Estimating the Reliability of Inferences Based on cDNA Microarray Ratio Data

Michael Bittner
NIH/NHGRI
Building 49, Room 4A52
9000 Rockville Pike
Bethesda, MD 20892 USA
telephone: 1-301-496-7980
fax: 1-301-402-3241
email: mbittner@nhgri.nih.gov
prestype: Platform
presenter: Michael Bittner

Y. Chen1, Y. Jiang1, E. Dougherty2, S. Kim2, Z. Yakhini3, A. Ben-Dor3, N. Sampas3, M. Radmacher4, R. Simon4, M. D. Gubitoso5, M. Brun6, J. Barrera6, C. Gooden, A. Glatfelter, P. Meltzer1, J. Trent1, and M. Bittner1

1Cancer Genetics Branch, National Human Genome Research Institute, NIH, Bethesda, Maryland 20892
2Department of Electrical Engineering, Texas A & M University, College Station, Texas 77843
3Chemical and Biological Systems Department, Agilent Laboratories
4National Cancer Institute, DCTDC, NIH, Bethesda, Maryland 20852
5Department of Computer Science, University of São Paulo
6 Institute of Mathematics and Statistics, University of São Paulo

The ratio data from cDNA microarray experiments is used in a variety of analyses that compare the patterns of expression across a series of samples. These analytical tools are used to group either the samples, on the basis of overall similarities in expression pattern, or the genes themselves, on the basis of their individual similarities in expression pattern. Both the strength and breadth of the inferences that can be made in these analyses ultimately depend on the accuracy and reproducibility of the initial measurements. In this presentation we will show studies of the strength of inferences made on the basis of ratio data from a variety of perspectives. Using repeated measurement of the same samples and external validation it has been possible to directly examine accuracy and reproducibility. We have used these estimates to determine the magnitude of effect they would have in typical analyses. Using perturbation and randomization of actual data, it has been possible to demonstrate that observed consistencies of gene expression associated with sample type are extremely unlikely to arise from error in the measurements or from the inherent biological order existing in all individual samples. Simulation studies of the effects of sample variance on gene by gene clustering provides a way to find forms of class discovery tools most likely to correctly identify genes with similar expression patterns at the levels of variance typical of microarray results.



  Abstract List


Abstracts * Speakers * Organizers * Home


Genetic Meetings