AlOmari Ahmad Khaleel, Almansour Khaled
Department of Biomedical Technology, College of Applied Medical Sciences, Prince Sattam Bin Abdulaziz University, Alkharj, Riyadh, Saudi Arabia.
Department of Pharmaceutics, College of Pharmacy, University of Hail, Hail, Saudi Arabia.
Sci Rep. 2025 Apr 26;15(1):14694. doi: 10.1038/s41598-025-99823-x.
A methodology based on Principal Component Analysis (PCA) and machine learning (ML) regression was developed in this study for predicting 5-aminosalicylic acid drug release from polysaccharide-coated formulation. The Raman method was used for collection of spectral data which were then used as inputs to the ML models for estimation of drug release. For ML modeling, we examined the predictive accuracy of three machine learning models-Elastic Net (EN), Group Ridge Regression (GRR), and Multilayer Perceptron (MLP)-for forecasting the release behavior of samples. The dataset, consisting of 155 data points with over 1500 spectral features, underwent preprocessing involving normalization, Principal Component Analysis (PCA) for dimensionality reduction, and outlier detection using Cook's Distance. Model hyperparameters were tuned using the Slime Mould Algorithm (SMA), and each model's performance was evaluated through k-fold cross-validation (k = 3). Assessment metrics, such as the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE), emphasize the MLP model's exceptional performance. On the test set, MLP achieved an R² of 0.9989, notably higher than EN's R² of 0.9760 and GRR's R² of 0.7137. Additionally, MLP exhibited remarkably low test RMSE and MAE values at 0.0084 and 0.0067, respectively, in comparison to EN's RMSE of 0.0342 and MAE of 0.0267, as well as GRR's RMSE of 0.0907 and MAE of 0.0744. Parity plots and learning curves further validate MLP's predictive reliability, demonstrating close alignment between actual and predicted values and efficient learning with minimal overfitting. Consequently, the MLP model emerges as the most effective approach for this predictive task, offering a robust tool for accurately modeling complex spectral data. These findings underscore the robustness of the MLP model, providing a reliable and efficient approach for predicting drug release in polysaccharide-coated formulations, with implications for advancing colonic drug delivery systems.
本研究开发了一种基于主成分分析（PCA）和机器学习（ML）回归的方法，用于预测多糖包衣制剂中5-氨基水杨酸药物的释放。拉曼方法用于收集光谱数据，然后将其用作ML模型的输入，以估计药物释放。对于ML建模，我们检查了三种机器学习模型——弹性网络（EN）、分组岭回归（GRR）和多层感知器（MLP）——预测样品释放行为的预测准确性。该数据集由155个数据点组成，具有1500多个光谱特征，经过预处理，包括归一化、用于降维的主成分分析（PCA）以及使用库克距离进行异常值检测。使用黏菌算法（SMA）调整模型超参数，并通过k折交叉验证（k = 3）评估每个模型的性能。评估指标，如决定系数（R²）、均方根误差（RMSE）和平均绝对误差（MAE），突出了MLP模型的卓越性能。在测试集上，MLP的R²达到0.9989，显著高于EN的R² 0.9760和GRR的R² 0.7137。此外，与EN的RMSE 0.0342和MAE 0.0267以及GRR的RMSE 0.0907和MAE 0.0744相比，MLP在测试集上的RMSE和MAE值分别低至0.0084和0.0067。残差图和学习曲线进一步验证了MLP的预测可靠性，表明实际值和预测值之间紧密一致，且学习效率高，过拟合最小。因此，MLP模型成为该预测任务最有效的方法，为准确建模复杂光谱数据提供了一个强大的工具。这些发现强调了MLP模型的稳健性，为预测多糖包衣制剂中的药物释放提供了一种可靠且高效的方法，对推进结肠给药系统具有重要意义。
