Scholarlink

<p><i>LigPCDS (Ligand Point Cloud Data Set)</i> is the first dataset of chemically labeled 3D point clouds of protein ligands. 3D images and structures of ligands were derived from X-ray protein crystallography experimental datasets deposited at the Protein Data Bank. The 3D point cloud format allowed for a computer-comprehensive representation of the ligand’s experimental data, enabling the interpretation of the ligand’s chemical structure using a building block-like labeling approach. For constructing <i>LigPCDS</i>, the images of the ligands were interpolated from their difference electron density map into a 3D grid-like structure, filtered around their atomic spheres, and stored in point clouds. The density value was used as a single feature. Chemical vocabularies, based on atoms and their cyclic structural arrangements, were designed and used to pointwise label these 3D representations of the ligands. The proposed imaging and labeling approaches were validated by training semantic segmentation deep learning models on a stratified dataset from <i>LigPCDS</i>, which could recover the protein ligand’s chemical structure with good performance. <i>LigPCDS</i> can be used to achieve solutions for building known and yet unknown protein ligands (small organic molecules) from experimental X-ray protein crystallography, in silico ligand screening, drug design, and to understand protein function in basic biology.</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

Labeled dataset of X-ray protein ligand images in 3D point cloud and validated deep learning models

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月31日

查看全文
  • Cristina F. Bazzano,
  • Luiz F. G. Alves,
  • Guilherme P. Telles,
  • Daniela B. B. Trivella

摘要

LigPCDS (Ligand Point Cloud Data Set) is the first dataset of chemically labeled 3D point clouds of protein ligands. 3D images and structures of ligands were derived from X-ray protein crystallography experimental datasets deposited at the Protein Data Bank. The 3D point cloud format allowed for a computer-comprehensive representation of the ligand’s experimental data, enabling the interpretation of the ligand’s chemical structure using a building block-like labeling approach. For constructing LigPCDS, the images of the ligands were interpolated from their difference electron density map into a 3D grid-like structure, filtered around their atomic spheres, and stored in point clouds. The density value was used as a single feature. Chemical vocabularies, based on atoms and their cyclic structural arrangements, were designed and used to pointwise label these 3D representations of the ligands. The proposed imaging and labeling approaches were validated by training semantic segmentation deep learning models on a stratified dataset from LigPCDS, which could recover the protein ligand’s chemical structure with good performance. LigPCDS can be used to achieve solutions for building known and yet unknown protein ligands (small organic molecules) from experimental X-ray protein crystallography, in silico ligand screening, drug design, and to understand protein function in basic biology.