AI-Driven Optimization of Curcumin Nanoparticles for Antidiabetic Drug Delivery

Curcumin, the principal bioactive polyphenol of Curcuma longa, exhibits broad pharmacological properties including anti-hyperglycemic, antioxidant, and anti-inflammatory activities. However, its clinical translation is severely impeded by poor aqueous solubility (< 0.1 μg/mL), rapid systemic metabolism, and negligible oral bioavailability (< 1%). Nanoparticle-based drug delivery systems offer a mechanistically sound strategy to overcome these limitations; nonetheless, conventional trial-and-error formulation methods are inherently time-consuming, resource-intensive, and incapable of capturing complex nonlinear interactions among critical formulation variables. This study aimed to develop, characterize, and AI-optimize PLGA-based curcumin nanoparticles (Cur-NPs) for antidiabetic applications using a hybrid artificial neural network (ANN)–random forest (RF) modeling framework integrated with Bayesian optimization. Curcumin-loaded PLGA nanoparticles were fabricated via the nanoprecipitation–solvent evaporation method. A Box–Behnken experimental design (33 runs) was employed to generate the training dataset, encompassing five independent formulation variables. Two machine learning architectures—a multilayer perceptron ANN and a Random Forest ensemble—were trained, validated, and evaluated against hold-out test sets. Optimized formulations were characterized by dynamic light scattering (DLS), transmission electron microscopy (TEM), zeta potential analysis, Fourier-transform infrared spectroscopy (FTIR), and differential scanning calorimetry (DSC). In vitro drug release kinetics were modeled using Higuchi and Korsmeyer–Peppas equations. Antidiabetic efficacy was assessed via α-glucosidase inhibition assays. The AI-optimized Cur-NPs exhibited a particle size of 128.6 ± 6.2 nm, PDI of 0.18 ± 0.02, zeta potential of −28.7 ± 1.8 mV, and entrapment efficiency of 87.3 ± 2.9%. The ANN model achieved an R² of 0.961 and RMSE of 5.32 nm for particle size prediction on the test set. Drug release followed Korsmeyer– Peppas kinetics (n = 0.62), indicating an anomalous diffusion–erosion mechanism. The optimized nanoparticles demonstrated a 1.77-fold improvement in α-glucosidase inhibitory activity (IC50 = 43.7 ± 3.2 μg/mL) compared to non-optimized formulations (IC50 = 89.4 ± 5.1 μg/mL). The hybrid AI-DoE optimization strategy significantly enhanced nanoparticle performance compared to conventional formulation approaches, demonstrating the transformative potential of machine learning integration in pharmaceutical nanotechnology for antidiabetic applications.

Перевод пока недоступен