Multiclass discovery in array data

BACKGROUND: A routine goal in the analysis of microarray data is to identify genes with expression levels that correlate with known classes of experiments. In a growing number of array data sets, it has been shown that there is an over-abundance of genes that discriminate between known classes as compared to expectations for random classes. Therefore, one can search for novel classes in array data by looking for partitions of experiments for which there are an over-abundance of discriminatory genes. We have previously used such an approach in a breast cancer study. RESULTS: We describe the implementation of an unsupervised classification method for class discovery in microarray data. The method allows for discovery of more than two classes. We applied our method on two published microarray data sets: small round blue cell tumors and breast tumors. The method predicts relevant classes in the data sets with high success rates. CONCLUSIONS: We conclude that the proposed method is accurate and efficient in finding biologically relevant classes in microarray data. Additionally, the method is useful for quality control of microarray experiments. We have made the method available as a computer program.