Bayesian Optimization for Efficient Multiobjective Formulation Development of Biologics.

Area of Science:

• Biopharmaceutical formulation development
• Protein biophysics
• Computational chemistry

Background:

• Biologics often face developability challenges impacting therapeutic translation.
• Formulation design is complex, requiring simultaneous optimization of multiple biophysical properties and excipient interactions.
• Traditional methods struggle with high-order complexities and local optima in formulation design.

Purpose of the Study:

• To develop an efficient machine learning-based method for optimizing biologic formulation.
• To concurrently optimize multiple critical biophysical properties of monoclonal antibodies.
• To ensure practical applicability by incorporating formulation constraints like osmolality and pH.

Main Methods:

• Combined Bayesian optimization and high-throughput experimental screening.
• Modeling nonlinear relationships and interactions among formulation features.
• Optimization of melting temperature (Tm), diffusion interaction parameter (kD), and air-water interface stability.

Main Results:

• Identified highly optimized formulation conditions in only 33 experiments.
• Demonstrated the method's ability to account for formulation constraints (osmolality, pH).
• Provided insights into excipient influence and highlighted trade-offs between conflicting properties (e.g., pH effects on Tm and kD).

Conclusions:

• Machine learning, specifically Bayesian optimization with screening, significantly reduces experiments needed for formulation optimization.
• The developed method efficiently navigates complex design spaces to improve biologic developability.
• The approach offers practical insights into formulation components and property trade-offs for successful therapeutic translation.