Bayesian Optimization for Efficient Multiobjective Formulation Development of Biologics

Biologics, including emerging engineered formats, can often exhibit poor developability profiles, complicating their translation into successful therapeutics. While formulation design can substantially mitigate some developability issues, it represents a highly complex optimization challenge due to the need to simultaneously improve multiple biophysical properties, navigate a vast design space, and account for nonlinear or synergistic interactions among excipients. Traditional design of experiments methods can reduce experimental effort but are limited by difficulties in managing high-order complexities and a propensity to become trapped in local optima. In response, machine learning techniques combined with (high-throughput) screenings have emerged as powerful strategies to overcome these limitations, dramatically reducing the number of required experiments. The ability of these models to capture nonlinear relationships and interactions among multiple features enables efficient navigation in a high-dimensional design space. We present a combined Bayesian optimization and experimental screening method that concurrently optimizes three key biophysical properties of a monoclonal antibody─melting temperature <i>T</i>_m, diffusion interaction parameter <i>k</i>_D, and stability against air-water interfaces. We demonstrate its effectiveness through the identification of highly optimized formulation conditions in just 33 experiments. Furthermore, our approach can account for essential formulation constraints such as osmolality and pH, ensuring practical applicability. We show that beyond optimization, our method provides valuable insights into the influence of individual excipients on each biophysical property across formulations. Furthermore, it highlights the need to balance trade-offs between conflicting properties, such as the opposing effects of pH on <i>T</i>_m and <i>k</i>_D.