A0751
Title: Incorporating gene ontology and disease ontology into Bayesian feature selection for cancer subtypes
Authors: Thierry Chekouo - University of Minnesota (United States) [presenting]
Abstract: Kidney and lung cancers are among the deadliest diseases and have multiple subtypes. A Bayesian variable selection method is proposed for genomic selection and survival prediction, incorporating gene ontology and disease ontology information with application to the cancer genome atlas's kidney and lung cancer repositories. Instead of treating each gene equally without regard to their molecular functions as in conventional Bayesian variable selection methods, an algorithm and a Markov chain-like prior are proposed for the inclusion probabilities of genes that account for the functional similarities of genes. The aim is to select genes with functional similarity, even if they are highly correlated or coexpressed. Also, as a pan-cancer model, the regression model of different tumor types is linked so that the tumor types with small sample sizes can borrow information from other regressions and the statistical power of regressions with small sample sizes can be better enhanced. A prior is proposed to correlate the inclusion statuses of the same gene in different regressions, and the strength of correlation depends on the disease ontology semantic similarity between corresponding tumor types. The proposed method can address some obstacles facing other pan-cancer models with improved variable selection and prediction performances, as demonstrated in the simulation studies.
