CASTER-DTA: equivariant graph neural networks for predicting drug-target affinity

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Accurate drug-target affinity (DTA) prediction is crucial for efficient drug design and screening.
• While 3D protein structure prediction tools like AlphaFold are accessible, scalable DTA prediction methods often underutilize this structural information.
• Existing methods lack robust integration of protein structural data for improved binding affinity predictions.

Purpose of the Study:

• To develop a novel deep learning framework, CASTER-DTA, that effectively utilizes 3D protein structure information for accurate drug-target affinity prediction.
• To enhance the interpretability of DTA prediction models by incorporating attention mechanisms between protein and drug components.
• To create a comprehensive resource of predicted binding affinities for FDA-approved drugs against the human proteome.

Main Methods:

• Utilized an equivariant graph neural network (GNN) for robust protein representation learning from 3D structures.
• Employed a standard GNN for learning molecular representations of drugs.
• Integrated an attention-based mechanism between protein residues and drug atoms to capture key interactions and improve interpretability.

Main Results:

• CASTER-DTA achieved state-of-the-art performance improvements across multiple benchmark datasets for DTA prediction.
• The model provided novel insights into drug-target interactions, demonstrating its utility beyond simple affinity prediction.
• Generated a large-scale dataset of binding affinities for all FDA-approved drugs against the human proteome.

Conclusions:

• CASTER-DTA represents a significant advancement in predicting drug-target binding affinity by effectively integrating 3D structural information.
• The developed model and associated large-scale dataset offer valuable resources for drug discovery and development.
• A publicly accessible web server allows researchers to predict binding affinities for custom protein-drug pairs.