Cancer molecular pattern discovery by subspace consensus kernel classification

Cancer molecular pattern efficient discovery is essential in the molecular diagnostics. The characteristics of the gene/protein expression data are challenging traditional unsupervised classification algorithms. In this work, we describe a subspace consensus kernel clustering algorithm based on the projected gradient nonnegative matrix factorization (PG-NMF). The algorithm is a consensus kernel hierarchical clustering (CKHC) method in the subspace generated by the PG-NMF. It integrates convergence-soundness parts-based learning, subspace and kernel space clustering in the microarray and proteomics data classification. We first integrated subspace methods and kernel methods by following our framework of the input space, subspace and kernel space clustering. We demonstrate more effective classification results from our algorithm by comparison with those of the classic NMF, sparse-NMF classifications and supervised classifications (KNN and SVM) for the four benchmark cancer datasets. Our algorithm can generate a family of classification algorithms in machine learning by selecting different transforms to generate subspaces and different kernel clustering algorithms to cluster data.