Identification of characteristic methylation sites in gastric cancer using genomics-based machine learning

Abstract

Objective: To construct a prediction model of gastric cancer related methylation using machine learning algorithms based on genomic data. Methods: The gene mutation data, gene expression data and methylation chip data of gastric cancer were downloaded from The Caner Genome Atlas database, feature selection was conducted, and support vector machine (radial basis function), random forest and error back propagation (BP) neural network models were constructed; the model was verified in the new data set. Results: Among the three machine learning models, BP neural network had the highest test efficiency (F1 score=0.89,Kappa=0.66, area under curve=0.93). Conclusion: Machine learning algorithms, particularly BP neural network, can be used to take advantages of the genomic data for discovering molecular markers, and to help identify characteristic methylation sites of gastric cancer.