基于文本卷积神经网络模型的抗菌药物发现

Text-Convolutional Neural Network-based Discovery of Antibacterial Agents

目的 基于文本卷积神经网络(Text-Convolutional Neural Network, Text-CNN)算法,构建抗金黄色葡萄球菌(Staphylococcus aureus)活性的预测模型,通过虚拟筛选发现具有抑制S.aureus活性的苗头化合物。方法 从ChEMBL数据库中收集并整理了26327个标注有S.aureus活性数据的化合物,通过随机采样建立10组训练集和测试集,采用Text-CNN算法建立10个模型,通过模型评估选择性能最佳的模型,对该模型进行Y-随机化检验和应用域分析。使用该模型虚拟筛选内部化合物库,确定潜在的抗菌化合物,并采用微量肉汤稀释法测定化合物的抗S.aureus活性。结果 名为Text-CNN3的机器学习模型具有良好的分类性能,该模型对于测试集的马修斯相关系数为0.573,ROC曲线下面积为0.881。基于该模型的虚拟筛选和抗菌活性测试,发现了两个抗菌活性化合物Y5和Y7,其对S.aureus的最低抑菌浓度(minimal inhibitory concentration, MIC)分别为8和4μg·mL^(-1)。结论 本研究建立的Text-CNN3模型可有效发现抗S.aureus化合物,所发现的苗头化合物Y5和Y7有进一步研究的意义和价值。

OBJECTIVE To build a text-convolutional neural network(Text-CNN)-based prediction model for anti-Staphylococcus aureus(S.aureus)activity and identify anti-S.aureus hits by virtual screening.METHODS A dataset containing 26327 compounds annotated with S.aureus activity data was collected and curated from the ChEMBL database.Ten pairs of training and test sets were generated by random partition for 10 times and then 10 models were built using the Text-CNN algorithm.The best-performing model was determined by model evaluation and further studied by Y-randomization test and applicability domain analysis.Following that,the best-performing model was used to virtually screen the in-house chemical library,by which the potential antibacterial agents were determined.The micro-broth dilution method was used to test anti-S.aureus activity of the potential hits.RESULTS The machine-learning model(named Text-CNN3)performed well in classification.Evaluated on the test set,its Mathews correlation coefficient was 0.573 and the area under the ROC curve was 0.881.With this model for virtual screening as well as antibacterial screening,compounds Y5 and Y7 were identified as antibacterial compounds,with minimum inhibitory concentrations(MIC)of 8 and 4μg·mL^(-1),respectively.CONCLUSION The Text-CNN3 model in this study is effective to identify anti-S.aureus compounds,while the antibacterial hits Y5 and Y7 are worthy of further study.