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

Pharmaceutical formulation development often requires the simultaneous optimisation of multiple, interdependent properties, such as solubility, gelation temperature, and particle size, where improving one attribute can compromise another. Traditionally, this challenge is addressed through trial-and-error experimentation guided by expert intuition or through empirical design-of-experiment (DoE) approaches, both of which become inefficient as the number of input parameters and competing objectives increase. Here we present a machine learning-driven strategy for the efficient development of complex pharmaceutical formulations, applied to an example in-situ thermoresponsive sertraline hydrochloride nasal gel. Using multi-objective Bayesian optimisation (MOBO), we achieved optimal trade-offs between key formulation attributes whilst preparing and characterising only 22 samples, making the resource efficiency competitive to even simple DoE approaches. MOBO not only accelerates the search for feasible solutions but also adapts dynamically to noisy, nonlinear response surfaces, quantitatively mimicking expert decision-making. This noise-aware, constraint-enabled framework represents a step toward data-driven formulation pipelines, where machine learning algorithms guide experimental design. By enabling near-optimal solutions in far fewer experiments, MOBO has the potential to transform pharmaceutical R&D from an empirical bottleneck into an agile, efficient process.