One promising use of microarray data is disease classification. Earlier studies of microarray data (Golub et al.) suggest that ``predicting cancer classes, independent of previous biological knowledge, is feasible. Since (Golub et al.), there has been a lot of work on sample classification on microarray data, but no standards have been established for the analysis. The goal of the proposed thesis is to develop analysis tools for microarray data that generate biologically meaningful outputs. In particular, focus is given to development of supervised sample classification methods.
The simplest form of the sample classification problem is the binary classification problem. Medical applications of binary classifiers abound. We have developed a binary sample classifier called Independently Consistent Expression Discriminator (ICED). ICED builds a voting scheme from the training data. To this end a number of features that are informative in distinguishing between the two classes are selected. The novelty of ICED is that it searches for features whose values are consistent in one class and are not consistent at the same level in the other class. Empirical evaluation of ICED shows that simple methods as ICED can perform well, even outperform current classifiers on most counts, and ICED seems to discover biologically relevant genes.
Although ICED achieved relatively high prediction performance empirically, its superiority over the existing methods needs rigorous statistical proof. Also, its most significant limitation is that it is a binary classifier. To study classification of samples into subtypes or comparison of several diseases in molecular level, multi-class classification mechanism is needed. Another important problem is how to integrate expression data with other relevant biological information to improve the performance of the methods. These will be the research goals for future study.