A point cloud-based deep learning strategy for protein-ligand binding affinity prediction.

Area of Science:

• Computational chemistry and structural biology
• Artificial intelligence in drug discovery
• Machine learning for molecular interactions

Background:

• Protein-ligand binding affinity prediction is crucial for drug discovery.
• Existing methods often require complex feature engineering.
• Deep learning offers a promising avenue for automated feature learning.

Purpose of the Study:

• To apply PointNet and PointTransformer for protein-ligand binding affinity prediction.
• To evaluate the performance of these 3D point cloud-based deep learning models.
• To explore the interpretability of learned features for molecular interactions.

Main Methods:

• Generated 3D point clouds from the PDBbind-2016 database.
• Trained PointNet and PointTransformer models on refined and extended datasets.
• Validated models using the CASF-2016 benchmark and adapted features for XGBoost.

Main Results:

• Achieved high prediction accuracy with Pearson correlation coefficients of 0.795 (PointNet) and 0.833 (PointTransformer) on the extended dataset.
• Demonstrated the models' ability to learn key protein-ligand interaction features, enabling visualization of important atoms.
• XGBoost models incorporating PointTransformer features reached an average Rp of 0.827, comparable to state-of-the-art methods.

Conclusions:

• 3D point clouds from PDBbind are effective for training and evaluating deep learning models.
• PointNet and PointTransformer demonstrate strong capabilities in learning atomic features of protein-ligand interactions.
• These deep learning approaches hold significant potential for advancing drug discovery and molecular interaction studies.