Source-linked AI summary
Network medicine framework reveals generic herb-symptom effectiveness of Traditional Chinese Medicine
Xiao Gan, Zixin Shu, Xinyan Wang, Dengying Yan, Jun Li, Shany ofaim, Réka Albert, Xiaodong Li, Baoyan Liu, Xuezhong Zhou, Albert-László Barabási
TL;DR
TCM lacks a general mechanistic account of how herbs treat symptoms beyond evidence on selected herbs or prescriptions. The study builds a symptom-centered network medicine framework on the human protein interactome and finds that symptom-module and herb-target proximity predict symptom relations and treatment effectiveness, including previously unrecognized candidates.
Problem
The mechanistic basis and general principles of TCM herb effectiveness remain largely unknown because prior approaches focus mainly on individual herbs or prescriptions.
Method
The study maps symptom-associated genes and herb targets onto the human protein interactome and analyzes their network proximity using multiple datasets and clinical validation.
Results
The framework linked shorter symptom-module distances with symptom similarity and co-occurrence, while herb-symptom proximity predicted effectiveness and identified candidates supported by hospital data.
Takeaways & Limitations
The framework provides a system-level platform for understanding TCM and identifying herb discovery or repurposing candidates for follow-up testing.
Takeaways & Limitations
The hospital dataset was limited to liver-cirrhosis inpatients, lacked negative herb-effect samples, and supported propensity-score matching only on a limited subset.
Abstract
from arXiv · showhide
Traditional Chinese medicine (TCM) relies on natural medical products to treat symptoms and diseases. While clinical data have demonstrated the effectiveness of selected TCM-based treatments, the mechanistic root of how TCM herbs treat diseases remains largely unknown. More importantly, current approaches focus on single herbs or prescriptions, missing the high-level general principles of TCM. To uncover the mechanistic nature of TCM on a system level, in this work we establish a generic network medicine framework for TCM from the human protein interactome. Applying our framework reveals a network pattern between symptoms (diseases) and herbs in TCM. We first observe that genes associated with a symptom are not distributed randomly in the interactome, but cluster into localized modules; furthermore, a short network distance between two symptom modules is indicative of the symptoms' co-occurrence and similarity. Next, we show that the network proximity of a herb's targets to a symptom module is predictive of the herb's effectiveness in treating the symptom. We validate our framework with real-world hospital patient data by showing that (1) shorter network distance between symptoms of inpatients correlates with higher relative risk (co-occurrence), and (2) herb-symptom network proximity is indicative of patients' symptom recovery rate after herbal treatment. Finally, we identified novel herb-symptom pairs in which the herb's effectiveness in treating the symptom is predicted by network and confirmed in hospital data, but previously unknown to the TCM community. These predictions highlight our framework's potential in creating herb discovery or repurposing opportunities. In conclusion, network medicine offers a powerful novel platform to understand the mechanism of traditional medicine and to predict novel herbal treatment against diseases.
I. Introduction
The study develops a symptom-centered network medicine framework for TCM by mapping symptom-associated genes and herb targets onto the human protein interactome. It finds that symptom-module proximity relates to symptom similarity and co-occurrence, while herb-symptom proximity predicts treatment effectiveness and novel candidates.
- Motivation: TCM diagnosis and treatment are modeled through patient symptoms, connecting symptom phenotypes with modern biological and clinical data.The framework uses symptoms rather than diseases because TCM diagnosis and herbal therapies are based on symptom phenotypes.
- Symptom network: Genes associated with 108 of 174 analyzed symptoms formed significantly large connected components, supporting localized symptom modules in the interactome.Symptoms with fewer than 20 associated genes were excluded because their gene data were considered too incomplete.
- Symptom network: Shorter distances between symptom modules correlated with greater co-occurrence and biological similarity across disease and Gene Ontology analyses.The study used 147,978 symptom-disease associations and found negative correlations between network distance and both co-disease count and GO semantic similarity.
- Herb-symptom network: Herb-symptom proximity was quantified using multiple target datasets and eight network pipelines, with lower proximity values indicating closer network relations.The proximity distance averages distances from herb targets to closest symptom-associated genes, while the z-score compares proximity against random expectation.
- Herb-symptom network: Known indicated herb-symptom pairs consistently showed greater proximity than non-indicated pairs across all eight pipelines.The evaluation used expert-curated SymMap indications recognized by the 2015 Chinese Pharmacopoeia.
- Clinical validation: Hospital data supported the framework: symptom distance correlated negatively with relative risk, prescribed herbs were more proximal, and proximal pairs predicted recovery benefits.Baizhu improved poor-appetite recovery from 72.51% in controls to 79.53% in matched treated patients, with P-value = 0.0316; Chaihu improved abdomen-distention recovery from 83.73% to 88.71%, with P-values = 0.0458.
VI. Discussion
The study establishes a network medicine framework for TCM by mapping symptom-associated genes and herb targets onto the human protein interactome. Hospital data support its predictive relations and reveal previously unrecognized herb-symptom pairs, although data quality and scope constrain validation.
- The framework maps symptom-associated genes and herb targets onto the human protein interactome to analyze their topological relations.
- Shorter symptom-module distances correlate with symptom co-occurrence, while herb-symptom proximity predicts treatment effectiveness in hospital data.
- The study identifies herb-symptom pairs predicted effective by network proximity and proven effective in hospital data but not recognized by the TCM community.
- Compared with direct target-disease overlap, the whole-interactome approach allows effectiveness through network neighborhoods rather than requiring direct hits on symptom genes.
- Hospital validation is limited to liver cirrhosis inpatients, with incomplete noisy data, limited analytical scale, and no negative herb-effect samples.
- Studying co-prescribed herb combinations is proposed as a future direction because co-prescribed herbs tend to be close in the protein interactome.
Symptom-gene association data
The study uses symptom-gene associations integrated from established disease-gene resources and refines them using the dual-phenotype concept.
- Symptom data come from Symmap, which integrates disease-gene associations from DisGeNet and MalaCards.
- The dataset contains 110,407 associations spanning 11,362 diseases and 13,271 genes represented through UMLS concept codes.
- The study uses dual phenotypes, such as obesity, fever, and insomnia, to obtain higher-quality symptom-gene associations.
Herb, chemical and target data
Herb data combine direct target information with chemical-composition data from complementary TCM databases to represent herbs and their protein targets.
- Herb data come from the updated HIT 2.0 database and the TCMIO collection of TCMSP, TCMID, and TCM-ID.
- HIT provides direct herb-target data, whereas the other databases represent each herb as an assembly of chemicals.
- Chemical composition data are used to obtain protein targets for herbs represented through TCM databases.
Human Protein Interactome
The human protein interactome is assembled from experimentally validated physical, signaling, kinase, and regulatory interactions.
- The interactome is drawn from prior work on predicting COVID-treating drugs.
- Its interactions include binary PPIs from yeast two-hybrid experiments and three-dimensional protein structures.
- The network also includes affinity-purification mass spectrometry, kinase-substrate, signaling, and regulatory interactions.
Hospital data - Clinical symptom-herb associations
The study used electronic medical records from liver cirrhosis inpatients at a Traditional Chinese Medicine hospital, extracting symptoms and their trajectories from admission and discharge records.
- Electronic medical records were collected for liver cirrhosis inpatient cases at Hubei Provincial Hospital of Traditional Chinese Medicine in Wuhan.The records included patients’ full clinical profiles.
- Text-mining methods extracted clinical symptoms and their trajectories, including symptom recovery, from admission and discharge records.
Hospital data - Symptom terminology mapping and processing
Researchers manually mapped Chinese clinical terms for symptoms and herbs to English terms in symptom-gene associations, linking clinical and genetic data.
- Trained medical researchers manually mapped Chinese symptom and herb terms to English symptom-gene association terms.The mapping was intended to ensure highly accurate terminology alignment.
- 315 English symptom terms with associated genes were mapped to 92 Chinese symptom terms in the clinical data.Multiple UMLS codes could therefore merge into one TCM symptom, such as two codes mapped to fever.
Hospital data - Propensity Score Matching
Propensity Score Matching was used to reduce bias from patient baseline information when evaluating herb-treatment outcomes. For each herb-symptom pair, treated and untreated patients were matched while adjusting for baseline characteristics and common comorbidities.
- Propensity Score Matching was applied to the clinical dataset to account for covariates predicting receipt of herb treatment.The method was used to remove bias from patients’ basic information in treatment-outcome analyses.
- For the Baizhu-fatigue pair, fatigue patients receiving Baizhu during hospitalization formed the case group, while untreated fatigue patients formed the control group.
- Matching adjusted for baseline characteristics such as age and sex and for high-incidence comorbidities including hypertension and diabetes.Other controlled comorbidities included esophageal and gastric varices, abdominal effusion, and hypoproteinemia.
LCC and LCC z-score
The framework evaluates network organization and relations among symptom-associated genes, herb targets, and symptom modules using LCC statistics, network separation, semantic similarity, and proximity metrics.
- LCC and LCC z-score: The LCC z-score compares an observed node-set LCC size with a degree-preserving random expectation, standardized by the randomization variability.An LCC z-score above 1.6 indicates a significantly larger connected component than expected randomly.
- Network separation: Network separation compares within-set mean shortest-path distances with the mean distance between two node sets in the interactome.Negative separation indicates a shared network neighborhood, whereas positive separation indicates topological separation.
- Symptom similarity: Symptom biological similarity is computed as the average Gene Ontology semantic similarity across all gene pairs associated with two symptoms.
- Network proximity: Network proximity distance averages each herb target’s shortest distance to its closest symptom-associated gene.The metric uses symptom-associated genes S, herb targets T, and shortest path lengths between them.
- Network proximity: The proximity z-score standardizes proximity distance against randomized protein groups matched for set size and node degree.Lower distance or z-score values indicate closer network proximity; z=0 is neutral, z<0 is more proximal than random, and z>0 is more distant.
- Network proximity: When direct herb targets are unavailable, second-order herb-symptom distances are calculated as either the average or minimum of first-order chemical-symptom distances.Combining four herb-target mapping methods with two proximity measures produces eight herb-symptom proximity pipelines.