A machine learning approach to identify active polysorbate 20 degrading hydrolases in biopharmaceutical formulations

Polysorbate degradation by host cell-derived hydrolases presents a critical challenge in biopharmaceutical formulations. It can lead to fatty acid release, particle formation and reduced product stability. Mass spectrometry-based host cell protein (HCP) analysis is widely used for HCP identification, but detection becomes challenging in formulations where monoclonal antibodies are present in large excess. In such cases, hydrolases can remain undetected, despite being enzymatically active at trace levels. In this study, we demonstrate that individual CHO-derived hydrolases generate distinct polysorbate degradation fingerprints that can be detected by reverse phase ultra performance liquid chromatography coupled to mass spectrometry (RP-UPLC-MS) and classified using supervised machine learning. Models were trained on single time point fingerprints comprising approximately 50 measurements for five hydrolases (CES1F, CES2C, LPLA2, PPT1 and PAF-AH). Evaluated algorithms included Logistic Regression, Random Forest, Gradient Boosting, Support Vector Classifier, AdaBoost, and Artificial Neural Networks. Seven out of eight models achieved 100 % accuracy on the test set, confirming that enzyme-specific information is preserved in single measurements in the presence of individual enzymes, independent of enzyme concentration or degradation time. External validation using an independently prepared hydrolase spike sample confirmed the robustness of the models. Prediction confidence was high at early degradation stages and decreased at late stages, as enzyme-specific degradation fingerprints became more similar. This work presents an activity-based classification framework for the functional identification of polysorbate degrading hydrolases. The approach supports downstream monitoring and risk-based mitigation strategies by identifying the enzymes that drive polysorbate hydrolysis under formulation conditions.