Title: Network modeling of large-scale genomic data
Abstract: Genome-wide network models of multi 'omics cancer data are popular tools for studying and revealing both unique and shared mechanisms across malignancies. In this talk we will discuss some "pan-can" (multi-cancer) models based on SICS (sparse inverse covariance selection). We will present some statistical methods for selecting network complexities and how one can use concepts of data depth to produce robust network estimates. Time permitting we will present some recent results on multiresolution network models (MR-SICS). While our SICS model is a highly useful tool for network modeling, there are important questions that warrant further study. First, the large number of co-linear variables creates instability of network estimation. Secondly, while model estimation is based on highly optimized solvers, improvements of scalability are needed to handle future data sets. Our multi-resolution method, MR-SICS, builds on a nested latent model formulation of network components. At each level of resolution of the model, the parameters for comparative inference are relatively small, substantially improving estimation stability and interpretability over standard SICS models.
Download the flyer here