Inferring gene regulatory networks from data requires the development of algorithms devoted to structure extraction. When time-course data is available, gene interactions may be modeled by a Bayesian Network (BN). Given a structure, that models the conditional independence between genes, we can tune the parameters in a way that maximize the likelihood of the observed data. The structure that best fit the observed data reflects the real gene network's connections. Well known learning algorithms (greedy search and simulated annealing) devoted to BN structure learning have been used in literature. We enhanced the fundamental step of structure learning by means of a classical evolutionary algorithm, named GA (Genetic algorithm), to evolve a set of candidate BN structures and found the model that best fits data, without prior knowledge of such structure. In the context of genetic algorithms, we proposed various initialization and evolutionary strategies suitable for the task. We tested our choices using simulated data drawn from a gene simulator, which has been used in the literature for benchmarking [Yu et al. (2002)]. We assessed the inferred models against this reference, calculating the performance indicators used for network reconstruction. The performances of the different evolutionary algorithms have been compared against the traditional search algorithms used so far (greedy search and simulated annealing). Finally we individuated as best candidate an evolutionary approach enhanced by Crossover-Two Point and Selection Roulette Wheel for the learning of gene regulatory networks with BN. We show that this approach outperforms classical structure learning methods in elucidating the original model of the simulated dataset. Finally we tested the GA approach on a real dataset where it reach 62% of recovered connections (sensitivity) and 64% of direct connections (precision), outperforming the other algorithms.