Hybrid attentional memory network for computational drug repositioning

Area of Science:

• Computational drug discovery
• Pharmacology
• Bioinformatics

Background:

• Drug repositioning is crucial for identifying new therapeutic uses of existing drugs.
• Existing computational drug repositioning models are often limited to either neighborhood-based or latent factor-based collaborative filtering (CF) approaches.
• The cold start problem remains a significant challenge, hindering the predictive capabilities of current models.

Purpose of the Study:

• To develop a hybrid model that integrates the strengths of both neighborhood-based and latent factor-based CF models for improved drug repositioning.
• To address the cold start problem in computational drug repositioning.
• To enhance the accuracy of predicting drug-disease associations.

Main Methods:

• Proposed the hybrid attentional memory network (HAMN), a deep architecture combining two classes of CF models non-linearly.
• Integrated a memory unit and attention mechanism to generate a neighborhood contribution representation for local drug-disease associations.
• Employed a variant autoencoder to extract latent factors of drugs and diseases, capturing global information and utilizing ancillary data to mitigate the cold start problem.

Main Results:

• The HAMN model effectively combines local and global information from drug-disease associations.
• Ancillary information of drugs and diseases was utilized to alleviate the cold start problem.
• Experimental results on two datasets demonstrated that HAMN outperforms existing models in predicting drug-disease associations, as indicated by AUC, AUPR, and HR metrics.

Conclusions:

• The HAMN model offers a novel solution for improving the prediction accuracy of drug-disease associations.
• This approach provides a new perspective for pharmaceutical researchers in drug development.
• The study highlights the potential of hybrid deep learning architectures in computational drug repositioning.