Research On The Generation Method Of Drug-like Molecule Based On Deep Learning

Drugs are playing an increasingly important role in humanity’s long struggle against disease,especially the global pandemic of novel coronavirus infection in recent years.Research and development of new drugs is expensive and time-consuming.The main challenge of traditional drug discovery process is that the search space of chemical molecules is huge and discrete,and the high-throughput screening and virtual screening methods require a lot of computing resources.With the wide application of deep learning,a large number of researchers have tried to introduce deep learning technology into the field of molecular generation.De novo molecular design combined with deep neural networks has become an important approach for drug research and development.However,the process still faces many challenges.Drug molecules,for example,are often converted to sparse,one-hot coding forms during their generation,ignoring the interactions between atoms.Considering the chemical rules of drug molecule generation,the molecular graph coding based on topological structure is more reasonable.In addition,the sequence structure of drug molecules is much more complex than that of ordinary small molecules,and the process of molecule generation requires a longer semantic learning ability of the model.To solve these two problems,based on the deep learning technology,this paper creates the generation model of two kinds of drug molecules,solves the problem of molecular generation process from different perspectives,and develops powerful generation tools.The main work of this thesis is as follows:1.Molecular graph generation model based on generative adversarial network.In order to solve the problem of model collapse in the "single-step" molecular graph generation process of the generative adversarial network model,firstly,regularization method is used to constrain the gradient of data in the environment space to alleviate the disappearance of training gradient.Secondly,a new punishment strategy is proposed to regulate the discriminator and maintain the game balance between "generation" and "identification".Finally,Y-networks are added to obtain parallel input variables,preventing the generation of monotonic values.In order to prove the effectiveness of the proposed method,the corresponding small molecule data were collected for experimental verification.At the same time,the transfer learning mechanism is introduced to generate dopamine receptor D2 active molecules with ideal pharmacological properties.2.Generation of focused drug molecule library using recurrent neural network.In order to further improve the semantic structure ability of the model to extract drug molecules,we proposed a molecular generation model based on nested recurrent neural network.In this model,we abandon the traditional stacking method,introduce nested long-short time memory neural network,and use sequence coding to further strengthen the balance between the distribution of elements and chemical bonds of generated molecules.Furthermore,the proposed method was used to generate drug molecules with chemical structure similarity to SARS-CoV-2 main protease inhibitors,and the active drug molecule library was constructed based on machine learning model.The results of the molecular docking experiment showed that the drug molecules generated by the model had the potential to be the main protease inhibitors of SARS-CoV-2.