CASTER-DTA: Equivariant Graph Neural Networks for Predicting Drug-Target Affinity

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Accurate prediction of ligand-protein binding affinity is crucial for drug design and development.
• Advancements in protein structure prediction (e.g., AlphaFold) offer new opportunities for structure-based drug design.
• Existing structure-based methods often underutilize 3D protein structural information.

Purpose of the Study:

• To develop a novel computational method for predicting drug-target affinity (DTA) that fully leverages 3D protein structural information.
• To improve the accuracy and interpretability of DTA prediction models.
• To establish a new benchmark for structure-based DTA prediction.

Main Methods:

• Developed CASTER-DTA (Cross-Attention with Structural Target Equivariant Representations for Drug-Target Affinity), a joint architecture combining SE(3)-equivariant graph neural networks for protein representation and standard graph neural networks for ligand representation.
• Incorporated an attention-based mechanism for cross-interaction between protein residues and ligand atoms to enhance interpretability.
• Utilized SE(3)-equivariant graph neural networks to learn robust protein representations from 3D structural data.

Main Results:

• CASTER-DTA achieved state-of-the-art performance in predicting drug-target affinity on the Davis and KIBA benchmark datasets.
• The use of SE(3)-equivariant graph neural networks significantly improved protein representation learning for DTA prediction.
• The attention mechanism provided insights into protein-ligand interactions, enhancing model interpretability.

Conclusions:

• CASTER-DTA demonstrates the effectiveness of SE(3)-equivariant graph neural networks and cross-attention for accurate and interpretable drug-target affinity prediction.
• The proposed method outperforms existing approaches without relying on external information like protein language model embeddings.
• This work paves the way for more sophisticated structure-based drug design and screening tools.