nico8771/zinc_clean
收藏资源简介:
该数据集名为nico8771/zinc_clean — cleaned ZINC-250k,是一个经过清洗的ZINC-250k化学分子数据集。每个数据行代表一个分子,以规范SMILES(简化分子线性输入规范)格式表示,并包含通过RDKit重新计算的目标值(logP、qed、SAS),同时保留了带电基团(如N+和O-)。数据集来源于torch_molecule ZINC-250k,并经过严格的预处理流程:包括使用RDKit解析(丢弃不可解析的分子)、标准化(移除立体化学信息并进行净化,保留电荷)、Kekulize处理(基于原子词汇表,丢弃不在词汇表中的原子)以及往返检查(确保SMILES可重构为规范单片段分子)。键使用4个类别(无键、单键、双键、三键)。在构建过程中,丢弃了2,006个分子(其中1个因往返检查失败,2,005个因原子不在词汇表中),最终保留了247,449个分子。数据集适用于化学、分子、图生成和流匹配等研究领域,规模在10万到100万之间。
This dataset, named nico8771/zinc_clean — cleaned ZINC-250k, is a cleaned version of the ZINC-250k chemical molecule dataset. Each row represents a molecule as canonical SMILES with RDKit-recomputed targets (logP, qed, SAS), while preserving charged groups (e.g., N+ and O-). It is sourced from torch_molecule ZINC-250k and undergoes a rigorous preprocessing pipeline: parsing with RDKit (dropping unparseable molecules), standardization (removing stereochemistry and sanitizing while preserving charges), Kekulization (based on an atom vocabulary, dropping atoms outside the vocab), and a round-trip check (ensuring SMILES can be reconstructed into a canonical single-fragment molecule). Bonds use 4 classes (none, single, double, triple). During construction, 2,006 molecules were dropped (1 due to round-trip failure, 2,005 due to atoms outside the vocabulary), resulting in 247,449 kept molecules. The dataset is suitable for research in chemistry, molecules, graph generation, and flow matching, with a size category between 100K and 1M.




