Additional Neural Matrix Factorization model for computational drug repositioning.

Area of Science:

• Computational biology
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Computational drug repositioning offers reduced costs and timelines for discovering new drug applications.
• Recommendation systems are increasingly used for drug repositioning, but face challenges in learning ability and data sparsity.
• Existing models struggle with higher-level drug-disease associations and are prone to overfitting without considering negative samples.

Purpose of the Study:

• To propose a novel computational drug repositioning method called Additional Neural Matrix Factorization (ANMF).
• To enhance drug and disease representation by incorporating drug-drug and disease-disease similarities.
• To improve the learning ability and generalization of drug repositioning models while minimizing overfitting.

Main Methods:

• Utilized drug-drug and disease-disease similarities to address data sparsity.
• Employed a variant autoencoder to extract hidden features of drugs and diseases.
• Integrated Generalized Matrix Factorization (GMF) for collaborative filtering and applied negative sampling to mitigate overfitting.

Main Results:

• The ANMF model demonstrated superior performance compared to state-of-the-art methods on the Gottlieb and C dataset.
• Effectively overcame data sparsity by enhancing feature representation.
• Showcased improved learning ability and reduced overfitting through integrated techniques.

Conclusions:

• The ANMF model shows significant promise for predicting new therapeutic indications for existing drugs.
• Addresses key challenges in computational drug repositioning, particularly in predicting and selecting novel applications.
• Provides a robust framework for advancing drug repositioning research and development.