A Local-Network Guided Linear Discriminant Analysis for Classifying Lung Cancer Subtypes using Individual Genome-Wide Methylation Profiles

Abstract

Accurate and efficient prediction of lung cancer subtypes is clinically important for early diagnosis and prevention. Predictions can be made using individual genomic profiles and other patient-level covariates, such as smoking status. With the ultrahigh-dimensional genomic profiles, the most predictive biomarkers need to be first selected. Most of the current machine learning techniques only select biomarkers that are strongly correlated with the outcome disease. However, many biomarkers, even though have marginally weak correlations with the outcome disease, may execute a strong predictive effect on the disease status. In this paper, we employee an ultrahigh-dimensional classification method, which incorporates the weak signals into predictions, to predict lung cancer subtypes using individual genome-wide DNA methylation profiles. The results show that the prediction accuracy is significantly improved when the predictive weak signals are included. Our approach also detects the predictive local gene networks along with the weak signal detection. The local gene networks detected may shed lights on the cancer developing and progression mechanisms.