GRN topology identification using likelihood maximization and relative expression level variations

Structure identification is investigated in this paper for a gene regulatory network (GRN) using knock out/down steady state experimental data. Through incorporating sparsity of a large scale GRN, estimates are derived respectively for the wild-type expression level of a gene and the variance of its measurement errors by means of likelihood maximization. Using these estimates, relative expression level variations (RELV) of a gene are further estimated that are due to gene knock out/down experiments. An algorithm is suggested through normalizing and modifying the magnitude of this RELV to identify direct causal regulations of a GRN. Computation results with the Size 100 sub-challenges of both DREAM3 and DREAM4 show that, compared with some well known Z-score based methods, prediction performances are substantially improved by the suggested method, especially the AUPR specification. Moreover, this method can even outperform the best team of both DREAM3 and DREAM4.