ABSTRACT

Detecting direct dependencies/associations between variables in a network based on the observed data is of great importance in the studies of various biological networks and relationships. Through multiscale association analysis, we show why conditional mutual information (CMI)/partial correlation (PC) suffers from an underestimation problem for networks with strong correlations, and we further resolve this issue with a new measure, partial association (PA), derived from multiscale conditional mutual information. Linear and nonlinear versions of PA correspond to PC and CMI respectively, and they are shown to be able to accurately quantify direct dependencies or construct biological networks from both theoretical and computational viewpoints.