Prediction of Lung Cancer Survival Based on Multiomic Data

Abstract

Lung cancer is the leading cause of cancer death among both men and women, which mainly results from low effectiveness of the screening programs and late occurrence of symptoms, that are usually associated with advanced disease stages. Lung cancer shows high heterogeneity which was many times associated with its molecular background, providing the possibility to utilize machine learning approaches to aid both the diagnosis as well as the development of personalized treatments. In this work we utilize multiple -omics datasets in order to assess their usefulness for predicting 2 year survival of lung adenocarcinoma using clinical data of 267 patients. By utilizing mRNA and microRNA expression levels, positions of somatic mutations, changes in the DNA copy number and DNA methylation levels we developed multiple single and multiple omics-based classifiers. We also tested various data aggregation and feature selection techniques, showing their influence on the classification accuracy manifested by the area under ROC curve (AUC). The results of our study show not only that molecular data can be effectively used to predict 2 year survival in lung adenocarcinoma (AUC = 0.85), but also that information on gene expression changes, methylation and mutations provides much better predictors than copy number changes and data from microRNA studies. We were also able to show the classification performance obtained using different dimensionality reduction methods on the most problematic copy number variation dataset, concluding that gene and gene set aggregation provides the best classification results.