Biclustering of Linear Patterns in Gene Expression Data

• Activity Number: 176
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Christine Ho*+ and Qinghui Gao and Haiyan Huang and Jessica Jingyi Li and Yingmin Jia
• Companies: University of California at Berkeley and Beihang University of Aeronautics and Astronautics and University of California at Berkeley and University of California at Berkeley and Beihang University of Aeronautics and Astronautics
• Keywords: biclustering ; gene expression ; contrast optimization ; biological networks ; genomics

Abstract

Identifying a bicluster, or a gene expression submatrix wherein the genes express similar behavior over the columns, is useful for discovering novel functional gene interactions. In this article, we introduce a new algorithm for finding biClusters with Linear Patterns (CLiP). Instead of relying solely on the Pearson correlation coefficient to assess profile shape, we introduce a similarity measure that takes into account both shape and variance. In the context of noisy data, as with microarray, the inclusion of a variance component more effectively captures these linear subprofiles than correlation alone. Further, we define a fitness function that considers contrast between a bicluster and its complementary sets. In so doing, we avoid a priori determination of the bicluster size. We employ a combination of greedy search and the genetic algorithms in optimization, and incorporate resampling for more robust discovery. When applied to both real and simulation datasets, our method is superior to existing methods. When applied to RNA-seq fly and worm time-course data from modENCODE, we uncover a set of similarly expressed genes suggesting maternal dependence.