Multi-Objective Bayesian Optimization for Data-Efficient Bioprocess Development

Process optimization for Chinese hamster ovary (CHO) cell culture remains a challenge in biopharmaceutical development because multiple interacting parameters jointly influence productivity and product quality attributes. Traditional design-of-experiments (DoE) methods, while systematic, become impractically expensive when extended across multiple parameters and clones. To address this challenge, we developed a multi-objective Bayesian Optimization (BO) framework that identifies optimal process conditions efficiently in grouped recommendations, which is well suited for experimental workflows in bioprocess development. The model integrates continuous variables such as pH, DO, temperature, and feed rate with categorical identifiers to enable knowledge transfer across clones and scales, optimizing titer, glycan profile, and charge variants. We validated the framework through in-silico benchmarks on analytic functions, retrospective cross-validation on historical CHO datasets, and forward experimental validation in small-scale bioreactors. Across these tests, our algorithm consistently outperformed Latin Hypercube Sampling (LHS) and Random Search baselines, achieving superior performance under a limited experimental budget. The framework improved titer by up to 37% under single-objective optimization. In the multi-objective setting, it increased titer by 25% while simultaneously reducing overall glycan-profile error by a factor of seven, demonstrating the ability to optimize multiple biologically coupled objectives simultaneously. Through comprehensive in-silico and experimental validation, this study establishes a framework that enables adaptive, AI-guided process development and improves decision-making across multiple objectives, clones, and scales while minimizing experimental runs in process development and optimization workflows.