Unexpected links reflect the noise in networks

Abstract

Gene regulatory networks are commonly used for modeling biological processes and revealing underlying molecular mechanisms. The reconstruction of gene regulatory networks from observational data is a challenging task, especially, considering the large number of involved players (e.g. genes) and much fewer biological replicates available for analysis. Herein, we proposed a new statistical method of estimating the number of erroneous edges that strongly enhances the commonly used inference approaches. This method is based on special relationship between correlation and causality, and allows to identify and to remove approximately half of erroneous edges. Using the mathematical model of Bayesian networks and positive correlation inequalities we established a mathematical foundation for our method. Analyzing real biological datasets, we found a strong correlation between the results of our method and the commonly used false discovery rate (FDR) technique. Furthermore, the simulation analysis demonstrates that in large networks, our new method provides a more precise estimation of the proportion of erroneous links than FDR.