Modeling and optimization of bioproduct formation with purple phototrophic bacteria using machine learning

Municipal and industrial wastewater, along with organic waste, can be transformed into valuable bioproducts using purple phototrophic bacteria. This study compares the performance of three machine learning models (Random Forest, XGBoost, and CatBoost) in predicting and optimizing the formation of key bioproducts: polyhydroxybutyrate, polyhydroxyvalerate, 5-aminolevulinic acid, coenzyme Q10, carotenoids, bacteriochlorophylls, and biomass. The models were trained on a dataset compiled from previous studies, using input variables such as reaction time, concentration of organic matter, ethanol, bicarbonate, levulinic acid, ferric citrate, mineral medium, and N, C/N ratio, illumination conditions (continuous or intermittent), operation mode (batch or semicontinuous), and volume exchange percentage. Bayesian optimization was applied to train and tune the models. Performance was assessed using R², Pearson correlation, RMSE, and MAPE. CatBoost outperformed the others, showing higher predictive correlation and lower error. It was subsequently used for further optimization. Feature importance analysis identified reaction time, mineral medium concentration, and volume exchange percentage as key drivers of bioproduct synthesis. The Particle Swarm Optimization algorithm was applied to determine optimal conditions for each target compound. Under the conditions studied, predicted maximum yields were: 569 mg polyhydroxybutyrate/L, 45 mg polyhydroxyvalerate/L, 79 µmol 5-aminolevulinic acid/L, 13 mg coenzyme Q10/g dw, 7 mg carotenoids/g dw, 17 mg bacteriochlorophylls/g dw, and 2040 mg biomass/L. Optimization suggests that operating as a sequencing batch reactor and employing discontinuous illumination for most targets, along with a reduced mineral medium concentration, is beneficial. Results highlight that each bioproduct requires distinct operational settings, supporting the idea of clustering target compounds.