An electronic nose was used for on-line gas phase monitoring of key metabolites in a Saccharomyces cerevisiae cultivation. The metabolites were either non-volatile or present at very low concentrations and therefore not detectable in the gas phase by the sensors in the electronic nose. It was found that it is still possible to make a prediction based on the off-gas emission. Artificial neural networks (ANNs) were trained using data acquired by the gas sensors and reference data obtained from on-line HPLC analyses, from a total of six cultivations to estimate concentrations of the metabolites glucose, glycerol, acetate and acetaldehyde. The ANNs were subsequently validated on an independent set of cultivation data resulting in a prediction accuracy described by the root mean square error (RMSE) of 0.13 (in the range 0-7.33), 0.015 (0.08-0.15), 0.012 (0-0.20) and 0.004 (0-0.11) g L-1, respectively. Data from a cultivation with higher initial glucose concentration were added to the original data and the extended set was used for training an ANN to determine concentration variables at higher concentration ranges than in the first study. The RMSE was 1.2 (0-9.31), 0.016 (0.09-0.20), 0.026 (0-0.19) and 0.010 (0-0.15) g L-1, respectively, when validating the ANNs.