Scholarlink

(GNN) Graph neural network-based methods have demonstrated remarkable potential in forecasting the quantum mechanical properties of molecules by modeling complex interactions. However, existing approaches often overlook many-body interactions, limiting their predictive accuracy. The UA-HGNN leverages a neural message-passing method enhanced with global molecule depictions and focus mechanisms, ensuring robust predictions invariant to translation, rotation, and permutation. Notably, an uncertainty estimation framework is integrated into UA-HGNN, providing confidence intervals for predicted quantum properties. Visualization of attention weights as attention scores for 1-body and 2-body predictions is also presented in this very research. This research discusses the development of a novel framework called the Uncertainty-Aware Heterogeneous Graph Neural Network (UA-HGNN) for accurately forecasting the quantum machine-driven properties of particles. The framework leverages heterogeneous molecular graphs (HMGs) to explicitly capture many-body interactions and incorporates an uncertainty estimation framework to provide confidence intervals for predicted quantum properties. The UA-HGNN model integrates attention mechanisms, a neural message-passing method and global molecule depictions to ensure robust predictions invariant to translation, rotation, and permutation. The research methodology involves dataset acquisition, preprocessing, model design, training, and evaluation, with a focus on predicting molecular properties using graph neural network architecture.

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

Uncertainty-Aware Molecular Property Prediction Using Heterogeneous Molecular Graph Neural Networks

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Midhilesh Momidi,
  • Priyanka S. Chauhan,
  • Adityaram Komaraneni,
  • Surya Prakash Ghattamaneni,
  • Kamal Upreti,
  • Nishant Kumar

摘要

(GNN) Graph neural network-based methods have demonstrated remarkable potential in forecasting the quantum mechanical properties of molecules by modeling complex interactions. However, existing approaches often overlook many-body interactions, limiting their predictive accuracy. The UA-HGNN leverages a neural message-passing method enhanced with global molecule depictions and focus mechanisms, ensuring robust predictions invariant to translation, rotation, and permutation. Notably, an uncertainty estimation framework is integrated into UA-HGNN, providing confidence intervals for predicted quantum properties. Visualization of attention weights as attention scores for 1-body and 2-body predictions is also presented in this very research. This research discusses the development of a novel framework called the Uncertainty-Aware Heterogeneous Graph Neural Network (UA-HGNN) for accurately forecasting the quantum machine-driven properties of particles. The framework leverages heterogeneous molecular graphs (HMGs) to explicitly capture many-body interactions and incorporates an uncertainty estimation framework to provide confidence intervals for predicted quantum properties. The UA-HGNN model integrates attention mechanisms, a neural message-passing method and global molecule depictions to ensure robust predictions invariant to translation, rotation, and permutation. The research methodology involves dataset acquisition, preprocessing, model design, training, and evaluation, with a focus on predicting molecular properties using graph neural network architecture.