Sequential active learning for medium optimization in mAb production

Monoclonal antibodies (mAbs) are key therapeutics for diseases like cancer and autoimmunity. The production of mAbs relies on cell culture, in which the culture medium for high productivity and activity is essential. Despite the traditional manual and advanced computational methodologies for medium optimization, it remains challenging to incorporate biological insights gained during cell culture experimentation into the optimization process. To address this issue, an active learning strategy that sequentially integrates machine learning predictions with experimental observations of biological meaningfulness was developed in the present study. Medium design and prediction were conducted with the combination of the design of experiment and two different machine learning models, to optimize the culture medium for Chinese hamster ovary (CHO) cells producing increased immunoglobulin G (IgG) titer. Using this approach, we iteratively adjusted the concentrations of 44 components in a serum-free medium and achieved a significant improvement in IgG monoclonal antibody production. Biological insights such as osmolality control and amino acid composition, which were not initially considered, were progressively incorporated into the data-driven optimization process. The proposed strategy is practical and effective, even under limited experimental resources, and offers a new direction for rational medium design in biopharmaceutical manufacturing.