Optimizing microbial cellulase fermentation with FERM-AI: risk-aware machine learning framework for fermentation and contamination prediction

Abstract

Stability, yield, and interactions between variables in the process, combined with the risk of contamination, pose a challenge for optimization of microbial fermentation processes. In this work, FERM-AI, a novel risk-aware machine learning surrogate framework for benchmark evaluation of cellulase fermentation, is presented, comprising of yield prediction, contamination risk assessment and an interactive decision-support interface. The framework was tested with a benchmark set of 1,000 runs of fermentation, consisting of 20 fermentation runs from experiments and 980 synthetic fermentation runs to develop the methods. Biologically relevant process optimality, stress responses and interaction effects were captured using domain-informed feature engineering, and features were then selected and expanded using polynomial feature expansion. Cellulase yield prediction was analysed using multiple regression models such as Ridge Regression, Bayesian Ridge Regression, Huber Regression, Partial Least Squares (PLS) Regression, Gradient Boosting, Extra Trees, XGBoost and ensemble models; and contamination risk was modelled using imbalance-aware classification with SMOTE and ensemble learning. The Huber Regressor outperformed the other algorithms with an R 2 of 0.9998, while PLS Regression performed well with an R 2 of 0.9937, further validating the strength of the proposed feature representation. The soft voting ensemble classifier was 83.5% accurate and had an AUC-ROC of 0.9045 for the contamination prediction. The integrated framework is a proof-of-concept showing the potential of simultaneously modelling productivity and contamination risks, which can aid in the evaluation of fermentation operating conditions in a risk-aware fashion. The data used in the benchmark is synthetic, so the predictive results are meant to show the validation of the proposed surrogate modelling framework, and not that the predictive ability of the framework is of industrial scale. The framework will be validated in the future with bigger fermentation datasets obtained from experiments and expanded to the real-world fermentation process optimization.