A Nonlinear Mixture Model based Unsupervised Variable Selection in Genomics and Proteomics

Typical scenarios occurring in genomics and proteomics involve small number of samples and large number of variables. Thus, variable selection is necessary for creating disease prediction models robust to overfitting. We propose an unsupervised variable selection method based on sparseness constrained decomposition of a sample. Decomposition is based on nonlinear mixture model comprised of test sample and a reference sample representing negative (healthy) class. Geometry of the model enables automatic selection of component comprised of disease related variables. Proposed unsupervised variable selection method is compared with 3 supervised and 1 unsupervised variable selection methods on two-class problems using 3 genomic and 2 proteomic data sets. Obtained results suggest that proposed method could perform better than supervised methods on unseen data of the same cancer type.