Machine Learning-Assisted Optimization of Surfactin Titer Produced by Bacillus subtilis MTCC 2423 from Minimum Fermentation Data with Experimental Validation

Optimizing surfactin production from a wild-type Bacillus sp., is challenging due to complex microbial growth kinetics, nonlinear interactions among process variables, and the lack of mechanistic models to describe surfactin production. Therefore, the present study develops data-driven machine learning models to optimize surfactin production from Bacillus subtilis MTCC 2423 under limited fermentation data. First, 25 batch fermentation experiments were performed to obtain surfactin titers under varied conditions of temperature, agitation speed, and inoculum volume, generated using a central composite design (CCD) integrated with maximin Latin hypercube sampling (LHS). Subsequently, machine learning models, viz, multivariable least-squares regression (MVLSR), Gaussian process regression (GPR), support vector regression (SVR), and random forest (RF), were trained and evaluated using the nested leave-one-out cross-validation (LOOCV) to achieve robust prediction of the surfactin titer. Hyperparameter tuning of the models was performed using a genetic algorithm (GA) and grid search. Both the GPR and SVR outperformed the other models and demonstrated comparable predictive performance on the unseen data, producing the LOOCV R2 of 0.89 and 0.86, and LOOCV RMSE of 0.6 and 0.67, respectively. Subsequently, the mean response surface plots were constructed using the GPR model to illustrate the effects of nonlinear interactions among the input features on the surfactin titer. Finally, a GPR-based constrained optimization yielded an optimal surfactin titer of 5.72 g L–1, compared to the base-case value of 1.49 g L–1, while identifying the optimal process variables. The data-driven optimization results closely agreed with the experimentally validated surfactin titer of 5.75 ± 0.09 g L–1, demonstrating high model reliability and predictive accuracy. Our experimental design approach, integrated with machine-learning-driven optimization using limited experimental data, can be adopted across various bioprocess sectors to achieve substantial improvements in surfactin titer.