Gene regulatory networks inference from gene expression data is an important problem in systems biology field, in which the main goal is to comprehend the global molecular mechanisms underlying diseases for the development of medical treatments and drugs. This problem involves the estimation of the gene dependencies and the regulatory functions governing these interactions to provide a model that explains the dataset (usually obtained from gene expression data) on which the estimation relies. However, such problem is considered an open problem, since it is difficult to obtain a satisfactory estimation of the dependencies given a very limited number of samples subject to experimental noises. Several gene networks inference methods exist in the literature, including those based on genetic algorithms, which codify whole networks as possible solutions (chromosomes). Given the huge search space of possible networks, genetic algorithms are suitable for the task, even though it is still hard to achieve good networks that explain the data by codifying whole networks as solutions. The objective of this work is the proposal of a method based on genetic algorithms to infer gene networks, whose main idea consists in applying one genetic algorithm for each gene independently, instead of applying a unique genetic algorithm to determine the whole network as usually done in the literature. Besides, the method involves the application of a network inference method to generate the initial populations to serve as more promising starting points for the genetic algorithms than random populations. To guide the genetic algorithms, we propose the use of Akaike information criterion (AIC) as fitness function. Results obtained from inference of artificial Boolean networks show that AIC correlates very well with popular topological similarity metrics even in cases with small number of samples. Besides, the benefit of applying one genetic algorithm per gene starting from initial populations defined by a network inference technique is evident according to the results. Comparative analysis involving a recently proposed genetic algorithm method for the same purpose is presented, showing that our method achieves superior performance.