Improved Feature Selection by Incorporating Gene Similarity into the LASSO

Introduction

An important goal of the healthcare industry is personalized medicine, which has the potential to transform healthcare practice. One aspect of the personalized medicine focuses on customizing treatments based on the genetic profile of a patient. In order to achieve this goal, researchers are looking for biomarkers, such as the expression of a gene or group of genes that correlate with treatment outcomes. Biologists use microarrays to read the gene expression of a biospecimen, see (Dubitzky, Granzow, Downes, & Berrar, 2009) for an introduction to microarrays. Some groups are using newer techniques such as RNA-seq to read gene expression (Marioni, Mason, Mane, Stephens, & Gilad, 2008). In this paper, we focus on microarrays; however, we believe our method could be applied in a similar manner to RNA-Seq data. The result of a microarray experiment is a gene expression profile. We represent a gene expression profile by , whereis the number of genes and corresponds to the expression of gene, or feature, . A series of microarray experiments yields a matrix where each row represents a single microarray experiment. In this paper, we assume each row of is a biospecimen from a different patient. The gene expression matrix can be used for supervised or unsupervised learning. However, we focus exclusively on supervised learning. In supervised learning, we have a vector where each corresponds to a row of . If is continuous then this is a regression problem, but if represents an element from a set of categories or labels then this is called a classification problem. A representation of these notational concepts can be seen.