Finding Important Genes

• Activity Number: 356
• Type: Contributed
• Date/Time: Tuesday, July 31, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Chamont Wang*+ and Jana Gevertz and Sudhir Nayak and Chaur-Chin Chen and Leonardo Auslender
• Companies: College of New Jersey and The College of New Jersey and The College of New Jersey and National Tsing Hua University and TD Bank
• Keywords: Microarray data ; simulation ; stochastic grading boosting

Abstract

Microarrays are a powerful tool for measuring the change in relative expression levels of thousands of genes in a single experiment. A multitude of statistical techniques have been applied to microarray data to determine which genes are in order to identify genes implicated in a particular disease state or to predict the state of a cell based on its gene expression profile. In this manuscript, we begin to develop a platform for rigorously testing the statistical and machine learning approaches used to analyze microarray data. Particularly, we look at two well-studied microarray data sets, analyze the data using a number of statistical techniques, and rigorously investigate the statistical and biological significance of the findings. One surprising finding is that many statistical tools achieved high prediction accuracies using different important genes, raising the possibility that some of the genes successfully used to classify the data sets may not actually be implicated in cancer. To further explore this phenomenon, simulated microarray data was developed and classified as diseased or normal based on a prescribed formula. Remarkably, when the genes implicated in the diseas