Analyzing gene expression data for childhood medulloblastoma survival with artificial neural networks

One of the major problems facing gene expression researchers is how to reduce the high dimensionality of gene expression data in the face of small sample sizes in comparison to the number of genes measured. We report on the application of a single layer neural network for reducing the number of originally measured genes from over 7000 to 64 through repeated application of thresholds on the weights linking genes to the class values of samples. This results in a small but informative gene set for the domain of brain cancer that can be further analyzed through the application of symbolic machine learning techniques (See5) and cluster analysis (cluster and tree view).

I. Introduction

Embryonal central nervous system (CNS) tumours are a diverse group of diseases that constitute the most common group of malignant brain tumours in childhood. Brain tumors are classified according to histology, but tumor location and extent of spread are important factors that affect treatment and prognosis. Immunohistochemical analysis, cytogenetic and molecular genetic findings, as well as measures of mitotic activity, have been used in tumor diagnosis and classification (e.g. [1], [2], [3], [4], [5], [6]). Each year, nearly 3,000 children in the U.S. are diagnosed with brain tumors. Almost half will die within five years, making it the most fatal cancer among children. If a child does survive a brain tumor, the long-term effects can be significant and can include neurological disabilities, retardation and psychological problems [7]. Beyond surgery, successful treatments for childhood brain tumors are rare.