PBMarsNet: A Multivariate Adaptive Regression Splines Based Method to Reconstruct Gene Regulatory Networks

Gene Regulatory Network (GRN) is a directed graph which describes the regulations between genes. The problem of reconstructing GRNs has been studied for decades. Most of existing methods infer the GRNs from gene expression data. Previous studies use random forest, partial least squares or other feature selection techniques to solve it. In this paper, we propose a Multivariate Adaptive Regression Splines (MARS) based method to estimate the feature importance and reconstruct the GRNs. MARS can catch the nonlinear relationships between genes. To avoid the overfitting and make the estimation robust, we apply an ensemble model of MARS based on bootstrap and weighted features by PMI (Part mutual information), called PBMarsNet. The results show that PBMarsNet performs better than the state-of-the-art methods.