Research On Targeted Drug Design Method Based On Gated Recurrent Unit And Fine-tuning Strategy

Traditional drug research and development faces many challenges,which is a process with high investment,long cycle,and high risk.The return rate of new drug development is decreasing year by year.With the continuous progress of science and technology and the massive growth of biomedical data,deep learning technology,with its powerful learning and representation capabilities,generates drug molecules with desired properties,showing a huge application prospect in the field of drug generation.With the aid of computer-aided drug screening methods,not only can the diversity of generated molecules be increased,but also the generated drug molecules can be evaluated for screening before biochemical experiments,effectively improving the efficiency of drug discovery,which greatly shortens the cycle of discovering new drugs,giving great hope for accelerating the development of new drugs.Using a deep learning algorithm,this paper proposes a targeted drug generation method,GFTP,based on a gated recurrent unit and a Top P sampling strategy.This method is based on a gated recurrent unit GRU（gate recurrent unit）,which trains a large number of drug molecules obtained from the Chembl database,learns the corresponding rules between drug molecules and compounds,uses a fine-tuning strategy to train the active compounds targeting specific target,learns the molecular characteristics of the active compounds,and uses a Top P sampling strategy for molecular sampling to reduce the number of drug molecules that cannot be parsed and that already exist,to improve the efficiency and accuracy of model generation.In addition,in this thesis,the novel coronavirus 3CL^pro protease is selected for algorithm verification.After the generation of new targeted drug molecules,the COVIDVS model is used to score and screen the generated molecules.Finally,molecular docking experiments are carried out on the molecules with high scores to verify their effects on 3CL^pro protease.Good experimental results have been obtained.In order to improve the effectiveness of the above methods,a new targeted drug design method GCTP based on molecular expansion strategy and child-tuning fine-tuning strategy is proposed.In this method,based on the non-singularity characteristics of drugs,a molecular expansion strategy use multiple different forms of entities to represent the same drug,to solve the problem of poor model performance due to the small number of drug molecules targeted at a specific target;The child-tuning strategy is used to solve problems such as poor performance and instability of the model due to size mismatches between the data and parameters of the pre training model and the fine-tuning data.The method proposed in this thesis can effectively learn the characteristics of active compounds targeting specific targets and generate new targeted drug molecules.Results show that the generated molecules have lower average binding energy than existing compounds.