Molecular Generation Optimization Based On Deep Learning And Its Application

Due to the diversification and complexity of diseases and the frequent occurrence of drug resistance,the need for drugs is increasing.Small molecule drugs are the focus of drug research and development because of many advantages.For the research and development of small molecule drugs,one strategy is to find drugs with new characteristics from the drugs already on the market.This method can save time and cost to some extent,but it may not achieve the expected effect.Therefore,for a certain disease,de novo molecular design is another important strategy.For example,the COVID-19 continues to be a global pandemic,and its high infectivity and pathogenicity seriously threaten people's lives and health.In addition to developing vaccines,we urgently need to find effective drugs to treat the disease.However,the traditional research and development of new drugs often face problems such as high investment and long time.In recent years,due to the excellent result of deep learning in many fields,researchers try to introduce deep learning into the field of de novo molecular design to improve efficiency and have achieved a lot of results.However,there are still many challenges.For example,many target-specific generative models mainly focus on the known inhibitors and thus produce similar molecules.However,derivatives of these known inhibitors are likely to be ineffective.Considering the cost of chemical synthesis and experimental validation,the low false-positive rate of generative molecules is very important.In addition,the generated molecules usually need to satisfy multiple properties to achieve the desired effect.Multi-objective optimization is another difficulty of molecular generation.The data of multi-objective optimization is often less,which may seriously affect the performance of the deep learning model.Given the above two problems,the main research contents of this paper are as follows:1.A molecular generation method based on unbalanced dataset is proposed:firstly,a new generation model architecture of generator-filter-predictor is proposed.The method uses transfer learning to optimize the properties of generated molecules.In order to reduce the false positive rate of generated molecules,we used unbalanced data to train the predictor model so that the model can learn the knowledge of a large number of negative samples.So the performance of the predictor is very important.In order to alleviate the impact of unbalanced datasets,a two-stage training process of pretraining-finetuning is designed for the predictor.Then,we collected relevant datasets and carried out molecular generation experiments to verify the effectiveness of the proposed method.2.A molecular generation method based on multi-objective optimization is proposed:Aiming at the possible lack of data in multi-objective optimization,a multi-task model based on graph attention mechanism is proposed,which can efficiently model multiple properties of molecules.Then,using the multitasking model as a predictor,the molecular generation model of multi-objective optimization is trained based on two different methods:reinforcement learning and substructure splicing.Further,relevant experiments are carried out by using the above methods in order to optimize the two properties of the generated molecules at the same time.