The Application of Machine Learning in the Development of Co-Amorphous Dry Powder Inhalation

Besides improving drug solubility and stability, co-amorphous systems (COAMS) have recently been reported to enhance the pulmonary delivery efficiency of dry powder inhalation (DPI), offering a novel approach to the development of DPI. Conventional drug formulation development utilizes trial-and-error experiments, which require a laborious workload and resources. Moreover, the correlation of forming COAMS to enhanced pulmonary deposition remains underrepresented. Therefore, we proposed applying multiple machine learning (ML) models to the development of co-amorphous DPIs. In this study, we first constructed the database of COAMS through literature mining and then preprocessed the dataset with a molecular representation method. Subsequently, we successfully developed and evaluated the predictive performance of multiple ML models (i.e., logistic regression, random forests, XGBoost, LightGBM, and support vector machines) for forming a co-amorphous system. The five ML models' performance varied, yet all achieved satisfactory predictive accuracy (ACC) of around 0.80 in the testing subset. Specifically, LightGBM exhibited the highest ACC of 0.790 in cross-validation and 0.845 in the testing subset. In addition, SHapley Additive ex Planations (SHAP) analysis revealed that several molecular features (i.e., API_EState_VSA10, Co-former_BCUT2D_MRHI) are critical for models' prediction. More importantly, we conducted experimental validation by using salbutamol sulfate and indomethacin as model drugs to prepare co-amorphous DPIs based on the fine-tuned LightGBM model. The co-amorphous DPIs showed satisfactory aerodynamic performance with fine particle fractions of 41.87%-69.30%. In conclusion, we successfully demonstrated the feasibility of ML for guiding the formation of co-amorphous DPIs, further facilitating the development process in the future.