Real-world and computational identification of herbal candidates associated with adverse event patterns in glucagon-like peptide-1 therapy for obesity Authors: Junkyu Park, Sujin Shin, Youngmin Kim, Jaiwha Seo, Byeongkil Kim, Kyungjin Lee Published: 2026 ABSTRACT Glucagon-like peptide-1 receptor agonists (GLP-1 RAs) are widely prescribed for obesity management; however, adverse events (AEs) remain a clinical concern. This study presents a computational pharmacovigilance framework integrating real-world safety data with graph-based modeling to characterize AE patterns associated with GLP-1 RA therapy and to explore herb–AE associations in an exploratory manner. We conducted a cross-sectional analysis of AE reports from the food and drug administration adverse event reporting system (2015–2025). Clinical characteristics included outcomes, reporting frequency, demographics, time-to-onset, and subgroup distributions. Signal detection employed disproportionality metrics and Bayesian approaches. Herb–compound–target–AE networks were constructed using HERB 2.0 and the Comparative Toxicogenomics Database, incorporating drug-likeness and pharmacokinetic filtering. Graph convolutional networks were applied to model herb–AE associations, followed by literature-based contextual evaluation. Among 142,705 GLP-1 RA AE reports, 4,090 involved obesity indications. Gastrointestinal events predominated, with 76% of reports involving female patients and onset clustering within 0–30 days. Semaglutide demonstrated a distinct onset distribution, including a higher proportion of late-onset cases (≥ 360 days). Strong signals were detected for biliary, pancreatic, renal, and coagulation events, with semaglutide-associated pairs showing reporting odds ratios > 10, whereas tirzepatide exhibited negative log-transformed reporting odds ratios for several gastrointestinal events. Network analysis and graph convolutional network modeling prioritized established medicinal herbs including Liquorice Root, Mulberry Leaf, Dahurian Angelica Root, Danshen Root, and Ginkgo Leaf as top candidates following degree debiasing and exclusion of non-herbal database entries. The graph convolutional network achieved area under the receiver operating characteristic curve/area under the precision–recall curve values of 0.798/0.841 (validation) and 0.666/0.719 (test), indicating moderate predictive performance within a sparse pharmacovigilance context. These findings describe real-world AE reporting patterns associated with GLP-1 receptor agonists and present a hypothesis-generating computational framework for prioritizing herb–AE signal associations. The results are exploratory and should be interpreted in light of the inherent limitations of spontaneous reporting systems and computational modeling frameworks, and require independent experimental and clinical validation prior to any clinical application.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-56002-w. Introduction In 2021, an estimated 2.11 billion adults worldwide were classified as overweight or obese, and this prevalence is projected to continue rising globally, underscoring obesity as a major public health problem1. Lifestyle interventions, pharmacotherapy, and bariatric surgery are common treatments for obesity, but their long-term effectiveness and accessibility are limited. These limitations highlight the need for improved strategies to enhance long-term treatment tolerability and persistence. Glucagon-like peptide-1 receptor agonists (GLP-1 RAs), originally developed for type 2 diabetes owing to their potent glycemic control, have subsequently gained FDA authorization for obesity treatment. These agents act on the GLP-1 receptor, stimulate insulin release, inhibit glucagon secretion, and promote satiety to support weight reduction2. Real-world cohorts also demonstrate clinically meaningful weight loss with GLP-1 RAs, and recent studies report improvements in patient-reported quality of life with once-weekly semaglutide in routine practice3,4. In addition to glucose and weight control, GLP-1 RAs have demonstrated cardiovascular and renal protective effects in large clinical outcome trials5. They also improve key cardiometabolic risk factors, such as blood pressure and lipid levels, and reduce the risk of major adverse cardiovascular events, supporting their role in comprehensive risk reduction strategies6. Recent real-world evidence indicates a rapid global expansion of newer GLP-1 RA-based weight-loss therapies, although their effectiveness in routine care remains moderate compared with trials, partly owing to treatment discontinuation driven by adverse events (AEs)7. GLP-1 RAs are associated with adverse events, most commonly gastrointestinal (GI) symptoms, such as nausea, vomiting, diarrhea, constipation, and impaired gastric emptying8. More serious but less frequent events include pancreatitis and gallbladder disorders9,10. Although such AEs have been investigated in diabetic populations through clinical trials and postmarketing reports, little is known about indication-specific patterns in obese patients treated primarily for weight reduction, as opposed to those receiving therapy for diabetes control. Pharmacovigilance databases, such as the Food and Drug Administration Adverse Event Reporting System (FAERS), are particularly valuable in this context as they capture spontaneous reports from routine clinical practice and, therefore, reflect real-world heterogeneity beyond the scope of controlled trials. FAERS has been extensively utilized for real-world pharmacovigilance analyses to detect and characterize drug-associated adverse event signals at scale11–13. Consensus recommendations emphasize dose titration and symptomatic management as primary approaches for managing GLP-1 RA–related gastrointestinal (GI) AEs14. However, persistent symptoms may affect long-term treatment adherence, prompting interest in complementary strategies that could support tolerability. Herbal medicines have historically been used for gastrointestinal and metabolic symptoms, and several natural products have been investigated for their potential relevance to drug-related adverse effects. For example, ginger has shown efficacy in the prevention and treatment of drug-induced nausea and vomiting10. Furthermore, studies have reported that natural extracts such as Gardenia jasminoides and Jatropha dioica exhibit protective effects against drug-induced hepatotoxicity14,15. Nevertheless, a systematic, data-driven evaluation of herbal candidates in the context of GLP-1 RA–associated adverse event patterns has not been conducted. Several databases have been developed to systematically catalogue drug-associated AE information, including the Comparative Toxicogenomics Database (CTD), which integrates chemical–gene–disease relationships; SIDER, which compiles drug side effects from package inserts; MetaADEDB 2.0, which integrates multiple AE data sources; and PersADE, a recently developed resource purpose-built for compound–target–AE relationships. In the present study, CTD was selected as the primary target–AE data source given its broad coverage of phytochemical–gene–disease associations. In this study, we analyzed the real-world safety profile of GLP-1 receptor agonists in obese patients using FAERS data. The primary objective was to characterize adverse event patterns in obesity-specific GLP-1 RA use. The secondary objective was to computationally prioritize herbal candidates based on integrated pharmacovigilance, network pharmacology, and graph-based modeling. By integrating real-world signal detection with computational modeling, this study provides a structured, hypothesis-generating framework for exploring herb–adverse event relationships associated with GLP-1 RA therapy. An overview of the analytical workflow is presented in Fig. 1.Fig. 1Workflow of analysis pipeline. BCPNN, Bayesian confidence propagation neural network; CTD, comparative toxicogenomics database; DEMO, demographic information; DRUG, drug information; FAERS, FDA adverse event reporting system; GCN, graph convolu-tional network; GLP-1 RA, glucagon-like peptide-1 receptor agonist; INDI, indication file; MGPS, multi-item gamma Poisson shrinker; PRR, proportional reporting ratio; REAC, reaction file; ROR, reporting odds ratio. Workflow of analysis pipeline. BCPNN, Bayesian confidence propagation neural network; CTD, comparative toxicogenomics database; DEMO, demographic information; DRUG, drug information; FAERS, FDA adverse event reporting system; GCN, graph convolu-tional network; GLP-1 RA, glucagon-like peptide-1 receptor agonist; INDI, indication file; MGPS, multi-item gamma Poisson shrinker; PRR, proportional reporting ratio; REAC, reaction file; ROR, reporting odds ratio. Materials and methods Reporting guideline and ethical approval This study was conducted in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) checklist for cross-sectional studies, as recommended by the EQUATOR Network16. The analysis was based exclusively on publicly available, de-identified data obtained from the FAERS. As the dataset contains no personally identifiable information and involves no direct interaction with human participants, this study did not constitute human subjects research. Therefore, institutional review board approval and informed consent were not required. The completed STROBE checklist is provided in the Supporting Information. Clinical trial number: not applicable, as this study did not involve a clinical trial. FAERS data sources and processes We retrieved AE reports from the FAERS spanning from the first quarter of 2015, when GLP-1 RAs were first approved for obesity and overweight, to the second quarter of 2025. Duplicate removal was based on keeping the latest record for each CASEID using CASEVERSION or FDA_DT and deleting identical combinations of DRUGNAME with drug role in DRUG and PT in REAC within each case; dose, route, and strength were not used. A total of 16,938,583 reports were identified from the demographic (DEMO) dataset. After removal of 2,411,798 duplicate records, 14,526,785 unique cases remained. Based on cross-referencing information from the DRUG and REAC files, 142,705 reports were found to be associated with GLP-1 RAs. Among these, we filtered reports containing indications (INDI) related to obesity, being overweight, or weight reduction (7,532 cases). Records missing key fields (DRUGNAME, PT, or FDA_DT) were excluded; no data imputation was performed. After excluding irrelevant or incomplete records, a final set of 4,090 cases of GLP-1 RA-associated AEs specifically linked to obesity was obtained. Among these cases, liraglutide accounted for 1,184 reports, semaglutide for 2,236 reports, and the more recently approved tirzepatide for 670 reports. Onset information was derived from the therapy (THER) file, which provides the start and end dates of drug therapy. FAERS is a spontaneous reporting system in which reports are voluntary and may contain incomplete, inaccurate, or duplicated information; therefore, the database cannot be used to establish causality or estimate the incidence of adverse events. Data mining algorithms This study applied data-mining algorithms to the FAERS database to analyze the clinical characteristics of AEs and perform signal detection. For clinical characterization, we first quantified the number of cases and incidence (outcome and incidence) for each drug–AE pair. The time to onset was calculated from the date of drug initiation to the onset of AE, and distributions were compared between drugs using the Kolmogorov–Smirnov (KS) test. Demographic characteristics, including sex, age, and weight, were summarized and subgroup analyses were conducted by stratifying the patients according to sex, age, and weight. In addition, for major AEs, disproportionality analyses were performed separately using generic drug and product names to evaluate the potential differences in reporting patterns. For signal detection, disproportionality analysis and Bayesian statistical models were employed. Disproportionality analyses included the reporting odds ratio (ROR) and proportional reporting ratio (PRR), whereas Bayesian methods included the Bayesian Confidence Propagation Neural Network (BCPNN) and Multi-item Gamma Poisson Shrinker (MGPS). ROR-based analyses were also used as the primary approach in subgroup assessments. ROR and PRR were evaluated with χ2 statistics for significance, whereas BCPNN used the information component (IC) and MGPS employed the empirical Bayes geometric mean (EBGM). A signal was considered significant if it met commonly applied thresholds (e.g., ROR point estimate > 3, χ2 > 4, IC lower bound > 0, EB05 > 2). Detailed equations for ROR, PRR, BCPNN, and MGPS were adopted from previously published studies17. Herb–compound-target–AE network construction and scoring Data on herb–compound–target relationships were obtained from the HERB 2.0 database (http://herb.ac.cn/, accessed on 2025-08-04), whereas target–AE associations were retrieved from the CTD (http://ctdbase.org, accessed on 2025-08-04). Adverse events that could not be retrieved from the CTD were excluded. Consequently, the herb–compound–target network comprised 2,620 herbs, 2,726 compounds, and 1,894 targets, whereas the final herb–compound–target–AE network, restricted to nodes with at least one CTD-linked AE association, comprised 2,435 herbs, 1,991 compounds, and 151 targets. In line with commonly adopted network pharmacology criteria, only compounds with available SMILES information, oral bioavailability ≥ 30%, drug-likeness ≥ 0.18, and a molecular weight between 180 and 550 Da were considered pharmacologically relevant and included in the analysis. Networks were constructed by integrating herb–compound, compound–target, and target–AE relationships. Specifically, compound–target edges were merged with target–AE edges to generate compound–target–AE relationships, which were then linked with compound–herb edges to build complete herb–compound–target–AE pathways. The resulting networks were analyzed using Cytoscape version 3.10.3, and the numbers of nodes and edges were quantified18. Centrality metrics (degrees) for the target genes were assessed using the CytoNCA plugin19. For target–AE mapping, only targets with documented therapeutic evidence in the CTD and those with the highest inference scores were retained, and a maximum of 30 targets per AE were allowed to minimize bias. KEGG enrichment analysis was performed for 268 AE-associated targets using ShinyGO 0.82 in conjunction with the KEGG database20. For each herb, we quantified the number of unique AEs (n_ae), total number of paths linking the herbs to AEs (n_paths), and number of distinct target genes (n_genes). The herbs were ranked according to these network-derived metrics. To identify robust candidates, the following two complementary scoring strategies were applied: scores from the “common” and “strong” edge sets were aggregated, and all paths from both sets were pooled with duplicate entries removed before recalculating herb scores. The final herbal recommendations were based on AE coverage and the number of herb–compound–target–AE pathways, with the top five herbs reported separately for common AEs, strong-signal AEs, and the union of both AE sets. This AI-driven ranking framework integrates heterogeneous network information into unified herb-level scores, thereby enabling the systematic and reproducible identification of candidate herbs with the potential to mitigate drug-related adverse events. To assess the directional specificity of the CTD-derived target–AE associations, the active compounds of the top-5 prioritized herbs were cross-validated against PersADE (https://persade.idrblab.net), a purpose-built compound–ADE database21. Compounds were matched by name and the resulting associations were filtered to retain only those overlapping with the GLP-1 RA–associated AE list identified in the FAERS analysis. GCN construction for herb candidate prioritization A heterogeneous graph was constructed with four node types (ingredients, herbs, genes, and adverse events, AEs) and three relationship types (ingredient-herb, ingredient–gene, and gene-AE). Gene–AE edges were partitioned by AE signal strength rather than randomly: union-set gene–AE edges were used for training, common-signal gene–AE edges for validation, and strong-signal gene–AE edges for testing. This stratified split was designed to evaluate the model’s capacity to generalize from broad signal associations to clinically stronger signals. Ingredient–herb and ingredient–gene edges from all three sets were provided to the encoder during both training and inference to maximize structural coverage. Each node type was initialized with a 128-dimensional embedding. Relation-specific linear layers propagate messages between nodes and perform degree-normalized mean pooling of aggregated incoming messages. The aggregated messages were combined with the original node features and then updated by a feed-forward transformation with ReLU activation and dropout (0.3). This procedure was repeated across the two layers to generate refined node representations. Edge weights for ingredient–gene relationships were adjusted based on evidence levels reported in HERB 2.0, assigning greater weight to associations supported by curated experimental or clinical evidence. This weighting scheme was applied to reflect confidence differences across heterogeneous data sources. Predictions were generated by computing the dot product between node embeddings, primarily modeling gene–AE associations. When available, herb–AE associations were incorporated as an auxiliary multitask objective with a fixed loss weight of 0.5, formulated as: total loss = BCE(gene–AE) + 0.5 × BCE(herb–AE). This weighting was set to allow the auxiliary task to provide regularization while preserving the primacy of the gene–AE prediction objective. Training employed binary cross-entropy loss with global negative sampling at a 1:1 positive-to-negative ratio. Negative samples were generated by randomly sampling unobserved gene–AE pairs while preserving node type constraints. The batch size was set to 4096. The optimization employed Adam with a learning rate of 0.001 and a weight decay of 1e-5. Early stopping was applied based on the validation of the PR-AUC with a patience of 10 epochs and a maximum of 50 epochs. After the initial training phase, model convergence was monitored using validation PR-AUC, and the best-performing model was selected based on early stopping. Model performance was evaluated using AUPRC, AUROC, and Hits@k (k = 10, 50, 100) as well as per-AE metrics and bootstrap confidence intervals. Herb–AE score matrices were generated to prioritize herbs within the common, strong signal, and overall AE groups. To correct for degree bias, each herb’s aggregated score was divided by log(1+d), where d denotes the number of ingredient connections associated with that herb, thereby penalizing herbs with disproportionately high network connectivity. Literature validation of selected herbs To complement the AI-driven herbal recommendations with clinical and mechanistic evidence, published studies were searched using PubMed, Web of Science, and Google Scholar. Searches combined each herb name with the associated adverse events, and clinical studies, RCT, and meta-analyses were reviewed. Searches used each herb’s scientific and common names combined with FAERS-derived adverse event terms, covering publications from inception through August 28, 2025 in English. Inclusion criteria were peer-reviewed primary clinical evidence and meta-analyses, while narrative reviews, traditional descriptions without clinical data, in vitro or animal-only studies, and non-scientific sources were excluded. Results Outcome events and incidence When comparing outcome distributions across GLP-1 RAs, most AEs were categorized as “other” or nonserious outcomes, although the proportion of hospitalization was higher for semaglutide than for liraglutide or tirzepatide (Fig. 2A). The most frequently reported AEs were GI symptoms, including nausea, vomiting, diarrhea, and constipation, with nausea showing the highest frequency and incidence among the three drugs. Dehydration, impaired gastric emptying, intestinal obstruction, and gallbladder-related events (e.g., cholelithiasis and cholecystitis) were among the most frequently reported AEs (Fig. 2B). The most common report overall was on off-label use, whereas off-label use and product use in unapproved indications were excluded from the disproportionality analysis because they were considered to be related to prescribing behavior rather than drug- associated AEs. Similarly, decreased appetite was excluded because it may have an intended effect in the context of obesity treatment. In the ROR heat map stratified by generic drugs, semaglutide showed relatively strong GI-related signals, particularly for impaired gastric emptying and intestinal obstruction, both of which exhibited significantly positive log(ROR) values. In contrast, tirzepatide exhibited negative log(ROR) values for these events (Fig. 2C). When stratified by product name, certain semaglutide formulations, such as Ozempic and Rybelsus, displayed strong GI-related signals, which were most prominent for impaired gastric emptying and intestinal obstruction, whereas liraglutide formulations showed relatively lower ROR values (Fig. 2D).Fig. 2Outcome and incidence of GLP-1 receptor agonist-associated adverse events (AEs). (A) Distribution of outcome categories across GLP-1 RA drug classes. (B) Top 10 most frequently reported AEs with corresponding case counts and incidence percentages. (C) Heatmap of reporting odds ratios (RORs) for gastrointestinal (GI)-related AEs stratified by GLP-1 RA generic drugs. (D) Heatmap of RORs for GI-related AEs stratified by GLP-1 RA products. Outcome and incidence of GLP-1 receptor agonist-associated adverse events (AEs). (A) Distribution of outcome categories across GLP-1 RA drug classes. (B) Top 10 most frequently reported AEs with corresponding case counts and incidence percentages. (C) Heatmap of reporting odds ratios (RORs) for gastrointestinal (GI)-related AEs stratified by GLP-1 RA generic drugs. (D) Heatmap of RORs for GI-related AEs stratified by GLP-1 RA products. Drug-stratified demographic and reporting characteristics Drug-stratified demographic analysis revealed significant differences across liraglutide, semaglutide, and tirzepatide in all four demographic categories (Fig. 3). Female patients predominated across all three drugs, accounting for 84.8%, 77.1%, and 81.2% of reports for liraglutide, semaglutide, and tirzepatide, respectively; these proportions differed significantly across drugs (χ2 = 88.98, p < 0.0001; Fig. 3A). Age distributions also differed significantly (Kruskal–Wallis H = 71.45, p < 0.0001), with median ages of 49 years (IQR 41–58) for liraglutide, 52 years (IQR 42–63) for semaglutide, and 51 years (IQR 41–61) for tirzepatide (Fig. 3B). Body weight distributions showed a statistically significant though modest difference across drugs (H = 7.71, p = 0.021), with median weights of 94.3 kg (IQR 81.0–112.0), 92.2 kg (IQR 79.0–110.3), and 91.8 kg (IQR 80.0–108.0) for liraglutide, semaglutide, and tirzepatide, respectively (Fig. 3C). Geographic distribution was predominantly U.S.-based across all three drugs; however, tirzepatide showed a markedly higher proportion of GB-based reports compared with liraglutide and semaglutide, reflecting its more recent approval timeline and differential global uptake (χ2 = 1789.05, p < 0.0001; Fig. 3D).Fig. 3Drug-stratified demographic characteristics of GLP-1 RA–associated adverse event reports. (A) Sex distribution by drug, presented as stacked bar charts showing the proportion of female (F) and male (M) reporters; group differences were assessed using the χ2 test. (B) Age distribution by drug, presented as box plots (median, IQR, and range, excluding outliers); group differences were assessed using the Kruskal–Wallis test. (C) Body weight distribution by drug, presented as box plots (median, IQR, and range, excluding outliers); group differences were assessed using the Kruskal–Wallis test. (D) Reporter country distribution by drug, presented as stacked bar charts showing the proportional contribution of the top five reporting countries; group differences were assessed using the χ2 test. Sample sizes are indicated above each bar in panel A. Drug-stratified demographic characteristics of GLP-1 RA–associated adverse event reports. (A) Sex distribution by drug, presented as stacked bar charts showing the proportion of female (F) and male (M) reporters; group differences were assessed using the χ2 test. (B) Age distribution by drug, presented as box plots (median, IQR, and range, excluding outliers); group differences were assessed using the Kruskal–Wallis test. (C) Body weight distribution by drug, presented as box plots (median, IQR, and range, excluding outliers); group differences were assessed using the Kruskal–Wallis test. (D) Reporter country distribution by drug, presented as stacked bar charts showing the proportional contribution of the top five reporting countries; group differences were assessed using the χ2 test. Sample sizes are indicated above each bar in panel A. Time-to-onset of AEs Overall, AE onset was concentrated within the early period after drug administration (0–30 days), although reports continued to occur beyond 1 year (Fig. 4A). When comparing the cumulative drug distributions, liraglutide and tirzepatide tended to show earlier AE onset, whereas semaglutide exhibited a delayed onset pattern (Fig. 4B). In the onset time bucket analysis, all three drugs showed the highest frequency of reports within 0–30 days, whereas semaglutide demonstrated a notable proportion of long-term cases reported after 360 days (Fig. 4C, D). Results from the KS test confirmed significant differences in onset distributions across drug pairs: liraglutide vs. semaglutide (stat = 0.307, p < 0.001), liraglutide vs. tirzepatide (stat = 0.114, p < 0.001), and semaglutide vs. tirzepatide (stat = 0.357, p < 0.001). Among these, semaglutide showed the most divergent distribution profile compared to the other GLP-1 RAs.Fig. 4Time-to-onset analysis of GLP-1 RA–associated adverse events. (A) Overall distribution of AE onset time. (B) Cumulative distribution of onset time by drug. (C) Case distribution by onset time buckets. (D) Proportional distribution of onset time buckets within each drug. Time-to-onset analysis of GLP-1 RA–associated adverse events. (A) Overall distribution of AE onset time. (B) Cumulative distribution of onset time by drug. (C) Case distribution by onset time buckets. (D) Proportional distribution of onset time buckets within each drug. Strong signal analysis of AEs The disproportionality analysis results that met all criteria (ROR, PRR, BCPNN, and EBGM) are presented in Fig. 5 and Table 1. For liraglutide, strong signals were detected in pancreatic events (pancreatitis, obstructive pancreatitis, and pancreatic carcinoma metastasis), cholangiocarcinoma, and cholelithiasis. For semaglutide, strong signals were observed for gallbladder and biliary events (cholelithiasis, gallbladder rupture, and cholecystitis), coagulation, and vascular events (disseminated intravascular coagulation, gastrointestinal wall thickening, and hepatobiliary procedural complications), acute kidney injury, and noncardiac chest pain. Notably, some events showed ROR values exceeding 20, and the χ2 test along with Bayesian-based indices (IC, EBGM) consistently met the thresholds. In contrast, tirzepatide did not yield a strong signal, which may be attributed to the relatively small number of reported cases.Fig. 5Strong signal analysis of GLP-1 RA adverse events. Forest plot of drug–AE pairs that satisfied all criteria across ROR, PRR, BCPNN, and EBGM. Each dot indicates the reporting odds ratio (ROR) with 95% confidence interval (log scale). Blue: liraglutide; green: semaglutide.Table 1Disproportionality analysis of strong drug-adverse event signals for GLP-1 receptor agonists.DrugAERORROR_lower95ROR_upper95PRRChi2IC_BCPNNEBGM_MGPSLIRAGLUTIDEObstructive pancreatitis28.557.6892.4122.83296.334.083.92LIRAGLUTIDECholangiocarcinoma22.584.37116.7622.3229.142.832.22LIRAGLUTIDECholelithiasis9.924.9120.067.0361.32.743.42LIRAGLUTIDEPancreatitis7.983.717.196.2839.382.582.86LIRAGLUTIDEPancreatic carcinoma metastatic6.782.3419.646.716.672.092.05SEMAGLUTIDENoncardiac chest pain54.786.1491.7654.0141.683.542.13SEMAGLUTIDECholecystitis (all)19.031.8775.1914.86604.613.023.85SEMAGLUTIDEGallbladder rupture10.734.6624.710.3948.222.663.07SEMAGLUTIDEGastrointestinal wall thickening10.282.6235.3410.0273.572.392.68SEMAGLUTIDEHepatobiliary procedural complication10.282.6235.3410.0273.572.392.68SEMAGLUTIDEDisseminated intravascular coagulation6.632.4817.746.5518.871.922.18SEMAGLUTIDESuspected counterfeit product6.323.1712.586.1635.921.872.67LIRAGLUTIDECholelithiasis9.924.9120.067.0361.32.743.42LIRAGLUTIDEPancreatitis7.983.717.196.2839.382.582.86SEMAGLUTIDEBile duct stone6.043.4710.875.68133.081.963.13SEMAGLUTIDECholelithiasis5.422.3212.644.4819.22.122.22SEMAGLUTIDEAcute kidney injury3.382.045.63.2225.31.482.29ROR, reporting odds ratio; PRR, proportional reporting ratio; CI, confidence interval; Chi2, chi-square test statistic; IC_BCPNN, information component using the Bayesian confidence propagation neural network; EBGM_MGPS, empirical Bayes geometric mean using the multi-item gamma-Poisson shrinker. Strong signal analysis of GLP-1 RA adverse events. Forest plot of drug–AE pairs that satisfied all criteria across ROR, PRR, BCPNN, and EBGM. Each dot indicates the reporting odds ratio (ROR) with 95% confidence interval (log scale). Blue: liraglutide; green: semaglutide. Table 1Disproportionality analysis of strong drug-adverse event signals for GLP-1 receptor agonists.DrugAERORROR_lower95ROR_upper95PRRChi2IC_BCPNNEBGM_MGPSLIRAGLUTIDEObstructive pancreatitis28.557.6892.4122.83296.334.083.92LIRAGLUTIDECholangiocarcinoma22.584.37116.7622.3229.142.832.22LIRAGLUTIDECholelithiasis9.924.9120.067.0361.32.743.42LIRAGLUTIDEPancreatitis7.983.717.196.2839.382.582.86LIRAGLUTIDEPancreatic carcinoma metastatic6.782.3419.646.716.672.092.05SEMAGLUTIDENoncardiac chest pain54.786.1491.7654.0141.683.542.13SEMAGLUTIDECholecystitis (all)19.031.8775.1914.86604.613.023.85SEMAGLUTIDEGallbladder rupture10.734.6624.710.3948.222.663.07SEMAGLUTIDEGastrointestinal wall thickening10.282.6235.3410.0273.572.392.68SEMAGLUTIDEHepatobiliary procedural complication10.282.6235.3410.0273.572.392.68SEMAGLUTIDEDisseminated intravascular coagulation6.632.4817.746.5518.871.922.18SEMAGLUTIDESuspected counterfeit product6.323.1712.586.1635.921.872.67LIRAGLUTIDECholelithiasis9.924.9120.067.0361.32.743.42LIRAGLUTIDEPancreatitis7.983.717.196.2839.382.582.86SEMAGLUTIDEBile duct stone6.043.4710.875.68133.081.963.13SEMAGLUTIDECholelithiasis5.422.3212.644.4819.22.122.22SEMAGLUTIDEAcute kidney injury3.382.045.63.2225.31.482.29ROR, reporting odds ratio; PRR, proportional reporting ratio; CI, confidence interval; Chi2, chi-square test statistic; IC_BCPNN, information component using the Bayesian confidence propagation neural network; EBGM_MGPS, empirical Bayes geometric mean using the multi-item gamma-Poisson shrinker. Disproportionality analysis of strong drug-adverse event signals for GLP-1 receptor agonists. ROR, reporting odds ratio; PRR, proportional reporting ratio; CI, confidence interval; Chi2, chi-square test statistic; IC_BCPNN, information component using the Bayesian confidence propagation neural network; EBGM_MGPS, empirical Bayes geometric mean using the multi-item gamma-Poisson shrinker. Subgroup analysis of AEs Figure 6 shows the subgroup analysis of common gastrointestinal AEs stratified by sex, age, and weight. For nausea, males exhibited predominantly positive log(ROR) values, especially in the < 60 kg and 60–79 kg groups across several age categories (20–29, 40–49, and ≥ 70 years), indicating a higher reporting risk compared with other male subgroups. In contrast, females showed largely neutral or negative signals, with only a few isolated subgroups showing mild elevations. For impaired gastric emptying, males demonstrated heterogeneous signals, with strong associations observed only in specific strata such as 60–69 years with < 60 kg. By comparison, females displayed consistently strong positive log(ROR) values across multiple weight groups, particularly in the 40–59 and ≥ 70 year age groups where the association exceeded log(ROR) 2. These results suggest that nausea is more strongly associated with low-weight males, whereas impaired gastric emptying is more consistently associated with females, especially middle-aged and older women.Fig. 6Subgroup analysis of common gastrointestinal adverse events associated with GLP-1 RAs. (A) Heatmap of ROR for nausea in males by age × weight. (B) Heatmap of ROR for nausea in females by age × weight. (C) Heatmap of ROR for impaired gastric emptying in males by age × weight. (D) Heatmap of ROR for impaired gastric emptying in females by age × weight. Subgroup analysis of common gastrointestinal adverse events associated with GLP-1 RAs. (A) Heatmap of ROR for nausea in males by age × weight. (B) Heatmap of ROR for nausea in females by age × weight. (C) Heatmap of ROR for impaired gastric emptying in males by age × weight. (D) Heatmap of ROR for impaired gastric emptying in females by age × weight. Herb–compound-target–AE network analysis The integrated herb–compound–target dataset comprised 2620 herbs, 2726 compounds, and 1894 unique targets connected by 13,421 herb–compound edges and 31,652 compound–target edges (Fig. 7A). The AE–target network included 15 AEs, 268 targets, and 434 edges, of which seven were classified as common AEs (131 genes) and eight as strong AEs (173 genes) (Fig. 7B). Hub gene analysis identified CASP3, IL1B, TNF, IL6, TP53, and ABCB1 as the central nodes (Fig. 7C), whereas Kyoto Encyclopedia of (KEGG) enrichment revealed dominant pathways related to inflammation, apoptosis, coagulation, and metabolic regulation, including PI3K-Akt, TNF, apoptosis, and HIF-1 signaling (Fig. 7D). These findings are consistent with the clinical profile of strong-signal AEs such as pancreatitis, cholelithiasis, acute kidney injury, and disseminated intravascular coagulation. Table 2 presents the baseline herb rankings according to network-derived metrics. Across the common, strong, and union AE sets, Licorice (Glycyrrhizae Radix et Rhizoma), Corydalis (Corydalis Rhizoma), Danshen (Salvia miltiorrhiza Radix et Rhizoma), and Mulberry Leaf (Mori Folium) consistently appeared among the top-ranked herbs, whereas Peucedanum Root (Peucedani Radix) and Rosewood (Dalbergiae Odoriferae Lignum) emerged as notable candidates. Collectively, these herbs demonstrated a broad network coverage, suggesting their potential as candidates for mitigating GLP-1 RA–associated AEs.Fig. 7Network analysis for herb candidate prioritization. (A) Overview of herb–compound–target network. (B) Target–AE network (red nodes: strong-signal AEs, green nodes: common AEs). (C) Top 10 hub genes by degree centrality in the target–AE network. (D) KEGG pathway enrichment analysis of AE-related targets. KEGG pathway information was obtained from the KEGG database (Kanehisa Laboratories, https://www.kegg.jp/).Table 2Network-based baseline herb recommendation ranking.GroupRankHerb idNameLatin namen_AEn-pathsn_genesCommon1HERB001779LicoriceGlycyrrhizae Radix et Rhizoma769747Common2HERB006353CorydalisCorydalis Rhizoma753750Common3HERB001193DanshenSalviae Miltiorrhizae Radix et Rhizoma749145Common4HERB004834Mulberry LeafMori Folium748355Common5HERB004453Peucedanum RootPeucedani Radix747051Strong1HERB001779LicoriceGlycyrrhizae Radix et Rhizoma889352Strong2HERB001193DanshenSalviae Miltiorrhizae Radix et Rhizoma855645Strong3HERB006353CorydalisCorydalis Rhizoma852450Strong4HERB002846RosewoodDalbergiae Odoriferae Lignum850644Strong5HERB004834Mulberry LeafMori Folium848159Union1HERB001779LicoriceGlycyrrhizae Radix et Rhizoma15159073Union2HERB006353CorydalisCorydalis Rhizoma15106173Union3HERB001193DanshenSalviae Miltiorrhizae Radix et Rhizoma15104766Union4HERB004834Mulberry LeafMori Folium1596484Union5HERB004453Peucedanum RootPeucedani Radix1589673n_AE, number of associated adverse events; n_paths, number of network paths connecting herb to AE via intermediate targets; n_genes, number of unique genes involved in herb–AE associations. Herbs are ranked separately for common AEs, strong-signal AEs, and the union of both categories, based on network embedding and link prediction. Network analysis for herb candidate prioritization. (A) Overview of herb–compound–target network. (B) Target–AE network (red nodes: strong-signal AEs, green nodes: common AEs). (C) Top 10 hub genes by degree centrality in the target–AE network. (D) KEGG pathway enrichment analysis of AE-related targets. KEGG pathway information was obtained from the KEGG database (Kanehisa Laboratories, https://www.kegg.jp/). Table 2Network-based baseline herb recommendation ranking.GroupRankHerb idNameLatin namen_AEn-pathsn_genesCommon1HERB001779LicoriceGlycyrrhizae Radix et Rhizoma769747Common2HERB006353CorydalisCorydalis Rhizoma753750Common3HERB001193DanshenSalviae Miltiorrhizae Radix et Rhizoma749145Common4HERB004834Mulberry LeafMori Folium748355Common5HERB004453Peucedanum RootPeucedani Radix747051Strong1HERB001779LicoriceGlycyrrhizae Radix et Rhizoma889352Strong2HERB001193DanshenSalviae Miltiorrhizae Radix et Rhizoma855645Strong3HERB006353CorydalisCorydalis Rhizoma852450Strong4HERB002846RosewoodDalbergiae Odoriferae Lignum850644Strong5HERB004834Mulberry LeafMori Folium848159Union1HERB001779LicoriceGlycyrrhizae Radix et Rhizoma15159073Union2HERB006353CorydalisCorydalis Rhizoma15106173Union3HERB001193DanshenSalviae Miltiorrhizae Radix et Rhizoma15104766Union4HERB004834Mulberry LeafMori Folium1596484Union5HERB004453Peucedanum RootPeucedani Radix1589673n_AE, number of associated adverse events; n_paths, number of network paths connecting herb to AE via intermediate targets; n_genes, number of unique genes involved in herb–AE associations. Herbs are ranked separately for common AEs, strong-signal AEs, and the union of both categories, based on network embedding and link prediction. Network-based baseline herb recommendation ranking. n_AE, number of associated adverse events; n_paths, number of network paths connecting herb to AE via intermediate targets; n_genes, number of unique genes involved in herb–AE associations. Herbs are ranked separately for common AEs, strong-signal AEs, and the union of both categories, based on network embedding and link prediction. Graph convolutional network (GCN) model performance and herb candidate prioritization The GCN model trained on the herb–compound–target–gene–AE network achieved an AUROC of 0.798 and area under the precision-recall curve (AUPRC) of 0.841 in the validation set and AUROC = 0.666 and AUPRC = 0.719 in the test set, indicating stable predictive performance despite data imbalance (Fig. 8). The top-ranked candidates predicted by the model are listed in Table 3. Following degree debiasing and exclusion of non-herbal database entries, the rankings were substantially revised; Honey, Cow Milk, and Sweetcane Culm were replaced by established medicinal herbs including Liquorice Root, Mulberry Leaf, and Danshen Root. Before-and-after ranking comparisons are provided in Supplementary Table 1. Across the common, strong, and union AE sets, Liquorice Root, Mulberry Leaf, Danshen Root, Dahurian Angelica Root, and Ginkgo Leaf consistently appeared among the leading nodes following degree debiasing and non-herbal entry exclusion. These results highlight the potential of network-based approaches for identifying candidates relevant to the mitigation of GLP-1 RA–associated adverse events. Cross-validation of the top-5 herb active compounds against PersADE confirmed four compound–ADE associations: Quercetin with diarrhea, pancreatitis, and acute kidney injury, and Salvianolic acid B with acute kidney injury (Supplementary Table 2). Although PersADE coverage of herbal constituents was limited, these confirmed associations provide additional support for the prioritization of Ginkgo Leaf and Danshen Root as candidate herbs for GLP-1 RA–associated AE mitigation. The full set of model training and hyperparameters is summarized in Supplementary Table 3.Fig. 8Performance of the GCN model trained on the herb–compound–target gene–adverse event network for herb candidate prioritization. (A, B) Validation ROC and PR curves (AUROC = 0.798, AUPRC = 0.841). (C, D) Test ROC and PR curves (AUROC = 0.666, AUPRC = 0.719).Table 3GCN-based herb recommendation ranking.GroupRankNameLatin nameGroup_scoreCommon1Danshen RootSalviae Miltiorrhizae Radix1.9995Common2Liquorice RootGlycyrrhizae Radix et Rhizoma1.9995Common3MyrrhMyrrha1.9990Common4Corydalis TuberCorydalis Rhizoma1.9983Common5RosewoodDalbergiae Odoriferae Lignum1.9973Strong1Mulberry LeafMori Folium1.9693Strong2Dahurian Angelica RootAngelicae Dahuricae Radix1.9684Strong3Liquorice RootGlycyrrhizae Radix et Rhizoma1.9678Strong4European Verbena HerbVerbenae Herba1.9676Strong5White Peony RootPaeoniae Radix Alba1.9672Union1Liquorice RootGlycyrrhizae Radix et Rhizoma1.9826Union2Mulberry LeafMori Folium1.9821Union3Dahurian Angelica RootAngelicae Dahuricae Radix1.9807Union4Danshen RootSalviae Miltiorrhizae Radix1.9805Union5Ginkgo LeafGinkgo Folium1.9801group_score, recommendation score across AE group. Herbs are ranked separately for common AEs, strong-signal AEs, and the union of both categories according to network embedding and link prediction scores. Performance of the GCN model trained on the herb–compound–target gene–adverse event network for herb candidate prioritization. (A, B) Validation ROC and PR curves (AUROC = 0.798, AUPRC = 0.841). (C, D) Test ROC and PR curves (AUROC = 0.666, AUPRC = 0.719). Table 3GCN-based herb recommendation ranking.GroupRankNameLatin nameGroup_scoreCommon1Danshen RootSalviae Miltiorrhizae Radix1.9995Common2Liquorice RootGlycyrrhizae Radix et Rhizoma1.9995Common3MyrrhMyrrha1.9990Common4Corydalis TuberCorydalis Rhizoma1.9983Common5RosewoodDalbergiae Odoriferae Lignum1.9973Strong1Mulberry LeafMori Folium1.9693Strong2Dahurian Angelica RootAngelicae Dahuricae Radix1.9684Strong3Liquorice RootGlycyrrhizae Radix et Rhizoma1.9678Strong4European Verbena HerbVerbenae Herba1.9676Strong5White Peony RootPaeoniae Radix Alba1.9672Union1Liquorice RootGlycyrrhizae Radix et Rhizoma1.9826Union2Mulberry LeafMori Folium1.9821Union3Dahurian Angelica RootAngelicae Dahuricae Radix1.9807Union4Danshen RootSalviae Miltiorrhizae Radix1.9805Union5Ginkgo LeafGinkgo Folium1.9801group_score, recommendation score across AE group. Herbs are ranked separately for common AEs, strong-signal AEs, and the union of both categories according to network embedding and link prediction scores. GCN-based herb recommendation ranking. group_score, recommendation score across AE group. Herbs are ranked separately for common AEs, strong-signal AEs, and the union of both categories according to network embedding and link prediction scores. Literature validation of herb candidate prioritization To validate the herbal recommendations derived from network analysis and GCN modeling, we systematically reviewed clinical evidence focusing on GLP-1 RA–related AEs. Table 4 summarizes the available randomized trials, meta-analyses, and early phase clinical studies evaluating herbal preparations in relation to gastrointestinal AEs, such as constipation, nausea, vomiting, diarrhea, delayed gastric emptying, and dysphagia.Table 4Literature evidence of herb recommendations for AEs.HerbAEStudy designIntervention and durationEfficacySafetyReferenceSugarcane Culm (Saccharum officinarum L.)ConstipationRCT, double-blind; 92 children (4–12 yrs, ROME III FC)1 mL/kg/day oral syrup, 1 month↑ Defecation frequency; ↓ stool retention, hard stool, fecal mass; effect comparable to PEGNo serological side effects; no body weight changes35Licorice (part of formulation)Nausea, vomitingMeta-analysis of 27 RCTs (n = 1843, colorectal cancer, oxaliplatin-based regimens)Oxaliplatin ± TM, with/without anti-emetics. 2005–213↓ CINV risk ratio 0.65 [0.59–0.71]; six specific herbs consistently associated with benefit; effects supported by experimental studies (anti-nausea, GI motility regulation, gastroprotective, antioxidant)No major safety concerns reported; integrative use considered safe in included trials36Licorice (part of PHY906 formulation)Diarrhea, nausea, vomitingPhase I, open-label, single institution (n = 24, advanced pancreatic & GI cancers)PHY906 800 mg BID (days 1–4) + capecitabine 1000–1750 mg/m2 BID (days 1–7) in 14-day cyclesMTD of capecitabine determined as 1500 mg/m2 BID; 1 partial response; 13 stable diseases > 6 weeksGenerally well tolerated; 1 DLT at lowest cohort (Grade 4 AST/ALT, Grade 3 hyponatremia); no DLTs at higher dose levels37HoneyNausea, delayed gastric emptyingRandomized controlled trial, single center (n = 55; laparoscopic cholecystectomy, ASA I–II)HS (diluted chestnut honey, n = 20) vs. CCB (n = 35); 800 mL evening before surgery + 400 mL 2 h preanesthesiaHS group: ↑ gastric motility (higher paracetamol absorption, 15–180 min); ↓ thirst, nausea, pain; ↑ stress/inflammation markers (cortisol, IL-6); ↓ appetite & well-being scoresNo major adverse events reported; overall well tolerated38Honey (Honey-lemon spray)Nausea, vomitingRandomized controlled trial (n = 46, head & neck cancer, radiotherapy 5 wks + 1 wk post)Honey-lemon spray vs. benzydamine hydrochloride sprayNo significant difference in mucositis score, pain, or QoL; OM occurrence rates lower with honey-lemon2 patients: mild nausea & throat burning; otherwise safe39Honey (Manuka honey)DysphagiaRandomized controlled trial (n = 163 lung cancer patients, concurrent chemoradiation)Supportive care vs. liquid honey 10 mL × 4/day vs. honey lozenge (10 mL equivalent) × 4/day, during treatmentNo significant difference in pain on swallowing (NRPS) or secondary endpoints; only reduced opioid use in liquid honey group at 4 weeksWell tolerated; no major safety issues reported40Honey syrupNausea, vomitingMeta-analysis of 10 RCTs (n = 3,982 women; low-risk singleton pregnancies, labor)Less-restrictive intake (including honey date syrup, carbohydrate drinks, low-residue diet) vs. restrictive intake (water/ice chips only), until deliveryLess-restrictive intake ↓ mean labor duration by ~ 16 min (95% CI -25 to -7); no differences in obstetric or neonatal outcomesNo increased incidence of vomiting, regurgitation, or aspiration; well tolerated41Honey and Sugarcane Culm are included in this table as they appeared in the original GCN rankings prior to degree debiasing and non-herbal entry exclusion. Literature evidence is retained for completeness and transparency; these entries are not considered as primary herbal candidates in the revised analysis. RCT, randomized controlled trial; TM, traditional medicine; CINV, chemotherapy-induced nausea and vomiting; OM, oral mucositis; PCS, patient-controlled sedation; NRPS, Numerical Rating Pain Scale; ASA, American Society of Anesthesiologists; BID, twice daily; DLT, dose-limiting toxicity; MTD, maximum tolerated dose; QoL, quality of life; CCB, commercial carbohydrate beverage; HS, honey solution. Table 4Literature evidence of herb recommendations for AEs.HerbAEStudy designIntervention and durationEfficacySafetyReferenceSugarcane Culm (Saccharum officinarum L.)ConstipationRCT, double-blind; 92 children (4–12 yrs, ROME III FC)1 mL/kg/day oral syrup, 1 month↑ Defecation frequency; ↓ stool retention, hard stool, fecal mass; effect comparable to PEGNo serological side effects; no body weight changes35Licorice (part of formulation)Nausea, vomitingMeta-analysis of 27 RCTs (n = 1843, colorectal cancer, oxaliplatin-based regimens)Oxaliplatin ± TM, with/without anti-emetics. 2005–213↓ CINV risk ratio 0.65 [0.59–0.71]; six specific herbs consistently associated with benefit; effects supported by experimental studies (anti-nausea, GI motility regulation, gastroprotective, antioxidant)No major safety concerns reported; integrative use considered safe in included trials36Licorice (part of PHY906 formulation)Diarrhea, nausea, vomitingPhase I, open-label, single institution (n = 24, advanced pancreatic & GI cancers)PHY906 800 mg BID (days 1–4) + capecitabine 1000–1750 mg/m2 BID (days 1–7) in 14-day cyclesMTD of capecitabine determined as 1500 mg/m2 BID; 1 partial response; 13 stable diseases > 6 weeksGenerally well tolerated; 1 DLT at lowest cohort (Grade 4 AST/ALT, Grade 3 hyponatremia); no DLTs at higher dose levels37HoneyNausea, delayed gastric emptyingRandomized controlled trial, single center (n = 55; laparoscopic cholecystectomy, ASA I–II)HS (diluted chestnut honey, n = 20) vs. CCB (n = 35); 800 mL evening before surgery + 400 mL 2 h preanesthesiaHS group: ↑ gastric motility (higher paracetamol absorption, 15–180 min); ↓ thirst, nausea, pain; ↑ stress/inflammation markers (cortisol, IL-6); ↓ appetite & well-being scoresNo major adverse events reported; overall well tolerated38Honey (Honey-lemon spray)Nausea, vomitingRandomized controlled trial (n = 46, head & neck cancer, radiotherapy 5 wks + 1 wk post)Honey-lemon spray vs. benzydamine hydrochloride sprayNo significant difference in mucositis score, pain, or QoL; OM occurrence rates lower with honey-lemon2 patients: mild nausea & throat burning; otherwise safe39Honey (Manuka honey)DysphagiaRandomized controlled trial (n = 163 lung cancer patients, concurrent chemoradiation)Supportive care vs. liquid honey 10 mL × 4/day vs. honey lozenge (10 mL equivalent) × 4/day, during treatmentNo significant difference in pain on swallowing (NRPS) or secondary endpoints; only reduced opioid use in liquid honey group at 4 weeksWell tolerated; no major safety issues reported40Honey syrupNausea, vomitingMeta-analysis of 10 RCTs (n = 3,982 women; low-risk singleton pregnancies, labor)Less-restrictive intake (including honey date syrup, carbohydrate drinks, low-residue diet) vs. restrictive intake (water/ice chips only), until deliveryLess-restrictive intake ↓ mean labor duration by ~ 16 min (95% CI -25 to -7); no differences in obstetric or neonatal outcomesNo increased incidence of vomiting, regurgitation, or aspiration; well tolerated41Honey and Sugarcane Culm are included in this table as they appeared in the original GCN rankings prior to degree debiasing and non-herbal entry exclusion. Literature evidence is retained for completeness and transparency; these entries are not considered as primary herbal candidates in the revised analysis. RCT, randomized controlled trial; TM, traditional medicine; CINV, chemotherapy-induced nausea and vomiting; OM, oral mucositis; PCS, patient-controlled sedation; NRPS, Numerical Rating Pain Scale; ASA, American Society of Anesthesiologists; BID, twice daily; DLT, dose-limiting toxicity; MTD, maximum tolerated dose; QoL, quality of life; CCB, commercial carbohydrate beverage; HS, honey solution. Literature evidence of herb recommendations for AEs. Honey and Sugarcane Culm are included in this table as they appeared in the original GCN rankings prior to degree debiasing and non-herbal entry exclusion. Literature evidence is retained for completeness and transparency; these entries are not considered as primary herbal candidates in the revised analysis. RCT, randomized controlled trial; TM, traditional medicine; CINV, chemotherapy-induced nausea and vomiting; OM, oral mucositis; PCS, patient-controlled sedation; NRPS, Numerical Rating Pain Scale; ASA, American Society of Anesthesiologists; BID, twice daily; DLT, dose-limiting toxicity; MTD, maximum tolerated dose; QoL, quality of life; CCB, commercial carbohydrate beverage; HS, honey solution. Discussion Use of GLP-1 RAs for obesity is rapidly increasing, and their AE profiles warrant close scrutiny as prescriptions expand beyond the diabetes population. This study is one of the few real-world pharmacovigilance analyses of GLP-1 RA–associated AEs in the context of obesity, leveraging the FAERS to evaluate clinical patterns, demographic features, and a hypothesis-generating framework for exploring potential mitigation strategies. The combination of AE profiling, network pharmacology, and GCN modeling distinguishes this study from earlier studies that primarily focused on single AEs or trial-derived data. Moreover, rather than emphasizing GLP-1 mimetic compounds, we explored herbal candidates as potential supportive options for GLP-1 RA–related side effects, offering a novel direction for pharmacovigilance research. Our analysis of the overall AE landscape confirmed that gastrointestinal events were the most frequent and represented class effects, consistent with the results of clinical trials and previous pharmacovigilance reports22. These symptoms include nausea, vomiting, diarrhea, and constipation, all of which arise from delayed gastric emptying and central satiety mechanisms intrinsic to GLP-1 signaling. The high frequency of these AEs mirrors previous findings in both diabetes and obesity cohorts. However, the present study emphasizes their persistence in real-world data specifically linked to weight loss indications23. The subgroup and time-to-onset analyses revealed further granularity. We observed that males, particularly those with lower body weight, reported nausea more frequently, whereas impaired gastric emptying was more consistently associated with middle-aged and older females. Onset analyses indicated that most AEs occurred within the first 30 days of treatment initiation, although semaglutide continued to show a substantial proportion of late-onset reports after 1 year. These findings underscore the need for clinicians to carefully monitor patients during the early stages of treatment while remaining aware of the potential for long-term complications. Importantly, the subgroup patterns suggest that demographic and anthropometric characteristics may inform future hypothesis-driven research on individualized risk patterns, an implication that has not been well established in previous research. Therefore, obesity-specific perspectives are crucial. Among more than 140,000 GLP-1 RA AE reports identified, only 2.9% were linked to obesity or weight loss indications. This highlights the relative novelty of obesity-specific prescriptions and the limitations of relying solely on diabetes-derived safety data. Existing pharmacovigilance studies have focused on different aspects of GLP-1 RA safety: misuse and abuse potential of semaglutide, disproportionate reporting of diabetic ketoacidosis, and mortality or serious adverse events associated with earlier GLP-1 RAs, such as liraglutide24–26. In contrast, our study offers a broad assessment of the AE landscape in obese populations, which is currently underrepresented in the literature. This distinction is important because obese patients differ from diabetic patients in terms of baseline comorbidities, concomitant medication use, and metabolic profiles, all of which can influence AE risks. Therefore, our findings provide context-specific insights that cannot be directly extrapolated to cohorts with diabetes. In the demographic analysis, adverse event reports were more frequent among females and middle-aged individuals. This pattern may reflect greater interest in obesity management among women and the rising prevalence of obesity in middle age, resulting in higher GLP-1 RA exposure in these groups. Disproportionality analyses revealed strong signals involving the pancreas, gallbladder, and kidney, consistent with previous FAERS-based studies that reported pancreatitis, biliary events, and renal complications among GLP-1 RA users. Thus, these associations are biologically plausible. GLP-1 receptor activation slows gastric emptying and alters biliary motility, which may predispose to gallbladder stasis and cholelithiasis27,28. The pancreatic effects may reflect increased enzyme secretion or inflammatory signaling triggered by the GLP-1 pathway29. Kidney-related events, including acute kidney injury, may arise from volume depletion due to gastrointestinal loss or a direct inflammatory response30. Our enrichment analysis further supported these observations, with pathways such as PI3K-Akt, TNF, apoptosis, and HIF-1 prominently represented among AE-related targets. These pathways converge on inflammation, cell death, and metabolic stress, which is consistent with observed clinical syndromes. Although rare, disseminated intravascular coagulation may also be explained by systemic inflammatory cascades linking immune dysregulation to coagulation abnormalities31. Taken together, the mechanistic plausibility of these strong signals emphasizes the importance of ongoing surveillance, even in patients without diabetes or other major comorbidities. Tirzepatide presented an interesting case because no strong signals were detected in our dataset. This aligns with clinical trial data, suggesting potentially different reporting patterns compared to older GLP-1 RAs32. In particular, the negative log(ROR) values for gastrointestinal events such as impaired gastric emptying and intestinal obstruction may indicate a lower risk of these AEs compared with semaglutide. However, the relative paucity of reports on tirzepatide in the FAERS dataset limits the confidence in this finding. Specifically, tirzepatide had 670 reports, compared with 1,184 for liraglutide and 2,236 for semaglutide, indicating that the smaller reporting volume may partly explain the absence of strong signals. Further postmarketing surveillance is needed as prescription volumes increase. Therefore, the absence of strong signals in this analysis should be interpreted cautiously and clinicians should remain attentive to emerging data as real-world exposure increases. In addition to signal detection, this study introduces a novel approach for exploring AE mitigation. By constructing an herb–compound–target–AE network using drug-likeness filters such as oral bioavailability, drug-likeness index, and molecular weight thresholds, we identified 2,435 herbs, 1,991 ingredients, 151 targets, and 15 AEs suitable for analysis. Notably, 36 targets overlapped between the common and strong AE sets, many of which were involved in inflammatory or immune responses. Enrichment analysis reinforced the conclusion that AE-associated genes clustered in inflammation- and apoptosis-related pathways. Network metrics identified well-studied herbs such as Licorice as central nodes, consistent with their widespread traditional use in East Asian medicine to moderate adverse effects. Although these findings highlight the biological plausibility of the network, they also reveal the limitations inherent to database-driven approaches: herbs with greater ethnopharmacological prominence are more likely to appear as top candidates, reflecting both biological activity and reporting bias. The GCN framework incorporates network structure to model herb–AE associations not only using node features but also the structure of local neighborhoods within the network33,34. This allowed the model to infer herb–AE associations more effectively than embedding alone. The validation performance was strong (AUPRC of approximately 0.84), whereas the test performance was moderate but meaningful, given the class imbalance (AUPRC of approximately 0.72). These results indicate that GCNs can be generalized beyond training data, although further refinement is required. Prior to degree debiasing, food-derived entries such as Honey and Cow’s Milk were ranked among the top candidates, reflecting a limitation of the HERB database. Following re-implementation of explicit degree debiasing and post-hoc exclusion of non-herbal entries, these artifacts were removed from the final rankings. Literature validation indicated that several top-ranked candidates, including Licorice, honey, and sugarcane, have varying degrees of clinical or safety-related evidence in supportive care contexts relevant to gastrointestinal or treatment-related adverse events35–41. While the GCN framework provides a structured approach for herb candidate prioritization, several factors beyond the scope of the current model may influence clinical outcomes. These include chemical interactions among co-occurring phytochemical constituents within a single herb, network topological characteristics such as hub node connectivity that may bias prioritization scores, individual genetic variation in drug-metabolizing enzyme activity, and real-world prescribing frequency, which may reflect accumulated clinical evidence not captured in the computational model. The current analysis should therefore be regarded as a hypothesis-generating signal prioritization tool rather than a definitive recommendation framework. However, a closer examination of the literature highlights the limitations of this study. Licorice has primarily been studied as a multi-herbal formulation, such as PHY906 or oxaliplatin-based TMs, rather than as a stand-alone intervention, and clinical evidence for the top-ranked herbs in the specific context of GLP-1 RA–associated AEs remains limited. In addition, clinical validation so far has focused mainly on gastrointestinal AEs, such as constipation and nausea/vomiting, whereas direct evidence for the mitigation of the “strong signal AEs identified in our FAERS analysis (e.g., pancreatitis and cholelithiasis) remains limited. Importantly, no clinical study has yet demonstrated clear benefit for these severe AEs, representing a major evidence gap that warrants focused investigation. These patterns suggest that the current evidence base is uneven, and further targeted studies are needed to strengthen its translational relevance. Previous pharmacovigilance studies on herbal medicine have mainly addressed their AEs and safety monitoring42,43. Other work has investigated pharmacovigilance signals of AEs arising from herb–drug interactions44. Clinically, interactions arise via metabolic inhibition or induction that alters drug exposure, pharmacodynamic antagonism or synergy that shifts therapeutic effect, and idiosyncratic reactions. In contrast, few studies have systematically investigated the use of herbal medicines to counteract the adverse effects of conventional drugs, highlighting the novelty of our integrative approach. From a clinical perspective, this underscores the importance of structured medication review and clear patient communication about herbal use. Cautious dose adjustment during co-administration and vigilant follow-up in routine care are also essential to ensure safe and effective integration. The potential clinical implications of the prioritized herbal candidates are twofold. First, the identification of herbs with established safety profiles, such as Licorice and Danshen, provided proof of the concept that network pharmacology and machine learning can converge on clinically meaningful candidates. Second, the emergence of herbs with limited clinical evidence in the GLP-1 RA context, such as Myrrh and European Verbena Herb, underscores the need for a cautious interpretation of database-derived predictions. While ethnopharmacological evidence suggests that many top-ranked herbs have long been used to alleviate adverse effects, herb–drug interactions remain a concern. Liquorice root contains glycyrrhetinic acid and glabridin, which have been shown to inhibit CYP2C9 and CYP3A4 in vitro, with species-specific variation in inhibitory potency45. Danshen root constituents including tanshinone I and dihydrotanshinone I are potent CYP1A2 inhibitors and moderate CYP2C9 inhibitors in human liver microsomes, with additional CYP3A4 induction via pregnane X receptor activation reported for lipophilic tanshinones46. Ginkgo leaf extract at doses above the recommended 240 mg/day has been associated with weak CYP2C19 induction and CYP3A4 inhibition, though clinically relevant interactions were not observed at standard doses in human pharmacokinetic studies47. Mulberry leaf exerts α-glucosidase inhibitory activity, which may produce pharmacodynamic synergism with the glucose-lowering effects of GLP-1 RAs, potentially increasing the risk of hypoglycemia48. Given that GLP-1 RAs are primarily eliminated via proteolytic degradation rather than hepatic CYP metabolism, direct pharmacokinetic interactions may be limited; however, concomitant effects on gastrointestinal motility, pancreatic function, and metabolic enzyme activity warrant prospective evaluation before clinical implementation. This underscores the need for rigorous preclinical and clinical validation before the recommendations can be translated into clinical practice. A prospective cohort study would be appropriate to evaluate real-world outcomes. This approach can assess both the effectiveness of herbal co-administration in reducing GLP-1 RA–related adverse events and its safety by monitoring symptom changes, discontinuation rates, and potential herb–drug interactions. Considering perspectives for clinical and assistive practice, herb-based adjunctive care may support GLP-1 RA therapy in obesity. Early gastrointestinal symptoms often reduce adherence and persistence, and safe, well-characterized herbs could help control these symptoms and reduce escalation through structured monitoring. Integration can occur within routine patient monitoring and education coordinated by physicians, pharmacists, and allied professionals. Programs should use standardized dosing windows and predefined stop rules, screen for herb–drug interactions before and during therapy, and record adverse-event and adherence metrics at fixed intervals. Evidence remains uneven and susceptible to database artifacts, so these points are hypothesis-generating rather than prescriptive. Implementation should proceed within protocolized pathways and await interaction-aware clinical trials to confirm effectiveness and safety. This study has several constraints inherent to spontaneous reporting systems and modeling frameworks. First, FAERS is subject to underreporting, as submission of adverse event reports is voluntary for healthcare professionals and consumers; consequently, the true incidence of adverse events associated with GLP-1 RAs is likely underestimated. Second, reporting bias may arise from factors such as media coverage, regulatory actions, and product labeling changes, which can disproportionately increase or decrease reporting for specific drug–event combinations over time. Third, FAERS lacks denominator data (i.e., the number of patients exposed to each drug), which precludes the calculation of true incidence rates and limits direct comparisons of event frequencies across drugs. Fourth, as a spontaneous reporting system, FAERS cannot establish causality between drug exposure and adverse events; disproportionality signals reflect reporting patterns and should be interpreted as hypothesis-generating associations rather than confirmed causal relationships. These limitations were considered throughout the analysis, and all findings should be interpreted accordingly. Addressing limitations and generalisability, this study has several constraints inherent to spontaneous reporting systems and modeling frameworks. When restricted to obesity, the relatively small number of reports may exaggerate some signals, and incomplete demographic and onset information could introduce bias in subgroup or latency analyses. The predominance of U.S.-based data also limits geographic representativeness. In addition, the network and GCN models depend on the completeness and consistency of public databases; compound–target information is not absolute and may influence predictive performance. Thus, the top-ranked herbs represent computational hypotheses, and supporting literature generally concerns symptomatic relief rather than GLP-1 RA–specific adverse events; non-herbal or mapping artifacts in source databases further warrant caution. The HERB 2.0 database includes food-origin entries that are not strictly classified as medicinal plants; degree debiasing and post-hoc exclusion were applied in the revised analysis to address this issue. The CTD database encompasses a broad range of chemical–gene–disease relationships including therapeutic, toxicological, and correlational associations, which are not exclusively restricted to adverse drug events. Although we attempted to cross-validate the target–AE associations of the top-5 herb compounds against PersADE, a purpose-built ADE database, the coverage of traditional herbal constituents in PersADE was insufficient for a complete replacement. Future studies should consider integrating ADE-dedicated resources as they expand their coverage of phytochemical compounds. Evidence remains limited, especially for strong-signal events such as pancreatitis and cholelithiasis. These factors restrict the generalisability of the findings, and the results alone are insufficient for direct extrapolation to non-U.S. populations or those with different reporting behaviors. To enhance external validity, interaction-aware prospective clinical studies are needed to confirm the effectiveness and safety of the predicted herbal candidates. Disproportionality signals identified in this study reflect reporting associations within the FAERS database and should not be interpreted as evidence of causality. In particular, signals for serious adverse events such as pancreatitis and cholelithiasis may be subject to confounding by indication, comedication, and underlying metabolic comorbidities, and require confirmation through prospective clinical studies or pharmacoepidemiological designs with appropriate comparator groups. Therefore, the present findings should be interpreted as exploratory and hypothesis-generating rather than confirmatory. Conclusions This study presents a real-world pharmacovigilance analysis of GLP-1 RA AEs in the context of obesity by integrating FAERS data with network pharmacology and graph convolutional modeling. Gastrointestinal AEs were confirmed as class effects, whereas strong signals involving the pancreas, gallbladder, and kidney highlight safety concerns in obesity populations beyond traditional diabetes cohorts. The subgroup and onset analyses revealed demographic variations that warrant further clinical investigation. Network embedding and GCN prediction identified candidate herbs, including Licorice, Corydalis, and Danshen, that were computationally prioritized in relation to the selected AEs. Following degree debiasing and exclusion of non-herbal database entries, established medicinal herbs including Liquorice Root, Mulberry Leaf, Dahurian Angelica Root, Danshen Root, and Ginkgo Leaf were prioritized as top candidates, and literature validation supported the plausibility of several of these herbs in relation to GLP-1 RA–associated AEs. Collectively, this integrative framework provides a foundation for identifying herbal candidates for further evaluation in the context of drug-related AEs and suggests future directions for clinical validation through prospective cohort and pragmatic interventional studies, with careful assessment of herb–drug interaction risks prior to clinical implementation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 1.