Recurrent Neural Networks for Forecasting Time-Varying Process Behavior in Mammalian Cell Culture

Abstract

Cell culture processes are the workhorse for the production of therapeutic protein biologics. With advances in process data acquisition and monitoring, there has been an increasing interest in developing cell culture process models for control, optimization, and scale-up. However, the kinetic behavior of cell culture processes is highly complex. As culture time progresses, cell metabolism may shift, and at times, similar culture conditions may give rise to very different time-varying process behavior. Hence, modeling complex metabolic shifts in biomanufacturing processes remains a major challenge. In this work, we systematically evaluated the application of recurrent neural networks (RNNs) for forecasting the time profiles of key process parameters, including glucose and lactate concentrations, viable cell density, and viability, using a comprehensive set of fed-batch biomanufacturing data. We compared the RNNs’ performance with that of traditional machine learning models and feedforward neural networks, included total base addition in the model input to embed secondary process information, and extended the RNN model to an encoder–decoder model that leverages the history of seed train profiles to enhance the prediction of process behavior at the production scale. Overall, the computational results highlight the potential of RNN-based models for predicting key process parameters in cell culture and demonstrate the impact of process history on cell culture performance in biologics biomanufacturing.