Stanojević Mark, Vračko Grobelšek Marjan, Sollner Dolenc Marija
University of Ljubljana, Faculty of Pharmacy, Aškerčeva cesta 7, 1000 Ljubljana, Slovenia; BiSafe d.o.o., V Kladeh 11c, 1000 Ljubljana, Slovenia.
National Institute of Chemistry, Hajdrihova 19, 1000 Ljubljana, Slovenia.
Chemosphere. 2021 Mar;267:129284. doi: 10.1016/j.chemosphere.2020.129284. Epub 2020 Dec 11.
Exposure to endocrine disrupting chemicals is an important public health concern although only a few endocrine disruption chemicals have been identified so far. To speed up their identification, in silico toxicological models appear to be the most appropriate, since the potential endocrine disruption of a large number of compounds can be estimated in a short time. In this study three in silico models (Endocrine disruptome software, VirtualToxLab and COSMOS KNIME) have been used. In silico predictions of the endocrine disruption potential of biocidal active substances have been made and predictions then compared with the available in vitro experimental binding affinities to androgen, estrogen, glucocorticoid and thyroid receptors. The chosen models had similar accuracies (around 60%), while differences were shown between the models in specificity and sensitivity. VirtualToxLab was the most balanced model. Additionally, three combined models were prepared and evaluated. As expected, the majority rule approach model was more accurate and balanced. However, the positive consensus rule model, that improved the specificity of predictions (≥80% for all studied nuclear receptors) was more applicable. This reduction of false positive predictions is especially useful in the search for positive (active) compounds. On the other hand, the novel negative consensus rule model improved the specificity of prediction (≥80% for all studied nuclear receptors), giving good predictions of negative (inactive) compounds that can be excluded from further testing. The results obtained by these combined models have great added value, since they can significantly reduce further experimental testing.
接触内分泌干扰化学物质是一个重要的公共卫生问题，尽管到目前为止仅识别出少数几种内分泌干扰化学物质。为了加快对它们的识别，计算机毒理学模型似乎是最合适的，因为可以在短时间内估计大量化合物的潜在内分泌干扰作用。在本研究中，使用了三种计算机模型（内分泌干扰组软件、虚拟毒理学实验室和COSMOS KNIME）。对杀生物活性物质的内分泌干扰潜力进行了计算机预测，并将预测结果与现有的体外实验中对雄激素、雌激素、糖皮质激素和甲状腺受体的结合亲和力进行了比较。所选模型具有相似的准确率（约60%），但在特异性和敏感性方面模型之间存在差异。虚拟毒理学实验室是最平衡的模型。此外，还制备并评估了三种组合模型。正如预期的那样，多数规则方法模型更准确且更平衡。然而，提高预测特异性（所有研究的核受体均≥80%）的阳性共识规则模型更适用。这种假阳性预测的减少在寻找阳性（活性）化合物时特别有用。另一方面，新型阴性共识规则模型提高了预测特异性（所有研究的核受体均≥80%），对可从进一步测试中排除的阴性（无活性）化合物给出了良好的预测。这些组合模型获得的结果具有很大的附加价值，因为它们可以显著减少进一步的实验测试。
