Comparing a polynomial DOE model and an ANN model for enhanced geranyl cinnamate biosynthesis with Novozym® 435 lipase

Central Composite Rotatable Design (CCRD), a type of factorial design of experiment is among the most traditional methods used for optimizing bioprocesses, but, in recent years artificial neural networks (ANNs) have emerged as a promising approach for data modeling in bioprocesses. A comparative study between CCRD and ANN modeling for data treatment in the optimization of geranyl cinnamate biosynthesis using Novozym® 435 lipase was conducted. The most effective ANN architecture identified for predicting and maximizing the enzymatic synthesis of geranyl cinnamate was a 3-3-1 neurons model utilizing logsig activation function in the hidden layer. The ANN was trained using all available experimental data and demonstrated a strong fit to the experimental data, coefficient of determination near 1 (R2 = 0.9948) and low Sum-squared Error (SSE = 43.06). The polynomial model (CCRD; R2 = 0.9806; SSE = 161.56) indicated the same optimal conditions as the ANN model, predicting a temperature of 90.2°C, molar ratio of 1:5.68, and enzyme concentration of 18.6% w/w. Comparison of R2 and SSE values due to lack of fit between both models suggests that ANN predictions closely align with experimental conversion values of geranyl cinnamate. Although the polynomial model is feasible to be applied in enzymatic synthesis, it may be less precise in its predictions than ANN models.