MEGDTA: multi-modal drug-target affinity prediction based on protein three-dimensional structure and ensemble graph...

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Drug development is expensive and slow.
• Accurate drug-target affinity (DTA) prediction is crucial for accelerating drug discovery.
• Existing DTA prediction models often overlook protein 3D structural features.

Purpose of the Study:

• To propose a multi-modal drug-target affinity prediction model (MEGDTA) that incorporates protein 3D structure.
• To improve the accuracy of DTA prediction by effectively extracting features from both drug molecules and target proteins.
• To leverage ensemble graph neural networks (GNNs) for enhanced feature extraction from protein structures.

Main Methods:

• Drugs are represented as molecular graphs and Morgan Fingerprints for feature extraction.
• Protein 3D structures are used to construct residue graphs.
• Long short-term memory (LSTM) networks and parallel GNNs process protein sequences and structures.
• A cross-attention mechanism fuses drug and protein features for final DTA prediction.

Main Results:

• MEGDTA effectively captures diverse features from drug and protein structures, particularly utilizing protein 3D information.
• The model demonstrated strong performance on benchmark datasets (Davis, KIBA, Metz).
• Comparative analysis showed MEGDTA outperforms existing models in key metrics like MSE and CI.

Conclusions:

• MEGDTA's effectiveness in DTA prediction is validated by its strong performance metrics.
• The model's ability to integrate protein 3D structure significantly enhances prediction accuracy.
• MEGDTA represents a promising advancement in computer-aided drug discovery.