基于MALDI-TOF MS平台结合机器学习算法鉴别三唑耐药热带念珠菌

Identification of triazole-resistant Candida tropicalis based on MALDI-TOF MS platform and machine learning algorithm

目的利用基质辅助激光解吸电离飞行时间质谱(MALDI-TOF MS)平台数据分析和机器学习算法快速鉴别三唑(氟康唑、伏立康唑、伊曲康唑)耐药和敏感的热带念珠菌。方法从临床各类标本中收集191株热带念珠菌,其中71株为三唑耐药热带念珠菌,120株为三唑敏感热带念珠菌。使用MALDI-TOF MS平台进行数据采集,并根据Mann-Whitney U-test及随机森林(RF)算法获得的重要性评分对耐药株及敏感株的质荷比特征进行分类和选择。利用RF算法及径向基函数核非线性支持向量机(RBF-SVM)构建分类模型,计算相同实验数据下RBF-SVM模型和RF模型的准确度、敏感度、特异度、F1值及受试工作者曲线下面积(AUC)以评估模型鉴别性能。结果所有菌株经过MALDI-TOF MS平台分析后共得到76个独特的质谱峰。其中,通过特征降维处理后选择6个峰3481、7549、6500、3048、6892、2596 m/z作为模型建立的特征峰。RBF-SVM模型和RF模型的准确度均为0.84,AUC分数分别为0.9305、0.9273。结论机器学习算法结合MALDI-TOF MS平台进行数据分析可作为一种快速区分三唑耐药热带念珠菌和三唑敏感菌株的方法。

Objective To rapidly identify triazole(fluconazole,voliconazole,iriconazole)drug resistance and sensitive Candida tropical using matrix-assisted laser desorption/ionization-time-of-flight mass spectrometry(MALDI-TOF MS)platform data analysis and machine learning algorithms.Methods A total of 191 Candida tropical were collected from various clinical specimens,71 of which were triazole-resistant Candida tropical and 120 were triazole-sensitive Candida tropical strains.Data acquisition was performed using the MALDI-TOF MS platform,and the mass and charge ratio features of resistant and susceptible strains were classified and selected based on the Mann-Whitney Rank-sum Test(Mann-Whitney U-test)and the importance score obtained by the Random Forest(RF)algorithm.The classification model was constructed using the RF algorithm and a nonlinear support vector machine with a radial basis function kernel(RBF-SVM),calculating the accuracy,sensitivity,specificity,F1 value and the area under the subject worker curve(AUC)of the RBF-SVM model under the same experimental data to evaluate the model discrimination performance.Results All strains obtained 76 unique mass spectrum peaks after analysis on the MALDI-TOF MS platform.Among them,six peaks 3481,7549,6500,3048,6892,2596 m/z were selected as the model feature peaks established by the feature dimensionality reduction treatment.The accuracy of both the RBF-SVM and RF models was 0.84,and the AUC scores were 0.9305 and 0.9273,respectively.Conclusion Machine learning algorithms combined with the MALDI-TOF MS platform for data analysis can serve as a possible method to rapidly distinguish triazole-resistant Candida tropical and triazole-sensitive strains.