Interaction model based on local protein substructures generalizes to the entire structural enzyme-ligand space

Area of Science:

• Computational chemistry
• Drug discovery
• Structural biology

Background:

• Chemogenomics aims to understand molecular recognition across the proteome for in silico drug discovery.
• Generalizable methods are needed to study cross-interactions between diverse proteins and molecules.
• Existing methods often struggle with proteins varying greatly in sequence, structure, and function.

Purpose of the Study:

• To present a general quantitative approach for predicting protein-ligand binding affinity.
• To develop a model that spans the entire structural enzyme-ligand space.
• To enable generalized predictions across a wide range of protein targets.

Main Methods:

• Trained a model on a dataset of enzyme-ligand complexes with available crystal structures from public databases.
• Characterized enzymes using local protein structure descriptors of the binding site.
• Described ligands using traditional quantitative structure-activity relationship (QSAR) descriptors.
• Evaluated the model on a curated test set with unknown binding modes, matching enzyme structures to a descriptor library.

Main Results:

• The model predicted binding affinities for external test set enzyme-ligand complexes with an r² of 0.53 and a Root Mean Square Error of Prediction (RMSEP) of 1.5.
• The training and test sets included enzymes from all major classes, covering a wide range of sequences and folds.
• Experimental binding affinities ranged from 0.5 to 11.9 (pKi).

Conclusions:

• The developed quantitative approach can generalize over a wide range of protein targets.
• The use of local descriptors facilitates the creation of predictive models for protein-ligand binding affinity.
• This strategy advances in silico drug discovery by improving the understanding of molecular recognition.