Comparison of data discretization methods for cross platform transfer of gene-expression based tumor subtyping classifier

Molecular stratification of tumors is essential for developing personalized therapies. While patient stratification strategies have been successful, computational methods to accurately translate and integrate gene signatures across different high-throughput platforms (e.g., microarray, RNA-seq) are currently lacking. We performed comparative evaluation of different data discretization and feature selection methods combined with state-of-the-art machine learning algorithms to derive platform-independent and accurate multi-gene signatures for classification of the four known subtypes of glioblastoma. Our results show that integrative application of feature selection and data discretization is crucial for successful platform transition and higher prediction accuracy of the derived molecular classifiers.