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

Type 2 diabetes mellitus has become a prevalent metabolic disease worldwide. Current synthetic PPARγ (Peroxisome proliferator-activated receptor γ)agonists are associated with adverse effects, highlighting the need to explore natural compounds with potential PPARγ agonistic activity. PPARγ is a vital therapeutic target for type 2 diabetes. To facilitate the discovery of natural PPARγ agonists, this study integrated graph neural network (GNN) prediction, molecular docking, ADMET evaluation, in vitro activity assessment, and molecular dynamics simulations for multistage virtual screening. Three GNN models, including GCN, GAT, and MPNN, were constructed and systematically compared, with GAT showing the best overall classification performance. The GAT model was subsequently applied to screen 4,112 natural products from the ZINC database, yielding 291 candidates for molecular docking. Following docking and ADMET evaluation, five compounds were prioritized for experimental assessment. Among them, glycyrrhisoflavone exhibited significant concentration-dependent PPARγ agonistic activity, with an EC 50 of 9.98 μM, whereas the other four candidates showed weak activity. Molecular docking revealed that glycyrrhisoflavone adopted a binding mode similar to that of rosiglitazone, while molecular dynamics simulations further supported the stability of the glycyrrhisoflavone–PPARγ complex. These findings demonstrate the utility of GNN-assisted multistage virtual screening for prioritizing natural PPARγ agonist candidates and identify glycyrrhisoflavone as a candidate for further investigation.