On-demand serum-free media formulations for human hematopoietic cell expansion using a high dimensional search algorithm

Abstract

Substitution of serum and other clinically incompatible reagents is requisite for controlling product quality in a therapeutic cell manufacturing process. However, substitution with chemically defined compounds creates a complex, large-scale optimization problem due to the large number of possible factors and dose levels, making conventional process optimization methods ineffective. We present a framework for high-dimensional optimization of serum-free formulations for the expansion of human hematopoietic cells. Our model-free approach utilizes evolutionary computing principles to drive an experiment-based feedback control platform. We validate this method by optimizing serum-free formulations for first, TF-1 cells and second, primary T-cells. For each cell type, we successfully identify a set of serum-free formulations that support cell expansions similar to the serum-containing conditions commonly used to culture these cells, by experimentally testing less than 1 × 10^-5 % of the total search space. We also demonstrate how this iterative search process can provide insights into factor interactions that contribute to supporting cell expansion.

Fig. 1

Schematic of the closed feedback optimization design leading to the improved optimization performance of the HD-DE strategy against a benchmark problem in silico. a All system inputs required by the algorithm (number of factors, dose levels, HD-DE parameters, decision-making criteria) were pre-defined. The algorithm-generated test formulations were compounded according to the corresponding recipe and cells cultured. The final cell count (system response) was input into the algorithm which analyzed the response and iterated (further optimized) or terminated. A single cycle from generating combinations through in vitro culture, analysis, evaluation, and decision to iterate or terminate made up a single generation (abbreviated to “G(n)” with n = number of generation) of the HD-DE-driven optimization process. For in silico simulations, a benchmark function and a normal random variable was used to generate the response in lieu of cell culture. The result of evaluating the benchmark function for each given formulation was treated as the system input equivalent to the biological response. b The classical Differential Evolution was ineffective in improving the overall performance against a high-dimensional benchmark problem such as the Rosenbrock function. Recognition of the variability (“Classic DE + var”) in the simulated data produced some improvement in the overall performance at the expense of more formulations being tested. Additionally, the introduction and utilization of information collated in the self-assembled data library (HD-DE strategy) enabled selective exploration and clearing of candidate formulations, further improving performance and efficiency of the overall optimization process. Data presented three-independent sets for Classic DE and Classic DE + var conditions, and eight-independent sets for HD-DE condition. Data represent mean ± standard error of the mean (SEM)

Fig. 2

In vitro results from three-independent experimental runs demonstrate robustness of the HD-DE optimization process for TF-1 cells. a The algorithm was able to identify optimized conditions that sustained cell expansion under serum-free formulations. Variation in overall performance underscored the presence of variability in biological systems. The performance was normalized to that of the known maximum score (PC of TF-1 cell culture). b Despite variability, the algorithm was able to keep the optimization cost under control as it utilized the information gathered. c PCA loading illustrated the degree to which factor doses were conserved in the candidate solution set formulations. (See Supplementary Table 1 for factor legend). Data represent mean ± SEM

Fig. 3

In vitro media optimization performance of 14-factor T-cell culture demonstrated applicability of HD-DE strategy to the optimization of primary cell culture formulations. a Ranking of individual formulations of the candidate solution set according to the score expressed as the cell expansion normalized to that of the PC condition for T-cells showing change through generations. A number of formulations that supported T-cell expansion at levels comparable to that of the PC condition were identified. b PCA biplot illustrate the distribution of identified formulations in the solution space and indicate the degree to which factors are conserved among the identified formulations. (See Supplementary Table 2 for factor legend)

Fig. 4

Characterization and comparison of the expanded T-cell population cultured using the top 5 serum-free formulations identified. a Comparison of the variation in cell expansion between donors for the variation of the media conditions. b Comparison of cell expansion capacity between two commercially available serum-free media formulations (PC and Xuri EM) with and without serum supplementation and the top 5 formulations identified (F1–F5). Cell expansion was compared across the three donor cells. p-values using ANOVA and post hoc Tukey’s multiple comparison tests listed in Supplementary Data 2. c The top 5 formulations analyzed for cell populations expressing T-cell markers CD3, CD4, and CD8. p-values using ANOVA and post hoc Tukey’s multiple comparison tests listed in Supplementary Table 3. d Analysis of the composition of the top 5 formulations (Supplementary Table 4) identified

Fig. 5

Multivariable analysis of the dataset obtained through HD-DE optimization in vitro. a The full test library of formulations obtained from all 3 runs of the HD-DE optimization performed on TF-1 cells (excluding G(1)) was used for post hoc analysis to elucidate the main effects, quadratic effects, and 2-factor interactions. The logworth statistical significance of the regression coefficient estimates are presented in a plot where the significant effects (false discovery rate (FDR)-corrected p-values <0.05) are colored while non-significant elements are grayed. b The data points that have significant effects (colored data points from (a)) are labeled with the corresponding factor or interaction. c The full test library of formulations obtained from the HD-DE optimization performed on T-cells (excluding G(1)) was used for post hoc analysis to elucidate the main effects, quadratic effects, and 2-factor interactions. The FDR-statistically significant effects (expressed as logworth p-values) are colored while non-significant elements are grayed. d The data points that have statistically significant effects (colored data points from (c)) are labeled with the corresponding factor or interaction. The square term suggest either a supra-additive relationship (positive term) between effect and dose or a saturation (negative term), and the cross-product terms refer to 2-factor interactions. (See Supplementary Table 2 for factor legend)