One-Time Optimization of Advanced T Cell Culture Media Using a Machine Learning Pipeline

The growing application of cell and gene therapies in humans leads to a need for cell type-optimized culture media. Design of Experiments (DoE) is a successful and well known tool for the development and optimization of cell culture media for bioprocessing. When optimizing culture media for primary cells used in cell and gene therapy, traditional DoE approaches that depend on interpretable models will not always provide reliable predictions due to high donor variability. Here we present the implementation of a machine learning pipeline into the DoE-based design of cell culture media to optimize T cell cultures in one experimental step (one-time optimization). We applied a definitive screening design from the DoE toolbox to screen 12 major media components, resulting in 25 (2k + 1) media formulations. T cells purified from a set of four human donors were cultured for 6 days and cell viability on day 3 and cell expansion on day 6 were recorded as response variables. These data were used as a training set in the machine learning pipeline. In the first step, individual models were created for each donor, evaluated and selected for each response variable, resulting in eight final statistical models (R ² > 0.92, RMSE < 1.5). These statistical models were used to predict T cell viability and expansion for 10⁵ random in silico-generated media formulations for each donor in a grid search approach. With the aim of identifying similar formulations in all donors, the 40 best performing media formulations of each response variable were pooled from all donors (n = 320) and subjected to unsupervised clustering using the k-means algorithm. The median of each media component in each cluster was defined as the cluster media formulation. When these formulations were tested in a new set of donor cells, they not only showed a higher T cell expansion than the reference medium, but also precisely matched the average expansion predicted from the donor models of the training set. In summary, we have shown that the introduction of a machine learning pipeline resulted in a one-time optimized T cell culture medium and is advantageous when working with heterogeneous biological material.