GCL-MSE: Graph Contrastive Learning with Mutual Similarity Enhancement for Drug Repositioning | JoVE Visualize

Area of Science:

Biomedical Informatics
Computational Biology
Pharmacology

Background:

Current data-driven drug repositioning models face challenges in capturing complex relationships within biomedical knowledge graphs for drug-disease association (DDA) mining.
Existing methods often fail to adequately represent the intricate semantic and topological connections essential for accurate DDA prediction.

Purpose of the Study:

To develop an advanced model, Graph Contrastive Learning with Mutual Similarity Enhancement (GCL-MSE), for more effective DDA mining.
To introduce a novel 'mutual similarity' concept, encompassing drug therapeutic domain similarity and disease pharmacological response similarity, to better capture complex biomedical relationships.
To enhance the accuracy and scope of drug repositioning through improved DDA prediction.

Main Methods:

Proposed GCL-MSE model incorporating a Mutual Similarity Enhancement (MSE) mechanism to fuse multiple similarity metrics and build a semantic relationship topology.
Developed an Adaptive Orthogonal Noise Contrastive Estimation Loss (AdaOrthoNCE) to disentangle biological relationships and optimize latent space representations.
Employed a three-channel graph convolutional model for topology-semantic co-representations and utilized AdaOrthoNCE for optimized embeddings, enabling cross-scale DDA prediction.

Main Results:

GCL-MSE significantly outperformed state-of-the-art models, achieving improvements of over 4.8% in AUROC and 25.5% in AUPRC.
The model demonstrated effectiveness in collaborative modeling by integrating pharmacological, therapeutic, and topological features.
Successfully predicted potential therapeutic roles for drugs, including lamotrigine for Alzheimer's disease and hydroxyurea for breast cancer, with further validation through molecular docking.

Conclusions:

GCL-MSE represents a significant advancement in DDA mining and drug repositioning by effectively integrating diverse data perspectives.
The novel mutual similarity concept and AdaOrthoNCE loss contribute to superior performance in capturing complex biomedical associations.
The model's predictive capabilities and validated drug repositioning suggestions highlight its potential for accelerating drug discovery and development.