Toxicity Prediction Of Small Molecules Based On Deep Learning Algorithms

In the field of drug discovery,more than 30 percent of promising pharmaceuticals have failed in human clinical trials because of their toxicity,despite of their promising pre-clinical studies in animal tests.It can be seen that the safety evaluation of drugs,especially early toxicity testing,plays a very important role in the development of new drugs.As reported,there are two main toxic signaling pathways: the stress response(SR)pathway and the nuclear receptor(NR)pathway.In this study,the antioxidant response element(ARE)response and Androgen Receptor(AR)response were selected as the research objects.The deep learning algorithms and traditional machine learning algorithms were used to establish the prediction models of these two types of toxicity responses.A series of predictive models on these two responses was obtained with satisfactory results.The main contents of the dissertation can be summarized in the following aspects:The first section presents the current research status of toxicity prediction and the basic principles of traditional machine learning algorithms and deep learning methods.In the second section,a series of predictive models based on multiple deep learning methods and traditional machine learning methods were constructed to predict the ARE responses of compounds by using Tox21 challenge data set.At first,two kinds of molecular features were calculated including traditional molecular structure descriptors and molecular fingerprints.In the development of models based on the traditional molecular descriptors,the Gini index was used to screen the structural descriptors closely related to the ARE response to prevent the over-fitting problems in traditional machine learning methods.However,in the development of models based on deep learning algorithm,due to the unique advantage of the deep learning algorithm,we didn’t need to perform feature selection,but introduced some effective model optimization strategies to construct the ARE response models.And the correlation between the molecular descriptors and their corresponding relationship on ARE response was also analyzed.For the development of model based on molecular fingerprints,all fingerprints were used to construct traditional machine learning and deep learning models without feature selection.The confusion matrix and the area under Receiver Operating Characteristic curve(ROC-AUC)were used to evaluate the reliability of the models.The Deep Neural Network(DNN)model based on fingerprint features give the best prediction ability,with the accuracy of 0.992,0.914 and 0.917 for the training set,test set and validation set,respectively.Consequently,this robust model can be adopted to predict the ARE response of molecules fast and accurately,which is of great significance for the evaluation of safety of compounds in the process of drug discovery and development.In the third part,the classification models to predict the compounds’ response of AR pathway were developed by using a variety of deep learning algorithms and other machine learning algorithms based on the AR response data in Tox21 challenge.The results show that the Recurrent Neural Network(RNN)model based on molecular fingerprint features has the best predictive ability.The predictive accuracy of this model is 0.997,0.980,and 0.976 for the training set,test set,and external validation set,respectively,indicating that the predicted results are very reliable.In addition,we compared our models with the best model of DeepTox pipline in Tox21 challenge in 2014.The results indicate our model is much better than the best model of DeepTox pipline,indicating that our constructed model is very satisfactory.The results obtained in this thesis can provide reliable models with good generalization ability for the prediction of compound toxic responses on the two kinds of pathways,and also indicate that deep learning methods are promising in small molecule toxicity prediction.