Machine learning-assisted analysis of serum metabolomics and network pharmacology reveals the effective compound from herbal formula against alcoholic liver injury.

Abstract

Background: The popularity of herbal formulas is increasing worldwide. Nevertheless, the effective compound is challenging to identify due to its intricate composition and multiple targets.

Methods: An integration machine learning-assisted approach was established, whereby the particular action mechanism and direct target were obtained through the correlation of compounds, targets, and metabolites. The association between a compound and an action pathway was selected from the shortest path of the "compound-target-pathway-disease" network, which was analyzed using the Floyd-Warshall algorithm. Subsequently, an investigation was conducted into the relationship between metabolites and action pathways, as well as targets, through the analysis of serum metabolomic profiling and the selection of metabolite biomarkers by random forest. In order to accurately identify the direct acting target as well as the most effective compound, the relationship between the compounds and their targets was investigated using a feature-based prediction model conducted by AdaBoost. The binding mode of the effective compound and the direct-acting target was verified by molecular docking, dynamics simulations, and western blotting. In this study, Baiji Wuweizi Granule (BWG) was employed to elucidate the effective compound against alcoholic liver injury (ALD).

Results: BWG exerted an influence on the serum metabolomic, resulting in the identification of seven potential biomarkers. Furthermore, six effective compounds and the PI3K-AKT signalling pathway were identified through a co-analysis with the shortest path from compound to ALD in the "compound-target-pathway-disease" network. It was postulated that the effective compounds would bind with key targets from the PI3K-AKT signaling pathway, as indicated by the prediction model of compound-target interaction (R² > 0.95). The dominant bonding type for the effective compounds and key targets was hydrogen bond. These results indicated that AKT1 was the notable target for BWG, and that 2,3,4,7-tetramethoxyphenanthrene was the marker compound for BWG against ALD. The present study provides evidence that the protective effect of BWG on ALD can be mediated by the PI3K-AKT signaling pathway.

Conclusions: Our findings demonstrate the value of a machine learning-assisted approach in identifying the key compound, target and pathway that underpin the efficacy of an herbal formula. This provides a foundation for future clinical and fundamental research.