Media Optimization

Reducing the cost of cell culture media is a key challenge to achieving commercial scale in cellular agriculture. The media provides essential nutrients for cell growth, and it represents a significant portion of the overall production cost. There are important cost-efficiency tradeoffs that must be balanced by carefully selecting from an almost infinite number of potential components. One major additional hurdle is the removal of Fetal Bovine Serum (FBS), which has traditionally been an important component but is expensive, ethically problematic, and introduces variability and contamination risks (Xu et al., 2014; Nikkhah et al., 2023).

The papers demonstrate a technical progression in applying artificial intelligence to the problem of media optimization. One theme is the use of AI in Design of Experiment (DOE), to reduce the number of costly trials that must be performed. Early approaches focused on developing robust hybrid frameworks capable of handling high-dimensional and non-linear design spaces. For instance, Cosenza & Block (2020) combined deep learning and genetic algorithms to create a general purpose algorithm for solving complex high-dimensional problems, such as those involved in selecting media ingredients and concentratins. In their later work, Cosenza et al. (2021) utilized a genetic algorithm sequential DOE scheme for optimizing muscle cell culture media, demonstrating improved efficiency over traditional DOE methods. These efforts show the capability of integrated AI/ML techniques to efficiently explore vast parameter spaces.

Another common technique is the application of Bayesian Optimization, particularly its multi-objective variants. Multiple papers showcase Bayesian Optimization’s ability to simultaneously minimize cost and maximize growth by iteratively refining on objective functions (Cosenza et al., 2022; Cosenza et al., 2023). These approaches use inexpensive computation to narrow the search space, reducing the number of expensive experiments that must be run.

Newer techniques like Deep Neural Networks (DNN) are also starting to be used. Yoshida et al. (2023) demonstrated their applicability to predicting important properties of cells cultured in different media formulations. In other instances, active learning strategies, sometimes combined with other techniques like gradient-boosting decision trees, are shown to fine-tune media components and shorten experimental timelines by intelligently selecting data for training and validation (Hashizume et al., 2022; Zhang et al., 2023). Serum-free-media and satellite-cell differentiation datasets for cultivated beef — the kind of data these methods train on — are catalogued in Datasets/Cow.md.