(667b) Panel of Gene-Pair Relative Expression Classifiers for Stratifying Various Cancer and Non-Cancer Diseases of the Human Brain

One of the main challenges in cancer therapy is the accurate diagnosis of tumor class and progression grade, which can be difficult because of morphologically indistinguishable tissue representation and inter-observer variability. In order to address this issue, cancer diagnosis is moving from classification based on microscopic observations of histologic specimens alone, to a method based on molecular criteria by using microarray gene expression profiling. One classification scheme based on gene expression microarrays is the k-Top Scoring Pairs (k-TSP) approach, a supervised machine-learning method based on the concept of finding gene-pairs that exhibit relative expression reversals, which can furthermore serve as classifiers to distinguish two different classes of disease. The goal of this study was to identify gene-pair relative expression decision rules that can distinguish seventeen different classes and progression grades of cancerous and non cancerous human brain diseases. By extending the binary classification k-TSP method into a series of multi-class decomposition schemes (1-vs-1, 1-vs-others) on an extensive collection of publicly available gene expression microarray data, we have accumulated all the candidate gene pairs from a training set into a panel of classifier markers for differentiating disease class and grade. The results show that our gene panel distinguishes each class and grade from the other sixteen classes at high accuracy using leave-one-out cross validation and also on a test set, proving to be a potential strategy for stratifying patients based on gene expression signatures. This panel of gene-pair relative expression classifiers is the first to be built for the stratification of such a large number of different brain diseases and grades, and can serve as a highly accurate diagnostic tool in combination with traditional histopathological techniques.