Machine learning-based Bayesian optimization facilitates ultrafiltration process design for efficient protein purification

The ultrafiltration (UF) process, known for its mild and efficient separation capabilities, is widely utilized for protein drug purification, concentration, and buffer exchange. However, conventional UF process design and optimization require significant labor, time, and protein consumption due to the complexity of multivariate influencing factors. In this study, machine learning (ML)-based Bayesian optimization was employed to recommend tailored membrane types and operational conditions based on feed solution properties and filtration objectives. Among six trained ML models, extreme gradient boosting (XGBoost) demonstrated the best predictive performance, achieving coefficients of determination (R-2) of 0.93, 0.90, and 0.98 for rejection rate, steady flux, and permeance variations, respectively, on test datasets. Key parameters, including transmembrane pressure (TMP), protein concentration, pH, and membrane pore size, were identified as dominant factors influencing UF performance. Dynamic feature importance analysis revealed stage-specific variations in critical features affecting permeance, enabling precise process adjustments. Experimental validation confirmed the model's accuracy in predicting rejection rates and permeance variations. Finally, Bayesian optimization was applied to identify optimal operational parameters, achieving a steady flux closely aligned with predicted values (error < 0.5 %) and significantly outperforming other conditions. The ML models developed in this work offer a powerful approach for designing and optimizing UF processes with reduced research and development costs.