图形神经网络与图形神经网络. 传统的QSAR:多标签分子气味预测的全面比较
Tengteng Wen1, Xianfa Cai1, Jincheng Li1
1Guangdong College of Medical Information Engineering, Pharmaceutical University, Guangzhou 510006, China.
Molecules (Basel, Switzerland)
|December 11, 2025
概括
图形神经网络 (GNN) 优于预测分子气味的传统方法. 这项研究强调了GNN.
科学领域:
- 计算化学是一种计算化学.
- 化学信息学 化学信息学
- 机器学习是机器学习.
背景情况:
- 分子气味预测对于香水设计,食品科学和安全至关重要.
- 传统的定量结构-活动关系 (QSAR) 方法使用手工制作的描述符.
- 图形神经网络 (GNN) 从分子结构提供直接学习.
研究的目的:
- 为了全面比较传统的QSAR和现代的GNN方法用于多标签的气味预测.
- 在GoodScent数据集上评估各种机器学习算法和GNN架构.
- 确定多标签化学品分类的最佳策略.
主要方法:
- 对23个模型配置进行系统评估,包括随机森林,SVM,GBDT,MLP,XGBoost,LightGBM和GNN (GCN,GAT,NNConv).
- 使用了GoodScent数据集,其中包含3304个分子和6种气味类型 (水果,绿色,甜味,花,木质,草本).
- 研究了传统方法的三个特征处理策略和三个GNN架构.
主要成果:
- 在多标签气味预测方面,GNN模型显著超过了传统方法.
- 图形卷积网络 (GCN) 获得了最高的宏观F1得分 (0.5193),比最好的传统方法 (MLP:0.4766) 有24.1%的相对改善.
- 发现值优化对于有效的多标签化学品分类至关重要.
结论:
- 在分子性质预测任务中,GNN是首选的方法.
- 研究结果为处理化学信息学中的类不平衡提供了关键的见解.
- 这项研究推进了机器学习在预测分子气味特征方面的应用.
相关概念视频
Olfaction
48.0K
The sense of smell is achieved through the activities of the olfactory system. It starts when an airborne odorant enters the nasal cavity and reaches olfactory epithelium (OE). The OE is protected by a thin layer of mucus, which also serves the purpose of dissolving more complex compounds into simpler chemical odorants. The size of the OE and the density of sensory neurons varies among species; in humans, the OE is only about 9-10 cm2.
The olfactory receptors are embedded in the cilia of the...
The olfactory receptors are embedded in the cilia of the...
48.0K
Physiology of Smell and Olfactory Pathway
12.1K
Humans detect odors with the help of specialized cells located in the upper part of the nasal cavity, called olfactory receptor neurons (ORNs). ORNs possess hair-like structures called cilia, which are receptive to sensations from the inhaled air. When an odorant molecule binds to a specific receptor on the cell of the cilia, it leads to a series of events that ultimately cause the ORN to send electrical signals to the olfactory bulb in the brain through the olfactory nerves.
The olfactory...
The olfactory...
12.1K
Predicting Molecular Geometry
44.5K
VSEPR Theory for Determination of Electron Pair Geometries
44.5K


