Drug Molecule Design Based On Deep Generative Model

Drug research and development has been facing the problems of high cost,long drug research and development cycle,and high failure rate,which seriously hinders the effi-ciency of drug research and development.In recent years,with the rapid improvement of computing power and the introduction of molecular data sets,deep generative mod-els have begun to be used in the design of de novo drug molecular design,that is,the design of new molecular compounds with specific properties.However,there are still two problems in the field of de novo drug molecular design:(1)Poor data processing and improper selection of deep generation models,resulting in a low proportion of new molecules generated by the model;(2)In the process of optimizing the specific properties of new molecules,due to the poor pre-training model,the specific property values of the new molecules generated by the model are not high enough and the number is small.In order to solve the above two problems,this paper proposes a Char Gated Recurrent Unit(CGRU)model based on a single-character vocabulary to solve the problem of the low proportion of new molecules generated by the deep generative model.On this basis,it combines transfer learning and reinforcement learning methods to solve the specific prop-erties of new molecules that the value is not high and the number is small.The main work of the thesis is as follows:1.In view of the low proportion of new molecules generated by the current deep gen-erative model,a CGRU model based on a single-character vocabulary is proposed,and the ability of the CGRU model to generate new molecules is verified on three molecular data sets.The problem of the low proportion of new molecules generated by the model mainly comes from two points:poor data processing and improper model selection.In order to solve the problem of poor data processing,this paper uses a single different character in the molecular data set as a "vocabulary" to build a vocabulary,and combines word em-bedding methods to convert molecules represented by strings into vectors;to solve the problem of improper selection of generative models This paper uses a three-layer GRU network to build a deep generative model.Experimental results show that on the three molecular data sets,compared with the existing deep generation model,the ratio of new molecules generated by the CGRU model has increased by 2.88%on average.2.Aiming at the problem of the low Penalized logP(plogp)value of the new molecules currently generated and the small number of molecules with high plogp values,this paper proposes a ZINC250K-CGRU Transfer Learning(Z-CGRU+TL)model.Use a specific data set to perform migration learning training on the pre-trained model(the model after training the CGRU using the ZINC250K data set).The experimental results show that the three highest plogp values of the new molecules generated by the trained Z-CGRU+TL model are 12.45,12.18 and 12.06,compared to the three highest plogp values of the new molecules generated by the existing model,respectively increased by 1.8%,7.9%and 9.1%.3.Aiming at the problem of generating a small number of new molecules with high drug-like properties,this paper proposes a ZINC250K-CGRU Reinforcement Learning(Z-CGRU+RL)model.This paper designs a reward function related to molecular drug properties,combines it with a pre-training model(the model after training CGRU using the ZINC250K data set)and conducts reinforcement learning training.The experimen-tal results show that compared with the existing model that can only generate two new molecules with a drug-like property value of 0.948,the number of new molecules with a drug-like value of 0.948 generated by the Z-CGRU+RL model has increased by 5 times.