Lhasa Limited, Granary Wharf House, 2 Canal Wharf, Leeds, LS11 5PS, UK.
Lhasa Limited, Granary Wharf House, 2 Canal Wharf, Leeds, LS11 5PS, UK.
Regul Toxicol Pharmacol. 2018 Jun;95:227-235. doi: 10.1016/j.yrtph.2018.03.015. Epub 2018 Mar 24.
A previously published fragmentation method for making reliable negative in silico predictions has been applied to the problem of predicting skin sensitisation in humans, making use of a dataset of over 2750 chemicals with publicly available skin sensitisation data from 18 in vivo assays. An assay hierarchy was designed to enable the classification of chemicals within this dataset as either sensitisers or non-sensitisers where data from more than one in vivo test was available. The negative prediction approach was validated internally, using a 5-fold cross-validation, and externally, against a proprietary dataset of approximately 1000 chemicals with in vivo reference data shared by members of the pharmaceutical, nutritional, and personal care industries. The negative predictivity for this proprietary dataset was high in all cases (>75%), and the model was also able to identify structural features that resulted in a lower accuracy or a higher uncertainty in the negative prediction, termed misclassified and unclassified features respectively. These features could serve as an aid for further expert assessment of the negative in silico prediction.
先前发表的一种用于进行可靠的计算机虚拟阴性预测的片段化方法已应用于预测人类皮肤致敏的问题，该方法利用了一个包含超过 2750 种化学物质的数据集，这些化学物质具有来自 18 种体内试验的公开可用的皮肤致敏数据。设计了一个试验层次结构，以便能够对该数据集中的化学物质进行分类，分为致敏剂或非致敏剂，只要有多个体内试验的数据可用。使用 5 倍交叉验证对内进行了阴性预测方法的验证，并针对制药、营养和个人护理行业成员共享的体内参考数据的大约 1000 种化学物质的专有数据集进行了外部验证。在所有情况下，该专有数据集的阴性预测率都很高（>75%），并且该模型还能够识别导致阴性预测准确性降低或不确定性增加的结构特征，分别称为错误分类和未分类特征。这些特征可以作为进一步对计算机虚拟阴性预测进行专家评估的辅助。
