A meta-learning framework using representation learning to predict drug-drug interaction.

Area of Science:

• Computational chemistry and pharmacology
• Bioinformatics and computational biology
• Machine learning in drug discovery

Background:

• Drug-drug interactions (DDIs) pose significant risks in polypharmacy, necessitating accurate prediction methods.
• Computational approaches for DDI prediction can reduce the extensive costs and time associated with in vitro experiments.
• Integrating diverse data sources (drug properties, gene, protein, disease, side effects) is crucial for effective DDI prediction.

Purpose of the Study:

• To develop and evaluate a novel semi-supervised learning framework for predicting drug-drug interactions.
• To leverage representation learning, positive-unlabeled (PU) learning, and meta-learning for enhanced DDI prediction.
• To create a robust system that integrates information from multiple data sources for more reliable DDI identification.

Main Methods:

• A semi-supervised learning framework integrating representation learning (Node2vec) and PU learning (bagging SVM).
• Feature networks were constructed using information from multiple drug-related data sources to learn meta-knowledge.
• A meta-classifier was designed to combine interaction probabilities derived from learned meta-knowledge.

Main Results:

• Representation learning (Node2vec) improved system performance by 22%.
• Positive-unlabeled (PU) learning (bagging SVM) enhanced performance by 12.7%.
• The developed meta-classifier demonstrated superior performance and reliability in predicting DDIs compared to base classifiers.

Conclusions:

• The proposed semi-supervised framework effectively predicts drug-drug interactions by integrating multiple data sources.
• The combination of representation learning and PU learning significantly boosts prediction accuracy.
• This computational approach offers a promising strategy for accelerating drug discovery and development by identifying potential DDIs early.