Sarah Zellnitz1, Elisabeth Fink1 & Amrit Paudel1,2
1Research Center Pharmaceutical Engineering, Inffeldgasse 13, Graz, 8010, Austria
2 Institute for Process and Particle Engineering, Graz University of Technology, Inffeldgasse 13, Graz, 8010, Austria
Summary
The treatment of chronically obstructive pulmonary disease (COPD), asthma and tuberculosis requires a multidrug therapy applying different active pharmaceutical ingredients (APIs). In order to reduce the drug dose variability while administering the combination products to patients, the generation of co-delivery systems like co-amorphous systems (COAMS) is suggested. Co-amorphous systems have recently gained considerable interest in the pharmaceutical field for tackling poor solubility of drugs. In order to facilitate and enhance the screening for COAMS for inhalation therapy, a machine learning (ML) approach is presented herein to predict the formation of potential COAMS. A training dataset of 254 systems was built based on available literature data on COAMS and non-COAMS. Based on the training dataset, a predictive ML model was built with a predictability of 80%. The resulting ML model was validated with 21 rows of validation data (not included in the training dataset), which resulted in a predictability of 76%. This model was then used to predict the formation of COAMS in API-API combinations of 37 common APIs used in the treatment of the aforementioned lung disease. The predicted class of these 666 possible API-API combinations was analysed and put in correlation with the original dataset. Positive predictions were weighted by their distance, with predicted co-amorphous API-API combinations further from the training data considered with more uncertainty. This provided a rating amongst all predicted COAMS from which the top 100 systems were taken into consideration. Taking conventional medication regimen into account, the number of promising combinations was further reduced. Promising predicted combinations for COAMS will be selected and tested experimentally in order to validate the prediction.
Key Message
Based on molecular descriptors of a training dataset a predictive machine learning (ML) model was established (predictability 80%) and validated with 21 additional rows of data (predictability 76%). This model could succesfully be used to predict the formation of promising co-amorphous API-API combinations to be used in the treatment of lung disease.