Abstract:
|
In genome-wide network analyses, it is of interest to estimate the mean differential expression level and also predict the differential expression level of a gene. We incorporate the network information via a covariance matrix based on an exponential diffusion kernel. We start with the sample covariance matrix based on expression data available on several other genes, imposing a positivity condition, and use it to estimate the diffusion parameter. We then use an empirical Bayes approach to make the desired inference about the expression levels of a gene. We benchmark our method by validating the estimated diffusion parameter using predictions on L1000 data on the HumanNet graph. We also compare our method to methods which use a covariance structure based only on the network such as one based on a random walk on the network, and evaluate the advantage of using a covariance matrix based on both the expression data and the network structure.
|