 | Generative Model Based Analysis of Cancer Associated Gene Expression Matrices
|
| | Mattias Wahde1) and Zoltan Szallasi2) |
| 1) Div. of Mechatronics Chalmers University of Technology
2) Department of Pharmacology, Uniformed Services University of the Health Sciences |
| | One of the main aims of analyzing cancer associated gene expression matrices is to identify a subset of genes that is consistently mis-regulated in a given type of tumor samples. Such a subset of genes forms, together with an appropriate function, a separator that can distinguish between normal and tumor samples. Separators can appear accidentally due to the high level of gene expression diversity detected in cancer. Various statistical methods can be used to estimate whether the appearance of a given separator is due to chance. However, the accuracy of all these tests will depend on the null hypothesis provided by the data structure. In this paper we are introducing generative models in order to simulate random, discrete gene expression matrices that retain the key features of massively parallel measurements in cancer. These include the number of changeable genes and the level of gene co-regulation as reflected in their pair-wise mutual information content. We show that the probability of the chance appearance of separators can be underestimated by many orders of magnitude if random and independent selection of mis-regulated genes is assumed instead of using the generative model outlined in this paper. |
| | | |
