Scholarlink

<p>In biomedical research, particularly within the drug discovery process, achieving rotation-equivariance in machine learning models is highly desirable. We propose an integrated architecture that leverages SO(3) spherical harmonics to encode geometric structures within convolutional neural networks, thereby attaining both translational and rotational equivariance. By incorporating these spherical tensor representations, our model learns feature representations with discrete directional dimensions, ensuring that outputs remain equivariant under various rotational transformations. Furthermore, we utilize gaussian radial basis functions to address distance-related issues and integrate Clebsch-Gordan coefficients for angular momentum coupling. This approach enables the convolutional neural networks to make predictions that effectively capture rotational symmetry and the coupling of angular and atomic characteristics in three-dimensional space. Consequently, our model enhances the accuracy of protein-ligand affinity predictions, improving both physical plausibility and generalization capability. Comparative analysis with state-of-the-art methodologies underscores the effectiveness of our model components, and the substantial improvement in prediction accuracy attests to the robustness of our algorithm for affinity prediction.</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

A geometric rotation-equivariant spherical convolutional and gaussian radial basis network for predicting protein-ligand binding affinity

  • 期刊论文

  • 发表日期：

    2026年7月28日

查看全文
  • Bin Wan,
  • Gaili Li,
  • Ruisheng Zhang

摘要

In biomedical research, particularly within the drug discovery process, achieving rotation-equivariance in machine learning models is highly desirable. We propose an integrated architecture that leverages SO(3) spherical harmonics to encode geometric structures within convolutional neural networks, thereby attaining both translational and rotational equivariance. By incorporating these spherical tensor representations, our model learns feature representations with discrete directional dimensions, ensuring that outputs remain equivariant under various rotational transformations. Furthermore, we utilize gaussian radial basis functions to address distance-related issues and integrate Clebsch-Gordan coefficients for angular momentum coupling. This approach enables the convolutional neural networks to make predictions that effectively capture rotational symmetry and the coupling of angular and atomic characteristics in three-dimensional space. Consequently, our model enhances the accuracy of protein-ligand affinity predictions, improving both physical plausibility and generalization capability. Comparative analysis with state-of-the-art methodologies underscores the effectiveness of our model components, and the substantial improvement in prediction accuracy attests to the robustness of our algorithm for affinity prediction.