Machine Learning-Based Optimization for Enhanced Pullulan Recovery from Aureobasidium pullulans

The solvent extraction protocol significantly influences the yield and purity of the exopolysaccharide recovery from cell-free broth. In this study, methanol, ethanol, isopropanol, acetone, and PEG-6000 were compared for the amount of pullulan recovery (PR), sucrose equivalent (SE), and protein impurity (PI) of the precipitate. The PR (7.0 g L-1) and SE (0.45 g g(-1)) from acetone had been significantly better than others. Quadratic and GA-optimized artificial neural network (ANN) models, developed from a central composite design, accurately predicted PR (R-2: 0.996-0.998), SE (R-2: 0.961-0.985), and PI (R-2: 0.952-0.984) based on the pH, incubation time, and solvent-to-broth volume (S/B) ratio. Individually optimized PR (10.53 g L-1), SE (0.64 g g(-1)), and PI (1.05 mg g(-1)) from quadratic models are comparable to those obtained from ANN models. Multiobjective optimization with equal weighting suggests a moderate PR (quadratic model: 8.14 g L-1, ANN: 6.77 g L-1) while maintaining SE and PI close to their optimal values. However, the relative importance of these objectives should be ascertained based on the intended application.