Bayseian genomic models for the incorporation of pathway topology knowledge into association studies

Pathway topology and relationships between genes have the potential to provide information for modeling effects of mRNA gene expression on complex traits. For example, researchers may wish to incorporate the prior belief that "hub" genes (genes with many neighbors) are more likely to influence the trait. In this paper, we propose and compare six Bayesian pathway-based prior models to incorporate pathway topology information into association analyses. Including prior information regarding the relationships among genes in a pathway was effective in somewhat improving detection rates for genes associated with complex traits. Through an extensive set of simulations, we found that when hub (central) effects are expected, the diagonal degree model is preferred; when spoke (edge) effects are expected, the spatial power model is preferred. When there is no prior knowledge about the location of the effect genes in the pathway (e.g., hub versus spoke model), it is worthwhile to apply multiple models, as the model with the best DIC is not always the one with the best detection rate. We also applied the models to pharmacogenomic studies for the drugs gemcitabine and 6-mercaptopurine and found that the diagonal degree model identified an association between 6-mercaptopurine response and expression of the gene SLC28A3, which was not detectable using the model including no pathway information. These results demonstrate the value of incorporating pathway information into association analyses.