This paper illustrates the potentiality of Bayesian neural networks to model the concentration of the antibiotic cephalosporin in a fermentator from the estimate of the control variables. We show that our models give satisfactory results together with an estimate of the uncertainty associated to each prediction, allowing a potential operator to deal with anomalies during the process. The determination of the relevance of the input allows also having a better understanding of the quality of the feature vector fed in to the network as well as an interpretation of the physics underlying the biochemical process.