Machine learning strengthened formulation design of pharmaceutical suspensions.

Abstract

Many different formulation strategies have been investigated to oppose suboptimal treatment of long-term or chronic conditions, one of which are the nano- and microsuspensions prepared as long-acting injectables to prolong the release of an active pharmaceutical compound for a defined period of time by regulating the size of particles by milling. Typically, surfactant and/or polymers are added in the dispersion medium of the suspension during processing for stabilization purposes. However, current formulation investigations with milling are heavily based on prior expertise and trial-and-error approaches. Various interacting parameters such as the milling bead size, stabilizer type and concentration have confounded the investigation of milling process. The present study systematically exploited statistical and machine learning (ML) strategies to understand the relationship between suspension characteristics and formulation parameters under full-factorial milling experiments. Stabilizer concentration was identified as a significant factor (p < 0.001) for median suspension diameter (D₅₀). A formulation stability classification ML model with high prediction accuracy (0.91) and F1-score (0.91) under 10-fold cross-validation was constructed based on 72 formulation datapoints. Model interpretation through Shapley additive explanations (SHAP) revealed the prominent impact of stabilizer concentration and milling bead size on formulation stability. The present work demonstrated the potential to achieve a deeper understanding of the design and optimization of nano- and microsuspensions through explainable ML modelling on formulation screening data.