Machine learning-based optimization of a multi-step ion exchange chromatography for ternary protein separation

Ion-exchange chromatography is an essential but complicated step in the biopharmaceutical downstream process, with multiple factors affecting the separation efficiency. Model-based optimization can help expedite process developments with limited time and resource investments. To address the mechanistic model's high computational complexity, a machine learning (ML)-based optimization framework was introduced. Specifically, Gaussian Process Regression models were utilized as substitutes for the mechanistic model in calculating the constraints and objective function. To further reduce the required sampling, feasibility and optimization stages were incorporated into the framework, with a penalty introduced at each stage to guide the search. This MLbased framework was applied to a case study to separate a protein mixture. Compared to the results obtained through genetic algorithm, this approach enhanced productivity by 50.1 % and reduced computation time by 70.8 % simultaneously. The effects of peak cutting criteria on the optimal results were examined, followed by a detailed analysis of design space.