Nonlinear response surface model based on artificial neural networks for growth of Saccharomyces cerevisiae

Response surface models (RSMs) are used to predict microbial growth as a function of environmental conditions. This study aimed to (1) develop an artificial neural network (ANN) model for predicting the combined effect of pH, and ethanol [E] and fructose [F] concentrations on the growth of Saccharomyces cerevisiae, (2) compare the prediction accuracies of ANN and traditional RSM and (3) determine the relative importance of the three environmental factors in affecting growth of S. cerevisiae. Data used to construct the network were obtained from a literature study in which RSM were developed using linear regression techniques for estimating the growth rate constant (mu(max)). A total of 59 data sets representing the growth of S. cerevisiae in relation to pH, [E] and [F] was utilized. The agreement between experimental and ANN-predicted values for the growth parameter (mu(max)) was assessed quantitatively using three statistical descriptors: the Mean of the Absolute values of the Relative Error (MARE), coefficient-of-determination (COD) and sum-of-squared-errors (SSE). The developed ANN-based RSM predicted the mu(max) of S. cerevisiae with higher accuracy than regression-based RSM. The differentiability of the neural network provided means for studying the sensitivity of the model output (ln mu(max)) to each input parameter averaged over the entire database. The mu(max) was found to be most sensitive to pH, followed by [E] and [F] concentrations.