Computer Science · Computational Theory and Mathematics
RealVS: Toward Enhancing the Precision of Top Hits\nin Ligand-Based Virtual Screening of Drug Leads from Large Compound\nDatabases
Yueming Yin (11535740), Haifeng Hu (823067), Zhen Yang (107412), Huajian Xu (5008034), Jiansheng Wu (193208)
See on the mapAbstract
Accurate\nmodeling of compound bioactivities is essential for the\nvirtual screening of drug leads. In real-world scenarios, pharmacists\ntend to choose from the top-<i>k</i> hit compounds ranked\nby predicted bioactivities from a large database with interest to\ncontinue wet experiments for drug discovery. Significant improvement\nof the precision of the top hits in ligand-based virtual screening\nof drug leads is more valuable than conventional schemes for accurately\npredicting the bioactivities of all compounds from a large database.\nHere, we proposed a new method, RealVS, to significantly improve the\ntop hits’ precision and learn interpretable key substructures\nassociated with compound bioactivities. The features of RealVS involve\nthe following points. (1) Abundant transferable information from the\nsource domain was introduced for alleviating the insufficiency of\ninactive ligands associated with drug targets. (2) The adversarial\ndomain alignment was adopted to fit the distribution of generated\nfeatures of compounds from the training data set and that from the\nscreening database for greater model generalization ability. (3) A\nnovel objective function was proposed to simultaneously optimize the\nclassification loss, regression loss, and adversarial loss, where\nmost inactive ligands tend to be screened out before activity regression\nprediction. (4) Graph attention networks were adopted for learning\nkey substructures associated with ligand bioactivities for better\nmodel interpretability. The results on a large number of benchmark\ndata sets show that our method has significantly improved the precision\nof top hits under various <i>k</i> values in ligand-based\nvirtual screening of drug leads from large compound databases, which\nis of great value in real-world scenarios. The web server of RealVS\nis freely available at noveldelta.com/RealVS for academic purposes, where virtual screening of hits from large\ncompound databases is accessible.