deepDR

deepDR applies network-based deep learning to predict drug-disease associations for in silico drug repositioning.

Key Features:

• Integration of heterogeneous networks: Integrates ten biological networks comprising one drug-disease network, one drug-side-effect network, one drug-target network, and seven drug-drug networks to represent multi-relational biomedical data.
• Multi-modal deep autoencoder: Employs a multi-modal deep autoencoder to extract high-level features and learn intricate patterns from the integrated networks.
• Variational autoencoder for prediction: Uses a variational autoencoder to encode and decode drug-disease pairs to infer candidate approved drugs for new indications.
• Performance metrics: Achieves an area under the receiver operating characteristic curve (AUROC) of 0.908 for predicting drug-disease associations, outperforming conventional network-based and machine learning approaches.
• Validation against clinical data: Validated predictions using the ClinicalTrials.gov database with a validation AUROC of 0.826.

Scientific Applications:

• Drug repositioning: Accelerates identification of new therapeutic uses for existing approved drugs, aiming to reduce time and cost relative to traditional drug development.
• Disease-specific candidate identification: Applied to identify novel approved drugs for complex diseases, including Alzheimer's disease (examples: risperidone and aripiprazole) and Parkinson's disease (examples: methylphenidate and pergolide).

Methodology:

Combine multiple biological networks, extract drug representations using a multi-modal deep autoencoder, and predict drug-disease associations with a variational autoencoder.