Abstract
Metabolic diseases-including type 2 diabetes, obesity, non-alcoholic fatty liver disease, and certain cancers-pose major global public health challenges. These conditions share common mechanisms such as insulin resistance, chronic inflammation, and oxidative stress. Although medical advances have improved disease management, current treatments remain suboptimal. Natural medicines have gained increasing interest due to their safety, bioactivity, and diverse mechanisms. This study targets adenosine receptors (ARs), key regulators in glucose metabolism, lipid homeostasis, and cellular stress. As members of the G protein-coupled receptor (GPCR) family, ARs include four subtypes-A1, A2A, A2B, and A3-each with distinct pharmacological profiles. We developed a multimodal computational strategy to design natural drug candidates that simultaneously target A1 and A2A, using A2A-selective ligands as controls to explore subtype selectivity. To mitigate toxicity, we incorporated a filtering criterion for low hERG channel affinity. A random forest-based QSAR model was constructed using SMILES representations to predict compound activity. A stacked LSTM neural network was applied to generate plant-derived molecules, while reinforcement learning and Pareto optimization enabled multi-objective refinement. Evolutionary operations-crossover, mutation, and selection-were further introduced to enhance molecular diversity and performance. The proposed framework successfully generated compounds with high target selectivity, low toxicity, and it has good drug-likeness and synthetic accessibility. This work presents a robust and intelligent strategy for natural drug discovery in metabolic diseases and underscores the promising synergy between botanical medicine and artificial intelligence in therapeutic innovation.