Molecular Structure Prediction Based On Graph Neural Networks

With the development of deep learning methods,its application in the field of exploring molecules has become more and more extensive,but most researches are devoted to the prediction of molecular properties using deep learning methods.In fact,the structure of a molecule is inextricably linked to its properties,and more studies have shown that using the spatial information of a molecule can more effectively predict its properties.Moreover,the molecular structure also has an important impact on the formation of molecules and the interaction between molecules,which means that the molecular structure is very important for the discovery of new molecules.Therefore,exploring the molecular structure is of great significance to the fields of biochemistry,material science and drug discovery.For the huge potential chemical space,it is often time-consuming and laborious to use the traditional experimental methods or based on density functional theory to obtain molecular structures.Therefore,this thesis mainly proposes two methods based on graph convolutional neural networks to predict molecular structures.As existing methods are not suitable for downstream tasks such as predicting molecular properties,a new molecular distance matrix prediction model based on graph convolutional neural network is proposed(Distance Matrix Graph Convolutional Networks,DMGCN).The Simplified Molecular Input Line Entry System(SMILES)processes the constructed graph to predict the pairwise atomic distances in the molecule,namely the molecular distance matrix,for the purpose of determining the molecular structure.In order to show the effect of the model,the model is compared with the method in RDkit for calculating the three-dimensional conformation of molecules and Deeper GCN-DAGNN＿dist which is a deep learning method,and the mean absolute error is superior to them.Although the first method has achieved good prediction results,its application scope is relatively limited.In view of this,the second method proposes a molecular structure prediction model based on graph convolutional neural network(Molecular Structure Graph Convolutional Networks,MSGCN).This model can predict molecular structure by predicting atomic coordinates.Also in comparative experiments,the method in RDKit achieves the best results among the three methods,followed by MSGCN,and finally Deeper GCNDAGNN＿coord which is a deep learning method.Although the method in RDKit achieves the best results,it is not possible to generate any molecular conformation,so the method in RDKit and MSGCN have their own advantages.Finally,in order to facilitate the use of the two methods proposed in this thesis,a molecular structure prediction system is designed and implemented,and the model trained in this thesis is embedded into the system.The system can accept user input,return prediction results,and provide users with the function of downloading prediction results.