Genomics, High-Dimension Optimization, and GPUs

• Activity Number: 214
• Type: Invited
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: International Indian Statistical Association
• Author(s): Kenneth Lange*+
• Companies: University of California at Los Angeles
• Keywords: Genetics ; computational statistics ; optimization ; graphic processing units

Abstract

Expression and genotyping chips and especially ultra-fast DNA sequencing generate more data than scientists can possibly analyze with traditional statistical models. Current genomics problems include: identifying different splice forms from RNA-Seq data, testing for association between rare variants and common diseases, clarifying population histories and ethnic admixture, haplotyping, and reconstructing the microbial communities that colonize particular hosts or environmental niches. This talk will stress the constraints and compromises entailed in modeling modern genomics data. Most relevant models involve high-dimensional optimization, and unless algorithms are carefully tailored to specific problems, the numerical hurdles are insurmountable. Graphics processing units (GPUs) offer one avenue to success. These cheap, fast devices thrive on parallel algorithms that perform the same operations on different subsets of the data. One can achieve parallelization by separating parameters via either block coordinate descent or EM and MM (majorize-maximize) algorithms. This talk will feature a few genomic examples illustrating these general principles.