Metabolomics data fusion between near infrared spectroscopy and high-resolution mass spectrometry: A synergetic approach to boost performance or induce confusion

In general, data fusion can improve the classification performance of the model, but little attention is paid to the influence of the data fusion on the spatial distribution of the modeling samples. In this paper, the effect of data fusion on sample spatial distribution was studied through integrating NIR data and UHPLC-HRMS data for sulfur-fumigated Chinese herb medicine. Twelve samples collected from four different geographical origins were sulfur fumigated in the lab, and then metabolomics analysis was conducted using NIR and UHPLC-LTQ-Orbitrap mass spectrometer. First of all, the discriminating power of each technique was respectively examined based on PCA analysis. Secondly, combining NIR and UHPLC-HRMS data sets together with or without variable selection was parallelly compared. The results demonstrated that the discriminable ability was remarkably improved after data fusion, indicating data fusion could visualize variable selection and enhance group separation. Samples in the margin between two classes of samples may increase the experience error but has positive effect on the separation direction. Besides, an interesting feature extraction could obtain better discriminable effect than common data fusion. This study firstly provided a new path to employ a comprehensive analytical approach for discriminating SF Chinese herb medicines to simultaneously benefit from the advantages of several technologies.