SELFormer-guided discovery of xanthohumol and cirsilineol as multi-target natural therapeutics for type 2 diabetes: computational prediction and experimental validation

Type 2 diabetes mellitus (T2DM) requires multi-target therapeutic approaches addressing both insulin resistance and insulin secretion deficits. Although natural compounds are promising multi-target candidates, systematic identification of their polypharmacological profiles remains challenging. The objective of this study was to establish a computational framework for identifying natural compounds with multi-target therapeutic potential against T2DM through integrated structure-activity analysis and experimental validation. We developed an SELFormer deep learning model to predict natural compound activities against six T2DM-related proteins including glucagon-like peptide-1 receptor (GLP1R), kinesin family member-11 (KIF11) for insulin secretion and insulin receptor (INSR), peroxisome proliferator-activated receptor-gamma (PPARG), fibroblast growth factor receptor-1 (FGFR1) and insulin-like growth factor-1 receptor (IGF1R) for insulin resistance. Uniform Manifold Approximation and Projection (UMAP) for dimension reduction clustering characterized chemical space distributions and molecular docking validated multi-target binding. Selected compounds were experimentally validated using 3T3-L1 adipocytes and mouse insulinoma (MIN6) pancreatic β-cells. The SELFormer model achieved R² = 0.937 and RMSE = 0.331 on the training dataset and R² = 0.918 and RMSE = 0.353 on the testing dataset, with minimal overfitting (ΔR² = 0.019). Among approximately two million screened compounds, 35 natural compounds demonstrated high predicted activity (pIC₅₀ > 7), clustering into eight distinct chemical families. Multi-target network analysis and molecular docking identified curcumin, xanthohumol, hesperetin, (-)-epicatechin, and cirsilineol as lead candidates with favorable binding energies ranging from -7 to -10 kcal mol^-1 across the six targets. Food source analysis identified strawberries, grapes, and tea as rich dietary sources of these bioactive compounds. In 3T3-L1 adipocytes, all five compounds significantly enhanced insulin-stimulated glucose uptake at 10 μM, achieving efficacy comparable to that of metformin. In MIN6 cells, xanthohumol and cirsilineol increased glucose-stimulated insulin secretion to levels comparable to exendin-4, while curcumin, hesperetin, and (-)-epicatechin produced modest but significant increases. In conclusion, this integrated computational and experimental framework identified food-derived natural compounds with validated dual-pathway therapeutic activity against T2DM, providing a systematic and reproducible methodology for multi-target drug discovery in complex metabolic disorders.