SAnDReS 2.0: Development of machine-learning models to explore the scoring function space.

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning

Background:

• Classical scoring functions often lack accuracy in predicting protein-ligand binding affinity.
• Machine learning models offer improved predictive performance for specific protein systems when trained on structural and affinity data.

Purpose of the Study:

• To introduce SAnDReS, a novel methodology for developing machine learning models to predict binding affinity.
• To explore the scoring function space by integrating AutoDock Vina with various regression methods.

Main Methods:

• SAnDReS combines AutoDock Vina 1.2 with 54 Scikit-Learn regression methods.
• Machine learning models are generated using crystal, docked, and AlphaFold-predicted protein-ligand structures.
• The performance of SAnDReS-generated models is evaluated through three case studies.

Main Results:

• SAnDReS-generated models demonstrated superior performance compared to classical scoring functions in all case studies.
• The predictive accuracy of SAnDReS models was comparable or better than existing machine learning models like K_DEEP, CSM-lig, and Δ_VinaRF₂₀.

Conclusions:

• SAnDReS provides an effective approach for developing accurate binding affinity prediction models.
• The methodology enables exploration of scoring function space for enhanced drug discovery and molecular modeling.