Stability-Based Comparison of Class Discovery Methods for DNA Copy Number Profiles

MOTIVATION: Array-CGH can be used to determine DNA copy number, imbalances in which are a fundamental factor in the genesis and progression of tumors. The discovery of classes with similar patterns of array-CGH profiles therefore adds to our understanding of cancer and the treatment of patients. Various input data representations for array-CGH, dissimilarity measures between tumor samples and clustering algorithms may be used for this purpose. The choice between procedures is often difficult. An evaluation procedure is therefore required to select the best class discovery method (combination of one input data representation, one dissimilarity measure and one clustering algorithm) for array-CGH. Robustness of the resulting classes is a common requirement, but no stability-based comparison of class discovery methods for array-CGH profiles has ever been reported. RESULTS: We applied several class discovery methods and evaluated the stability of their solutions, with a modified version of Bertoni's [Image: see text]-based test [1]. Our version relaxes the assumption of independency required by original Bertoni's [Image: see text]-based test. We conclude that Minimal Regions of alteration (a concept introduced by [2]) for input data representation, sim [3] or agree [4] for dissimilarity measure and the use of average group distance in the clustering algorithm produce the most robust classes of array-CGH profiles. AVAILABILITY: The software is available from http://bioinfo.curie.fr/projects/cgh-clustering. It has also been partly integrated into "Visualization and analysis of array-CGH"(VAMP)[5]. The data sets used are publicly available from ACTuDB [6].