A discretized differential evolution algorithm for the inference of Gene Regulatory Networks

Feature selection is a very important procedure in the pattern recognition research field. Feature selection algorithms are particularly important in the inference of Gene Regulatory Networks (GRNs) from its gene expression profiles, which usually involves data with a large number of variables and small number of observations. This work presents the inference of GRNs from two different feature selection approaches: classical (deterministic) SFS and SFFS search algorithms, and (meta-heuristic) Discretized Differential Evolution (DDE) algorithm. Several Artificial Gene Networks (AGNs) were probabilistically generated for testing. They had different sizes and topologies following three models: Erdös-Rényi uniformly-random, Watts-Strogatz small-world, and Barabási-Albert scale-free. Results confirm that, as the number of candidate predicting genes increase, the performance of the compared methods decrease, revealing the exponential characteristic of the problem. Notwithstanding, for all network sizes and topologies studies, the DDE algorithm achieved better results than the other approaches, especially for GRNs with hubs.