Determining the most suitable correlation estimators for gene network inference algorithms

Gene network inference (GNI) algorithms have a significant role in bioinformatics research area. Those algorithms provide us to explore the vast amount of the interactions among the molecules in the cell. Application areas of GNI algorithms are very wide; such as discovering a genome-wide interactions and associations among the genes and gene-products (e.g. proteins); determining the main target of a drug in pharmacological studies; observing the interaction of the regulator and regulated genes, and so on. GNI or in other words reverse engineering of gene networks is the process of predicting the interactions of the genes by using microarray gene expression values. GNI process is a challenging process because of the current very large-scale biological datasets and the noise caused by the experimental and computational processes. In almost all GNI algorithms the main process is to estimate the association scores among the variables of the dataset. If this step is not correctly fulfilled then the ultimate inference process becomes erroneous for whichever the GNI algorithm is used. Therefore, this is the most crucial process of any GNI algorithms. Nonetheless, there is no commonly accepted estimator to compute association scores that is used with the current GNI algorithms. In this study, we investigate almost all of the available estimators that might be used in the GNI algorithms. Estimators are classified according to different point of views and presented to the researchers by this study. Among the reviewed estimators, we determine the most suitable and the best performing estimators for various GNI algorithms. We achieved this by comparing the inference performance of the estimators over different GNI algorithms, which are selected as representatives of many others. For that purpose, from the literature, we first reviewed correlation-based estimators and then entropy-based and finally direct MI estimator approaches to estimate Mutual Information; in the context whether they are used in genomics datasets or not. So far we reviewed 27 different estimators and identified 12 of the promising methods and their derivatives for effective usage in GNI. We expect this study to assist many researchers before using those estimators in their own GNI studies. We compared the estimators on three popular GNI algorithms; those are called Accurate Cellular Networks (ARACNE), Relevance Networks (RelNet) and Conservative Causal Core Network (C3NET). At the end of the study, we prepared an R software package which includes the determined estimators. Since we provide the most prominent estimators out of the study, any of the algorithm that wants to use our proposed estimators, needs to reassess the inference performance before replacing their estimators. Ultimately, this study aimed to be a reference guide for the usage of the estimators in GNI algorithms.