Enhancing Response Surface Methodology for Bioprocesses Development Using Interpretable Machine Learning

Abstract Response surface methodology (RSM) has been a cornerstone for experimental bioprocess optimization, yet its reliance on polynomial equations often fails to capture complex nonlinearities and requires extensive experimental data to attempt capturing them. This study evaluates the enhancement of RSM through machine learning (ML) techniques, specifically Gaussian Processes and Support Vector Machines, across three biological systems: succinic acid fermentation, biobutanol production, and continuous ethanol cultivation. To mitigate the “black-box” nature of ML, explainability measures such as Mutual Information (MI) and Local Interpretable Model-agnostic Explanations (LIME) were integrated to assess feature importance and ensure biological relevance. Results demonstrate that ML models consistently outperform traditional polynomials in predictive accuracy, even when having access to less experimental data for calibration and a reduced space of features, particularly on validation data sets where errors were significantly reduced (e.g., reaching MAPE lower than 1% in simulated experimental scenarios). Furthermore, the application of Model-Based Design of Experiments using Latin Hypercube Designs achieved optimal conditions with up to 50% fewer experimental runs compared to traditional factorial approaches. These findings suggest that coupling ML with explainability tools improves the reliability of finding optimal conditions for bioprocess development while also significantly enhancing operational efficiency by reducing experimental runs, time, and costs.