Use of Sparse Linear Discriminant Analysis in Classification and in Testing Gene Pathways

• Activity Number: 223
• Type: Topic Contributed
• Date/Time: Tuesday, August 5, 2008 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Computing
• Author(s): Michael C. Wu*+ and Lingsong Zhang and Xihong Lin
• Companies: Harvard University and Harvard University and Harvard University
• Keywords: Linear discriminant analysis ; Variable selection ; Classification ; Gene pathways ; Proteomics

Abstract

Linear discriminant analysis (LDA) has been effectively applied in many settings such as proteomics/genomics. We consider two-class sparse LDA (sLDA) which imposes an L1-constraint on the discriminant direction to incorporate variable selection and remove the contribution of noise variables. Our formulation leads to a piecewise linear solution path and we exploit this property to develop a fast computational algorithm for computing the entire regularized solution path. We study two applications. First, sLDA has high classification accuracy when predicting mercury exposure from proteomic profiles. Second, sLDA leads to a novel approach for testing differential gene pathway activity when many genes are contributing only noise and is applied to a microarray experiment studying metal fume exposure. The environmental health implications of our analyses are interesting and will be discussed.