Drug Discovery And Optimization Based On Virtual Screening And Deep Generative Models

Drug discovery and development is faced with many challenges,including high cost,long time and high risk,which also leads to the decline of the rate of return of drug development year by year.Based on computational chemistry and combining molecular mechanics,quantum mechanics and structural biology,computer aided drug design has developed a series of computational simulation analysis methods,which have been widely used in drug discovery and optimization of lead compounds and compound ADME/T prediction,etc.Based on massive accumulation of omics data,artificial intelligence technology has attracted wide attention in the field of drug research and development due to its powerful characterization,learning and classification capabilities,and has been widely applied in different stages of drug research and development.Virtual screening is a traditional method of computer aided drug design,which has many successful cases in drug research and development.With the continuous improvement of computing methods,virtual screening methods have also been developed.We summarize the classification and development of virtual filtering methods in the first section of the first chapter.Artificial intelligence technology has been widely used in many fields of new drug research and development,such as protein structure prediction,molecular design and optimization,target prediction,chemical reaction prediction,drug properties prediction,etc.The goal of drug research and development is to find compounds with ideal pharmacological properties.Deep generative models have shown great application prospects in de novo molecular design and have attracted extensive attention in academia and industry.In the second and third sections of the first chapter,the development and application of artificial intelligence technology in drug design are summarized,with emphasis on the method,development and application of deep generation model.As a new tool in the field of drug design,deep generative models can directly generate new molecules with desired properties through powerful characterization learning ability,which has great application potential.One of the hallmarks of cancer cells is abnormal energy metabolism.Even under aerobic conditions,cancer cells still supply energy through glycolysis,a phenomenon known as the Warburg effect.GAPDH is the key rate-limiting enzyme involved in the sixth step of the glycolysis process.A large number of studies have shown that inhibition of GAPDH can prevent the further development of cancer,and it’s of great significance and research value for anti-tumor therapy by inhibiting GAPDH enzyme activity.In the second chapter,a structure-based virtual screening model was established,and an active compound,DC-5163,with an IC₅₀ of 176.3 n M,was screened through the model.Through molecular simulation,we proposed the possible binding mode of this molecule to GAPDH.A series of compounds were purchased from commercial libraries for structure-activity relationship analysis using similarity search method.Through cell proliferation inhibition experiment,we found that DC-5163 can inhibit the proliferation of cancer cells,and has no effect on normal cells.By inhibiting GAPDH enzyme activity in cancer cells,DC-5163 blocks the process of glycolysis of cancer cells and induces apoptosis of cancer cells,thus achieving the anti-tumor therapeutic effect.Schizophrenia is a complex psychiatric disease affecting approximately 1%of the world’s population,and it requires drugs targeting multiple GPCRs to modulate therapeutically complex neuropsychiatric functions.Conventional drug discovery methods have some limitations,such as high throughput screening workload,high cost,limited virtual screening library,etc.,making it difficult to achieve reasonable multi-target design.In the third chapter,based on artificial intelligence technology and big data,we established an automatic design platform for multitargeted antipsychotic drugs.We first established an activity prediction platform for designed molecules based on multi-task neural network,and then established a molecular generation model based on recurrent neural network for the automatic design of the multi-target GPCR molecular library.Then the generated molecules were evaluated and filtered by their properties like predictive activity,synthetic accessibility,drug-likness,etc.,so as to form an iterative cycle of generation-virtual screening,and continuously improve the molecular activity.Finally,the candidate compound 8 was obtained by the platform.Compound 8exhibites multi-target equilibrium activity and dose-dependent reversal of PCP-induced hyperspontaneous activity in mice,which is a promising candidate compound for schizophrenia.Molecular generative models are attracting great attention as a promising in silico molecular design tool for assisting drug discovery.Dysregulation of DDR1 is associated with a number of human diseases,including inflammatory diseases（atherosclerosis,osteoarthritis,and organ fibrosis）,pancreatic,gastric,and non-small cell lung cancer,etc..In our previous studies on FGFR,based on the kinase selectivity profile of compound 1,we found that compound 1 exhibits weak inhibitory activity against DDR1.Through molecular simulation studies,we found that compound 1formed three key hydrogen bonds with the hinge region of DDR1 protein,and there is much more room for the tail and side chain of the compound to optimize and expand.In the fourth chapter,we optimized the molecular structure of compound 1 based on deep generative model.A molecular generative model based on Seq2Seq model was established,when the scaffold was input to the model,the model decorated the first attachment point in the scaffold SMILES string,the generated decoration was then joined back to the scaffold and the half-built molecule was fed back in the generative model.The process was repeated until all attachment points were decorated,so as to generate a scaffold-based virtual molecular library.Subsequently,virtual screening of the molecular library based on the kinase spectrum prediction model and molecular docking was performed.Compound 2 was finally screened,which exhibited excellent kinase selectivity and inhibited the expression of proinflammatory factors and DDR1autophosphorylation in cells.Compound 2 showed significant therapeutic protection from DSS-induced inflammatory bowel disease in vivo,and the efficacy was comparable to that of the positive drug.