ARTIFICIAL INTELLIGENCE-DRIVEN FORMULATION OPTIMIZATION OF POLYMERIC NANOPARTICLES CONTAINING ANTIDIABETIC AGENTS

Abstract

Background: Diabetes mellitus remains a principal global health burden, and the oral bioavailability of many antidiabetic agents including pioglitazone is limited by erratic gastrointestinal absorption, hepatic first-pass metabolism, and suboptimal physicochemical properties. Polymeric nanoparticles fabricated from poly(lactic-co-glycolic acid) (PLGA) offer a robust platform to overcome these barriers, yet conventional Design of Experiments (DoE) alone cannot capture the nonlinear interplay among critical formulation variables.

Objective: To develop PLGA-based pioglitazone nanoparticles (PGZ-NPs) and employ an ensemble of artificial intelligence (AI) algorithms artificial neural networks (ANN), random forest (RF), support vector machine (SVM), and gradient boosting (GB) for predictive modeling and multi-response optimization of particle size, polydispersity index (PDI), zeta potential, and entrapment efficiency (EE).

Methods: A Box–Behnken design (BBD) was adopted within a Quality by Design (QbD) framework. Three independent variables PLGA concentration (X₁), PVA surfactant concentration (X₂), and homogenization speed (X₃) were systematically varied across 17 experimental runs. The resulting dataset was partitioned 80:20 for AI model training and external validation. Hyperparameter optimization was executed via grid search with five-fold cross-validation. Physicochemical characterization encompassed dynamic light scattering, zeta potential measurement, FTIR, DSC, XRD, SEM, and TEM analyses. Drug release was evaluated in phosphate-buffered saline (pH 6.8) over 72 h, and stability was assessed per ICH Q1A(R2) guidelines.

Results: Among the four AI paradigms evaluated, the gradient boosting model achieved the highest predictive accuracy for particle size (R² = 0.986, RMSE = 4.2 nm) and entrapment efficiency (R² = 0.979, RMSE = 1.3%). The AI-optimized formulation (F-Opt) exhibited a mean particle size of 187.4 ± 5.6 nm, PDI of 0.138 ± 0.011, zeta potential of −28.7 ± 1.4 mV, and EE of 87.3 ± 2.1%. Cumulative drug release reached 89.6% over 72 h following Korsmeyer–Peppas kinetics (n = 0.52), indicating anomalous diffusion. Stability studies confirmed retained quality attributes across three months at 25 °C/60% RH.

Conclusion: AI-driven multi-model optimization substantially outperforms conventional polynomial response-surface methodology in predicting and controlling nanoparticle quality attributes. The gradient boosting algorithm emerged as the most reliable predictive tool, generating an optimized PGZ-NP formulation with superior colloidal stability and extended drug release. This work establishes a transferable AI-assisted workflow for pharmaceutical nanoparticle development.