Machine learning-enhanced ¹NMR for rapid amino acid quantification in monoclonal antibody production

In monoclonal antibody (mAb) production, fluctuations in amino acid supply significantly impact cell growth, productivity, and product quality. Quantitative proton nuclear magnetic resonance (¹H NMR) -based amino acid monitoring has been widely adopted during mAb production due to its non-destructive nature, simple sample preparation, and high reproducibility. However, the complexity of the composition of biological samples leads to the fact that their quantitative analysis by hydrogen spectroscopy requires more time for analysts. The breakthrough of machine learning (ML) technology provides a new idea to solve the above bottleneck problem. Considering the important role of amino acids in cellular metabolism, this study establishes a general quantitative model with excellent generalization ability based on ¹H NMR spectra for analysis of 12 key amino acids across diverse mAb processes. Initially, three individual ¹H NMR datasets from distinct bioprocesses were modeled separately using PLSR, SVR, and Lasso Regression, with investigation of potential model performance improvement through whale optimization algorithm (WOA)-based feature selection. Model performance was evaluated comprehensively using regression coefficients (R), root mean square error (RMSE), and residual prediction deviation (RPD), and the WOA-Lasso combination consistently outperformed other methods. Employing this algorithmic framework, we constructed a generalized model for mixed biomanufacturing processes that achieved desirable prediction performance on the validation set, with R values ≥ 0.90, RPD values ≥ 2.3, and RMSE ranging from 0.244 to 1.173. This study innovatively developed a highly efficient and generalized rapid quantitative ¹H NMR method for amino acids using ML, demonstrating significant application value for optimizing biopharmaceutical processes and enhancing product quality.