Enhancing Generalizability in Protein-Ligand Binding Affinity Prediction with Multimodal Contrastive Learning

Area of Science:

• Computational Chemistry
• Structural Biology
• Drug Discovery

Background:

• Accurate prediction of protein-ligand binding affinity is crucial for drug discovery.
• Current machine learning methods often struggle with generalization due to single-modal representations of interactions.

Purpose of the Study:

• To develop a graph neural network-based scoring function that improves generalization ability in protein-ligand binding affinity prediction.
• To enhance the comprehensive understanding of protein-ligand interactions through multi-modal representations.

Main Methods:

• Utilized a graph neural network architecture for scoring function development.
• Implemented a triplet contrastive learning loss function.
• Integrated three-dimensional complex representations with fused two-dimensional ligand and coarse-grained pocket representations.

Main Results:

• The proposed model demonstrated superior generalization capabilities on multiple external datasets.
• Achieved better performance compared to existing deep learning-based scoring functions.
• Validated the effectiveness of the multi-modal representation and contrastive learning approach.

Conclusions:

• The developed scoring function shows significant promise as a tool in drug discovery.
• The training framework is adaptable for other biophysical and biochemical prediction tasks, such as protein-protein interactions and mutation effects.