Biology-aware machine learning enables efficient cell culture media optimization

Cell culture is a basic technology widely used in fields such as pharmaceutical production, regenerative medicine, food science, and materials engineering. A key ingredient in successful cell culture is the Culture Media A solution containing essential nutrients that support cell growth. Therefore, it is essential to optimize the medium for a particular application. Recently, machine learning has become a powerful tool for efficient media optimization. However, the experimental data used to train such models often show biological variability generated by variations in cell behavior and noise from the experimental procedure or instrument. This variability can significantly reduce the prediction accuracy of machine learning models.

In this study, the researchers developed a machine learning model that explicitly describes biological variability and applied it to identify the optimal formulation of serum-free media. CHO-K1 cells (derived from Chinese hamster ovaries) were cultured in various media and cell concentrations were measured to quantify biological variability. The researchers integrated data on medium composition, biological variation and cell density into a machine learning framework that combines multiple algorithms. They further adopted active learning in iterative cycles of model training and experimental validation.

As a result, they have successfully developed serum-free media that achieves approximately 1.6 times higher cell density than commercial products. Because the medium was specifically optimized for CHO-K1 cells, this study demonstrated the ability of the model to capture the unique nutritional needs of individual cell types. These findings are expected to help develop more efficient media for pharmaceutical manufacturing and regenerative medicine. Given that biological variation is inherent to biological experiments, the proposed approach retains wide applicability across a diverse range of biological research.

This work was supported by JSPS Kakenhi Grant Numbers 21K19815 and 25K22838 (to BWY) and JP25KJ0680 (Th).