Predicting Glass-Forming Ability of Pharmaceutical Compounds by Using Machine Learning Technologies.

Area of Science:

• Pharmaceutical Sciences
• Computational Chemistry
• Materials Science

Background:

• Low aqueous solubility is a major hurdle in drug development, leading to poor absorption and bioavailability.
• Amorphization enhances solubility but amorphous drugs are thermodynamically unstable and prone to recrystallization.
• Glass-forming ability (GFA) quantifies the tendency of a substance to form a stable amorphous glass.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting the GFA of drug molecules.
• To identify effective molecular representations and ML algorithms for GFA prediction.
• To demonstrate the potential of in silico screening for stable amorphous drug development.

Main Methods:

• Developed and compared random forest (RF), XGBoost, and support vector machine (SVM) ML models.
• Utilized two molecular representations: 2D descriptors and Extended-connectivity fingerprints (ECFP).
• Trained and tested models on a dataset of 171 drug molecules to predict GFA.

Main Results:

• The 2D-RF model achieved the highest performance with accuracy (0.857), AUC (0.850), and F1 score (0.828) on the testing set.
• Feature importance analysis confirmed the model's interpretability and alignment with existing scientific literature.
• Successfully predicted GFA for 171 drug molecules, identifying potential stable glass formers.

Conclusions:

• Machine learning models, particularly 2D-RF, can accurately predict the glass-forming ability of drug molecules.
• In silico screening using ML offers a promising strategy for identifying stable amorphous drugs.
• This approach can accelerate the development of amorphous drugs with improved solubility and bioavailability.