Scholarlink

Accurate prediction of molecular properties, including physicochemical characteristics, biological activity, and ADME/T profiles, remains a critical challenge in AI-driven drug discovery. We present RGFP-Net (ResGAT Fingerprint NET), a novel architecture that integrates graph neural networks with residual connections and molecular fingerprint features for precise molecular property prediction. We conducted systematic evaluations on six classification tasks using the MoleculeNet benchmark datasets. Results show that RGFP-Net achieves superior performance on BACE, HIV, and Tox21, with up to 5% AUC improvement over state-of-the-art baselines, while maintaining competitive results on BBBP, ClinTox, and SIDER. The model demonstrates strong capabilities in single-label regression, multi-label classification, and multi-label prediction with missing labels. Furthermore, RGFP-Net provides extensibility by supporting variants such as GIN and GCN, offering a flexible framework for future research in computational drug discovery.

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

RGFP-Net: Graph Neural Networks for Molecular Property Prediction in Drug Discovery

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kefan Ni,
  • Jiayi Li,
  • Zihang Zhang,
  • Zhenyu Lei,
  • Jiake Wang,
  • Xiangmei Li,
  • Shangce Gao

摘要

Accurate prediction of molecular properties, including physicochemical characteristics, biological activity, and ADME/T profiles, remains a critical challenge in AI-driven drug discovery. We present RGFP-Net (ResGAT Fingerprint NET), a novel architecture that integrates graph neural networks with residual connections and molecular fingerprint features for precise molecular property prediction. We conducted systematic evaluations on six classification tasks using the MoleculeNet benchmark datasets. Results show that RGFP-Net achieves superior performance on BACE, HIV, and Tox21, with up to 5% AUC improvement over state-of-the-art baselines, while maintaining competitive results on BBBP, ClinTox, and SIDER. The model demonstrates strong capabilities in single-label regression, multi-label classification, and multi-label prediction with missing labels. Furthermore, RGFP-Net provides extensibility by supporting variants such as GIN and GCN, offering a flexible framework for future research in computational drug discovery.