Random Forests Extract Interaction Information from Next-Generation Sequencing Data

• Activity Number: 221
• Type: Invited
• Date/Time: Monday, August 10, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Memorial
• Author(s): Bin Yu* and Sumanta Basu and Ben Brown
• Companies: UC Berkeley and UC Berkeley/Lawrence Berkeley National Laboratory and Lawrence Berkeley National Laboratory/UC Berkeley
• Keywords: Random forests ; interaction ; RNA-seq ; ChiP-seq ; ENCODE3

Abstract

Identification of information passage mechanisms between nascent transcripts and local chromatin landscape poses considerable statistical and computational challenges for next generation high-throughput sequencing data (e.g. RNA-seq, ChiP-seq, fRIP-seq etc. produced by ENCODE3 consortium). This is due to the high-dimensionality of the datasets and inherent heterogeneity of the human transcriptome. In this talk, we use random forests (RFs) to develop a robust and scalable method for feature selection and detection of potentially higher-order interactions among high-dimensional predictors. We propose a two-step strategy based on node selection and random intersection to identify important features and their combinations without imposing hierarchical restrictions on the interactions. We investigate the properties of the proposed method via extensive simulation studies with biologically motivated generative models with favorable outcomes. Applying our proposed method to high-throughput sequencing data mentioned above, we attempt at understanding the role of chromatin marks in alternative splicing in human with promising results.