Research On Pharmaceutical Patent Text Analysis Method Based On Reinforcement Learning

In recent years,the pharmaceutical industry has developed rapidly and the demand for the pharmaceutical market has grew continuously under the medical system reform policy.Such backgournd is beneficial for Chinese pharmaceutical companies.However,the relevant data shows that the development of pharmaceutical companies has fallen into a dilemma recently due to time-consuming and money-consuming in the process of drug research and development.Pharmaceutical patents are one of the main channels for pharmaceutical researchers to obtain drug information,and it is a heavy workload to extract the required information from large quantities of text.At the same time,in pharmaceutical industry,the development of new drugs usually requires a large number of experiments or the experience of experts to change the chemical properties of the compounds.However,these highly rely on artificial experience and tend to be inefficient.Therefore,how to effectively control the time cost of drug patent analysis and the experimental cost of drug improvement are of great significance for pharmaceutical companies to develop new drugs and improve their economic benefits.In response to such problems,this study proposes a directed molecular design model based on layered reinforcement learning.The idea of hierarchical reinforcement learning introduced in the model is Option Framework.It can stratify the complex state space and effectively optimize the chemical properties of compounds.In the action option module,this study designs a learning criterion based on the Q-learning method.By continuously updating the action-value function,the optimal action in the action option set(adding,deleting,replacing operations)is selected.Following the rules of SMILES encoding,the problem of SMILES grammar errors in text-based deep reinforcement learning method is overcome.In the state option module,this study proposes the updated strategy of the compound based on the three-step time difference algorithm.The current compound is also updated by the expectation of the next three time steps to improve the operational efficiency of the model.However,changing the molecular main frame structure will affect the molecule.With the consideration of this problem,this study designs a model reward function based on water solubility and structural similarity,which can guide the model to produce the requiredphysical or biological characteristics of the structure.To facilitate researchers to search and analyze patents,a medicines patent analysis system software is developed based on the algorithm proposed in this study and the requirements of pharmaceutical companies.The front-end,back-end,and database are done with React scaffolding,MyBatis-based Spring Boot framework,and MySQL,respectively.The system adopts the B/S architecture,with the introduction and implementation of five modules: user login,basic information query,result refining,SMILES main structure display,and molecular property optimization.The system application developed in this study is still in the experimental simulation stage.Using the system developed in this study for analysis can not only quickly query drug patent information but also optimize the chemical properties of molecules in drug patents,thereby shortening the drug development cycle.