)-integrin alpha-7 (ITGA7) gene pairs can provide strong clinically significant interaction effects on the survival of breast cancer patients. Our technique has the potential to be a powerful tool for classification, prediction, and prognosis of cancer and other complex diseases.