| Cancer is one of the most lethal diseases threatening human health. Traditional medicine is facing challenges in meeting the needs of diagnosis and treatment in personal medicine. With the advances of DNA microarray technology, it becomes possible to monitor thousands of genes at the expression level across the genome. Many biomarkers selection approaches have been proposed and satisfying classification results achieved. Among them, supervised feature selection methods select biomarkers mainly based on classification quality and filtering methods detect differential expressed genes, however microarray data are notorious for the curse of dimension and it is desirable that the selected gene set is biologically relevant, consequently robust to the noise, and highly advantageous to build reliable classification models. In this study, we propose a strategy that identifies biomarkers primarily in terms of their clinical outcome relevance, and then adjust the candidate biomarkers based on the classification quality. Technically, Supervised Singular Value Decomposition (SSVD) and a Random Forest based method are used to correlate genes with cancer diagnostic outcomes and fine tune the selected gene set in terms of classification accuracy respectively. The advantages are that rather than identifying genes from tens of thousands of candidates with so few samples, genes are sorted into different groups which are dramatically small in number and are biological function relevant. More importantly, the relationship of genes and clinical outcomes are visualized and Genes are essentially identified in terms of their relevance with clinical diagnostic outcomes. The Random Forest method is then proposed to sort the pre-selected candidate gene. The approach was test on 3 broadly used public data sets. Graphically, the identified gene set shows a clear association with clinical cancer outcomes. Comparative study and statistical analysis shows the proposed methods compares favorably to the other two typical gene selection methods in terms of classification over 4 classifiers. More importantly, the identified genes demonstrate closer relationship with clinical outcomes, are less variable and comparatively invariant to the external influences. Ontology study and literature research shows that many of the identified genes are also found to be biologically relevant to the states of cancers. |
