Statistics Colloquium : Dr. Deepak Ayyala
Augusta University
Friday, November 9, 2018 · 11 AM - 12 PM
Title : Penalized multivariate count models for genomic data
Abstract: Genomic and metagenomic experiments yield high dimensional count data which are extremely sparse, making parameter estimation very difficult. Features which are differentially expressed under two or more conditions are hard to detect due to the high sparsity. In this talk, I will present two penalized models for multivariate count data using Dirichlet-Multinomial distribution. In the first model, a `2-based penalty function is designed to perform simultaneous gene selection and cell-type detection in single-cell RNA-seq experiments. We implemented a fast Newton-Raphson algorithm by avoiding large matrix inversions. The computation cost is reduced to a quadratic rate with respect to number of variables. In the second model, we proposed a new penalty function to address feature selection in metagenomic studies where the scale of parameters are extremely different. The proposed penalty function equally weighs both highly abundant and sparsely abundant taxa. 1