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

Jiamu Ma, Wei Peng, Xiao Xu, Ruijuan Dong, Xin Deng, Feng Zhang, Mengyu Sun, Mingxia Li, Wei Liu, Jianling Yao, Yu Cao, Lisha Ying, Yuqing Yang, Yongqi Yang, Xiaopeng Wu, Gaimei SheChinese 2025 · DOI

TL;DR

This paper introduces a novel machine learning-assisted approach to untangle the complex mechanisms of traditional Chinese medicine (TCM) formulas. By integrating serum metabolomics and network pharmacology, they identified a key compound and its target pathway responsible for the protective effects of Baiji Wuweizi Granule (BWG) against alcoholic liver injury (ALD). This work offers a powerful framework for understanding and developing TCM therapies.

What Problem Does This Paper Address?

Traditional Chinese Medicine (TCM) formulas are gaining global recognition for their therapeutic potential, but their intricate composition with numerous active compounds and multiple targets makes it incredibly challenging to pinpoint the exact mechanisms of action. This "black box" problem hinders the development and standardization of these remedies. Specifically, identifying the key active compounds, their direct targets, and the underlying pathways responsible for a formula's efficacy against diseases like alcoholic liver injury (ALD) is a significant hurdle.

Methods

The authors developed an integrated machine learning-assisted approach to tackle this complexity. Their methodology can be broken down into several key steps:

  • Serum Metabolomics: They analyzed serum samples from mice to identify differential metabolites and potential biomarkers associated with ALD and the effects of BWG. This involved untargeted metabolomics followed by statistical analysis (PCA, PLS-DA) and biomarker selection using random forest.
  • Network Pharmacology (CTPDN): They constructed a "compound-target-pathway-disease" network (CTPDN) by collecting nodes (compounds, targets, pathways, disease) from various databases and literature. Edges representing interactions were weighted using molecular docking, protein-protein interaction (PPI) analysis, and pathway enrichment data.
  • Shortest Path Analysis: The Floyd-Warshall algorithm was employed on the CTPDN to identify the shortest weighted path from compounds to the disease, thereby ranking potential effective compounds and their associated targets and pathways.
  • Compound-Target Interaction (CTI) Prediction: A feature-based machine learning model (AdaBoost) was developed to predict the interaction between compounds and their targets. This model utilized molecular fingerprints (MACCS, ECFP6) and ADMET properties of compounds, along with protein sequence and property descriptors.
  • Validation: Molecular docking, molecular dynamics simulations, and western blotting were used to experimentally validate the predicted interactions and the expression levels of key targets.

Datasets

  • Public Databases: PubChem, HMDB, MassBank, Genecards, OMIM, Drugbank, TCMSP, Uniprot, STRING, KEGG, GO, RCSB Protein Data Bank, and PubChem were used for collecting compound, target, and pathway information.
  • Experimental Data: Serum metabolomics data from mouse models of ALD treated with BWG.
  • BWG Compound Information: UHPLC-MS/MS data for identifying compounds within BWG.

Key Insights and Ideas

  • Integrated Approach is Key: The study strongly advocates for the synergistic use of metabolomics and network pharmacology, powered by machine learning, to dissect complex herbal medicine mechanisms.
  • Shortest Path as a Ranking Metric: Using the Floyd-Warshall algorithm on a weighted CTPDN provides a robust way to rank compounds based on their "distance" to the disease, offering a more nuanced view than simple node centrality.
  • Feature Engineering for CTI: The success of the AdaBoost model in predicting compound-target interactions highlights the importance of combining diverse features, including molecular fingerprints and ADMET properties, for accurate predictions.
  • AKT1 as a Central Target: The consistent identification of AKT1 as a key target across multiple analytical steps (CTPDN, CTI prediction, western blotting) underscores its critical role in BWG's therapeutic effect on ALD.
  • 2,3,4,7-tetramethoxyphenanthrene as a Marker Compound: This specific compound emerged as a strong candidate for BWG's efficacy against ALD, supported by its binding affinity and molecular dynamics simulations.

Key Results

  • Identification of Biomarkers: Seven potential serum biomarkers were identified for ALD.
  • Effective Compounds and Pathway: Six effective compounds and the PI3K-AKT signaling pathway were identified as crucial for BWG's action against ALD.
  • High Prediction Accuracy: The AdaBoost model achieved high performance metrics (e.g., Precision=0.96, AUC=0.96 in training; 1.00 in testing) for predicting compound-target interactions.
  • Key Target Validation: Western blotting confirmed that BWG significantly modulated the expression of AKT1, pAKT1, Catalase, and STAT1, aligning with the predicted targets.
  • Marker Compound Identified: 2,3,4,7-tetramethoxyphenanthrene was identified as a key marker compound of BWG for treating ALD.

Limitations and Future Work

  • Experimental Validation Scope: While molecular docking, dynamics, and western blotting were used, the authors acknowledge the need for broader experimental validation, including surface plasmon resonance and cellular thermal shift assays, to strengthen the evidence for effective compounds and targets.
  • Polysaccharide Role: The study did not fully address the role of polysaccharides in BWG, despite their potentially high content and pharmacological activity. Future work could involve adapting screening criteria to include polysaccharides in the CTPDN.
  • Generalizability: The authors note that they haven't yet expanded the application range of their CTI prediction model to fully confirm its generality.
  • In Vivo vs. In Silico: The study relies heavily on computational predictions, and further in vivo studies are needed to fully elucidate the complex interactions and mechanisms.

Summary updated

Abstract

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.
BibTeX
@inproceedings{ma2025machine,
  author = {Jiamu Ma and Wei Peng and Xiao Xu and Ruijuan Dong and Xin Deng and Feng Zhang and Mengyu Sun and Mingxia Li and Wei Liu and Jianling Yao and Yu Cao and Lisha Ying and Yuqing Yang and Yongqi Yang and Xiaopeng Wu and Gaimei She},
  title = {{Machine learning-assisted analysis of serum metabolomics and network pharmacology reveals the effective compound from herbal formula against alcoholic liver injury}},
  booktitle = {Chinese},
  year = {2025},
  doi = {10.1186/s13020-025-01094-1},
}
Figure