A Structure-Aware Multimodal Framework for Drug-Target Interaction Prediction via Heterogeneous Graph Learning

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Predicting drug-target interactions is crucial for efficient drug discovery.
• Existing deep learning methods often fail to capture atom-residue-level interactions.
• Integrating diverse data sources can enhance prediction accuracy.

Purpose of the Study:

• To develop a multimodal deep learning framework, PHGDTI, for accurate drug-target interaction prediction.
• To incorporate both sequence and structural information at the atom-residue level.
• To address limitations of current methods in capturing complex molecular relationships.

Main Methods:

• Utilized Mol2Vec and TAPE for sequence embedding, refined with self-attention.
• Developed a graph encoder for drug atom, protein residue, and heterogeneous atom-residue graphs.
• Employed graph attention and SAGPooling for information propagation and representation learning.

Main Results:

• PHGDTI demonstrated superior performance compared to existing methods on the Davis kinase and GalaxyDB datasets.
• Ablation studies confirmed the significant contribution of heterogeneous graph modeling.
• The framework effectively fuses sequence and structural features for precise affinity estimation.

Conclusions:

• PHGDTI offers a powerful approach for drug-target interaction prediction by leveraging multimodal data and heterogeneous graph learning.
• The model's ability to capture atom-residue relationships represents a significant advancement.
• This framework has the potential to accelerate the drug discovery pipeline.