Dynamic optimization of monoclonal antibody production through reinforcement learning on hybrid models

Monoclonal antibodies (mAbs) are biopharmaceuticals primarily used for the treatment of severe diseases such as, among others, cancer. Their industrial production is carried out through fed batch cultures of mammalian cells, which are strongly influenced by process operating conditions, such as the nutrient feeding strategy. Identifying the optimal feeding policy requires a large experimental effort and long timelines, drastically increasing the total drug development cost and time-to-market. However, digital models and autonomous policy optimization techniques are extremely beneficial to limit the required experimental burden and reduce development costs and timelines. In this Thesis, the optimal nutrient feeding policy is identified through reinforcement learning and a hybrid model of cell culture for the production of mAbs. Specifically, the reinforcement learning agent learns the optimal feeding policy by interacting with the hybrid process model. The proposed solution is tested on a simulated process developed by (Kontoravdi et al., 2010). The identified optimal feeding policy resulted in a 32.8% increase in the mAbs concentration at harvest (from 1579 mg/L to 2097 mg/L) compared with the ten (10) historical cultures used for model training. The large mAbs increase and the low data requirement prove that this approach represents a viable strategy to identify improved operating strategies while reducing the need for extensive experimental campaigns.