Research On Molecular Properties Prediction Based On Graph Contrastive Learning

In recent years,the COVID-19 epidemic has spread on a large scale around the world,threatening human health and affecting social stability.Therefore,in response to the emergence of similar viruses,human beings urgently need to develop corresponding drugs and vaccines.Molecular property prediction has always been an important task in the field of drug design and development.The traditional new drug research and development cycle and high cost.Therefore,how to effectively and accurately predict the properties of molecules has become a common concern of researchers.Due to the excellent application of machine learning and deep learning,they have become important tools in many fields,including various algorithms for molecular property prediction.Graph neural networks have excellent computing power.They can not only capture the complex structure of molecules,but also abstract them.Therefore,they can not only be used for large-scale datasets,but also capture the characteristics of molecules more accurately.This is far superior to traditional machine learning algorithms.However,the deep learning algorithms of the graph depend on the data with a large number of labels.In reality,it is difficult to determine the characteristics of the molecule.It takes a lot of manpower,material and financial resources to mark the molecular dataset.As contrastive learning shines in the fields of computer vision and other fields,it is increasingly attracted to the attention of researchers.Recently,researchers have tried to apply the method of contrastive learning to graph data,and have achieved good results.However,the existing graph contrastive learning methods rely heavily on data augmentation and require manual selection of augmentation methods for each dataset,or through some empirical methods.In order to avoid manual selection of data augmentation methods and allow the model to automatically select negative samples during the training phase,inspired by graph contrastive learning,the research content of this dissertation is as follows:1.We propose a molecular property prediction model(Au Co)based on automated graph contrastive learning.First of all,we optimize the predefined sampling distribution through the Bayesian method and then automatically select the enhancement scheme.Further,we use a scoring function to evaluate the difficulty of negative samples,and automatically select negative samples in order from easy to hard during each training process,and then use a graph neural network encoder to extract more robust molecular feature representations.The experimental results show that compared with the existing graph contrastive learning methods,the prediction accuracy is improved.2.We propose a molecular property prediction model(Au Co-CEL)based on automated graph contrastive learning and ensemble learning classifiers.We consider that many of the current methods use a single classifier to predict the properties of the molecules.In order to improve the accuracy of the classification,we use the ideas of ensemble learning,integrate several random forests through the Adaboost algorithm,and finally obtain a strong classifier to replace the tradition single classifier.Finally,we use Au Co to learn feature representations for graph data,and then combine with ensemble learning-based classifiers for graph classification experiments.The experimental results show that our method has achieved relatively ideal results,which proves that our method is effective.Finally,in this dissertation,we review past research on molecular property prediction and suggest how it can be improved in the future.