Machine Learning and Multi-Omics Reveal Toxicity Signals in Adverse Reactions to a Traditional Chinese Medicine

Evodia rutaecarpa, a fruit long used in traditional Chinese medicine to warm the interior, dispel cold and relieve pain, occupies an unusual position in modern pharmacology. It is classified as a low-toxicity herb, yet clinical experience shows that some patients develop adverse reactions while taking formulas that contain it, and those reactions are often subtle, variable from person to person, and easy to overlook. A new study published in Frontiers in Pharmacology has now combined untargeted metabolomics, oral microbiome sequencing and an ensemble of machine learning algorithms to map the biological fingerprints of these reactions in real patients, and to build a quantitative formula that estimates an individual’s likelihood of experiencing them.

The research team enrolled 86 patients with gastric diseases who were receiving compound herbal formulas containing Evodia rutaecarpa. Based on whether adverse reactions appeared during treatment, the cohort was divided into a control group of 50 patients and an adverse reaction group of 36 patients. From each participant, the investigators collected three biological matrices: serum, urine and saliva. Serum and urine underwent untargeted metabolomics profiling, while saliva was subjected to 16S rRNA gene sequencing to characterize the oral bacterial community. This design allowed the researchers to search for toxicity signals across both chemical and microbial dimensions simultaneously, rather than relying on any single type of measurement.

Clinically, the picture that emerged was dominated by the digestive tract. Adverse events associated with Evodia rutaecarpa predominantly presented as upper-gastrointestinal symptoms, a pattern consistent with the herb’s known irritant potential at mucosal surfaces. But the omics data revealed that the perturbations extended well beyond the gut. Patients in the adverse reaction group showed disturbed amino acid metabolism, suggesting a systemic shift in how nitrogen and protein building blocks were being processed. Their neuroactive molecule profiles were also shifted, hinting that the herb’s bioactive alkaloids may be interacting with neurotransmitter-related pathways in susceptible individuals.

A third systemic signal involved bile. The metabolomics analysis identified altered bile secretion signatures among patients who experienced adverse reactions, pointing to possible disruption of enterohepatic circulation, the recycling loop that shuttles bile acids between liver and intestine. Because bile acids act not only as digestive detergents but also as signaling molecules that regulate gut barrier integrity and inflammation, this finding offers a plausible mechanistic thread connecting the herb’s chemical constituents to the gastrointestinal symptoms that dominated the clinical presentation. Alongside these metabolic changes, the salivary microbiome data showed dysbiosis, a disturbance in the composition of the oral bacterial community, which the authors flag as a tentative correlative trend rather than a proven cause.

To convert these high-dimensional omics observations into a usable risk tool, the team deployed seven different machine learning algorithms, including logistic regression, random forest and LightGBM, and validated model performance using leave-one-out cross-validation, a rigorous scheme in which each patient is held out in turn and predicted by a model trained on all others. The algorithms were used both to build classifiers that distinguish adverse reaction cases from controls and to screen the feature space for the molecular and microbial variables that carried the most discriminative information. This dual use of machine learning, as predictor and as feature selector, is increasingly common in pharmacogenomics because it helps prioritize candidates for experimental validation.

The screening process yielded multiple model-associated classification features drawn from the multi-omics datasets. In serum, three metabolites showed statistical trends linked to lower model prediction outputs for adverse events: 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) correlated with reduced predictive values, while arachidoyl ethanolamide, an endocannabinoid-family lipid, associated with elevated model prediction scores. The directionality of these associations matters: metabolites that push predictions downward may represent protective or homeostatic states, whereas those that push predictions upward may mark vulnerability to the herb’s toxic effects.

The microbial features were equally specific. Higher relative abundances of four salivary taxa, the genera Granulicatella and Alysiella, the species Leptotrichia hongkongensis and the genus Streptococcus, corresponded to higher predictive probabilities from the classification model. Oral bacteria are increasingly recognized as active participants in drug metabolism, capable of transforming xenobiotics before they reach the gut and liver, so the presence of specific taxa as risk markers raises the possibility that individual differences in oral microbial chemistry contribute to the inter-individual variability that has long made Evodia rutaecarpa’s safety profile hard to pin down.

The culmination of the analysis is an AUC-weighted multi-omics joint formula, a single equation that combines the discriminative features from serum, urine and saliva, with each feature’s contribution weighted by the predictive performance of the model that selected it. The area under the receiver operating characteristic curve, or AUC, quantifies how well a classifier separates true positives from true negatives across all decision thresholds, so weighting by AUC gives better-performing models proportionally more influence over the final score. When applied to a new patient, the formula produces a continuous score: values at or above 0.597 indicate a higher model-derived predictive tendency toward adverse reactions, while values below that threshold indicate lower predictive possibility.

In total, the study identifies ten omics-derived discriminative features and packages them into the joint prediction formula, enabling quantitative estimation of individual-level risk for Evodia rutaecarpa-related adverse reactions. The authors present this as a preliminary reference framework rather than a finished clinical decision rule, and the correlative nature of the omics trends means that experimental validation of the underlying mechanisms remains an open task. Even so, the work demonstrates a concrete path from patient samples to a personalized safety score, something that has been largely absent from the safety assessment of traditional Chinese medicines.

The broader significance lies in what the approach could mean for low-toxicity herbal therapies generally. Herbs like Evodia rutaecarpa are used by millions, yet their adverse reactions are frequently underestimated precisely because they are mild, delayed and variable. By showing that a combination of metabolomics, microbiome sequencing and machine learning can stratify risk at the individual level using noninvasive matrices such as urine and saliva, the study offers a template for safety assessment, risk stratification and personalized application across the wider class of low-toxicity traditional Chinese medicines. If validated in larger and more diverse cohorts, the strategy could shift herbal pharmacovigilance from reactive case reporting toward proactive, biology-driven prediction.

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