非洲传统医学的计算评估:机器学习模型能否预测萱草叶小金梅草(Hypoxis hemerocallidea)植物化学成分的抗逆转录病毒作用机制OA
Computational assessment of African traditional medicine:Can machine learning models predict antiretroviral mechanisms of action in Hypoxis hemerocallidea phytochemicals?
背景:在众多发展中国家,高达80%的人口依赖传统医学,包括萱草叶小金梅草(Hypoxis hemerocallidea,俗称非洲马铃薯)等植物.尽管科学证据有限,该植物仍在南部非洲被广泛用于人类免疫缺陷病毒(HIV)感染的辅助治疗.目的:研究采用人工智能(AI)与计算化学法,预测萱草叶小金梅草植物化学成分的抗逆转录病毒作用机制,为非洲民族药理学与现代科学搭建关联桥梁.方法:基于7个ChEMBL数据集训练直方图梯度提升(HGB)与极端梯度提升(XGB)两种机器学习(ML)模型,用于预测5类抗逆转录病毒作用机制,包括逆转录酶抑制剂(RTI)、蛋白酶抑制剂(PI)、HIV包膜糖蛋白160(gp160)进入抑制剂、CC趋化因子受体5(CCR5)抑制剂及整合酶链转移抑制剂(INSTI).采用扩展连接性指纹4(ECFP4)生成分子描述符,执行严格的特征选择流程(阈值<0.01),通过Optuna进行超参数优化,以10折交叉验证评价模型性能.利用构建的定量构效关系(QSAR)模型与机器学习模型,筛选萱草叶小金梅草中的47种主要植物化学成分.当化合物平均预测pIC50>6且预测值变异系数<5%时,判定为具有潜在活性.结果:模型表现出优异的预测性能,其中对CCR5抑制剂的预测效果最佳.HGB模型在测试集上的R2为0.91,XGB模型可达0.94.原花青素A2(Procyanidin A2)被鉴定为潜在的强效CCR5抑制剂;其余化合物均未达到其余抗逆转录病毒机制的活性判定标准(平均pIC50≥6;变异系数<5%).结论:该研究通过整合人工智能与民族药理学方法,借助机器学习识别了萱草叶小金梅草主要植物化学成分的潜在抗逆转录病毒作用机制,为传统药用经验提供了分子层面的科学依据,同时筛选得到潜在生物活性候选化合物.上述计算预测结果仍需后续实验验证.
Background:In many developing countries,up to 80%of the population depends on traditional medicine.This includes the use of plants like Hypoxis hemerocallidea(African potato),which is widely utilized in Southern Africa for human immunodeficiency virus(HIV)treatment,despite limited scientific evidence to support its efficacy. Objective:The objective of this study was to bridge African ethnopharmacology and modern science by employing artificial intelligence(AI)and computational chemistry to predict antiretroviral mechanisms of action among H.hemerocallidea phytochemicals. Methods:The machine learning(ML)models,namely HistGradientBoosting(HGB)and XGBoost(XGB)were trained on seven ChEMBL datasets to predict five antiretroviral mechanisms of action,including reverse tran-scriptase inhibitor(RTI),protease inhibitor(PI),the envelope glycoprotein 160 of the human immunodeficiency virus(gp160)entry inhibitors,C-C chemokine receptor type 5(CCR5)inhibitors,and integrase strand trans-fer inhibitor(INSTI).Molecular descriptors were generated using extended connectivity fingerprints 4(ECFP4).A rigorous feature selection procedure(threshold<0.01)was applied,and hyperparameter optimization was performed using Optuna,with model performance evaluated through 10-fold cross-validation.The resulting quantitative structure-activity relationship(QSAR)and ML models screened 47 main phytochemicals from H.hemerocallidea.Compounds were considered potentially active when the average predicted pIC50 exceeded 6 and the coefficient of variation among predictions was below 5%. Results:The models demonstrated excellent predictive performance,especially for CCR5 inhibitors.HGB achieved an R2 score of 0.91 on the testing set,while XGB achieved 0.94.Procyanidin A2 was identified as a potent potential CCR5 inhibitor.No other compounds met the activity criteria(mean pIC50 ≥ 6;CV<5%)for any of the remaining antiretroviral mechanisms. Conclusion:By integrating AI and ethnopharmacology,this study applied machine learning to identify potential antiretroviral mechanisms of action among the main phytochemicals of H.hemerocallidea,providing a molecular rationale for traditional medicine and predicting potential bioactive candidates.Experimental validation remains necessary to confirm these computational findings.
Hamza Age Daudo
Department of Pharmacy,Lurio University,Marrere Neighborhood,Street No.4250,Km 2.3,Nampula 3100,Mozambique
Hypoxis hemerocallidea机器学习定量构效关系抗逆转录病毒民族药理学
Hypoxis hemerocallideaMachine learningQSARAntiretroviralEthnopharmacology
《临床传统医学和药理学(英文)》 2026 (2)
70-77,8
评论