Machine Learning and Multi-Omics Reveal Toxicity Signals in Adverse
Evodia rutaecarpa is a fruit that has long been used in traditional Chinese medicine to warm the body, remove cold and relieve pain. In modern medicine, it has an unusual status. It is considered a low-toxicity herb, meaning it is thought to have a relatively low risk of harm. However, clinical experience shows that some people have unwanted reactions when taking herbal formulas containing it. These reactions can be mild, differ from person to person and be easy to miss. A new study published in Frontiers in Pharmacology combined metabolomics, oral microbiome sequencing and several machine learning methods to identify biological signs of these reactions in real patients. The researchers also developed a mathematical formula to estimate how likely each person is to experience them.
The research team studied 86 patients with stomach diseases who were taking herbal formulas containing Evodia rutaecarpa. Based on whether they developed unwanted reactions during treatment, the researchers divided them into two groups: 50 patients without adverse reactions and 36 patients with adverse reactions. The team collected three types of biological samples from each person: blood serum, urine and saliva. They used untargeted metabolomics to study the many small molecules in serum and urine. They also analyzed saliva using 16S rRNA gene sequencing to identify the types of bacteria living in the mouth. This approach allowed the researchers to look for possible signs of toxicity in both body chemicals and microbes instead of relying on just one type of measurement.
Most of the unwanted reactions affected the digestive system, especially the upper part of the gastrointestinal tract. This fits with the herb's known potential to irritate the moist tissues lining the body. However, the biological data showed changes beyond the digestive system. Patients who experienced adverse reactions had changes in amino acid metabolism, suggesting that their bodies processed nitrogen and the building blocks of proteins differently. Their levels of molecules linked to nervous system activity also differed, suggesting that the herb's active alkaloids might affect pathways related to neurotransmitters in people who are more sensitive to them.
A third important signal involved bile. The metabolomics results showed changes in patterns related to bile secretion in patients who experienced adverse reactions. This suggests that the enterohepatic circulation, the cycle in which bile acids travel between the liver and intestines and are recycled, may have been disrupted. Bile acids do more than help the body digest food. They also act as signaling molecules that help control the gut's protective barrier and inflammation. These findings offer a possible explanation for how the herb's chemical ingredients might be linked to the digestive symptoms seen in patients. The saliva analysis also found dysbiosis, meaning a change in the balance of bacteria in the mouth. The researchers described this as a possible association, not proof that the bacteria caused the reactions.
To turn these complex biological measurements into a practical tool for estimating risk, the team tested seven machine learning algorithms, including logistic regression, random forest and LightGBM. They checked how well the models worked using leave-one-out cross-validation. In this method, the researchers leave out one patient at a time, train the model on all the other patients and then test its prediction for the patient who was left out. The algorithms were used for two purposes: to build models that could distinguish patients with adverse reactions from those without them, and to identify which chemical and microbial measurements were most useful for making those predictions. Using machine learning both to predict outcomes and to select important features is becoming more common in pharmacogenomics, the study of how genes and biology influence responses to medicines. It helps researchers choose which findings to investigate further in experiments.
The screening process identified several features that helped the models distinguish between the two patient groups. In serum, three metabolites showed statistical trends associated with lower model predictions of adverse reactions: the phospholipid PE (P-18:1/22:0), hydroxyisocaproic acid and ureidoisobutyric acid. In urine, 1,3-dihydro-(2H)-indol-2-one and the monoglyceride MG (0:0/20:4/0:0) were associated with lower predicted values. Arachidoyl ethanolamide, a lipid belonging to the endocannabinoid family, was associated with higher model prediction scores. The direction of these relationships is important. Metabolites linked to lower predictions might reflect protective effects or normal body regulation, while those linked to higher predictions might indicate greater vulnerability to the herb's harmful effects.
The researchers also identified specific microbial features. Higher relative amounts of four bacterial groups in saliva were linked to higher predicted probabilities of adverse reactions: the genera Granulicatella and Alysiella, the species Leptotrichia hongkongensis and the genus Streptococcus. Scientists increasingly recognize that bacteria in the mouth can affect how medicines and other foreign substances are processed. These bacteria may change such substances before they reach the intestines and liver. Finding particular bacteria among the risk markers raises the possibility that differences in the chemical activity of oral microbes help explain why Evodia rutaecarpa affects some people differently from others, a variation that has long made its safety difficult to assess.
The final result was a combined multi-omics prediction formula weighted by AUC. This single formula brings together the features identified in serum, urine and saliva. Each feature's contribution is weighted according to the predictive performance of the model that selected it. The area under the receiver operating characteristic curve, or AUC, measures how well a classification model distinguishes true positive results from true negative results across different decision thresholds. Giving more weight to models with higher AUC values means that better-performing models have a greater influence on the final score. When the formula is used for a new patient, it produces a score. A score of 0.597 or higher indicates a higher model-predicted likelihood of an adverse reaction, while a score below 0.597 indicates a lower predicted likelihood.
Overall, the study identifies ten biological features from the multi-omics data that can help distinguish between patients with and without adverse reactions. The researchers combined these features into one formula to estimate each person's risk of experiencing reactions linked to Evodia rutaecarpa. They describe this as an early framework for future research, not a fully developed rule for making clinical decisions. Because the findings show associations rather than proven causes, further experiments are needed to confirm the biological processes involved. Even so, the study shows a practical way to use samples from patients to create a personalized safety score, something that has largely been missing from safety assessments of traditional Chinese medicines.
The wider importance of this work is its potential to improve the safety assessment of other herbal treatments considered low in toxicity. Millions of people use herbs such as Evodia rutaecarpa, but their harmful effects may be underestimated because reactions are often mild, delayed or different from person to person. The study shows how combining metabolomics, microbiome sequencing and machine learning could help estimate individual risk using samples such as urine and saliva, which can be collected without invasive procedures. This approach could provide a model for assessing safety, grouping people by risk and tailoring the use of other traditional Chinese medicines considered low in toxicity. If future studies confirm the findings in larger groups of people with more varied backgrounds, this strategy could help shift herbal medicine safety monitoring from responding to reported problems to predicting risks based on biological evidence.
Subject of Research: Multi-omics and machine learning analysis of adverse reaction toxicity to the traditional Chinese medicine Evodia rutaecarpa
Article Title: Multi-omics and machine learning based on clinical samples: revealing the toxicity characteristics in adverse reactions to Evodia rutaecarpa
Article References: Tan, X., Xie, J., Guo, X., Wu, L., Shi, S., Fu, L., Zhang, W., Zeng, L., Ma, X., Zeng, J., & Zhao, Y. (2026). Multi-omics and machine learning based on clinical samples: revealing the toxicity characteristics in adverse reactions to Evodia rutaecarpa. Frontiers in Pharmacology, 17, Article 1901889. https://doi.org/10.3389/fphar.2026.1901889
Image Credits: AI Generated
DOI: 10.3389/fphar.2026.1901889
Keywords: Evodia rutaecarpa, multi-omics, machine learning, metabolomics, 16S rRNA sequencing, adverse drug reactions, traditional Chinese medicine, drug safety, risk prediction, personalized medicine, oral microbiota, pharmacology