Gene regulatory network inference from expression and genetic data using tree-based methods

One of the pressing open problems of computational systems biology is the elucidation of the topology of genetic regulatory networks (GRNs). In our previous work, we have proposed a method, called GENIE3, for the unsupervised inference of gene regulatory networks from expression data. This method decomposes the prediction of a regulatory network between p genes into p different regression problems and solves each of these problems using tree-based ensemble methods such as Random Forests. After a brief presentation of the method and a discussion of its performance when applied on steady-state and time series expression data, we will discuss its adaptation in the context of systems genetics data.
The idea of systems genetics is to exploit the natural variations that exist between the DNA sequences of related individuals and that can represent the randomized and multifactorial perturbations necessary to recover GRNs. We propose two new methods, called GENIE3-SG-joint and GENIE3- SG-sep, that incorporate information about genetic markers into the original GENIE3 method. The first one builds a single joint model incorporating both expression and genetic markers, while the second one builds two separate models and aggregate their predictions a posteriori. Experiments on the artificial data of the DREAM5 Systems Genetics challenge and of the more recent StatSeq benchmark show that both methods can benefit strongly from genetic data, with however a significant advantage to the GENIE3-SG-sep method, which outperforms by a large extent the official best performing method of the DREAM5 challenge data.