Bayesian Hierarchical Models for Proteomic-Pathway Analysis

Sammanfattning

Contemporary techniques for estimating the effect of a treatment on biological pathways use a two-step method which fails to capture the complex interdependencies of biological systems. We propose a fully Bayesian approach which can account for these intricacies, by using hierarchical modelling and modern probabilistic computational methods, in order to draw more robust and credible inferences. Using proteomics data collected from brains of mice treated with either cocaine or saline, we specify regression-based models for observed protein intensities. The selected model uses a regularised horseshoe shrinkage prior on treatment effects, and hierarchically accounts for variation across proteins, mice, and brain regions. The hierarchical design is shown to improve predictive performance through partial pooling, and the shrinkage prior allows us to single out the pathways most affected by treatment. Compared to a two-step method, our approach finds smaller and more region-specific pathways to be most affected by treatment. Additionally, we can quantify the treatment effect on each pathway, and find that the proteins in the strongest and most credibly affected pathways have absolute intensity values shifted by up to one standard deviation due to treatment.