De Novo Molecular Design Method Based On Deep Learning And Neuroinflammatory Model

Deep learning is a new research direction in the field of artificial intelligence.Deep learning models can obtain their inherent rules by learning massive amounts of data,and are widely used in the computer-aided drug design(CADD)process.Under the background of knowledge automation,CADD is an important means to realize the automation of drug design.The de novo molecular design method is currently the most popular method in CADD.This method can shorten the development cycle and improve the efficiency of research and development.Chronic inflammation mediated by microglia is involved in the pathological process of many chronic neurodegenerative diseases,Therefore,the use of deep learning to construct a de novo molecule generation model based on inhibiting microglia activation has important theoretical and practical significance.This thesis aims to generate new molecules that satisfy semantic validity,drug-like properties and activity(which can inhibit the activation of microglia),namely,to generate new lead molecules.First,a de novo molecule generation model is constructed based on the deep learning network,which aims to generate molecules that satisfy semantic validity and drug-like properties.In order to determine whether a molecule is an active molecule,its activity needs to be judged.Therefore,an activity classification model or regression model must be constructed.Since the number of active molecules currently available is not enough to construct an activity regression model,only an activity classification model can be constructed.By comparing the different combinations of feature selection methods and classification models,it is determined that using the LASSO feature selection method combined with the SVM classification model is the best;and through the fusion of different molecular characterization forms,it is determined that the FCFP6 and BRICS substructure codes are used for the best classification results after fusion;On this basis,this thesis proposes a classification model for inhibiting microglia activation based on molecular characterization forms fusion and feature selection.This activity classification model can accurately determine the activity of molecules.Based on the activity classification model,aiming at the problem of low generation efficiency of new lead molecules,a cyclic finetune model based on inhibiting the activation of microglia was proposed.By adding the newly generated lead molecules after each transfer learning to the target domain,and reusing the model parameters,the model can fully acquire the attribute knowledge of the lead molecules in the target domain.Compared with the existing transfer learning methods,the experimental results verify the effectiveness of the proposed cyclic finetune model.Finally,to solve the problem of low quality of new lead molecule generation,this thesis proposes a reinforcement learning model for the quality of new lead molecules that inhibits the activation of microglia.Use the activity probability value as the activity quality,the qed value as the drug-like quality,and select the policy gradient method as the model strengthening method.By introducing bias in the sequence return,the weight of the high return molecule is increased,construct singleattribute reinforcement models and multi-attribute reinforcement models based on biased policy gradients,the experimental results verify the effectiveness of the proposed reinforcement model.