Fast-tracking complex formulation development with multi-objective Bayesian optimisation.

Area of Science:

• Pharmaceutical Science
• Computational Chemistry
• Materials Science

Background:

• Pharmaceutical formulation development involves optimizing multiple interdependent properties, which is challenging with traditional methods like trial-and-error or design-of-experiment (DoE).
• Increasing numbers of parameters and competing objectives make empirical approaches inefficient and time-consuming.

Purpose of the Study:

• To present a machine learning-driven strategy for efficient pharmaceutical formulation development.
• To demonstrate the application of multi-objective Bayesian optimisation (MOBO) for optimizing complex formulations.

Main Methods:

• Utilized multi-objective Bayesian optimisation (MOBO), a machine learning strategy.
• Applied MOBO to develop an in-situ thermoresponsive sertraline hydrochloride nasal gel.
• Focused on optimizing interdependent formulation attributes like solubility, gelation temperature, and particle size.

Main Results:

• Achieved optimal trade-offs between key formulation attributes using MOBO.
• Required characterization of only 22 samples, demonstrating significant resource efficiency compared to traditional DoE.
• MOBO adapted dynamically to noisy, nonlinear response surfaces, mimicking expert decision-making.

Conclusions:

• The noise-aware, constraint-enabled MOBO framework advances data-driven formulation pipelines in pharmaceutical R&D.
• MOBO accelerates the search for feasible solutions, potentially transforming pharmaceutical R&D into a more agile and efficient process.
• This machine learning approach enables near-optimal solutions in significantly fewer experiments.