B1321
Title: Network-based clustering of pancancer data accounting for clinical covariates
Authors: Fritz Bayer - ETH Zurich (Switzerland) [presenting]
Giusi Moffa - University of Basel (Switzerland)
Niko Beerenwinkel - ETH Zurich (Switzerland)
Jack Kuipers - ETH Zurich (Switzerland)
Abstract: Cancer progresses in diverse ways leading to a heterogeneous landscape of mutations within and across cancer types. This heterogeneity is a considerable challenge for precision medicine, and the task of leveraging genomic data to predict survival and treatment outcomes. We focus on learning the diverse probabilistic relationships among mutations and clinical covariates. We propose a novel network-based clustering method that allows us to learn distinct mutational patterns while accounting for covariate effects. Using probabilistic graphical models, we cluster the mutations and covariates based on their distinct probabilistic relationships. Since the covariates should not drive the clustering of mutational patterns but are necessary to accurately model the mutations, we propose a covariate-adjusted clustering framework. Our framework allows us to detach the effects of covariates on the clustering, by exploiting causal relationships among the variables. Over a broad range of simulations, we demonstrate that our method outperforms standard clustering methods in correlated data. We apply our method to a large-scale genomic dataset, including the mutational profiles and clinical covariates of 8085 patients, where we identify novel clusters based on mutational patterns. These clusters are significantly predictive of survival beyond clinical information and could serve as biomarkers for targeted treatment.