Abstract
In the drug discovery process, traditional structure-based virtual screening methods, such as molecular docking, are often limited by empirical scoring functions. Recent studies have demonstrated that target-specific scoring functions developed using machine learning approaches can enhance the accuracy of virtual screening. Furthermore, the extrapolation performance of these scoring functions is crucial for their broader applicability. Therefore, the study tried combining molecular graph and convolutional neural networks as a way to improve the extrapolation ability of target-specific scoring functions in the face of data expanded within a certain range of chemical space. Taking cGAS and kRAS proteins as examples, through rigorous data screening and feature extraction, the study constructed multiple supervised learning models containing traditional machine learning models, and deep learning models like graph convolutional networks. The results show that compared with the generic scoring functions, these target-specific scoring functions showed significant superiority. In addition, the target-specific scoring functions also exhibit remarkable robustness and accuracy in determining whether a molecule is active. This indicates that the graph convolutional network can be generalized to the prediction of heterogeneous data based on the complex patterns of molecular protein binding that have been learned. The comprehensive performance evaluation of different target-specific scoring functions shows that they hold significant potential for applications in structure-based virtual screening. In particular, graph convolutional networks was demonstrated to greatly improve the screening efficiency and accuracy of target-specific scoring functions for targets such as cGAS and kRAS.