Structural Knowledge Is What Matters in Protein-Ligand Binding Affinity Prediction

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning in drug discovery

Background:

• Binding affinity prediction estimates drug-protein interaction strength, crucial for drug discovery.
• Computational methods accelerate compound screening, reducing costly lab experiments.
• Machine learning (ML) approaches are gaining traction over traditional physics-based methods for binding affinity prediction due to computational efficiency.

Purpose of the Study:

• To identify key binary properties of ML models that correlate with higher predictive accuracy (Pearson coefficients).
• To analyze the impact of structural knowledge, 3D information, and drug-protein relationships on model performance.

Main Methods:

• Analysis of diverse ML architectures used in binding affinity prediction.
• Statistical evaluation (t-test) to determine the significance of binary input features.
• Training and testing models on established benchmarks like PDBBind 2016 and CASF 2016.

Main Results:

• Representing the protein (or parts of it) as a graph significantly improves prediction accuracy.
• Including the binding pocket and drug-protein interactions as input features enhances model performance.
• Incorporating atomic distances and chemical bond types into the model positively impacts Pearson coefficients.

Conclusions:

• Specific structural and interaction-based features are critical for high-performance binding affinity prediction using ML.
• Graph-based protein representation, pocket information, and detailed atomic features are key drivers of predictive success.