CAMDA '00 Conference Report William R. Greco Director, Biomathematics/Biostatistics Resource Roswell Park Cancer Institute Elm and Carlton Streets Buffalo, NY 14263 William.Greco@,RoswellPark.org The Duke Bioinformatics Shared Resource sponsored a conference on the Critical Assessment of Techniques for Microarray Data Analysis (CAMDA '00), which took place at Duke University from December 18-19, 2000. Kim Johnson from Duke was the conference chair. Microarray technology is the hottest new approach for quantitatively measuring gene expression (amount of mRNA which codes for a specific protein, and which has been transcribed from DNA for a specific gene) for thousands of genes simultaneously. It has been found that gene expression patterns can be used as fingerprints to differentiate cancer subtypes, and thus microarray technology has become a powerful new tool for molecular pathologists. One possible disadvantage of this technology, from the point of view of laboratory scientists, is that the many thousands of data points that are generated from each experimental unit (e.g., tumor specimen or cultured cell population) must be carefully recorded, organized and statistically analyzed, in order to yield useful information. This disadvantage is however, seen as a challenge and opportunity for both computer scientists and statisticians to develop new methodologies for the statistical analysis and graphical presentation of microarray data. This challenge which crosses the boundaries of classical Bioinformatics, Machine Learning and Applied Statistics, was the main focus of this meeting. The format of the symposium was quite unusual. The meeting was organized around the reanalysis of miccroarray data from two previously-published data sets: Spelhnan et al, Comprehensive identification of cell-cycle-regulated genes of the yeast Saccharomyces cerevisiae by microarray hybridization. Mol. Biol. Cell 9: 3273-3297, 1998; and Golub et al, Molecular classification of cancer: class discovery and class prediction by gene expression monitoring. Science 286(5439): 531-537, 1999. The original data sets were available over the World Wide Web, and meeting participants were invited to reanalyze either or both data sets, with any data analysis approach that they liked, prior to the time of the meeting. Abstracts on the reanalysis of data were due about 5 weeks prior to the meeting, and acceptances of abstracts were sent out 4 weeks prior to the meeting. There were 14 oral presentations of reanalyses, and 33 poster presentations. The approaches used for reanalyses of the microarray data spanned a very wide spectrum, and included classical and avant-garde classification, clustering, machine learning, visualization and statistical techniques. The conference web site, http://bioinformatics.duke.edu/camd_,_a/ includes the agenda, the two data sets, descriptions of the two data sets, and other supplementary information. Most participants presented analyses of only the data set from Golub et al (1999). In this; paper, the training set included data from 38 bone marrow samples, one sample from each of 27 patients with acute lymphoblastic leukemia (ALL) and I 1 patients with acute myeloid leukemia (AML). Each sample contributed a relative gene expression value for each of 6,817 human genes (Affymetnx GeneChips). The test set included data from 24 bone marrow and 10 peripheral blood samples, one sample from each of 34 leukemia patients, different from the patients who contributed data to the training set. The very surprising result of this contest was that all techniques worked extremely well; i.e., for the test data set, correct classifications of ALL or AML were made by almost all approaches for 32-33 of the 34 patients. The highly complex techniques did not perform materially better than the simplest of approaches. There was a high amount of redundancy among the gene expression data: i.e., many sets of a small number of genes; e.g., two,, were adequate to correctly predict the disease in the test set. Therefore, it was not clear whether any approach was superior, in a practical sense, to the other approaches. The participants were given the opportunity to vote for the "best" approach. The winner was Chris Stoeckert from the University of Pennsylvania. The title of his presentation was: "Using non-palametric methods in the context of multiple testing to determine differentially expressed genes". Opening remarks were given by John Weinstein from the National Cancer Institute, keynote addresses were given by Mike West from Dud~e University and Gavin Sherlock from Stanford University, and closing remarks were provided by Athel CornishBowden from CNRS in France. New array technologies are currently being developed in both academic laboratories and in Biotechnology companies. There is a great deal of financial interest in this area of Biotechnology. The opportunities for computer and other quantitative scientists in this area are vast. 26

/lp/association-for-computing-machinery/camda-00-conference-report-qbANt8do2B