Aim: Descriptors of molecules are important in the discovery of lead compounds. Most of these descriptors are used to represent molecular structures, although structural formulas are the most intuitive representation. Convolutional neural networks (ConvNets) are effective for managing intuitive information. Results/methodology: Convolutional neural networks (ConvNets) based on two-dimensional structural formulas were used for the preliminary screening of CDK4 inhibitors. After supervised learning of our homemade dataset, our models screened out ten approved drugs, including indocyanine green and candesartan cilexetil, with IC50 values of 2.0 and 5.2 μM, respectively. Conclusion: Depending only on intuitive information, the developed method was shown to be feasible, thus providing a new method of lead compound discovery.
Keywords: CDK4 inhibitors; convolutional neural networks; drug discovery; machine learning; virtual screening.