Predicting drug-drug interactions using multi-modal deep auto-encoders based network embedding and positive-unlabeled...

Area of Science:

• Pharmacology and Computational Biology
• Artificial Intelligence in Healthcare

Background:

• Drug-drug interactions (DDIs) pose significant risks to public health and patient safety.
• Current computational DDI prediction methods face challenges with incomplete, non-linear, and heterogeneous drug data.
• Integrating diverse drug features is vital for accurate DDI prediction models.

Purpose of the Study:

• To develop a robust method for predicting DDIs using multi-modal deep auto-encoders for drug representation learning.
• To address challenges of large-scale, noisy, and sparse data in DDI prediction.
• To enhance prediction accuracy by incorporating positive-unlabeled (PU) learning.

Main Methods:

• Proposed DDI-MDAE method utilizes multi-modal deep auto-encoders to learn unified drug representations from multiple feature networks.
• Employs four operators on learned drug embeddings to represent drug-drug pairs.
• Utilizes a random forest classifier, enhanced with PU learning, for DDI prediction.

Main Results:

• DDI-MDAE demonstrates significant effectiveness and improvement over existing benchmark methods for DDI prediction.
• The PU learning-enhanced model shows substantial gains in AUPR (7.1% on 3-CV, 6.2% on 5-CV) and F-measure (10.4% on 3-CV, 8.4% on 5-CV) compared to the original DDI-MDAE.
• Case studies confirm the practical utility of the DDI-MDAE approach.

Conclusions:

• DDI-MDAE effectively learns drug representations from heterogeneous data for accurate DDI prediction.
• The integration of PU learning further refines prediction accuracy, enhancing its clinical applicability.
• The proposed method offers a promising advancement in computational DDI prediction for improved patient safety.