Machine learning-accelerated 1H NMR quantification for bioprocess metabolite monitoring in monoclonal antibody production.

Monitoring of metabolite dynamics during cell culture process is crucial for ensuring consistent monoclonal antibody (mAb) yield and quality. While proton nuclear magnetic resonance (1H NMR) offers an ideal solution for process monitoring due to its rapid, non-destructive, and inherently quantitative nature, traditional manual spectral analysis methods are inefficient and heavily reliant on expert knowledge, limiting their industrial application. This study proposes a novel NMR spectral analysis strategy based on machine learning, successfully developing rapid quantitative models for 20 key metabolites in adalimumab biosimilar production process. By comprehensively comparing the prediction performance of partial least squares regression (PLSR), ridge regression, and light gradient boosting machine (LightGBM), we found that ridge regression and LightGBM outperformed PLSR in terms of prediction accuracy and feature interpretability. Although LightGBM showed slightly better accuracy for 11 metabolites, ridge regression demonstrated distinct advantages in quantifying key metabolites such as glucose, lactate, and most amino acids, making it more suitable for bioprocess monitoring. The obtained optimal models for each metabolite exhibited great model performance with regression coefficients (R2) above 0.92 and residual prediction deviation (RPD) values exceeding 3.6. This study provides a reliable tool for intelligent monitoring of mAb production process, which can help ensure bioprocess consistency, improve yield and product quality, and ultimately reduce cultivation costs.