Model-Guided Optimization of a Low-Serum Medium for Madin-Darby Canine Kidney Cells Using a Kolmogorov-Arnold Network and Bayesian Optimization.

Madin-Darby canine kidney (MDCK) cells are important hosts for cell culture-based viral vaccine production, but fetal bovine serum (FBS) supplementation increases cost, batch variability, adventitious-agent risk, and regulatory burden. This study developed a 2% FBS low-serum medium optimization workflow for MDCK cells by integrating design of experiments (DoE) with interpretable machine learning. A Plackett-Burman (PB) design screened 27 nutritional components and identified seven factors that significantly affected proliferation. L-glutamine, L-asparagine, and L-tyrosine showed positive effects and were further optimized using a three-factor Box-Behnken design (BBD). A Kolmogorov-Arnold network (KAN) modeled the nonlinear proliferation response, and its response function served as the objective function for Bayesian optimization (BO). After two KAN-BO validation cycles, the optimized formulation contained 292.30 mg/L L-glutamine, 54.36 mg/L L-tyrosine, and 11.34 mg/L L-asparagine. The predicted relative proliferation rate was 159.38%, in close agreement with the experimental value of 158.63 ± 1.70%. Compared with Baseline medium, Optimized medium enhanced MDCK cell proliferation while maintaining typical adherent epithelial-like morphology and high viability. In H1N1 BVR-26 production assays, Optimized medium achieved a peak titer of 3.50 ± 0.05 log