Researchers at Shandong University of Traditional Chinese Medicine and affiliated institutions have identified glycyrrhisoflavone as a potential natural agonist for the peroxisome proliferator-activated receptor gamma, or PPARγ, a key therapeutic target for type 2 diabetes. The study, published in Frontiers in Chemistry, utilized a multistage virtual screening pipeline that integrated graph neural network prediction, molecular docking, ADMET evaluation, in vitro activity assessment, and molecular dynamics simulations. This computational approach aimed to overcome the limitations of current synthetic PPARγ agonists, which are often associated with adverse effects such as fluid retention and weight gain. By focusing on natural products, the team sought to discover compounds with improved safety profiles and distinct binding characteristics.

The research team constructed three different graph neural network models to predict the biological activity of compounds against PPARγ: Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and Message-Passing Neural Networks (MPNN). These models were trained on a dataset of 4,029 compounds with known PPARγ agonist activity data derived from the ChEMBL database. The dataset was divided into training, validation, and test sets to ensure robust model evaluation. The GAT model demonstrated the best overall classification performance, achieving an accuracy of 0.809 and a Matthews correlation coefficient of 0.617, which outperformed the other two architectures in identifying active compounds.

Using the optimized GAT model, the researchers screened a library of 4,112 natural products from the ZINC database. This initial computational filter yielded 291 candidate compounds with activity predictive scores greater than 0.5. These candidates were then subjected to molecular docking studies against the PPARγ protein structure (PDB ID: 2F48) to evaluate their predicted binding affinities. The docking process identified 50 molecules with favorable predicted binding affinities, which were subsequently analyzed for their pharmacokinetic properties using the ADMET lab 3.0 platform. This step was crucial for assessing the potential oral bioavailability and safety of the candidates.

The ADMET analysis revealed that the five top-ranked candidates met the Lipinski rule of five, a guideline for evaluating the oral absorption potential of compounds. These molecules showed good predicted intestinal absorption and low blood-brain barrier penetration, which is desirable for drugs intended to target peripheral insulin-sensitive tissues rather than the central nervous system. The predicted hERG blocking values were also low, suggesting a reduced risk of cardiotoxicity. Based on these computational predictions, five specific compounds were selected for experimental validation: Hypocrellin B, Resistomycin, Glycyrrhisoflavone, Rottlerin, and Cowaxanthone.

In vitro assays using a firefly luciferase reporter gene system confirmed that only glycyrrhisoflavone exhibited significant concentration-dependent PPARγ agonistic activity. The compound achieved an EC50 value of 9.98 μM, indicating moderate potency, with a maximum activation efficacy of 100%. In contrast, the other four candidates showed weak activity, with maximum activation rates below 13%. Notably, Cowaxanthone displayed an inhibitory effect at high concentrations, further distinguishing glycyrrhisoflavone as the most promising lead compound among the tested natural products.

Molecular docking and visualization studies revealed that glycyrrhisoflavone adopts a binding mode in the PPARγ ligand-binding domain that is highly similar to that of rosiglitazone, a clinically approved full PPARγ agonist. Both compounds form conventional hydrogen bonds with the key residues ASN388 and TYR244, which are critical for ligand recognition in the binding pocket. Additionally, glycyrrhisoflavone forms cation-π interactions with the guanidine group of ARG253 and the amino group of LYS243. These interactions, along with hydrophobic contacts with MET251 and ILE205, suggest that glycyrrhisoflavone anchors itself in the same core region of the receptor as the synthetic drug.

To further investigate the stability and dynamic behavior of the complex, the researchers performed 100 ns molecular dynamics simulations using the Amber software package. The simulations compared the glycyrrhisoflavone-PPARγ complex with the rosiglitazone-PPARγ reference system. The results showed that while the rosiglitazone complex reached equilibrium quickly and remained stable, the glycyrrhisoflavone complex exhibited moderate dynamic fluctuations. These fluctuations included reversible conformational rearrangements, which may be indicative of a partial agonist mechanism rather than full receptor activation.

The molecular dynamics analysis also highlighted differences in the correlation of movements between the ligand-binding domain (LBD) and the DNA-binding domain (DBD). In the rosiglitazone complex, there was a strong positive correlation between these domains, supporting the stabilization of the activated conformation necessary for transcriptional activation. In contrast, the glycyrrhisoflavone complex showed a weakened positive correlation within the LBD and a significant negative correlation between the LBD and DBD. This distinct dynamic behavior suggests that glycyrrhisoflavone may induce a partially activated conformational state, potentially offering a mechanism for selective modulation of PPARγ signaling.

The study acknowledges several limitations, primarily that the ADMET properties and binding affinities are based on computational predictions and require experimental validation. The in vitro activity of glycyrrhisoflavone, while significant, is moderate compared to synthetic full agonists like rosiglitazone, which has an EC50 of 0.043 μM. The authors emphasize that further studies are needed to evaluate the metabolic stability, pharmacokinetic profile, and in vivo efficacy of glycyrrhisoflavone in diabetic animal models. Additionally, complementary biophysical techniques are recommended to further validate the binding mechanism and conformational dynamics observed in the simulations.

This research demonstrates the utility of integrating graph neural networks with traditional computational chemistry methods for the discovery of natural product drug candidates. The successful identification of glycyrrhisoflavone as a PPARγ agonist with a distinct binding mode and dynamic profile provides a new avenue for the development of safer therapeutic agents for type 2 diabetes. The multistage screening approach effectively narrowed the vast chemical space of natural products to a manageable number of candidates for experimental testing, highlighting the efficiency of AI-assisted drug discovery.

The findings suggest that natural products may offer unique pharmacological profiles that differ from synthetic drugs, potentially reducing the risk of adverse effects associated with full PPARγ activation. Glycyrrhisoflavone’s ability to form stable interactions with key receptor residues while exhibiting partial agonist-like dynamics makes it a compelling candidate for further preclinical evaluation. Future work will focus on optimizing the compound’s metabolic stability and confirming its therapeutic potential in vivo.

The study underscores the importance of combining data-driven prediction with rigorous experimental validation in modern drug discovery. By leveraging the structural insights provided by molecular docking and dynamics simulations, researchers can better understand the mechanisms of action of novel compounds. This integrated approach not only accelerates the identification of lead candidates but also provides a deeper mechanistic understanding that can guide subsequent structural optimization and development.

As the global burden of type 2 diabetes continues to rise, the need for novel therapeutic strategies with improved safety profiles is urgent. The identification of glycyrrhisoflavone as a natural PPARγ modulator represents a significant step forward in this area. The research team’s work provides a systematic framework for the discovery of natural agonists and establishes a foundation for the further development of glycyrrhisoflavone as a potential treatment for metabolic disorders.

The computational screening pipeline began with a dataset of 4,029 compounds sourced from the ChEMBL database, which contained experimentally determined PPARγ agonist activity data. To classify these molecules, the researchers applied a threshold of 1,000 nM for the half-maximal effective concentration, or EC50. Compounds with EC50 values at or below this threshold were labeled as active, while those exceeding it were labeled as inactive. This binary classification allowed the graph neural network models to learn the structural features associated with PPARγ binding. The dataset was randomly divided into training, validation, and test sets using an 8:1:1 ratio, ensuring that the models could be evaluated on unseen data to assess their generalization capabilities.

Among the three architectures tested, the Graph Attention Network, or GAT, demonstrated superior performance in identifying active compounds. While the Graph Convolutional Network, or GCN, achieved a slightly higher area under the curve value of 0.879 compared to the GAT’s 0.866, the GAT model exhibited a significantly higher recall rate of 0.798 versus 0.668 for the GCN. This higher recall indicates that the GAT model was more effective at correctly identifying true positive agonists, thereby reducing the risk of missing potentially active natural products during the initial screening phase. The Message-Passing Neural Network, or MPNN, performed the weakest across all metrics, with an accuracy of 0.774 and a Matthews correlation coefficient of 0.548, confirming that the attention mechanism in the GAT provided a distinct advantage in capturing complex molecular topology.

The molecular dynamics simulations provided further insight into the dynamic behavior of the glycyrrhisoflavone-PPARγ complex. The radius of gyration, a measure of structural compactness, showed an instantaneous jump of 0.5 Å at the 50 ns mark, indicating a brief conformational expansion of the complex. In the subsequent fluctuations, the root mean square deviation increased by 2 Å and 2.5 Å, while the radius of gyration increased by only 0.1 Å. This disparity suggests that the later changes were driven by internal atomic rearrangements rather than a significant expansion of the overall protein structure. By the end of the 100 ns simulation, both the glycyrrhisoflavone and rosiglitazone complexes reached a stable equilibrium, although the glycyrrhisoflavone complex retained a degree of dynamic flexibility that may underpin its partial agonist activity.

The study also evaluated the metabolic profiles of the five candidate compounds using the ADMET lab 3.0 platform. All five molecules adhered to the Lipinski rule of five, a heuristic for predicting oral absorption, with no more than one parameter exceeding the recommended limits. The predicted human intestinal absorption values ranged from 0.01 to 0.17, suggesting reasonable oral bioavailability despite low Caco-2 cell permeability values between -5.04 and -5.36. Regarding drug-drug interaction risks, the compounds displayed varied interactions with the CYP3A4 enzyme. Molecule 2F48-234-1 was identified as a strong substrate for CYP3A4, while 2F48-79-1 and 2F48-41-1 were strong inhibitors. The other two candidates were moderate inhibitors. These metabolic characteristics are critical for understanding how glycyrrhisoflavone might interact with other medications in a clinical setting, although these remain computational predictions requiring experimental confirmation.

Subject of Research: Identification of potential natural PPARγ agonists via graph neural networks and integrated computational approaches

Article Title: Identification of potential natural PPARγ agonists via graph neural networks and integrated computational approaches

Article References: Zong, S., He, Y., Feng, Y., Lu, G., Ma, R., Wan, T., Bu, F., & Cheng, B. (2026). Identification of potential natural PPARγ agonists via graph neural networks and integrated computational approaches. Frontiers in Chemistry, 14, Article 2003644. https://doi.org/10.3389/fchem.2026.2003644

Image Credits: AI Generated

DOI: 10.3389/fchem.2026.2003644

Keywords: Identification, potential, natural, PPAR, agonists, graph, neural, networks, integrated, computational, approaches, scientific research