ASFP

ASFP constructs customized scoring functions to improve protein–ligand docking predictions and binding affinity estimation for structure-based virtual screening.

Key Features:

• Descriptor Generation Module: Generates up to 3437 descriptors that model protein–ligand interactions.
• AI-Based SF Construction Module: Applies three distinct machine learning techniques to build target-specific scoring functions from the pre-generated descriptors.
• Online Prediction Module: Offers pre-built target-specific scoring functions and a generic scoring function for binding affinity prediction.

Scientific Applications:

• Target-specific virtual screening: Enhances structure-based virtual screening with target-specific scoring functions validated to an average ROC AUC of 0.841 across 32 targets.
• Binding affinity prediction: Enables binding affinity estimation with a generic scoring function that attains a Pearson correlation coefficient of 0.81 on the PDBbind v2016 core set.
• Structure-based drug discovery and structural biology: Facilitates development of tailored scoring functions to improve prediction of protein–ligand interactions in drug discovery and structural biology studies.

Methodology:

Generate 3437 protein–ligand interaction descriptors, train three machine-learning models to construct target-specific and a generic scoring function, and validate performance on benchmark datasets reporting ROC AUC and Pearson correlation (including PDBbind v2016 core set).