Potentially highly potent drugs for 2019-nCoV

Duc Duy Nguyen; Kaifu Gao; Jiahui Chen; Rui Wang; Guo-Wei Wei

The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.02.05.936013 doi: bioRxiv preprint resources used all over the world. A total of 1553 FDA-approved drugs are contained in the DrugBank. However, in the present work, a number of FDA-approved drugs encountered difficulties in docking with the target molecule. Therefore, the MathPose successfully created 3D protein-ligand complex structures for 1465 FDA-approved drugs and 2019-nCoV protease. MathDL, designed for predicting various druggable properties of 3D molecules [23] , is capable of efficiently and accurately encoding the high-dimensional biomolecular interactions into low-dimensional representations. Algebraic graph theory-based algorithms [25] , differential geometry, and algebraic topology methods [23] are applied to generate three mathematical representations of data in MathDL. These data representations can be integrated with well-designed deep learning models, such as gradient-boosted trees (GBTs) and convolutional neural networks (CNNs), for pose ranking and binding affinity predictions. In D3R Grand Challenges (https://drugdesigndata.org/about/grand-challenge), a worldwide competition series in computer-aided drug design, MathDL had been proved as the top performer in free energy prediction and ranking [22, 23] . Figure 4 illustrates the framework of the MathDL model, which combined the aforementioned mathematical representations with the CNN architecture for druggable properties predictions. The PDBbind 2018 general set [15] , along with the SARS 3CL protease related dataset is used in our training process. In this section, we briefly describe the algebraic topology representation used in the present work. Details can be found in the literature [23] .