机器学习算法构建附子线粒体功能高通量筛选模型及机制研究

Construction of a high-throughput screening model for mitochondrial function of Aconiti Lateralis Radix Praeparata by machine learning algorithm and mechanism analysis

构建线粒体功能高通量筛选机器学习模型,选择附子作为示例药物进行预测分析,选取模型筛选出来的乌头碱类化合物中得分最高的去氧乌头碱以及得分最低的苯甲酰新乌头原碱进行线粒体功能机制研究。收集来自PubChem、Tox21数据库中线粒体功能数据,采用随机森林和提升树2种算法分别进行建模;分别使用ECFP4和Mordred描述符进行训练,使用交叉验证检验,并采用平衡准确率、总体准确率对不同组合的模型进行性能评价以得到最好模型方法与模型参数。收集TCMSP数据库中附子化合物数据,经所构建的高通量机器学习模型预测筛选后,选取去氧乌头碱、苯甲酰新乌头原碱进行线粒体膜电位、活性氧(ROS)含量水平以及B淋巴细胞瘤-2(Bcl-2)、Bcl-2相关X蛋白(Bax)、过氧化物酶体增殖物活化受体γ协同活化因子1α(PGC-1α)蛋白表达测定。结果表明,使用提升树+Mordred算法构建的模型表现更好,交叉验证平衡准确率(BA)为0.825,测试集准确率为0.811;去氧乌头碱、苯甲酰新乌头原碱能够改变ROS含量(P<0.001)、线粒体膜电位(P<0.001)以及Bcl-2(P<0.001,P<0.01)、Bax蛋白表达(P<0.001);去氧乌头碱能够升高PGC-1α蛋白表达(P<0.01)。提升树+Mordred算法较随机森林+ECFP4算法构建的线粒体功能高通量模型表现更为准确,可为后续研究构建算法模型;去氧乌头碱与苯甲酰新乌头原碱均可影响线粒体功能,但得分更高的去氧乌头碱还能够特征性通过调节PGC-1α蛋白影响线粒体生物合成。

A high-throughput screening machine learning model for mitochondrial function was constructed,and compounds of Aco-niti Lateralis Radix Praeparata were predicted.Deoxyaconitine with the highest score and benzoylmesaconine with the lowest score among the compounds screened by the model were selected for mitochondrial mechanism analysis.Mitochondrial function data were collected from PubChem and Tox21 databases.Random forest and gradient boosted decision tree algorithms were separately used for mo-deling,and ECFP4(extended connectivity fingerprint,up to four bonds)and Mordred descriptors were employed for training,respectively.Cross-validation test was carried out,and balanced accuracy(BA)and overall accuracy were determined to evaluate the performance of different combinations of models and obtain the optimal algorithm and hyperparameters for modeling.The data of Aconiti Lateralis Radix Praeparata compounds in TCMSP database were collected,and after prediction and screening by the constructed high-throughput screening machine learning model,deoxyaconitine and benzoylmesaconine were selected to measure mitochondrial membrane potential,reactive oxygen species(ROS)level and protein expression of B-cell lymphoma 2(Bcl-2),Bcl-2-associated X protein(Bax)and peroxisome proliferator-activated receptor-γ-coactivator 1α(PGC-1α).The results showed that the model constructed using gradient boosted decision tree+Mordred algorithm performed better,with a cross-validation BA of 0.825 and a test set accuracy of 0.811.Deoxyaconitine and benzoylmesaconine changed the ROS level(P<0.001),mitochondrial membrane potential(P<0.001),and protein expression of Bcl-2(P<0.001,P<0.01)and Bax(P<0.001),and deoxyaconitine increased the expression of PGC-1αprotein(P<0.01).The high-throughput screening model for mitochondrial function constructed by gradient boosted decision tree+Mordred algorithm was more accurate than that by random forest+ECFP4 algorithm,which could be used to build an algorithm model for subsequent research.Deoxyaconitine and benzoylmesaconine affected mitochondrial function.However,deoxyaconitine with higher score also affected mitochondrial biosynthesis by regulating PGC-1αprotein.