Dynamic-GLEP: Method of GPCR Ligand Efficacy Prediction Leveraging MD-based Ensemble Docking and Transfer Learning
This repository is the offical implementation of Dynamic-GLEP
G protein-coupled receptors (GPCRs), the largest family of human membrane proteins, play pivotal roles in numerous physiological processes and represent an important class of drug targets. Accurately identifying the efficacy of GPCR ligands is crucial for the development of drugs targeting GPCRs. Many computational prediction methods, when balancing speed and accuracy, often overlook the dynamic nature of interactions between proteins and ligands, which is crucial for the recognition of ligand efficacy.
Here we present Dynamic-GLEP, an approach for predicting the efficacy of GPCR ligands, leveraging dynamic conformational information of complexes and transfer learning (TL) based on equivariant graph neural networks. The core of this method involves considering the dynamic processes during protein-ligand binding and characterizing the representative protein conformational ensemble in multiple dimensions. Considering the rotational-translational equivariance of the molecule and the dynamic features of GPCR ligand binding, we integrated the previously developed Equiscore based on equivariant graph neural networks and transfer learning methods to predict ligand efficacy. Subsequently, we tested Dynamic-GLEP using 5-HT1A receptor as a case study, and the results demonstrated that our method outperformed 11 other methods in external tests. Furthermore, in virtual screening scenario, Dynamic-GLEP achieved a higher enrichment of active molecules compared to other static models.
This code requires the installation of the following packages:
1.python == 3.7.0
2.numpy
3.pandas
4.scipy
5.scikit-learn
6.pytorch
7.tqdm
8.tap (pip install typed-argument-parser)
9.rdkit
Building of EquiScore+TL models are based on EquiScore (https://github.com/CAODH/EquiScore).
All packages can be installed using conda. Neural networks can be trained with CPU or GPU.