Deep Neural Networks For Drug Virtual Screening Large Compound Libraries

High performance ligand-based virtual screening(VS)models have been developed by using various computational methods,including the latest deep neural network(DNN)method.There are high expectations of exploring the advanced capability of DNN for improved VS performance,and this capability has been optimally achieved by using big 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 big data DNN VS models for large compound libraries.In this work,we developed ligand-based big data DNN VS models for the inhibitors of 9 anticancer targets using 0.5 M training compounds.The developed VS models were evaluated by 10-fold cross-validation and achieved 77.89-97.81% sensitivity,99.93-99.99% specificity,0.8206-0.9801 Matthews correlation coefficient,0.9878-0.9983 area under the curve,which outperformed the random forest models.Moreover,DNN VS models developed by the pre-2015 inhibitors identified 50% of post-2015 inhibitors with 0.01-0.09% false positive rate in screening 89 M PubChem compounds,which also outperformed previous works.Experimental assays of the selected virtual hits of the EGFR inhibitor model led to 9 novel EGFR inhibitors with 67.0-99.2% inhibition and 13 EGFR L858 R inhibitors with 53.0-98.4% inhibition at 10 ?M.Further experiment indicated that 6 EGFR inhibitors with 61.4-98.3% inhibition and 8 EGFR L858 R inhibitors with 63.8-99.2% inhibition at 1 ?M.Our results confirmed the usefulness of big data DNN as a ligand-based VS tool to screen large compound libraries.