Comparative Study of Mutual Information based Algorithms for the Inference of Gene Regulatory Networks using Transcriptomic Data

Resumo

The inference of gene regulatory networks from genome-wide expression data provides valuable insight to understand biological systems. Several learning algorithms based on Mutual Information (MI) measures have been applied to this aim with encouraging results. Two-way MI methods such as ARACNE or CLR focus on the discovery of direct regulatory interactions and are not able to detect more complex relationships between gene regulators and targets. Other methods have been proposed to address this issue based on three-way mutual information, sinergy or partial correlation indexes. They are able to identify more complex relationships between genes, allowing discovery of causal relationships. However, the performance and properties of the different algorithms remain difficult to assess in the context of real biological problems.We present an extensive empirical study of several MI learning algorithms for the inference of gene regulatory networks. To this aim we have considered two public available expression datasets (E. coli regulators and MYC oncogene regulation) for which the regulatory interactions are well studied. The algorithms have been compared considering several objective measures. Additionally, a biological analysis of the cofactors and the regulatory interactions have provided a deeper understanding of different methods.The experimental results suggest that the specific MI estimator considered does not have great impact on the results, though the best estimator is based on the empirical distribution. The results also suggest that Three-Way Mutual Information helps to discover more complex regulatory mechanisms that involve more than two genes and helps to reduce false positive and negative interactions