L analysis is performed. This paper describes a simple microarray data analysis technique for tumor classification ranking. In particular, we apply MDR, our recently developed Multi-Dimensional Ranking algorithm, for analyzing gene expression in various types of cancers including leukemia, lung, prostate, lymphoma, and breast cancers. These data have been used in previous cancer research studies [1,3,5-8,10]. They are publicly available and can be obtained from the Kent Ridge Biomedical Data Set Repository [11].ResultsWe analyze microarray data for 11 types and subtypes of tumors using MDR. The two Leukemia expression data sets are concerned with classification of acute lymphoblastic leukemia (ALL). Golub et al. [8] PubMed ID:https://www.ncbi.nlm.nih.gov/pubmed/26266977 used the ALL-AML Leukemia expressions to help discover a single diagnostic test to differentiate between two types of human acute leukemia: acute myeloid leukemia (AML) and ALL. The ALL-subtype expression data were used by Yeoh et al. [7] to identify six known prognostically important leukemia subtypes of ALL from pediatric ALL patients. These ALL subtypes include: BCR-ABL, E2A-PBX1, Hyperdiploid > 50 chromosomes, MALL, T-ALL, and TEL-AML1. Because different leukemia types and subtypes respond to chemotherapy differently, the ability to determine the classification of an ALL subtype for a new leukemia tissue sample can be valuable for cancer treatment. MDR has been designed for binary classification ranking. When dealing with multiple classes, we employ the “one against many” purchase Trichostatin A strategy (i.e., for each class, perform a binary classification between that class and all the other classes). As recommended in a recent study by Li and Liu [9], we divide the ALL-subtype expression data into six cases, each of which focuses on classification of a particular ALL subtype against all the other subtypes. The ALL classification model (or classifier) obtained from MDR contains two predictor genes, whereas the number of predictor genes in the classifiers for the six ALL subtypes ranges from one to nine. This is a huge reduction from the original number of genes, over 12,000, in the ALL-AML and ALL subtype expression data. The lung cancer data are analyzed to distinguish between malignant pleural mesothelioma (MPM) and adenocarcinoma (ADCA) [3], whereas the prostate and breast profile expression data are analyzed for tumor diagnosis and prognosis (e.g., “relapse” in patients who developed distance metastates within five years), respectively [5,6]. The lymphoma microarray data include gene expressions of diffuse large B-cell lymphoma, a subtype of non-Hodgkin’s lymphoma [1]. To indicate different stages of B-cell malignancies, gene expression patterns studied are of two types: the germinal centre B-like type and the activated Blike type. Patients with the germinal centre B-like type have had better survival rates. The classification models obtained by MDR from the lung, prostate, lymphoma and breast expression data contain five, six, three and eight predictor genes, respectively. The predictor genes in each of these MDR classification models, derived from the corresponding expression data sets, are shown in Table 1. While constructing the classification model, MDR reduces the number of data dimensions to a smaller number, selecting features relevant for classification ranking in the model. Unlike the machine learners that require an additional step for feature selection, MDR does not do featurePage 2 of(page number not for citation purposes)BM.