Application of FMO for Protein–ligand Binding Affinity Prediction

Abstract

The fragment molecular orbital (FMO) method has emerged as a powerful computational tool for structure-based drug design. Pair interaction energy decomposition analysis (PIEDA) enables detailed analysis of protein–ligand interactions, and many studies have shown that interaction energies can be used to predict protein–ligand binding affinities. However, the accuracy is insufficient for application to lead optimization. To increase the method’s accuracy, we introduce an ensemble FMO method in which molecular dynamics simulations are used to generate multiple protein–ligand complex structures, and FMO calculations are performed for ensembles of conformers with explicit water molecules. To assess the ensemble FMO method, we examined the correlations between experimental and calculated binding affinities of two systems, internal project A and Pim1 kinase. The correlations between experimental pIC₅₀ values and FMO-based interaction energies in vacuo calculated based on MM-optimized X-ray crystal structures were R² = 0.28 and R² = 0.53 for internal project A and Pim1, respectively. Using the ensemble FMO method, the correlation for internal project A improved (R² = 0.67), whereas the correlation for Pim1 was unchanged (R² = 0.53). If combinations of different PIEDA energies were allowed, the best correlations for internal project A and Pim1 were R² = 0.76 and R² = 0.62, respectively. We also discuss the application of a reinforcement learning method, Best Arm Identification, in which the performance of the ensemble FMO method was maximized by avoiding unpromising compounds in the early stages to allocate limited computational resources to more-promising compounds.