Xu Youjun, Pei Jianfeng, Lai Luhua
Center for Quantitative Biology, Academy for Advanced Interdisciplinary Studies, ‡Beijing National Laboratory for Molecular Sciences, State Key Laboratory for Structural Chemistry of Unstable and Stable Species, College of Chemistry and Molecular Engineering, and ¶Peking-Tsinghua Center for Life Sciences, Peking University , Beijing 100871, China.
J Chem Inf Model. 2017 Nov 27;57(11):2672-2685. doi: 10.1021/acs.jcim.7b00244. Epub 2017 Oct 27.
Median lethal death, LD, is a general indicator of compound acute oral toxicity (AOT). Various in silico methods were developed for AOT prediction to reduce costs and time. In this study, we developed an improved molecular graph encoding convolutional neural networks (MGE-CNN) architecture to construct three types of high-quality AOT models: regression model (deepAOT-R), multiclassification model (deepAOT-C), and multitask model (deepAOT-CR). These predictive models highly outperformed previously reported models. For the two external data sets containing 1673 (test set I) and 375 (test set II) compounds, the R and mean absolute errors (MAEs) of deepAOT-R on the test set I were 0.864 and 0.195, and the prediction accuracies of deepAOT-C were 95.5% and 96.3% on test sets I and II, respectively. The two external prediction accuracies of deepAOT-CR are 95.0% and 94.1%, while the R and MAE are 0.861 and 0.204 for test set I, respectively. We then performed forward and backward exploration of deepAOT models for deep fingerprints, which could support shallow machine learning methods more efficiently than traditional fingerprints or descriptors. We further performed automatic feature learning, a key essence of deep learning, to map the corresponding activation values into fragment space and derive AOT-related chemical substructures by reverse mining of the features. Our deep learning architecture for AOT is generally applicable in predicting and exploring other toxicity or property end points of chemical compounds. The two deepAOT models are freely available at http://repharma.pku.edu.cn/DLAOT/DLAOThome.php or http://www.pkumdl.cn/DLAOT/DLAOThome.php .
半数致死量（LD）是化合物急性经口毒性（AOT）的一个通用指标。为了降低成本和缩短时间，人们开发了各种计算机模拟方法来预测AOT。在本研究中，我们开发了一种改进的分子图编码卷积神经网络（MGE-CNN）架构，以构建三种类型的高质量AOT模型：回归模型（deepAOT-R）、多分类模型（deepAOT-C）和多任务模型（deepAOT-CR）。这些预测模型的性能大大优于先前报道的模型。对于包含1673种化合物（测试集I）和375种化合物（测试集II）的两个外部数据集，deepAOT-R在测试集I上的R值和平均绝对误差（MAE）分别为0.864和0.195，deepAOT-C在测试集I和II上的预测准确率分别为95.5%和96.3%。deepAOT-CR的两个外部预测准确率分别为95.0%和94.1%，而在测试集I上的R值和MAE分别为0.861和0.204。然后，我们对deepAOT模型进行了深度指纹的正向和反向探索，与传统指纹或描述符相比，深度指纹能够更有效地支持浅层机器学习方法。我们进一步进行了深度学习的关键要素——自动特征学习，将相应的激活值映射到片段空间，并通过特征的反向挖掘得出与AOT相关的化学子结构。我们用于AOT的深度学习架构通常适用于预测和探索化合物的其他毒性或性质终点。这两个deepAOT模型可在http://repharma.pku.edu.cn/DLAOT/DLAOThome.php 或http://www.pkumdl.cn/DLAOT/DLAOThome.php 免费获取。
