Abstract
Gene regulatory network (GRN) inference is a challenging problem that lends itself to a learning task. Both positive and negative examples are needed to perform supervised and semi-supervised learning. However, GRN datasets include only positive examples and/or unlabeled ones. Recently a growing interest is being devoted to the generation of negative examples from unlabeled data. Within this context, the authors propose to generate potential negative examples from the set of unlabeled ones and keep those that lead to the best classification accuracy when used with positive examples. A new proposed genetic algorithm for fixed-size subset selection has been combined with a support vector machine model for this purpose. The authors assessed the performance of the proposed approach using simulated and experimental datasets. Using simulated datasets, the proposed approach outperforms the other methods in most cases and improves the performance metrics when using balanced data. Experimental datasets show that the proposed approach allows finding the optimal solution for each transcription factor in this study.