Adaptive debiasing learning for drug repositioning.

Area of Science:

• Computational biology
• Pharmacology
• Artificial intelligence

Background:

• Drug repositioning accelerates pharmaceutical development by identifying new uses for existing drugs.
• Graph neural network (GNN) deep learning methods show promise for drug repositioning.
• Existing drug repositioning datasets exhibit node polarization, with popular and long-tail entities, potentially introducing bias.

Purpose of the Study:

• To analyze drug repositioning datasets for inherent biases.
• To propose a novel deep learning framework, DRDM, to mitigate data bias in drug repositioning.
• To enhance the representation of long-tail entities for novel drug discovery insights.

Main Methods:

• Analysis of three common drug repositioning datasets to identify node polarization.
• Development of DRDM, a deep learning framework with a debiasing mechanism.
• Implementation of dynamic weight adjustment to address biases towards popular entities.
• Integration of dual-view contrastive learning to improve model robustness.

Main Results:

• Consistent node polarization identified across analyzed drug repositioning datasets.
• DRDM effectively mitigates biases associated with popular entities, improving long-tail entity representation.
• Experimental results show DRDM's strong competitiveness against existing models.
• Case studies demonstrate DRDM's practical potential for drug discovery.

Conclusions:

• Node polarization is a critical characteristic of drug repositioning datasets that requires bias mitigation.
• DRDM offers an effective deep learning approach to address data bias and enhance drug repositioning.
• The proposed framework has the potential to uncover novel drug-target associations and accelerate drug discovery.