Abstract：
The antidepressant drug known as 5-HT reuptake inhibitor (5-HT-RI) was commonly detected in biological tissues and result in significant adverse health effects. Homology modeling was used to characterize the functionalities (efficacy and resistance), and the adverse outcome pathway was used to characterize its human health interferences (olfactory toxicity, neurotoxicity, and gut microbial interference). The convolutional neural network coupled with the gated recurrent unit (CNN-GRU) deep learning method was used to construct a comprehensive model of 5-HT-RI functionality and human health interference effects selectivity with small sample data. The architecture with 2 SE, 320 neuronal nodes and 6-folds cross-validation showed the best applicability. The results showed that the confidence interval of the constructed model reached 90 % indicating that the model had reliable prediction ability and generalization ability. Based on the CNN-GRU deep learning model, seven high-priority chemicals with a weak comprehensive effect, including D-VEN, (1R,4S)-SER, S-FLX, CTP, S-CTP, NEF, and VEN, were screened. Based on the molecular three-dimensional structure information, a comprehensive-effect three-dimensional quantitative structure-activity relationship (3D-QSAR) model was constructed to confirm the reliability of the constructed control list of 5-HT-RI high-priority chemicals. Analysis with the ranking of calculated values based on the molecular dynamics method and predicted values based on the CNN-GRU deep learning model, we found that the consistency of the three methods was above 85 %. Additionally, by analyzing the sensitivity, molecular electrostatic potential, polar surface area of the comprehensive-effect CNN-GRU deep learning model, and the electrostatic field of the 3D-QSAR models, we found that the significant effects of five key characteristics (DM, Qyy, Qxz, I, and BP), molecular electronegativity, and polarity significantly affected the high-priority degree of 5-HT-RI. In this study, we provided reasonable and reliable prediction tools and discussed theoretical methods for the risk assessment of functionality and human health interference of emerging pollutants such as 5-HT-RI.
摘要：
通常在生物组织中检测到称为5-HT再摄取抑制剂（5-HT-RI）的抗抑郁药物，并导致严重的不良健康影响。同源性建模用于表征功能(功效和抗性)，并使用不良结局途径来表征其对人类健康的干扰(嗅觉毒性，神经毒性，和肠道微生物干扰)。利用卷积神经网络与门控循环单元(CNN-GRU)深度学习方法,构建了5-HT-RI功能和人体健康干扰效应选择性的小样本数据综合模型。具有2SE的架构，320个神经元节点和6倍交叉验证显示最佳适用性。结果表明，所构建模型的置信区间达到90%，表明该模型具有可靠的预测能力和泛化能力。基于CNN-GRU深度学习模型,七种综合效应较弱的高优先级化学品，包括D-VEN,(1R,4S)-SER,S-FLX,CTP,S-CTP,NEF,VEN,被筛选。基于分子三维结构信息,构建了综合效应三维定量构效关系(3D-QSAR)模型,以证实构建的5-HT-RI高优先级化学品控制列表的可靠性。用基于分子动力学方法的计算值排序和基于CNN-GRU深度学习模型的预测值排序进行分析,我们发现三种方法的一致性在85%以上。此外,通过分析灵敏度，分子静电势，综合效应CNN-GRU深度学习模型的极表面积，以及3D-QSAR模型的静电场，我们发现五个关键特征(DM，Qyy,Qxz,I,和BP），分子电负性,极性显著影响5-HT-RI的高优先级。在这项研究中,为5-HT-RI等新兴污染物的功能和人体健康干扰风险评估提供了合理可靠的预测工具，并探讨了理论方法。
