Machine Learning-Guided Development of Anti-Tuberculosis Dry Powder for Inhalation Prepared by Co-Spray Drying

<b>Background/Objectives:</b> Tuberculosis (TB) remains a major global health threat. Current administration methods for anti-TB drugs, including oral or intravenous, suffer from systemic side effects, low lung distribution, and poor patient compliance. Dry powder inhalers (DPIs) offer a promising alternative. This study investigates the aerodynamic performance of co-spray-dried DPIs containing rifampin or pyrazinamide and amino acids by using machine learning. <b>Methods:</b> Firstly, 72 formulations were prepared by varying drug-amino acid combinations, molar ratios, and spray-drying parameters. Subsequently, the aerodynamic performance of all 72 formulations was evaluated using a Next Generation Impactor, and the solid-state characterizations of optimal DPIs were carried out. Finally, four machine learning (ML) models were successfully developed and were utilized to predict the fine particle dose (FPD), FPF, MMAD, and geometric standard deviation (GSD) of DPIs based on the high-quality in-house data above. <b>Results:</b> Key results showed that the aerodynamic performance of DPIs was highly dependent on the specific drug-amino acid combination, with rifampin-L-lysine acetate and pyrazinamide-L-leucine formulations achieving the highest fine particle fraction (FPF, 73.37%, 87.74%) and optimal mass median aerodynamic diameter (MMAD, 2.59 µm, 1.88 µm). Notably, XGBoost (v3.1.3) exhibited the best predictive performance, with R² values ranging from 0.894 to 0.991 in the testing set for the four prediction tasks. Meanwhile, SHapley Additive exPlanations (v0.50.0) was used for model interpretability analysis. The molecular weights and LogP of the drug and amino acid were identified as two of the most important features affecting the prediction of FPD, FPF, MMAD, and GSD. <b>Conclusions</b>: This work demonstrates the feasibility of ML in accelerating the development of inhalable spray-dried anti-TB drugs by enabling the prediction of DPI formulations.