Exploiting General Independence Criteria For Network Inference

Abstract

Inference of networks representing dependency relationships is a key tool for understanding data derived from biological systems. It has been shown that nonlinear relationships and non-Gaussian noise aid detection of directions of functional dependencies. In this study we explore how far generalised independence criteria for statistical independence proposed in the literature are better suited to the inference of networks compared to standard independence criteria based on linear relationships and Gaussian noise. We compare such criteria within the framework of the PC algorithm, a popular network inference algorithm for directed acyclic dependency graphs. We also propose and evaluate a method to apply unconditional independence criteria to assess conditional independence and a method to simulate data with desired properties from experimental data. Our main finding is that a recently proposed criterion based on distance covariance performs well compared to other independence criteria in terms of error rates, speed of computation, and need of fine-tuning parameters when applied to experimental biological datasets.