BAYESIAN HIERARCHICAL LINEAR MODELS FOR DIFFERENTIAL PROTEIN EXPRESSION ANALYSIS

Sammanfattning: It is evident that the study of proteins is crucial for a deeper understanding of how drug treatments affect the body. However, differential protein expression analysis, which can be described as the method of finding which proteins are affected by a treatment, faces some major challenges. First of all, because proteomics data typically comprise several thousand different proteins for just a small number of biological tissues, there are both problems concerning multiple comparisons and low statistical power. Secondly, proteomics data are prone to suffer high rates of missing values, which could bias the results. One approach to handle these issues, which is gaining popularity, is to apply Bayesian hierarchical modeling in order to pool information from the complete dataset of all proteins when making inferences for each protein individually. Yet, in practice, there seems to be essentially only one Bayesian hierarchical model that currently is being employed, which uses a conjugate prior for the error variances but has no prior for the coefficients or the missing values. Given this, the aim of the thesis is to investigate how the model can be improved by adding priors for the coefficients and the missing values. The results show that by adding a hierarchical prior for the coefficients prediction accuracy may be increased. Furthermore, the results show that by adding a prior for the missing values differently expressed proteins can be detected that otherwise would have been overlooked.