Integrating metabolomic data with machine learning approach for discovery of Q-markers from Jinqi Jiangtang preparation against type 2 diabetes

• Published in: Chinese medicine Vol. 16; no. 1; p. 30
• Main Authors: Yang, Lele, Xue, Yan, Wei, Jinchao, Dai, Qi, Li, Peng
• Format: Journal Article
• Language: English
• Published: England BioMed Central Ltd 19-03-2021; BioMed Central; BMC
• Subjects: Analysis; Artificial intelligence; Backpropagation artificial neural network; Jinqi Jiangtang; Machine learning; Mass spectrometry; Medicine, Botanic; Medicine, Herbal; Metabolomics; Neural networks; Q-markers; Rankings; Type 2 diabetes; Metabolomics; Backpropagation artificial neural network; Q-markers; Jinqi Jiangtang; Mass spectrometry; Machine learning
• ISSN: 1749-8546; 1749-8546
• DOI: 10.1186/s13020-021-00438-x

Jinqi Jiangtang (JQJT) has been widely used in clinical practice to prevent and treat type 2 diabetes. However, little research has been done to identify and classify its quality markers (Q-markers) associated with anti-diabetes bioactivity. In this study, a strategy combining mass spectrometry-based untargeted metabolomics with backpropagation artificial neural network (BP-ANN)-based machine learning approach was proposed to screen Q-markers from JQJT preparation.
This strategy mainly involved chemical profiling of herbal medicines, statistic processing of metabolomic datasets, detection of different anti-diabetes activities and establishment of BP-ANN model. The chemical features of seventy-eight batches of JQJT extracts were first profiled by using the untargeted UPLC-LTQ-Orbitrap metabolomic approach. The chemical features obtained which were associated with different anti-diabetes activities based on three modes of action were normalized, ranked, and then pre-selected by using ReliefF feature selection. BP-ANN model was then established and optimized to screen Q-markers based on mean impact value (MIV).
Optimized BP-ANN architecture was established with high accuracy of R > 0.9983 and relative low error of MSE < 0.0014, which showed better performance than that of partial least square (PLS) model (R
 < 0.5). Meanwhile, the BP-ANN model was subsequently applied to further screen potential bioactive components from the pre-selected chemical features by calculating their MIVs. With this machine learning model, 10 potential Q-markers with bioactivity were discovered from JQJT. The tested anti-diabetes bioactivities of 78 batches of JQJT could be accurately predicted.
This proposed artificial intelligence approach is desirable for quick and easy identification of Q-markers with bioactivity from JQJT preparation.