Drug Molecular Design Based On Neural Network

With the continuous development of society,human beings increasingly need to accelerate the development of new drugs to overcome complex and changeable viruses and diseases.The most important link in drug development is molecular design,that is,constructing a large number of potential candidate drug molecules and screening the most likely molecules for real experimental testing.Traditional molecular design schemes require the manual construction of candidate molecules with expert knowledge in the field,making molecular design a rate-determining step in drug development.In recent years,benefiting from the rapid development of artificial intelligence and computer hardware technology,neural network-based computational methods related to drug molecular design have been rapidly developed and widely used,and are considered to be the most promising solution to accelerate molecular design.Although many current deep generative models dedicated to molecular design can effectively generate molecules with discrete structures,these models still have problems such as low molecular quality and only considering a single molecular property,which makes the model unlikely to generate real drug molecules.In order to solve these problems in the molecular design of deep learning models,this paper proposes a molecular conditional constrained variational autoencoder model that can be used to generate molecules with multiple target properties,and based on the continuous molecular space learned by the model Uncertainty evolution algorithms were used to verify the performance of the model for generating molecules with multiple properties.Specifically,based on the variational autoencoder,this paper starts from the target chemical properties of molecules in specific applications,and uses gaussian distribution to approximate the distribution of molecular target properties to construct the constraints of variational autoencoder inference.To drive the variational autoencoder to make full use of the target characteristics of the data itself,so as to construct a molecular continuous space that is more in line with the target characteristics and easy to optimize.At the same time,in order to improve the expressiveness of the variational inference model,this paper proposes a block Gaussian variational probability model that is suitable for multiple molecular targets and has a stronger expressive ability and the corresponding inference model method and reparameter sampling techniques with Cholesky decomposition of the block variance matrix to construct an encoder.Furthermore,this paper considers that there are two types of special phenomena in the latent space generated by the model,that is,the continuous space constructed by the model is not complete,which leads to the inability to decode valid molecules for any continuous point and different two variables in the continuous space may be decoded to generate the same discrete molecule.A method to measure the uncertainty in continuous space is introduced to quantitatively express the magnitude of these two types of phenomena,and a continuous evolution operator based on the uncertainty is proposed.In the experimental session,the conditional variational autoencoder model was combined with bayesian optimization and an evolutionary algorithm aware of the continuous space uncertainty of the model,and its performance on single-objective and multi-objective problems in molecular design was verified.And the effectiveness of the proposed model and algorithm is demonstrated by comparing with multiple state-of-the-art models on benchmark problems of two classes of molecular design...