Development of Ligand-based Big Data Deep Neural Network Models for Virtual Screening of Large Compound Libraries

High-performance ligand-based virtual screening (VS) models have been developed using various computational methods, including the deep neural network (DNN) method. There are high expectations for exploration of the advanced capabilities of DNN to improve VS performance, and this capability has been optimally achieved using large data training datasets. However, their ability to screen large compound libraries has not been evaluated. There is a need for developing and evaluating ligand-based large data DNN VS models for large compound libraries. In this study, we developed ligand-based large data DNN VS models for inhibitors of six anticancer targets using 0.5 M training compounds. The developed VS models were evaluated by 10-fold cross-validation, achieving 77.9-97.8 % sensitivity, 99.9-100 % specificity, 0.82-0.98 Matthews correlation coefficient and 0.98-0.99 area under the curve, outperforming random forest models. Moreover, DNN VS models developed by pre-2015 inhibitors identified 50 % of post-2015 inhibitors with a 0.01-0.09 % false positive rate in screening 89 M PubChem compounds, also outperforming previous models. Experimental assays of the selected virtual hits of the EGFR inhibitor model led to reasonable novel structures of EGFR inhibitors. Our results confirmed the usefulness of the large data DNN model as a ligand-based VS tool to screen large compound libraries.