Mutual information estimation for transcriptional regulatory network inference

Mutual information-based network inference algorithms are an important tool in the reverse-engineering of transcriptional regulatory networks, but all rely on estimates of the mutual information between the expression of pairs of genes. Various methods exist to compute estimates of the mutual information, but none have been firmly established as optimal for network inference. The performance of 9 mutual information estimation methods are compared using three popular network inference algorithms: CLR, MRNET and ARACNE. The performance of the estimators is compared on one synthetic and two real datasets. For estimators that discretise data, the effect of discretisation parameters are also studied in detail. Implementations of 5 estimators are provided in parallelised C++ with an R interface. These are faster than alternative implementations, with reductions in computation time up to a factor of 3,500. Results The B-spline estimator consistently performs well on real and synthetic datasets. CLR was found to be the best performing inference algorithm, corroborating previous results indicating that it is the state of the art mutual inference algorithm. It is also found to be robust to the mutual information estimation method and their parameters. Furthermore, when using an estimator that discretises expression data, using N 1 / 3 bins for N samples gives the most accurate inferred network. This contradicts previous findings that suggested using N 1 / 2 bins.