A self-driving bioprocessing platform that combines machine learning with automated bioreactors has been developed by researchers at Merck and collaborators, offering a new approach to monoclonal antibody process optimization.
The platform uses an integrated software framework to autonomously design, execute, and adapt perfusion cell culture experiments in real time. It links Bayesian experimental design algorithms with a “cognitive” digital twin of the bioreactor process, enabling continuous learning and decision-making during cultivation.
Biopharmaceutical process development remains both time- and resource-intensive. Although high-throughput bioreactors and automation have increased experimental capacity, key decisions – such as adjusting feeding strategies or perfusion conditions – still rely heavily on manual input.
To address this, the researchers integrated a hybrid modeling framework with real-time experimental control. The platform combines Gaussian process–based models with Bayesian experimental design, continuously retraining on incoming data generated during the run. This enables the system to incorporate prior knowledge, update predictions as new data are acquired, and identify optimal operating conditions without predefined experimental plans.
The system was integrated with a parallel perfusion bioreactor platform, enabling automated control of key parameters such as feed rates, bleed rates, and culture conditions. Unlike conventional workflows – which are typically pre-planned with limited mid-run adjustment – the platform dynamically modified conditions in response to real-time culture performance.
As a proof of concept, the team applied the system to a 27-day monoclonal antibody production run, with the platform operating autonomously for 20 days. Over this period, the system met predefined performance targets, including increasing viable cell volume and maintaining cell viability within specified limits, while adapting process conditions in real time.
The results also demonstrated the system’s ability to transfer knowledge between experiments and cell lines, using historical data to inform new process development campaigns. This capability could reduce the number of experiments required to reach optimal conditions, particularly in early-stage development where process understanding is limited.
Although automation is already widely used in upstream bioprocessing, most existing systems focus on executing predefined protocols rather than making higher-level decisions during cultivation. The authors note that embedding predictive models and adaptive control into these platforms could help reduce reliance on manual intervention during long and complex perfusion runs.
Such approaches may be particularly relevant as biopharmaceutical companies face increasing pressure to shorten development timelines and improve process robustness when scaling from early-stage experiments to clinical manufacturing.
Newsletters
Receive the latest pharmaceutical news, personalities, education, and career development – weekly to your inbox.
