Nonlinear Sparse Component Analysis with a Reference: Variable Selection in Genomics and Proteomics

Many scenarios occurring in genomics and proteomics involve small number of labeled data and large number of variables. To create prediction models robust to overfitting variable selection is necessary. We propose variable selection method using nonlinear sparse component analysis with a reference representing either negative (healthy) or positive (cancer) class. Thereby, component comprised of cancer related variables is automatically inferred from the geometry of nonlinear mixture model with a reference. Proposed method is compared with 3 supervised and 2 unsupervised variable selection methods on two-class problems using 2 genomic and 2 proteomic datasets. Obtained results, which include analysis of biological relevance of selected genes, are comparable with those achieved by supervised methods. Thus, proposed method can possibly perform better on unseen data of the same cancer type.