Build an ML-guided platform for optimizing cell culture media, explicitly accounting for biological variability and experimental noise.

Use biology-aware active learning to overcome limitations of traditional ML in biological experiments.

Reformulate a complex 57-component serum-free medium for human cell culture.

Provide a robust framework that bridges computational modeling and wet-lab experimentation.

Cell culture technologies are widely used in academia and industry, yet optimizing culture media remains an art due to the complexity of cell-medium interactions. Machine learning has emerged as a promising solution, but it is hindered by biological fluctuations and experimental errors. To address these issues, we developed a medium optimization platform that integrates simplified and effective experimental manipulation, error-aware data processing for model training, predictive model construction to enhance accuracy and avoid local optimization, and an efficient optimization framework of active learning. Using this approach, we fine-tuned a 57-component serum-free medium for CHO-K1 cells, in which a total of 364 media were experimentally tested. The reformulated medium achieved approximately 60 % higher cell concentration than commercial alternatives. The improved cell culture is definitive toward CHO-K1, underscoring the platform's precision in targeted cell culture optimization. Our approach offers a robust tool for optimizing complex systems in cell culture and broader experimental studies, as well as in biomedical engineering applications.

Medium optimization

Error-aware data processing

Machine learning

Active learning

Experimental error

Biological fluctuation

Serum-free

© 2025 The Author(s). Published by Elsevier B.V.