CLMT: graph contrastive learning model for microbe-drug associations prediction with transformer

Area of Science:

• Computational Biology
• Bioinformatics
• Drug Discovery

Background:

• Accurate microbe-drug association prediction is crucial for drug development and disease diagnosis.
• Existing methods face challenges in modeling complex nonlinear relationships and long-range dependencies.
• Distinguishing subtle similarities between microbes and drugs remains a significant hurdle.

Purpose of the Study:

• To develop an advanced model for microbe-drug association prediction.
• To address limitations of current methods in capturing complex interactions and sparse data.
• To enhance the accuracy and generalizability of microbe-drug association predictions.

Main Methods:

• CLMT model integrating Graph Transformer network with an attention mechanism.
• Application of graph contrastive learning with node perturbation and edge dropout for data augmentation.
• Optimization using a contrastive loss to learn robust and generalizable embeddings.

Main Results:

• CLMT significantly outperforms existing methods in microbe-drug association prediction.
• Demonstrated state-of-the-art performance on MDAD, aBiofilm, and Drug Virus datasets.
• Achieved accuracy improvements of 4.3%, 3.5%, and 2.8% over the previous best model, respectively.

Conclusions:

• CLMT effectively mitigates data sparsity and captures nonlinear microbe-drug interactions.
• The model's architecture enhances the ability to model high-order dependencies and long-range associations.
• CLMT shows significant promise for real-world biomedical applications in drug development and diagnostics.