目的 在新型冠状病毒感染尚无特效药的情况下，根据潜在的治疗靶点，从中药活性成分中筛选出先导化合物为药物研发、中药组方等提供理论线索。方法 构建消息传递机制神经网络（message passing neural networks，MPNN）模型，以化合物描述符SMILES码为输入，以化合物对新型冠状病毒3CL蛋白酶抑制活性为输出，利用开源数据对模型进行训练和优化。结果 用优化后的模型从186味清热解毒中药所含的3 863个活性成分中筛选出101个潜在的抑制剂。其中龙胆素、桑辛素C、5-羟基-4-氧代戊酸甲酯等化合物预测活性较高，黄芩、苍耳子、昆布、芫荽、紫苏等中药含有的潜在抑制剂数量较多。结论 使用MPNN模型虚拟筛选出的抑制剂中约1/5已被其他的研究报道验证有效，证明了MPNN模型虚拟筛选结果的可靠性。此外，优化的神经网络模型微调后可用于分子其他属性的预测，在药物虚拟筛选领域有着广泛的应用前景。

Objective In the absence of specific drugs for severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infections, screening of lead compounds from active ingredients of traditional Chinese medicine (TCM) based on potential therapeutic targets can provide theoretical clues for drug research andm development and TCM formulations. Methods In this study, a message passing neural networks (MPNN) model was constructed, with the compound descriptor SMILES code as input and the inhibitory activity of compounds against SARS-CoV-2 3CL protease as output. The model was trained and optimized using open-source data. Results The optimized model successfully screened out 101 potential inhibitors from 3 863 active ingredients of 186 TCMs with heat-clearing and detoxifying effects. Among them, compounds such as gentianin, mycin C, methyl 5-hydroxy-4-oxopentanoate, and others exhibited the higher predicted activity, while TCMs like Huangqin (Scutellariae Radix), Cang’erzi (Xanthii Fructus), Kunbu (Laminariae Thallus), Yansui (Herba Coriandri Sativi), Zisu [Perilla frutescens (L.) Britt.] had a relatively higher number of potential inhibitors. Conclusion Approximately one-fifth of the inhibitors identified by MPNN model screening have been validated by other experimental or theoretical studies, demonstrating the reliability of the screening results. Additionally, the neural network model optimized in this study can be fine-tuned for predicting other molecular properties, indicating its extensive application prospects in drug virtual screening.

国家自然科学基金-河南联合基金项目（U2004196）；郑州轻工业大学博士基金项目（2011BSJJ012）