论文部分内容阅读
Understanding xenobiotic interactions with nuclear receptors (NRs) is important in the context of endocrine disruptors and environmental toxicity assessment.1 The related NRs are androgen receptor (AR), estrogen receptor (ER).Due to the limited experimental data of the three receptors binding chemicals, high cost and time consuming, in silico virtual screening endocrine disrupting chemicals is encouraged.In the current study, computational models for predicting AR, ER binders, were developed using k-nearest neighbor (kNN), C4.5 DT(C4.5 Decision Tree), na(i)ve Bayes (NB) and support vector machine (SVM) algorithms combining with seven fingerprints.Therefore, 35 models were built for each receptor.The algorithms and calculation of fingerprints were carried out by open source software tools.Five-fold cross validation of training sets determined the prediction overall accuracies of 0.65-0.85 for AR binders, 0.68-0.83 for ER binders.Then extemal validation was performed to test the predicting ability of the models.Taking the results of five-fold cross validation and external set validation into consideration, the best models for AR binders was PubChem Fingerprint-SVM with prediction accuracy of 0.84, for ER binders was Fingerprint-SVM of 0.86.Moreover, the combination of information gain and substructure fragment analysis2 was used to indentify several substructure alerts.Phenolic ring, aromatic ring, annelated ring, alcohol etc.were picked up to distinguish binders and non-binders of nuclear receptors.Thus, our study suggested that the machine learning methods and fingerprints are potentially useful to predict EDCs interacting with AR, ER.