|
Development of hybrid machine learning models for enhanced quality control and optimization of resveratrol-loaded polymeric nanoparticles |
|---|---|
| รหัสดีโอไอ | |
| Title | Development of hybrid machine learning models for enhanced quality control and optimization of resveratrol-loaded polymeric nanoparticles |
| Creator | Phuvamin Suriyaamporn |
| Contributor | Warut Pannakkong, Advisor |
| Publisher | Thammasat University |
| Publication Year | 2568 |
| Keyword | Resveratrol, Polymeric nanoparticles, Hybrid machine learning models, Genetic algorithm, Reinforcement learning |
| Abstract | The formulation of resveratrol-loaded polymeric nanoparticles (RES-PNPs) is hindered by resveratrol’s low solubility, instability, and rapid clearance. This study introduces an integrated hybrid machine-learning (ML), genetic-algorithm (GA), and reinforcement-learning (RL) framework to rationally design RES-PNPs and improve their anticancer performance. Hybrid ML models combining linear regression (LR), k-nearest neighbors (k-NN), and artificial neural networks (ANN) were constructed to predict particle size (PS), polydispersity index (PDI), zeta potential (ZP), and drug loading (%DL). The hybrid models substantially outperformed single learners (e.g., ANN RMSE: 69.19 for PS, 0.06 for PDI, 4.28 mV for ZP; k-NN RMSE: 6.69 for %DL), reducing error by 15–40% across endpoints. Final hybrid RMSE values were 55.12 (PS), 0.05 (PDI), 3.90 mV (ZP), and 5.21 (%DL), demonstrating strong predictive fidelity. Optimization employed a GA with a population of 100,000 individuals, 100 generations, and 10 RL-tuned episodes. Fitness was defined as f = –PS – PDI – ZP + %DL, with the best episodes achieving a terminal fitness of –1.025. RL adaptively refined GA parameters, converging on crossover = 0.10 and mutation = 0.18, which improved stability and reduced variance relative to fixed settings. The optimal solution identified (PAA = 0.30, GT = 0.13, P407 = 8.11, Hz = 11.56, Time = 12.50) produced PS = 80 nm, PDI = 0.31, ZP = –36.94 mV, and %DL = 68.02%, all within predefined feasibility limits. Experimental validation showed no significant differences between predicted and measured CQAs (p > 0.05). This AI-augmented workflow establishes a robust design space aligned with ICH Q10 and demonstrates a powerful strategy for intelligent nanomedicine formulation and optimization. |
