Alqarni Saad, Huwaimel Bader
Department of Pharmaceutical Chemistry, College of Pharmacy, University of Ha'il, Hail, 55473, Saudi Arabia.
Medical and Diagnostic Research Center, University of Ha'il, Hail, 55473, Saudi Arabia.
Sci Rep. 2025 Jul 1;15(1):20765. doi: 10.1038/s41598-025-06872-3.
This study employed multiple machine learning (ML) methods to model and predict key attributes of PLGA nanoparticles, specifically particle size and zeta potential. The predictions were based on input variables, including PLGA polymer type, PLGA concentration, anti-solvent type, and anti-solvent concentration. Advanced regression models, including Kernel Ridge Regression (KRR), Gaussian Process Regression (GPR), and Adaptive Neuro-Fuzzy Inference System (ANFIS), were applied to a dataset following rigorous preprocessing. This preprocessing involved Leave-One-Out encoding for categorical variables, Z-score-based outlier detection, and Min-Max normalization for numerical inputs. GPR outperformed the other models in predicting particle size and zeta potential, achieving the best test R scores of 0.9427 and 0.9841, respectively. Furthermore, GPR recorded the lowest total Mean Squared Error (MSE) for particle size (87.504 nm) and zeta potential (1.103 mV), with minimal Mean Absolute Percentage Errors (MAPE) of 3.76% and 2.31%, underscoring its precision and robustness. Cross-validation results further affirmed GPR's consistency, with a mean 0.9611 for zeta potential and R of 0.9588 for particle size and low standard deviations (0.0141 and 0.0083, respectively).
本研究采用多种机器学习（ML）方法对聚乳酸-羟基乙酸共聚物（PLGA）纳米颗粒的关键属性进行建模和预测，具体为粒径和zeta电位。预测基于输入变量，包括PLGA聚合物类型、PLGA浓度、抗溶剂类型和抗溶剂浓度。在经过严格预处理后，将包括核岭回归（KRR）、高斯过程回归（GPR）和自适应神经模糊推理系统（ANFIS）在内的先进回归模型应用于一个数据集。这种预处理包括对分类变量进行留一法编码、基于Z分数的异常值检测以及对数值输入进行最小-最大归一化。GPR在预测粒径和zeta电位方面优于其他模型，分别取得了0.9427和0.9841的最佳测试R分数。此外，GPR记录的粒径（87.504 nm）和zeta电位（1.103 mV）的总均方误差（MSE）最低，平均绝对百分比误差（MAPE）最小，分别为3.76%和2.31%，突出了其精度和稳健性。交叉验证结果进一步证实了GPR的一致性，zeta电位的均值为0.9611，粒径的R为0.9588，标准差较低（分别为0.0141和0.0083）。
