Molecular Simulations Of PD-1/PD-L1 Complex And De Novo Molecular Design Of Small Chemical Inhibitors

Programmed cell death protein 1（PD-1）/programmed cell death ligand 1（PD-L1）mediated signaling pathway involves many physiological processes such as cell proliferation,differentiation,signal transduction and programmed cell death.It has become a key target for the treatment of cancer and immune-related diseases.Although antibody drugs targeting PD-1/PD-L1 have achieved significant efficacies in the clinical treatments of a variety of cancers,they still have many shortcomings,such as lack of oral bioavailability,poor penetration of solid tumors and long half-life.Therefore,small chemical inhibitors targeting the PD-1/PD-L1 signaling pathway have become a hot spot in drug researches and developments.In this paper,molecular simulation and molecular design of PD-1/PD-L1 system have been studied in detail by means of molecular simulation,quantitative structure-activity relationship（QSAR）,deep learning theory based molecular generation network,which provides a theoretical reference for the developments of small-molecule inhibitors targeting PD-1/PD-L1 pathway.Main contents and results are as follows:（1）Prediction of small-molecule inhibitors targeting PD-1/PD-L1.A total of10,232 structures and actitities of small-molecule PD-1/PD-L1 inhibitors were collected from recent literature and patents,which were then applied to establish an optimal prediction model with seven key structural variables by partial least squares discriminant analysis（PLSDA）and stepwise multiple linear regression（SMLR）.The predictive accuracy（Acc）,sensitivity（Sen）and specificity（Spe）of this optimal model were 0.938,0.840 and 0.971 for 8,000 training samples,and 0.940,0.835 and 0.975 for2 232 test samples,respectively.The results showed that intramolecular hydrogen bond,amphoteric index,radius of gyration,non-bonded electrostatic energy,fractional van der Waals surface area of H-bond donors,octanol-water partition coefficient and molecular weight are the key structural characteristics for distinguish the targeted PD-1/PD-L1inhibitors.（2）Molecular simulation of PD-1/PD-L1 complex.The protein docking of PD-1/PD-L1 and PD-L1/PD-L1 were studied by rigid ZDOCK method and flexible Swarm Dock method,respectively.For the PD-1/PD-L1 system,both ZDOCK and Swarm Dock achieved excellent prediction performances with the root mean squared errors（RMSDs）ofα-carbon atoms and main-chain-heavy atoms of the TOP-2lowest-energy conformations less than 2.6（?）.The PD-L1/PD-L1 protein docking studies showed that the Swarm Dock performed slightly better than ZDOCK method,of which the main-chain-heavy-atom RMSD of the Top-1 lowest-energy conformation was only 1.7（?）.（3）Molecular docking of BMS inhibitors.The semi-flexible CDOCKER and Flexible Docking methods have been used to conduct molecular docking researches on a series of BMS inhibitors.The results showed that the Flexible Docking protocol with strong comformation search power can reproduce the near native binding conformations of BMS inhibitors.Besides,a strong linear correlation was observed between IC₅₀values and ligand-receptor interaction energies for these BMS inhibitors（R=0.81）.（4）Deep learning based de novo molecular design and screening of small-molecule PD-1/PD-L1 inhibitors.Based on the SMILES information of 1,048,576small molecules retrieved from Ch EMBL database and 2,222 known PD-1/PD-L1inhibitors,a molecular generation model of PD-1/PD-L1 inhibitors was successfully established by using generative GRU-based RNN network and transfer learning method.The results showed that the percentage of the valid molecules sampled by the pre-trained GRU generative model is above 94%.Then,the pretrained GRU network was refined by transfer learning of 2,222 known inhibitors,which was finally used for sampling a total of 20,000 SMILES strings.After removing repeated structures,a total of 2,763 novel structures were obtained,of which 2,175 compounds（78.72%）were predicted as positive samples by the established PLSDA model.Then,the 2,175 positive samples were further filtered by Flexible Docking mothod,which resulted in 431molecules with potential activities.Besides,4 representative moleculars with high docking scores were used for interaction mechanism researches.The results showed that hydrogen bond,π-πconjugation and hydrophobic interactions are the key interaction modes of the 4 potential inhibitors of PD-1/PD-L1 pathway.