(334ak) Smart Biomanufacturing: A Process Systems Engineering and Artificial Intelligence Approach

My core research interests consist of data and complex systems under the general theme of decision-making. How to make better decisions is a fundamental question for individuals, organizations, and societies, and there is no easy answer. Systems research, especially process systems engineering (PSE), with the help of artificial intelligence (AI), is the approach I plan to use to tackle some of the biomanufacturing decision-making challenges.

The production and sale of monoclonal antibody (mAb) therapeutics are a hundred-billion-dollar and expanding biopharmaceutical industry to treat cancer, immune disorder, cardiovascular disease, and inflammatory diseases. However, mAbs are currently produced commercially in fed-batch cultures using media recipes and predetermined operating protocols based primarily on heuristics. Such a heuristic-based approach often leads to suboptimal production outcomes and requires costly experimentation to improve existing protocols. On-line control of mAb productivity and product quality metrics is not widely adopted for several reasons. For instance, measurement of critical quality attributes (CQAs) such as glycan distribution (the result of glycosylation) is available only after the process ends, making it impossible to adopt conventional feedback control. Cell growth, metabolism, antibody formation, and glycosylation are complex, multi-scale dynamics that are difficult to model. While currently available models can predict mAb synthesis or glycosylation dynamics accurately under specific process configurations, adapting them to new processes remains challenging.

It is tempting to contemplate adopting a purely data-driven approach; however, the unique challenges associated with modeling biopharmaceutical processes make such an approach less likely to succeed. On the one hand, the biopharmaceutical industry is not data-rich because it is expensive to generate data, and existing data are usually sparse. As a result, models trained using these data may not extrapolate well outside the training data. On the other hand, domain knowledge and mechanistic understanding of biochemical processes can compensate for the lack of data. These two considerations argue for a hybrid approach to the development of a model that is suitable for predicting and for model-based control of CQAs. Meanwhile, PSE techniques such as Kalman filtering and AI techniques such as artificial neural networks can be used to estimate CQAs based on other measurements and signals available on-line. The development of high-fidelity, hybrid, adaptable models and state estimation techniques allows us to apply existing PSE tools to design more efficient biomanufacturing processes and to control them.