Premixed cocktail configurations for media blending altered cell culture outcomes discovered by active learning

Medium formulation is critical for successful mammalian cell culture, yet the influence of component blending during medium preparation remains underexplored. Here, we applied machine learning-guided active learning to investigate whether different patterns of medium premixes, designed as cocktail configurations, affect cell culture performance. As a first challenge, CHO-K1 cells were used, and all cocktail configurations were formulated for the eRDF medium nominal recipe. In configurations that combined 55 components into seven cocktails, we observed substantial variation in cell densities. Iterative modeling using a neural network and genetic algorithm-based selection identified cocktail configurations that significantly increased cell density beyond that of the commercial medium, with up to a 1.24-fold increase in independent validation experiments. Analysis of all 127 cocktail configurations revealed diverse optimal patterns and identified component pairs whose co-cocktail configurations either benefited or impaired culture performance. These findings demonstrate that medium preparation, specifically cocktail configuration, substantially influences cell culture outcomes and suggest that optimizing cocktail configurations is a promising strategy to improve medium design and cell culture efficiency.