Accurate characterization of ligand binding modes in proteins is a critical step to better understand the details of molecular interactions determining the binding affinity and further optimize the ligand structures in small molecule drug discovery. Computational approaches are widely used to circumvent the cost and time of experimental methods. However, it is still a daunting task to predict ligand binding poses in a receptor conformation that was determined without a ligand (i.e., apo structure) or with a structurally different ligand (i.e., non-cognate receptor). Here, we present an efficient methodology to reliably predict ligand binding poses in apo structures using both molecular dynamics (MD) simulations and deep learning model. Our benchmark performance evaluation shows that utilizing ligand stability and molecular mechanics generalized Born surface area energies during short 1-ns MD simulations outperforms docking scoring in distinguishing near-native ligand poses from the decoys. We also combine the MD simulations with a geometric deep learning models based on molecular graphs for an advanced sampling. The model focuses on the interactions between a binding site and ligand poses as well as intramolecule features represented by their chemical properties. This integration achieves improved efficiency in prediction accuracy, indicating performance complementarity between physics-based and data-driven approaches.

Prediction of ligand binding poses in apo structures by combining physics-based and data-driven approaches

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

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