Comparison of gene network inference algorithms with ROC and PR analysis

Abstract

Understanding the network of interactions between the genes of a single cell would lead to considerable advances in biotechnologies.<br />Hundreds of methods have already been proposed in order to mine the data generated by time-series microarray experiments and infer the underlying gene interaction network, but none has yet reached an acceptable level of reliability, and none seems to prevail. In this thesis we propose a procedure to rigorously evaluate and compare these methods. We first focus the treatment of experimental data and the design of benchmarking experiments, then introduce ROC and PR analysis, and discuss the comparability of the network inference algorithms.<br />As an application we use a dataset from the BOKU Vienna to benchmark some algorithms proposed for the language R, in particular those relying on mutual information.<br />We show that the definition of a class of problems is a critical step for the comparative benchmarking of these algorithms.