Latent Dirichlet Allocation Models to Enable Personalized Genomic Medicine

• Activity Number: 149
• Type: Invited
• Date/Time: Monday, July 30, 2012 : 10:30 AM to 12:20 PM
• Sponsor: WNAR
• Author(s): John A. Dawson and Christina Kendziorski*+
• Companies: University of Wisconsin-Madison and University of Wisconsin-Madison
• Keywords: latent Dirichlet allocation ; survival analysis ; topic models

Abstract

Genomic based studies of disease now involve highly diverse types of data collected on large groups of patients. A challenge facing statistical scientists is how best to combine the data, extract important features, and comprehensively characterize the ways in which the features affect an individual's disease course and treatment response. In this talk, I will review methods developed to address this challenge. Drawing an analogy from information retrieval, we consider each patient as a document, and data on each patient as text. Documents are constructed using data from multiple clinical sources and high-throughput assays. By introducing priors that accommodate known structure among genomic variables, we develop a latent Dirichlet allocation based model that allows for discovery of distinct topics (collections of genomic aberrations, clinical variables, and treatments) as well as determination of patient-specific mixtures over topics. Further model extensions provide for survival-related responses to supervise model fit. The approach facilitates data integration across multiple platforms and scales to enable powerful patient-specific inference.