H.Hong, H.Fang, Q.Xie, R.Perkins, W.Tong, NCTR, FDA, Jefferson AR
DNA microarray technology enables investigating the expression of thousands of genes in a single experiment. Accurate analysis and meaningful interpretation of massive information from this technology pose a unique challenge. Classification using the pattern recognition methods can be used to predict disease type and/or stage and chemical-induced toxicity based on the gene expression data from microarray experiment, which shows promising to correlate genotype with phenotype. The classification model development consists of two steps: model construction and validation. Most classification models based on microarray data in the literature are developed using only a small set of genes out of a large set of genes on a chip through a gene selection procedure first. The same set of genes is then used in both model development and validation steps, which has been proved inappropriate to validate the robustness of a model. In this poster, a novel classification method, named Decision Forest is presented, which combines the gene selection and model construction into a single step. The method offers a number of advantages over traditional classification approaches. More specifically, the cross-validation results of Decision Forest have better indication to the quality of a classification model.