Deep Learning-Based Prediction of Critical Parameters in CHO Cell Culture Process and Its Application in Monoclonal Antibody Production

Abstract

This study presents a deep learning method for predicting values in the CHO cell culture process for the production of monoclonal antibodies. The developed hybrid architecture combines convolutional neural networks (CNN) and short-term temporal (LSTM) networks to capture both spatial and temporal aspects of bioprocess data. The model was trained and validated using data collected from donor cultures, including 167 unsupervised processes over a 14-day cultivation period. Feature selection and engineering methods were used to identify critical parameters, while Bayesian optimisation was employed for hyperparameter tuning. The model achieves the best prediction with an R2 score of 0.956 and an RMSE of 0.082, demonstrating significant improvement over conventional models. Implementing the framework led to several improvements in process efficiency, including a 28.1% increase in product titer and a 39.5% reduction in variable costs. The model maintains good performance across different tasks, with exceptional results in predicting metabolic rate (R² > 0.932) and cell density (R² > 0.945). The ability of real-time forecasting leads to process control, resulting in a 19.1% improvement in overall process yield. This framework provides a robust framework for using expertise in bioprocess management, providing solutions for improving product quality and process performance in the biopharmaceutical manufacturing industry.