Multiple similarity drug-target interaction prediction with random walks and matrix factorization.

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Drug-target interactions (DTIs) are crucial for pharmaceutical development.
• Existing computational methods like random walk and matrix factorization have limitations in DTI prediction.
• These limitations include unsupervised embedding generation and distorted insights from linear similarity combinations.

Purpose of the Study:

• To propose a novel optimization framework, Multiple similarity DeepWalk-based Matrix Factorization (MDMF), for accurate DTI prediction.
• To address limitations of existing methods by employing a multi-layered network approach.
• To develop an ensemble method (MDMF2A) for optimizing DTI prediction performance.

Main Methods:

• Utilized a multi-layered network approach to integrate diverse drug and target similarities.
• Developed the MDMF framework, unifying embedding generation and interaction prediction.
• Created an ensemble method (MDMF2A) by integrating two MDMF model instantiations, optimizing for AUPR and AUC.

Main Results:

• The MDMF framework learns vector representations capturing higher-order proximity and local invariance.
• MDMF2A achieved statistically significant improvements over state-of-the-art approaches on real-world DTI datasets.
• Validation of highly ranked non-interacting pairs suggests potential for discovering novel DTIs.

Conclusions:

• The proposed MDMF and MDMF2A methods offer a powerful approach for accurate DTI prediction.
• These methods overcome limitations of previous techniques by effectively handling heterogeneous data.
• MDMF2A shows promise in identifying novel drug-target interactions, advancing drug discovery.