Cell culture media development approach using automation and machine learning

Here, we incorporate high-throughput cell culture, automation, and machine learning methodologies to improve media development. We sought to apply this method of media optimization as a wider scope alternative to complement Design of Experiments (DoE) focused rational media design. Furthermore, this workflow can identify important and impactful components early on during media development. We utilize CHO cell culture to demonstrate our process, but this process is being applied to workflows involving a number of cell types, adherent and suspension alike.

Media development is a time intensive process that can include iterative DoE and raw material evaluations. This improved process provides crucial insight into optimal formulations. Traditionally performed at the shake flask scale for CHO media, experiment scale is limited by the number of flasks that a scientist can manage at one time. Importantly, it is critical to understand the relationship between growth models in high throughput and flask formats as we interpret our data. We hypothesize that media screening can support the traditional media development workflow by generating more data, identifying key components affecting outcomes and synergies between components, and reducing the overall timelines of a media development project.

Automated liquid handling techniques were developed and applied to carry out high-throughput culture and performance evaluation of CHO cells. First, scale-down models in 96 or 24 deep well plates were developed to produce correlative results. By scaling down from flask culture, we aimed to screen hundreds of media in one experiment. Unique media are created by blending a panel of diverse base media, each with varied formulations. From 12 media, we created over 500 unique media. This is at least a 10-fold increase in the number of media than can be performed by one FTE with flasks.

Once a media screen is complete, top media are ranked and further tested. Moreover, by using machine learning, we were able to predict unique formulations that were not part of the initial screen. We first fit the media screen data into regression models that well predicted cell growth and productivity. The models were then adopted to computationally screen the media at large scale and predict candidate top media. In this way, the search space was further expanded by a thousand-fold. At the same time, as it is embedded in the models, we were able to glean insight into the specific impacts of components and synergies. If desired, follow-up experiments that focus on the most important components can be further optimized for increased gains.

Overall, this approach of media screening can expand the design space to identify unique media component interactions while significantly decreasing the FTE time and material costs.