Research On Drug Structure Generation Based On Generation Antagonism Network

With the deepening of the research on the structure of compounds and the continuous exploration of the application of compounds,a large-scale database of compounds has been formed,and a large number of compounds may have similar application characteristics.Taking drug application as an example,we know that a certain kind of drug compound can be used for the treatment or treatment of a certain kind of disease or symptom.When a new disease appears,we hope to find a new compound to deal with this problem,or we hope to find other alternative compounds besides the known application scheme.However,it is extremely difficult to find new molecules with required biological activities.In this case,artificial intelligence and generation model have been used in molecular design and compound optimization.In this paper,we propose a generation model based on generation antagonism network,combined with maximum likelihood and evolution to generate new compounds with drug properties.In this model,the maximum likelihood sampling reward is used as the generator goal to reduce the variance of intensive training.In addition,we also hope to strengthen the different indicators of drug characteristics,using genetic thought,so that the generator of each training can strengthen different targets when it generates offspring,so as to generate different targets of the offspring generator,so as to avoid discussing different indicators The problem of weighting in the process of reinforcement.The enhanced generation generator is filtered by the environment fitness function.At the end of each generation change,the generator with better fitness is left.Until the end of iteration,we get a variety of enhanced generators.In this way,we can get molecules with better comprehensive quality and ensure the diversity of data generated.A large number of molecular structures have been obtained through the generation model,but for the existing diseases,the scope is still very large,and it still needs a lot of resources to find the matching compound structure.So in this study,we use a semi supervised interaction network to explore and predict the potential interactions between drugs and diseases.We construct a graph based on the similarity of each drug molecular sequence and disease,and determine the number of interaction isomers between the nodes in the two graphs.Based on these two graphs and known interactions,we calculated the score of the drug disease relationship pair as the prediction result.Experiments show that the model can effectively predict the potential drug disease relationship.