GraphCL-DTA: A Graph Contrastive Learning With Molecular Semantics for Drug-Target Binding Affinity Prediction.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Drug-target binding affinity prediction is crucial for early drug discovery.
• Existing computational models struggle with representation learning from molecular graphs and ignore uniformity metrics.
• Previous methods often rely solely on supervised data, neglecting inherent molecular graph information.

Purpose of the Study:

• To introduce GraphCL-DTA, a novel graph contrastive learning framework for drug-target binding affinity prediction.
• To enhance drug and target representation learning by incorporating molecular semantics and optimizing uniformity.
• To improve the performance and generalization capability of computational models in drug discovery.

Main Methods:

• Developed a graph contrastive learning framework (GraphCL-DTA) utilizing embedding-space data augmentation to preserve molecular graph semantics.
• Introduced a new loss function to directly optimize the uniformity of drug and target representations.
• Validated the model on KIBA and Davis datasets, comparing performance against GraphDTA.

Main Results:

• GraphCL-DTA demonstrated improved drug-target binding affinity prediction accuracy, with relative improvements of 2.7% on KIBA and 4.5% on Davis compared to GraphDTA.
• The graph contrastive learning and uniformity function enhanced the quality of drug and target representations without additional supervised data.
• The proposed modules showed potential for improving generalization in other computational models.

Conclusions:

• GraphCL-DTA offers a more effective approach to drug-target binding affinity prediction by leveraging graph contrastive learning and representation uniformity.
• The framework provides a robust method for learning essential drug representations, outperforming previous models.
• The GraphCL-DTA modules can be integrated into existing computational models to enhance their predictive power and generalization.