Quantitative surface field analysis: learning causal models to predict ligand binding affinity and pose.

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Structure-activity relationship (SAR) data is crucial for understanding ligand-target interactions.
• Existing methods for modeling ligand binding pockets often rely on detailed structural information or fragmented approaches.

Purpose of the Study:

• To introduce the QuanSA (Quantitative Structure-Activity) method for inducing field-based models of ligand binding pockets using only SAR data.
• To address the challenge of mutual ligand alignment and optimize model parameters and ligand poses via machine learning.

Main Methods:

• QuanSA utilizes a learned scoring field, differing from fragment-based pocket construction.
• Multiple-instance machine learning is employed for parameter and pose optimization.
• The method was validated on sixteen diverse structure-activity data sets.

Main Results:

• QuanSA successfully induced physically meaningful models from SAR data alone.
• Models extrapolated from small datasets identified potent, novel ligands with high specificity.
• Combining QuanSA predictions with physics-based simulations showed synergistic benefits.

Conclusions:

• QuanSA offers a powerful, data-driven approach for drug discovery and lead optimization.
• The method provides valuable predictions including binding affinities, ligand strain, poses, novelty, and confidence.
• QuanSA is applicable for both identifying new leads and refining existing drug candidates.