Quantification of non-covalent interactions (NCI) can greatly facilitate the understanding of protein-ligand binding and thereby support drug design. Although molecular mechanics force field can be used to depict the strength of NCI, there is still an urgency to improve the accuracy to quantum mechanical (QM) level with acceptable time consumption. A previous study on hydrogen bond shows a clue of addressing this problem, demonstrating the linear relationship between hydrogen bond strength and electron density at bond critical point (BCP). However, for other types of NCI, such as pi-pi interactions, there are still gaps in quantifying the interaction strength with electron density characteristics. We hereby report our attempt to bridge the gap by using a machine learning model to quantitatively link NCI strength with electron density topology information at critical points including but not limited to BCP. In our work, GFN2-xTB is used to achieve computational electron density for over 1600 samples including NCI commonly seen in protein-ligand systems. Besides the electron density, other topology properties such as kinetic energy density at inter-molecular critical points including (3, -1), (3, +1), and (3, +3) are extracted as features for our machine learning model. Interaction energies at CCSD/CBS level are used as labels. With the relation coefficient over 0.8 between labels and predicted interaction energies, our approach can sever as a powerful tool to explore various NCI in protein-ligand systems at QM level with limited increase of time cost.

Achieving quantum mechanics level quantification of noncovalent interactions in protein-ligand complex with limited time cost by using machine learning model

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.