Scholarlink

<p>Traditional quantitative structure-property relationship (QSPR) methods rely on molecular descriptors to quantify molecular structures and establish correlations with physical properties. In this study, we propose an approach that incorporates complete molecular structures to refine traditional QSPR methods and improve predictive accuracy. The supercritical properties used for modeling are collected from the literature. Molecular structures are optimized using density functional theory, from which molecular descriptors are derived. Both the structures and descriptors serve as inputs to the models developed in this work. Three models are constructed: a traditional artificial neural network model, a ResNet model, and a convolutional neural network (CNN)-enhanced model. Comparison with the JOBACK method shows that the CNN-enhanced model achieves higher predictive accuracy, whereas the ResNet model, which relies solely on molecular structures, suffers from pronounced overfitting.</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

Enhancing quantitative structure-property relationship models by integrating complete molecular structure with deep learning

  • 期刊论文

  • 发表日期：

    2026年1月9日

查看全文
  • Bo Ouyang,
  • Dian Zhang,
  • Zhe Chen,
  • Zhao-Quan Wen,
  • Zheng-Hong Luo

摘要

Traditional quantitative structure-property relationship (QSPR) methods rely on molecular descriptors to quantify molecular structures and establish correlations with physical properties. In this study, we propose an approach that incorporates complete molecular structures to refine traditional QSPR methods and improve predictive accuracy. The supercritical properties used for modeling are collected from the literature. Molecular structures are optimized using density functional theory, from which molecular descriptors are derived. Both the structures and descriptors serve as inputs to the models developed in this work. Three models are constructed: a traditional artificial neural network model, a ResNet model, and a convolutional neural network (CNN)-enhanced model. Comparison with the JOBACK method shows that the CNN-enhanced model achieves higher predictive accuracy, whereas the ResNet model, which relies solely on molecular structures, suffers from pronounced overfitting.