Predicting combinative drug pairs via multiple classifier system with positive samples only.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in drug discovery

Background:

• Drug combinations offer synergistic effects for complex diseases, but identification via dose-response methods is costly.
• Existing supervised learning approaches for drug combination prediction suffer from inadequate heterogeneous feature utilization and bias from assuming unknown pairs as non-combinations.

Purpose of the Study:

• To develop a large-scale, cost-effective approach for predicting potential drug combinations.
• To address the inadequate feature integration and inherent bias in current predictive models.
• To improve the accuracy and reliability of drug combination prediction.

Main Methods:

• A Two-Layer Multiple Classifier System (TLMCS) was designed to integrate heterogeneous features: anatomical therapeutic chemical codes, drug-drug interactions, drug-target interactions, gene ontology, and side effects.
• One-class support vector machines were employed as member classifiers to avoid bias by training solely on positive (approved) drug combinations.
• Model performance was validated using 10-fold cross-validation (10-CV) and a novel prediction experiment.

Main Results:

• TLMCS outperformed three state-of-the-art methods in 10-CV, achieving an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.824 and an Area Under the Precision-Recall Curve (AUPRC) of 0.372.
• In a novel prediction task, 9 out of the top-20 predicted drug pairs were validated through literature review.
• Analysis of newly validated combinations revealed five types of drug combinations and three types of pathway-based drug relationships.

Conclusions:

• The proposed TLMCS offers an effective framework for integrating heterogeneous features in drug combination prediction.
• Training with only positive samples successfully mitigates the bias associated with classifying unknown pairs as negative.
• The predictive capability of TLMCS was demonstrated, revealing novel insights into drug combination types and their underlying pathway interactions.