IMA Complex Systems Seminar

3:30 Wednesday, October 8, 2003


Randomization test for array-based comparative genomic hybridization


Michael Newton

Departments of Statistics and Biostatistics and Medical Informatics

University of Wisconsin Madison

Madison, WI 53706


After low-level preprocessing, data from an array-based comparative genomic hybridization (aCGH) experiment may be viewed as a set of independent and identically distributed (iid) copies of the random vector X = (X_1, X_2, ..., X_n) where X_i ~ Bernoulli( p_i ), but in which there is dependence among the components owing to aspects of the tumor growth process that is being measured by X. Here i denotes a position in the genome and X_i indicates whether or not genomic damage occurs at position i in the sampled tumor. One statistical problem is to test the null hypotheses H_0: p_i = p for all i; that is, there are no aberration hot-spots in the genome. I will describe several attempts to implement such a test using conditional inference and three permutation methods. I conjecture that an elementary shuffling procedure provides a conservative hypothesis test. The approach is tested on a set of 60 aCGH profiles from cancer cell lines. Issues with hidden-Markov model-based inferences will also be discussed.