Hide/Show Apps

Novel model selection criteria on high dimensionalbiological networks

Bülbül, Gül Baha
Gaussian graphical model (GGM) is an useful tool to describe the undirected associ-ations among the genes in the sparse biological network. To infer such high dimen-sional biological networks, thel1-penalized maximum-likelihood estimation methodis used. This approach performs a variable selection procedure by using a regular-ization parameter which controls the sparsity in the network. Thus, a selection ofthe regularization parameter becomes crucial to define the true interactions in the bi-ological networks. In this sense, we suggest to combine some information-theoreticmeasures such as CAIC, CAICF and ICOMP with a penalized likelihood approachin order to yield the true graph. Also, loop-based multivariate adaptive regressionsplines (LMARS) can be presented as a nonparametric modelling technique which isgood at dealing with the problem of nonlinearity and collinearity in the data whichthe problems arise from high-dimensional networks. In this study, we interfere themodel selection procedure of LMARS by applying our measures to find the correctstructure, while it has been originally introduced with generalized cross validation asa model selection technique.