Comparison of Metabolomics Peak-Picking Parameter Optimization Algorithms Based on Chromatographic Peak Shape
Peak picking is one of the essential steps in non-targeted metabolomics data preprocessing based on liquid chromatography-mass spectrometry.Among various peak-picking algorithms,centWave algorithm based on continuous wavelet transform has been widely adopted in high-resolution mass spectrometry.In this study,the optimization effects of two centWave parameter optimization algorithms,IPO and centWave Sweep,were compared.Two datasets including metabolite standards and urine were used for comprehensive evaluation of these two algorithms with respect to three indicators:good peak shape ratio,reliable peak ratio,and repeatable peak ratio.To quickly and accurately distinguish good and bad peak shapes,three ensemble learning algorithms,random forest,adaboost and gradient boosting decision tree,were selected to establish a model for distinguishing chromatographic peak shape.Finally,according to the accuracy and F1 score,random forest was selected to establish a discrimination model(Accuracy 93.5%,F1 score 0.938).Compared with recommended parameters of XCMS Online,the proportion of reliable peaks and the proportion of repeatable peaks of two parameter optimization algorithms were improved in different datasets.However,when it came to the proportion of peaks with good shape,there was no significant difference between the optimized parameters and the parameters recommended by XCMS Online in different datasets.Furthermore,all three parameter settings resulted in relatively low proportions of peaks with good shape.The results indicated that the current parameter optimization algorithm was unable to improve the proportion of peaks with good shape.An excessive number of bad shape peaks could not only decrease the statistical power of analysis but also generate false positive results.Therefore,it was critical to perform additional confirmation of potential markers in the practical application of metabolomics researches.
国家科技期刊平台
NSTL
万方数据
