Scholarlink

Purpose <p>The quality of natural medicine granules is jointly determined by material properties and manufacturing process characteristics. Accurately predicting CGQAs (e.g., yield rate, friability) for delicate herbal medicines remains a critical challenge. This study aims to establish a robust decision-making framework for high-quality natural drug granule production using machine learning. The aim of this study was to develop a prediction model and identify relative important factors in the evaluation of yield rate and friability of pharmaceutical granules using MLP, SVM and RF techniques for the datasets of Natural medicine formula.</p> Methods <p> A database was constructed, covering 92 materials with 18 properties and 13 process parameters. Three machine learning algorithms—Support Vector Regression (SVR LibSVM)), Random Forests (RF), and Multi-Layer Perceptrons (MLP)—were developed and compared, utilizing KNIME 4.7.0 as a software tool. Optimization of the models was also conducted applying Parameter Optimization node for MLP, Nu-SVR, RF. Hyperparameter tuning was conducted via Bayesian optimization, and model performance was evaluated using R² (Coefficient of Determination) and MAE (Mean Absolute Error). Three feature datasets were compared: (I) Dimensionality reduction dataset (features with Spearman correlation coefficients greater than 0.2 for key particle quality attributes are treated as separate datasets), (II) original 22-dimensional, (III) ribbon density-excluded.</p> Results <p> MLP outperformed SVR (LibSVM) and RF across all datasets. On Dataset II, MLP achieved the highest R² value of 0.983 with the lowest MAE of 0.007, indicating excellent predictive capability. MLP achieved the highest average R² value of 0.852 with the lowest average MAE of 0.067. This study reveals that the MLP model is an optimal model for predicting yield rate and friability of pharmaceutical granules for the dry granulation process model for the datasets of Natural medicine formula.</p> Conclusion <p> Multi-source material-process data combined with optimized machine learning (especially MLP) enables accurate natural drug CGQAs prediction. Compared with single-component or binary component prediction models, this model is more accurate for predicting drugs with more complex component contents and has a wider range of applications. The findings of this study can help for better understanding of the dry granulation process of natural medicines.</p> Summary <p> This study addresses natural drug Critical Granule Quality Attributes (CGQAs) prediction by building a database (92 materials, 18 properties) and applying three machine learning algorithms (Support Vector Regression (SVR (LibSVM)), Random Forests (RF), and Multi-Layer Perceptrons (MLP)) under 13 process parameters. MLP models attained optimal performance on both dataset II (predicting yield rate with R² = 0.959) and dataset III (predicting friability with R² = 0.959). Among the three datasets, dataset II exhibited the best fit across all models, with superior prediction accuracy and stability relative to the other datasets. The approach supports Critical granule quality attributes predicting and intelligent control in natural drug granulation.</p> Graphical Abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Application of Machine Learning in Predicting Critical Granule Quality Attributes for Modeling of Relationship Between the Properties of Natural Medicine and Process Attributes

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Qi Dongli,
  • Zhao Yu,
  • Peng Peijin,
  • Wu Yiyao,
  • Yan Yuan,
  • Yang Yaguang,
  • An Guangtao,
  • Wang Yajing

摘要

Purpose

The quality of natural medicine granules is jointly determined by material properties and manufacturing process characteristics. Accurately predicting CGQAs (e.g., yield rate, friability) for delicate herbal medicines remains a critical challenge. This study aims to establish a robust decision-making framework for high-quality natural drug granule production using machine learning. The aim of this study was to develop a prediction model and identify relative important factors in the evaluation of yield rate and friability of pharmaceutical granules using MLP, SVM and RF techniques for the datasets of Natural medicine formula.

Methods

A database was constructed, covering 92 materials with 18 properties and 13 process parameters. Three machine learning algorithms—Support Vector Regression (SVR LibSVM)), Random Forests (RF), and Multi-Layer Perceptrons (MLP)—were developed and compared, utilizing KNIME 4.7.0 as a software tool. Optimization of the models was also conducted applying Parameter Optimization node for MLP, Nu-SVR, RF. Hyperparameter tuning was conducted via Bayesian optimization, and model performance was evaluated using R² (Coefficient of Determination) and MAE (Mean Absolute Error). Three feature datasets were compared: (I) Dimensionality reduction dataset (features with Spearman correlation coefficients greater than 0.2 for key particle quality attributes are treated as separate datasets), (II) original 22-dimensional, (III) ribbon density-excluded.

Results

MLP outperformed SVR (LibSVM) and RF across all datasets. On Dataset II, MLP achieved the highest R² value of 0.983 with the lowest MAE of 0.007, indicating excellent predictive capability. MLP achieved the highest average R² value of 0.852 with the lowest average MAE of 0.067. This study reveals that the MLP model is an optimal model for predicting yield rate and friability of pharmaceutical granules for the dry granulation process model for the datasets of Natural medicine formula.

Conclusion

Multi-source material-process data combined with optimized machine learning (especially MLP) enables accurate natural drug CGQAs prediction. Compared with single-component or binary component prediction models, this model is more accurate for predicting drugs with more complex component contents and has a wider range of applications. The findings of this study can help for better understanding of the dry granulation process of natural medicines.

Summary

This study addresses natural drug Critical Granule Quality Attributes (CGQAs) prediction by building a database (92 materials, 18 properties) and applying three machine learning algorithms (Support Vector Regression (SVR (LibSVM)), Random Forests (RF), and Multi-Layer Perceptrons (MLP)) under 13 process parameters. MLP models attained optimal performance on both dataset II (predicting yield rate with R² = 0.959) and dataset III (predicting friability with R² = 0.959). Among the three datasets, dataset II exhibited the best fit across all models, with superior prediction accuracy and stability relative to the other datasets. The approach supports Critical granule quality attributes predicting and intelligent control in natural drug granulation.

Graphical Abstract