Biology-aware machine learning enables efficient cell culture media optimization

Cell culture is a foundational technology widely used across fields such as pharmaceutical production, regenerative medicine, food science, and materials engineering. A critical component of successful cell culture is the culture medium-a solution containing essential nutrients that support cell growth. Therefore, optimizing the culture medium for specific applications is vital. Recently, machine learning has become a powerful tool for efficient media optimization. However, the experimental data used to train such models often exhibit biological variability caused by fluctuations in cell behavior and noise from experimental procedures or equipment. This variability can significantly reduce the predictive accuracy of machine learning models.

In this study, researchers developed a machine learning model that explicitly accounts for biological variability and applied it to identify optimal formulations for serum-free culture media. CHO-K1 cells (derived from Chinese hamster ovary) were cultured in various media, and cell concentrations were measured to quantify biological variability. The researchers integrated data on medium composition, biological variability, and cell density into a machine learning framework that combined multiple algorithms. They further employed active learning-an iterative cycle of model training and experimental validation.

As a result, they successfully developed a serum-free culture medium that achieved approximately 1.6-fold higher cell density compared to commercially available products. Since the medium was specifically optimized for CHO-K1 cells, the study demonstrated the model's ability to capture the unique nutritional needs of individual cell types. These findings are expected to aid in the development of more efficient culture media for pharmaceutical manufacturing and regenerative medicine. Given that biological variability is inherent to biological experiments, the proposed approach holds broad applicability across diverse areas of biological research.

This work was supported by the JSPS KAKENHI grant numbers 21K19815 and 25K22838 (to BWY) and JP25KJ0680 (to TH).