Computational toxicology is widely applied to screen tens and thousands of potential environmental endocrine disruptors (EDCs) for its great efficiency and rapid evaluation in recent years. Polychlorinated biphenyls (PCBs) with 209 congeners have not been comprehensively tested for their ability to interact with the thyroid receptor (TR), which is one of the most extensively studied targets related to the effects of EDCs. In this study, we aimed to determine the thyroid-disrupting mechanisms of 209 PCBs through the combination of a novel computational ternary classification model and luciferase reporter gene assay. The reporter gene assay was performed to examine the hormone activities of 22 PCBs via TR to classify their profiles as agonistic, antagonistic or inactive. Thus, six agonists, eleven antagonists and seven inactive chemicals against TR were identified in in vitro assays. Then, six relevant variables, including molecular structural descriptors and molecular docking scores, were selected for describing compounds. Machine learning methods (i.e., linear discriminant analysis (LDA) and support vector machines (SVM)) were used to build classification models. The optimal model was obtained by using SVM, with an accuracy of 88.24% in the training set, 80.0% in the test set and 75.0% in 10-fold cross-validation. The remaining 187 PCB congeners' hormone activities were predicted using the obtained models. Out of these PCBs, the SVM model predicted 38 agonists and 81 antagonists. The findings revealed that higher chlorinated PCBs tended to have TR-antagonistic activities, whereas lower chlorinated PCBs were agonists. This study provided a reference for further exploring PCBs' thyroid effect.

近年来，计算机毒理学被广泛应用于筛选成千上万种潜在的环境内分泌干扰物（EDC），以其出色的效率和快速的评估。尚未对具有209个同类物的多氯联苯（PCB）与甲状腺受体（TR）相互作用的能力进行全面测试，这是与EDC的作用有关的研究最广泛的目标之一。在这项研究中，我们旨在通过结合新型的计算三元分类模型和荧光素酶报告基因检测来确定209种多氯联苯的甲状腺破坏机制。进行该报告基因分析来检查的22个多氯联苯激素活性通过TR将其轮廓分类为激动，拮抗或不活跃。因此，在体外鉴定出六种激动剂，十一种拮抗剂和七种针对TR的非活性化学品分析。然后，选择了六个相关变量，包括分子结构描述符和分子对接分数来描述化合物。机器学习方法（即线性判别分析（LDA）和支持向量机（SVM））用于建立分类模型。通过使用SVM获得最佳模型，训练集的准确度为88.24％，测试集的准确度为80.0％，十倍交叉验证的准确度为75.0％。使用获得的模型预测了剩余的187种多氯联苯同源物的激素活性。在这些多氯联苯中，SVM模型预测了38种激动剂和81种拮抗剂。研究结果表明，较高的氯化多氯联苯往往具有TR拮抗活性，而较低的氯化多氯联苯是激动剂。该研究为进一步探讨多氯联苯的甲状腺作用提供了参考。