A consensus approach to predict regulatory interactions

Exploiting microarray gene expression data to predict regulatory interactions has become a key challenge in recent years, for which many network inference algorithms have been developed. Combining predictions of multiple algorithms qualitatively to produce a consensus network has been previously implemented. Here, we propose a quantitative consensus approach based on combining regulatory interactions using the Fisher's combined probability test. Edge significance values of different network inference algorithms were combined statistically to determine whether the edges should be included in a resulting consensus network. We validated and tested our approach with a variety of benchmark datasets, including data from the DREAM4 challenge. We have evaluated our algorithm against static and dynamic Bayesian networks and other individual networking methods. The results demonstrate that consensus networks predict many biological interactions with higher performance measures and outperform individual methods. We conclude that consensus networks are more robust and provide high confidence to predict regulatory interactions.