Introducing AutoML framework for drug-drug interaction prediction: application of AutoGluon

Area of Science:

• Computational chemistry
• Pharmacology
• Machine learning

Background:

• Drug-drug interactions (DDIs) are a significant safety concern in pharmacotherapy.
• Advanced artificial intelligence (AI) methods improve DDI prediction but require specialized expertise.
• Existing AI models for DDI prediction often involve complex development processes.

Purpose of the Study:

• To apply the AutoGluon AutoML framework for automated DDI prediction using molecular features.
• To simplify and streamline the development of DDI prediction models.
• To assess the performance of AutoML in identifying chemically meaningful predictors of DDIs.

Main Methods:

• Utilized a curated dataset of 100,000 drug pairs from DrugBank.
• Employed three molecular representations: 2D molecular descriptors, Morgan fingerprints, and a combination.
• Trained models using AutoGluon-Tabular without manual hyperparameter tuning.

Main Results:

• The model using only 2D molecular descriptors achieved the highest performance (84.4% accuracy, 0.916 AUC).
• This descriptor-based model outperformed models using fingerprints or a hybrid approach.
• Feature importance analysis highlighted physicochemical and topological descriptors (e.g., drug-likeness, electrotopological indices) as key DDI predictors.

Conclusions:

• AutoML, specifically AutoGluon, can effectively predict DDIs while reducing technical barriers.
• The framework extracts chemically relevant patterns and provides feature-level interpretability.
• This approach offers a scalable solution for DDI prediction, paving the way for future advancements with larger datasets and complex chemical representations.