Physics-Informed Neural Networks for Predictive Modeling and Optimization of pH-Triggered Drug Release from Quercetin-Loaded Nanocomposites

Abstract

Precise control of drug-release kinetics remains a major challenge in nanomedicine, particularly for pH-responsive delivery systems targeting the acidic tumor microenvironment. Here, we develop a physics-informed neural network (PINN) framework for predicting and optimizing pH-triggered drug release from quercetin-loaded nanocomposites. The model embeds Fickian diffusion and pH-dependent polymer erosion directly into the learning architecture, enabling mechanistically consistent predictions that incorporate nanocomposite morphology, surface chemistry, and material composition. Trained on 69 formulations, the PINN accurately reproduces cumulative release profiles across pH 5.4–7.4 (R² = 0.96, MAE = 2.8%). Global sensitivity analysis identifies pore architecture (optimal 8–12 nm), surface functionalization density, zeta potential, polymer molecular weight, hydrophilic–hydrophobic balance, and drug loading as the dominant determinants of release behavior. Coupling the PINN with Bayesian multiobjective optimization reveals Pareto-optimal designs that maximize release under acidic conditions while minimizing premature release at physiological pH. The best-performing formulation achieves 95.2% release at pH 5.4, less than 15% release at pH 7.4 over 72 h, and sustained delivery beyond 96 h. Mechanistic and biological studies corroborate the optimized release behavior and demonstrate effective tumor growth suppression with minimal systemic toxicity. Overall, this framework provides a robust strategy for the rational design of pH-selective drug-delivery systems with tunable release profiles.