BACKGROUND: Production of microbial enzymes in bioreactors is a complex process including such phenomena as metabolic networks and mass transport resistances. The use of neural networks (NNs) to infer the state of bioreactors may be an interesting option that may handle the nonlinear dynamics of biomass growth and protein production. RESULTS: Feedforward multilayer perceptron (MLP) NNs were used for identification of the cultivation phase of Bacillus megaterium to produce the enzyme penicillin G acylase (EC. 3.5.1.11). The following variables were used as input to the net: run time and carbon dioxide concentration in the exhausted gas. The NN output associates a numerical value to the metabolic state of the cultivation, close to 0 during the lag phase, close to 1 during the exponential phase and approximately 2 for the stationary phase. This is a non-conventional approach for pattern recognition. During the exponential phase, another MLP was used to infer cellular concentration. Time, carbon dioxide concentration and stirrer speed form an integrated net input vector. Cellular concentrations provided by the NN were used in a hybrid approach to estimate product concentrations of the enzyme. The model employed a first-order approximation. CONCLUSION: Results showed that the algorithm was able to infer accurate values of cellular and product concentrations up to the end of the exponential growth phase, where an industrial run should stop. (c) 2008 Society of Chemical Industry.