Efficient multi-class cancer diagnosis algorithm, using a global similarity pattern

Since different subtypes of a cancer respond differently to the same therapy, it is important to diagnose the cancer type of a patient correctly, and then customize the treatment for that patient. DNA microarrays have recently received a great deal of attention in cancer diagnosis. Given a microarray dataset for multiple subtypes of cancer, the proposed procedure sequentially combines a gene-rank algorithm for detecting significant genes, with a pattern-based classifier for diagnosing a query test sample. In detail, for each cancer subtype, genes are ranked to determine a characteristic pattern, and the classifier measures a similarity between the sample and its type, based on the selected top-ranked genes. The sample is then classified according to the subtype to which it is the most similar. This is different from the widely applied k-nearest neighbor approaches using local similarity patterns. The procedure utilizes reliable global patterns to classify the types in test samples. Empirical studies using public datasets show that the top-ranked genes in each subtype provide a clear means of discrimination, and the classifier uses a few significant genes to distinguish the types in the test samples correctly. The procedure is an excellent alternative to more complex approaches due to its simplicity, ease of use, and robustness.