Encoding Protein-Ligand Interactions: Binding Affinity Prediction with Multigraph-based Modeling and Graph Convolutional Network
Preprint
- 25 October 2023
- preprint
- Published by American Chemical Society (ACS) in ChemRxiv
Abstract
Machine learning models are increasingly harnessed to expedite drug discovery by effectively predicting the properties of small molecules, such as pKa, solubility, and binding affinity. While these models offer accelerated compound identification and optimization, the challenge arises when studying properties influenced by the interaction between a ligand and its corresponding protein. In response to this challenge, graph neural networks (GNNs) have emerged as a solution for integrating 3D structural data to enhance our understanding of protein-ligand interactions. Our novel model, InterGraph, introduces a unique approach to modeling protein-ligand interactions as topological multigraphs. This approach provides a comprehensive representation of the intricate spatial organization and connectivity patterns within these systems. By incorporating "interaction spheres" with varying edge densities, we capture the proximity-based influence of interactions, filtering out those occurring beyond 9 Å from the ligand, which are deemed irrelevant. Our model was trained using a ligand binding dataset from PDBbind and achieved an RMSE value of 1.34 when tested on a hold-out dataset. Our results highlight the efficacy of the multigraph in encoding the significance of close interactions, a critical factor in understanding binding affinity. On average, our model accurately predicts binding affinity values for various protein-ligand complexes and demonstrates heightened accuracy for hydrolase, lyase, and families of proteins involved in mediating protein-protein interactions. Furthermore, when compared to a set of complexes that underwent redocking calculations, the InterGraph method displayed sensitivity to the binding mode. In the context of drug discovery and development, protein-ligand interactions are pivotal in numerous biological processes. Understanding how these interactions affect biological processes is central to advancing pharmaceutical research. Binding affinity, a thermodynamic property characterizing the strength of such interactions, offers invaluable insights into these intricate processes.Keywords
All Related Versions
This publication has 26 references indexed in Scilit:
- PLANET: A Multi-objective Graph Neural Network Model for Protein–Ligand Binding Affinity PredictionJournal of Chemical Information and Modeling, 2023
- Are Deep Learning Structural Models Sufficiently Accurate for Virtual Screening? Application of Docking Algorithms to AlphaFold2 Predicted StructuresJournal of Chemical Information and Modeling, 2023
- Graph neural network approaches for drug-target interactionsCurrent Opinion in Structural Biology, 2022
- Synergy and Complementarity between Focused Machine Learning and Physics-Based Simulation in Affinity PredictionJournal of Chemical Information and Modeling, 2021
- InteractionGraphNet: A Novel and Efficient Deep Graph Representation Learning Framework for Accurate Protein–Ligand Interaction PredictionsJournal of Medicinal Chemistry, 2021
- AutoDock Vina 1.2.0: New Docking Methods, Expanded Force Field, and Python BindingsJournal of Chemical Information and Modeling, 2021
- Accurate Prediction of GPCR Ligand Binding Affinity with Free Energy PerturbationJournal of Chemical Information and Modeling, 2020
- Binding Affinity Prediction by Pairwise Function Based on Neural NetworkJournal of Chemical Information and Modeling, 2020
- Comparative Assessment of Scoring Functions: The CASF-2016 UpdateJournal of Chemical Information and Modeling, 2018
- AutoDock Vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreadingJournal of Computational Chemistry, 2009