AffiGrapher: Contrastive Heterogeneous Graph Learning with Aromatic Virtual Nodes for RNA-Small Molecule Binding...

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• RNA molecules are crucial drug targets due to their diverse structures and functions.
• Predicting RNA-small molecule binding affinity is difficult due to limited data and RNA conformational flexibility.

Purpose of the Study:

• To develop a robust computational method for predicting RNA-small molecule binding affinity.
• To enhance the accuracy and generalization of binding affinity predictions for RNA-targeted drug discovery.

Main Methods:

• Developed AffiGrapher, a physics-driven graph neural network integrating a physics-informed graph architecture with contrastive learning.
• Incorporated multiple RNA conformations to capture diverse structural information and improve prediction robustness.
• Evaluated model performance using rigorous cross-validation, cold-start settings, and multiscenario evaluations under structural uncertainty.

Main Results:

• AffiGrapher achieved state-of-the-art performance in binding affinity prediction across various validation schemes.
• Demonstrated strong generalization ability, even with predicted docking poses and unseen data.
• Showcased exceptional potential in virtual screening tasks for RNA-targeted drug discovery.

Conclusions:

• Integrating physics-based architectures with contrastive learning effectively addresses challenges in RNA-small molecule affinity prediction.
• AffiGrapher offers a powerful tool for accelerating RNA-targeted drug discovery.
• The method's robustness and generalization capabilities highlight its practical utility in computational drug design.