A gene network depicts the inter-regulatory relations among genes. Knowledge of the gene network is key to an understanding of biological processes. A Bayesian network, consisting of nodes and directed arcs, is a convenient vehicle to model gene networks. We described a nonlinear model for the rate of gene transcription. Levels of gene expression are continuous in the model. We employed a genetic algorithm to evolve the structure of a Bayesian network. Given a candidate structure, the best parameters are estimated by the downhill simplex algorithm. The methodology features a reconstruction resolution that is limited by data noise. We tested the implementation by artificial gene networks in simulations. We then applied the methodology to reconstruct the regulation network of 27 yeast cell cycle genes from a real microarray dataset. The result obtained is promising: 17 out of the 22 reconstructed regulations are consistent with experimental findings.