Understanding the ancient classic and famous prescriptions via the property of Chinese materia medica

Dandan Qin, He Zhang, Bin Du, Hui Wang, Ligang Liu, Yun WangFrontiers in Pharmacology 2025 · DOI

TL;DR

Researchers have developed a novel feature extraction model that represents Traditional Chinese Medicine (TCM) formulas not by their ingredients, but by their underlying pharmacological properties (Property Combinations of Chinese Materia Medica, or PCCMM). By treating formula reconstruction as a combinatorial optimization problem, the study demonstrates that these property-based "fingerprints" can effectively classify, reconstruct, and reveal hidden medication patterns in ancient classic prescriptions.

What Problem Does This Paper Address?

Traditional Chinese Medicine (TCM) relies on "Ancient Classic and Famous Prescriptions" (ACFPs)—formulas honed over centuries. However, analyzing these formulas systematically is difficult.

The primary challenge is data sparsity. There are hundreds of unique Chinese medicinal materials (CMMs), but official catalogs of classic formulas are relatively small. When researchers try to use standard data mining on herb names alone, the signals are often too weak to find meaningful patterns. Furthermore, herb names are just labels; they don't inherently describe how a medicine works.

The authors address this by shifting the focus from the "what" (the specific herb) to the "why" (its medicinal properties). By mapping herbs to a multidimensional space of properties, flavors, and meridian tropisms, they create a denser, more mathematically tractable representation of TCM knowledge.

Methods

The researchers introduced a framework consisting of a "forward" and "backward" process:

  1. Forward Feature Extraction (CMM to PCCMM): They mapped 604 herbs from the Chinese Pharmacopoeia to a set of PCCMM triplets. Each triplet combines three core attributes: Property (e.g., cold, warm), Flavor (e.g., bitter, pungent), and Meridian Tropism (the organ system it targets, like the lung or spleen). This reduced the "vocabulary" from 604 herbs to 254 unique property combinations.
  2. Backward Reconstruction (PCCMM to Formula): To prove that these properties actually capture the essence of a formula, the authors tried to "rebuild" the original prescriptions using only the property data. They formulated this as a constrained combinatorial optimization problem, inspired by the mathematical framework of compressive sensing. Essentially, they asked: "What is the smallest set of herbs that satisfies this specific combination of medicinal properties?"
  3. Analysis & Validation: They used Jaccard similarity coefficients to measure reconstruction accuracy and employed t-SNE (t-distributed Stochastic Neighbor Embedding) to visualize how well these property features could distinguish real formulas from random "pseudo-formulas."

Datasets

  • Chinese Pharmacopoeia (ChP) 2020 & 2015: Used to define the properties of 604 (and 607) medicinal materials.
  • Official ACFP Catalogs: A dataset of 178 classic prescriptions (93 from the first batch and 85 from the second batch) issued by the Chinese government.
  • Incompatible CMM Pairs: A list of 89 herb pairs that are traditionally forbidden from being used together (used as constraints in the optimization model).

Key Insights and Ideas

  • Quasi-Equivalent Representation: The most significant insight is that PCCMM acts as a "quasi-equivalent" for the formula. You can lose the specific herb names but keep the property profile, and you still have enough information to identify or even reconstruct the formula.
  • Functional Redundancy: The optimization model sometimes suggested different herbs than the original prescription but with identical properties. This highlights the "substitutability" of certain medicines in TCM, provided their pharmacological profile matches.
  • The "Even-Sweet-Spleen" Axis: Frequency analysis revealed that high-frequency property combinations in classic formulas predominantly map to the "even" property, "sweet" flavor, and "spleen" meridian. This suggests a systemic historical preference for tonic medicines that harmonize the digestive system.
  • Superiority over Nomenclature: In classification tasks (like identifying which "deficiency syndrome" a formula treats), the property-based model outperformed models based on herb names alone (0.55 vs 0.42 accuracy).

Key Results

  • High Reconstruction Accuracy: The model achieved an average Jaccard similarity coefficient above 0.8 when reconstructing formulas from their weighted property vectors.
  • Implicit Rule Mining: The PCCMM-based association rule analysis yielded significantly more "implicit compatibility rules" than herb-based analysis. For example, it identified the synergistic pairing of Jujubae Fructus (Dazao) and Zingiberis Rhizoma Recens (Shengjiang) as a statistically dominant pattern, which aligns perfectly with clinical tradition.
  • Effective Clustering: t-SNE visualizations showed a clear, almost linear boundary between legitimate classic formulas and random herb combinations when using property features, proving that classic formulas are not just random assortments but mathematically distinct "property clusters."

Limitations and Future Work

  • Lack of Dosage Data: The current model treats herbs as binary (present or absent). It does not yet account for the specific proportions or dosages of herbs, which are critical to TCM efficacy.
  • Chemical Complexity: The model relies on traditional property labels rather than modern chemical compositions or molecular targets.
  • Small Dataset: While the model covers 90% of the official ACFP catalog, the total number of samples (178) is still small for deep learning applications.

Future Directions: The authors suggest incorporating dosage data to create a "quantitative CMM–PCCMM network" and exploring how these property patterns correlate with modern nutritional and chemical data.

Summary updated

Abstract

The PCCMM-based feature extraction model provides a quasi-equivalent representation of TCM formulas, effectively capturing implicit medication patterns within ACFPs. PCCMM outperforms traditional CMM methods in formula reconstruction, classification, and medication pattern mining. This study offers novel insights and methodologies for systematically understanding TCM formulas, guiding clinical application, and facilitating the design and optimization of new TCM formulas.
BibTeX
@inproceedings{qin2025understanding,
  author = {Dandan Qin and He Zhang and Bin Du and Hui Wang and Ligang Liu and Yun Wang},
  title = {{Understanding the ancient classic and famous prescriptions via the property of Chinese materia medica}},
  booktitle = {Frontiers in Pharmacology},
  year = {2025},
  doi = {10.3389/fphar.2025.1551531},
}

Loading related content...

Figure