GLP-1 receptor agonists are a class of peptide-based drugs that mimic a gut hormone involved in appetite and blood sugar regulation. They have become among the most talked-about compounds in metabolic medicine. But like any pharmacologically active compound, they come with a side-effect profile that researchers are still characterizing in real-world use.
A 2026 paper published in Scientific Reports took a novel approach to that question. The research team combined ten years of voluntary adverse event reports from the FDA with a computational framework built around graph convolutional networks, a type of machine learning that excels at finding patterns in complex, interconnected data. The goal was not just to catalog side effects, but to ask whether commonly used medicinal herbs might share biological targets with those side effects, potentially amplifying or altering them.
The findings are explicitly hypothesis-generating. The authors are careful to note that their results require independent experimental and clinical validation before any clinical application. What the study does offer is a detailed map of who reported what, when, and a computational shortlist of herbs worth investigating further.
The data set and how it was built
The researchers drew on the FDA Adverse Event Reporting System, a voluntary database where healthcare providers, patients, and manufacturers submit reports of unexpected drug reactions. The analysis covered reports filed between 2015 and 2025, yielding 142,705 GLP-1 receptor agonist adverse event reports in total. Of those, 4,090 were filed specifically under an obesity indication, which was the subgroup the team focused on most closely.
Voluntary reporting systems have well-known limitations. Reports can be duplicated, incomplete, or filed by people with varying levels of clinical detail. The authors acknowledge these limitations directly and note that signal detection in such databases reflects reporting patterns rather than confirmed causal links. With that caveat in place, the sheer volume of reports does allow statistical patterns to emerge that would be invisible in a smaller clinical trial.
Who reported side effects and when
Across the obesity-indication subgroup, gastrointestinal events were the most frequently reported category. This aligns with what earlier clinical trials had already documented, since nausea, vomiting, and related complaints are well-established features of this drug class.
Demographically, 76 percent of reports came from female patients. The researchers note this without drawing a firm mechanistic conclusion, though it is consistent with broader patterns in voluntary pharmacovigilance data where women tend to report adverse events at higher rates.
Time to onset was another variable the study tracked carefully. The majority of reports clustered in the 0 to 30 day window after starting treatment. One compound in the class, identified by its generic name semaglutide, showed a distinct pattern: a higher proportion of its reports involved late-onset cases at 360 days or more. The researchers flag this as worth attention but stop short of explaining why it might differ, since the data cannot answer that question on its own.
Signals beyond the gut
Beyond gastrointestinal complaints, statistical signal detection pointed to several less commonly discussed adverse event categories. Strong signals were detected for biliary events, pancreatic events, renal events, and coagulation events. The team used both disproportionality metrics and Bayesian approaches, two standard pharmacovigilance methods that ask whether a drug-event pair appears more often in the database than chance alone would predict.
For semaglutide, reporting odds ratios exceeded 10 for several of these pairs, which in pharmacovigilance terms indicates a strong statistical signal even if it does not confirm causation. Another compound in the class, tirzepatide, showed negative log-transformed reporting odds ratios for several gastrointestinal event types, suggesting its reporting profile differs meaningfully from the rest of the class. The researchers do not speculate heavily on the mechanism but note the difference is worth investigating.
Building the herb-compound-target network
The second and more novel component of the study was a computational network analysis designed to ask whether medicinal herbs might interact with the same biological targets implicated in these adverse events. The team used two databases: HERB 2.0, which catalogs the active compounds found in traditional medicinal herbs, and the Comparative Toxicogenomics Database, which maps relationships between chemicals, genes, and disease outcomes.
From those sources, the researchers built a multi-layer network connecting herbs to their active compounds, those compounds to their known molecular targets, and those targets to adverse event categories. They then applied graph convolutional networks to model which herb-event associations were most likely to be meaningful, followed by a filtering step that removed entries lacking drug-likeness or pharmacokinetic plausibility.
After degree debiasing, a mathematical step to prevent highly connected nodes from dominating the results, five herbs emerged as the top candidates: Liquorice Root, Mulberry Leaf, Dahurian Angelica Root, Danshen Root, and Ginkgo Leaf. The model's predictive performance was moderate, with area under the curve values of 0.798 on validation data and 0.666 on test data. The authors describe this as reasonable given how sparse pharmacovigilance networks tend to be.
What the model does and does not say
It is worth being precise about what this finding means. The model is identifying herbs whose known active compounds share molecular targets with the biological pathways implicated in GLP-1 agonist adverse events. It is not claiming that taking any of these herbs alongside a GLP-1 drug will produce a specific outcome in a specific person.
The authors are emphatic on this point. They describe the entire analysis as hypothesis-generating and explicitly state that the results require independent experimental and clinical validation before any clinical application. The computational framework is designed to narrow down a very large search space of possible herb-drug interactions to a smaller list of candidates worth testing in controlled settings.
Liquorice Root, for example, contains compounds that interact with a range of enzyme systems. Ginkgo Leaf has a long research history around coagulation pathways. Whether those interactions would matter in the context of GLP-1 therapy at typical herbal supplement doses is a question the current study cannot answer. It can only say that the molecular connections exist and are worth examining.
Why this kind of research matters for peptide science
GLP-1 receptor agonists are peptide-derived compounds, meaning they belong to the same broad chemical family as many research peptides being studied for metabolic, appetite-related, and signaling functions. Understanding the side-effect landscape of the most clinically advanced members of that family helps researchers ask better questions about the mechanisms shared across the class.
The computational approach itself is also notable. Graph convolutional networks applied to pharmacovigilance data represent a relatively new methodology. If the approach can be validated in prospective studies, it could accelerate the identification of herb-drug interactions across many compound classes, not just GLP-1 agonists. The 2026 paper is an early demonstration of what that pipeline might look like.
For now, the literature suggests that real-world adverse event data, when combined with network-based molecular modeling, can generate plausible hypotheses faster than traditional methods. Whether those hypotheses hold up under experimental scrutiny is the next question on the research agenda.
