DeepMice: a novel protein-ligand molecular docking model based on multilevel mapping modules

Our study presents DeepMice, a novel artificial intelligence-based molecular docking framework designed to predict protein-ligand binding conformations with improved accuracy. DeepMice's scoring function utilizes a graph transformer network (GTN) as its backbone. It transforms residue-level representations into atomic-level representations, enhancing representation precision. A multilevel mapping module is incorporated to reduce the graph model's size and computational complexity. Subsequently, the mixture density network (MDN) is employed to further realize scoring prediction. In terms of conformational search, DeepMice employs a hybrid strategy combining global heuristic search and local gradient-based optimization. The process initiates with a global exploration using the Differential Evolution (DE) algorithm, followed by local refinement via the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm. This combined approach enhances conformational search efficiency. Performance tests on the DEKOIS2.0 and DUD-E datasets showed that DeepMice outperformed existing virtual screening technologies such as Glide SP and RTMScore in terms of area under the receiver operating characteristic curve (AUROC), boltzmann-enhanced discrimination of receiver operating characteristic (BEDROC), and enrichment factor (EF) values. In particular, DeepMice demonstrates advanced molecular docking capabilities in the CASF-2016 standard test set. In addition, DeepMice considers the multiscale structure of proteins, optimizing the conformation scoring process and improving docking efficiency. In summary, DeepMice is an efficient and accurate molecular docking model, which is expected to accelerate the process of new drug research and development. The program based on the DeepMice model, which is now freely available at https://www.deepmice.com , provides a powerful tool for drug discovery.