A Multimodal Drug-Target Affinity Prediction Framework with Pretrained Models and Hierarchical Graph Transformer

Area of Science:

• Computational chemistry
• Drug discovery and development
• Bioinformatics

Background:

• Drug-target affinity (DTA) prediction is essential for understanding drug-target interactions in drug discovery.
• Current DTA prediction methods struggle to capture global structural patterns from molecular graphs, often lacking 3D structural data, which limits accuracy and generalizability.

Purpose of the Study:

• To develop a novel multimodal framework, PMHGT-DTA, for accurate drug-target affinity prediction.
• To address the limitations of existing methods by incorporating 3D structural information and advanced graph representation learning.

Main Methods:

• Proposed a multimodal framework, PMHGT-DTA, integrating pretrained models and a hierarchical graph transformer (HGT).
• Utilized graph neural networks (GNNs) and transformers to represent local and global structural information in molecular graphs.
• Incorporated 3D drug conformation graphs and binding site-focused protein graphs, complemented by sequence features.
• Employed a cross-attention module to model drug-atom and protein-residue interactions for enhanced interpretability.

Main Results:

• PMHGT-DTA demonstrated superior performance compared to baseline methods on the Davis and KIBA benchmark datasets.
• The framework achieved high accuracy in both standard and real-world DTA prediction scenarios.
• The cross-attention mechanism provided interpretable insights into drug-target relationships.

Conclusions:

• The PMHGT-DTA framework effectively predicts drug-target affinity by integrating multimodal data and advanced graph transformer architectures.
• This approach enhances model accuracy and generalizability by leveraging 3D structural information.
• PMHGT-DTA shows significant potential to accelerate the drug discovery and development process.