Research And Implementation Of Compound Toxicity Prediction System Based On Molecular Heterogeneous Edge Feature Embedding

The toxicity of a compound,which is closely related to the complex and varied molecular structure of a compound,is often the main cause of failure throughout the research process of a drug from design to marketing.And because molecular structures usually involve multibody interactions and complex electron configurations,the toxicity of compounds is hard to predict.In early studies,multiple molecular fingerprints and molecular descriptors were often used as characteristics to build models and perform predictive tasks in combination with traditional machine learning algorithms,but the results were very limited.With the development of computer softhardware and artificial intelligence,it is found that the use of molecular graph can not only reduce the complexity of molecular structure representation of compounds,but also the artificial network model based on molecular graph has excellent performance.Therefore,in-depth study of the structure of molecular graph is a key entry point to solve the problem of toxicity prediction of compounds.Molecular graph can implicitly capture key interactions between nuclei and electrons in a molecule,giving insight into molecular geometry and properties to predict the toxicity of compounds.More importantly,if today’s computer technology is used to enable drug research and development,it can not only improve the efficiency and accuracy of drug toxicity screening,but also reduce the cost of animal experiments,which is of great significance for the industry.The existing graph neural network assumes that nodes are connected by only one type of edge.However,in the structure of molecular graphs,there is not only one type of edge between nodes due to the different types of chemical bonds.Such changes in the edge-level may improve the performance of the prediction model.Aiming at this theory,the experiment invented a new method to extract the heterogeneous edge features of molecules,and fused the features with the attention neural network to build a model.Firstly,the compound data were preprocessed,then the molecular graph structure was calculated and the molecular heterogeneous edge features were extracted.Finally,the features were weighted and embedded into the fully connected layer.After several iterations,the model was established.At the same time,in order to facilitate the comparison of the model with earlier studies,the experiment also established dozens of classification and regression models based on traditional machine learning algorithms using the same compound data.Based on the above experiments,this article developed a compound toxicity prediction system,which is equipped with innovative neural network models and traditional machine learning models.The system includes two major functional modules: user side and management side.On the user side,users can use toxicity prediction related functions,view modeling datasets,or download public data.On the management side,the super administrator can manage the modeling data set,common user information and user grouping information uniformly after logging in.All system is developed by the Django framework,the front-end page interaction uses H5 and Bootstrap,the back-end data processing uses Python,and the database uses MySQL.