Self-driving development of perfusion processes for monoclonal antibody production

Abstract

Abstract The development of autonomous agents in bioporcess development is crucial for advancing biopharma innovation, as it can significantly reduce the time and resources required to transition from product to process. While robotics and machine learning have greatly accelerated drug discovery and initial screening, the later stages of development have primarily benefited from experimental automation, lacking advanced computational tools for experimental planning and execution. For example, in the development of new monoclonal antibodies, the search for optimal upstream conditions (such as feeding strategy, pH, temperature, and media composition) is often conducted using sophisticated high-throughput (HT) mini-bioreactor systems, while the integration of machine learning tools for experimental design and operation in these systems have not matured accordingly. In this work, we introduce an integrated user-friendly software framework that combines a Bayesian experimental design algorithm, a cognitive digital twin of the cultivation system, and an advanced 24-parallel mini-bioreactor perfusion experimental setup. This results in an autonomous experimental machine capable of: (1) embedding existing process knowledge, (2) learning during experimentation, utilizing information from similar processes, (4) predicting future events, and (5) autonomously operating the parallel cultivation setup to achieve challenging objectives. As proof of concept, we present experimental results from 27-day-long cultivations operated by the autonomous software agent, which successfully achieved challenging goals such as increasing the viable cell volume (VCV) and maximizing the viability throughout the experiment.