Integrative Structural Learning of Mixed Graphical Models via Pseudo-likelihood

Markov random field is a common tool to characterize interactions among a fixed collection of variables. In recent biomedical research, there arise new concerns about the discovery of regulatory and co-expression relationships among different types of features across multiple biological classes. Consequently, we propose a data integration framework to jointly learn multiple mixed graphical models simultaneously. To address the common asymmetry problem in neighborhood selection, we construct a new estimator using regularized pseudo-likelihood, which produces symmetric and consistent estimates of network topologies. We demonstrate the practical merits of our method through learning synthetic networks as well as constructing gene regulatory networks from TCGA data.