On-line measurement of intra-cellular variables in a bioreactor is difficult through methods using instrumentation. However, close monitoring of these variables is essential in recombinant fermentations, which are sensitive to disturbances and spatial heterogeneity. Based on an earlier study, a recurrent neural network of the Elman type was applied to a fed-batch fermentation for beta -galactosidase by a temperature-sensitive strain of Escherichia coli. Simulated data of four intra-cellular concentrations in an imperfectly mixed bioreactor were used to train the network, and its predictive capability was tested with unseen data of the rDNA and beta -galactosidase concentrations. Over a wide range of mixing intensity and a fermentation period longer than normally required, the Elman network's performance was better than that reported for modified feedforward networks and it was unaffected by the hydrodynamics. This type of neural network may thus be employed as an online soft sensor for large scale recombinant fermentations.