Prediction of herbal compatibility for colorectal adenoma treatment based on graph neural networks

Limei Gu, Yinuo Ma, Shaopeng Liu, Qinchang Zhang, Qiang Zhang, Ping Ma, Huang Dongfang, Haibo Cheng, Yang Sun, Tingsheng LingChinese 2025 · DOI

TL;DR

Researchers have developed a heterogeneous Graph Neural Network (GNN) to quantify "herbal compatibility"—the synergistic potential of herb pairings—specifically for treating colorectal adenoma. By integrating molecular data with clinical records and validating findings through in vitro experiments, the study identifies high-potential prescriptions and pinpoints Maackiain as a key compound for inhibiting colorectal cancer cell growth.

What Problem Does This Paper Address?

Traditional Chinese Medicine (TCM) relies on complex multi-herb prescriptions designed to work synergistically. However, understanding why certain herbs work well together (compatibility) has historically been a "black box" based on clinical experience rather than quantifiable data.

Existing computational methods, like association rule mining, often fail to capture the non-linear, high-dimensional relationships between herbs, their chemical constituents, and biological targets. Furthermore, most models are generic; this paper addresses the need for disease-specific models, focusing on colorectal adenoma—a common precancerous lesion with a high risk of transforming into malignancy.

Methods

The authors built a multi-layered pipeline that bridges "dry lab" AI with "wet lab" validation:

  1. Heterogeneous Graph Construction: Instead of a simple herb-to-herb network, they built a graph containing three node types: Herbs (122), Small Molecules (234), and Molecular Targets (657), connected by over 24,000 edges.
  2. Graph Neural Network (GNN): They used a Graph Convolutional Network (GCN) with SAGEConv operators to aggregate features. This allows the model to learn the representation of an herb based not just on its own properties, but on the molecules it contains and the targets those molecules hit.
  3. Compatibility Prediction: A Multi-Layer Perceptron (MLP) was trained to predict the "Herbal Combination Potential Index" (HCPI) between herb pairs, using co-occurrence in expert-recommended prescriptions as the ground truth.
  4. Validation: The model's predictions were tested against 26,267 real-world clinical records. Finally, the top-predicted compounds were subjected to molecular docking and CCK-8 cell viability assays to prove biological efficacy.

Datasets

  • Training Data: 72 expert-recommended TCM prescriptions for colorectal adenoma gathered from clinical trials and literature.
  • Clinical Validation Data: 26,267 treatment records (2022–2023) from the Jiangsu Province Hospital of Chinese Medicine, filtered down to 3,210 records specifically related to colon polyps and tumors.
  • Biological Databases: TCMSP (herb-molecule-target relationships), STRING (protein-protein interactions), and TCGA (cancer gene expression data).

Key Insights and Ideas

  • Beyond Co-occurrence: Traditional stats look at how often herbs appear together. This GNN looks at the functional reason they are together by analyzing if their underlying molecular targets overlap or complement each other in biological pathways (like inflammatory response or cell apoptosis).
  • Clinical Alignment: The model found that "high-frequency" prescriptions (those used more than 10 times by doctors in the real world) had significantly higher HCPI scores than low-frequency ones. This suggests the GNN successfully captured the "tacit knowledge" of experienced clinicians.
  • Targeting CHRM3: Through network pharmacology, the study identified CHRM3 as a pivotal molecular target. Molecular docking showed that the compounds Oxyberberine and Maackiain have high binding affinity for this target, suggesting a specific mechanism for their anti-tumor effects.

Key Results

  • Model Performance: The GNN achieved a high degree of fit with an $R^2$ of 0.91 at the 60th training epoch, indicating strong predictive reliability for herb compatibility.
  • Core Prescription Identification: The model highlighted a high-potential combination involving Hedyotis Diffusae Herb, Atractylodes macrocephala Koidz., and Arum ternatum Thunb.
  • Experimental Proof: In cell assays (HCT116 and HT29 lines), Maackiain significantly inhibited cell viability in a dose-dependent manner. This provides a concrete biological link between the GNN's mathematical predictions and actual therapeutic potential.

Limitations and Future Work

The authors acknowledge several hurdles:

  • Static vs. Dynamic: The current model is static. TCM often involves adjusting prescriptions over time as a disease progresses; the model doesn't yet account for these temporal changes.
  • Data Completeness: The accuracy is highly dependent on the quality of herb-ingredient databases (like TCMSP), which may not capture every trace compound.
  • Personalization: Future iterations need to incorporate patient-specific variables—such as genetics, age, and comorbidities—to move toward true "precision TCM."
  • Synergy vs. Antagonism: While the model excels at finding synergistic pairs, it is currently limited in its ability to characterize antagonistic (negative) interactions between herbs.

Summary updated

Abstract

Colorectal adenoma is a common precancerous lesion with a high risk of malignant transformation. Traditional Chinese medicine and its complex prescriptions have shown promising efficacy in the treatment of adenomas; however, there remains a lack of systematic understanding regarding the compatibility patterns within these prescriptions, as well as an effective model for predicting therapeutic outcomes. In this study, we collected numerous TCM prescriptions and their components, recommended by experts for the treatment of colorectal adenoma, and developed a heterogeneous graph neural network model to predict the compatibility strength and probability among the herbs within these prescriptions. This model delineates the complex relationships among herbs, active compounds, and molecular targets, allowing for a quantification of the interactions and compatibility potential among the herbs. Using this model, we identified high-potential therapeutic prescriptions from clinical prescription records and identified their active components through network pharmacology. Through this approach, we aim to provide a theoretical foundation for the clinical TCM treatment of colorectal adenoma, foster the discovery of new prescriptions to optimize the therapeutic efficacy of TCM, and ultimately advance the field of cancer prevention and treatment based on traditional Chinese medicine.
BibTeX
@inproceedings{gu2025prediction,
  author = {Limei Gu and Yinuo Ma and Shaopeng Liu and Qinchang Zhang and Qiang Zhang and Ping Ma and Huang Dongfang and Haibo Cheng and Yang Sun and Tingsheng Ling},
  title = {{Prediction of herbal compatibility for colorectal adenoma treatment based on graph neural networks}},
  booktitle = {Chinese},
  year = {2025},
  doi = {10.1186/s13020-025-01082-5},
}
Figure