自主监督学习与化学意识的碎片化,以有效地预测分子性质
Ailin Xie1, Ziqiao Zhang1, Jihong Guan2
1Shanghai Key Lab of Intelligent Information Processing, and School of Computer Science, Fudan University, 200438 Shanghai, China.
Briefings in bioinformatics
|August 20, 2023
概括
本研究介绍了化学意识碎片化有效分子性质预测 (CAFE-MPP),一种新的自我监督学习方法. CAFE-MPP通过专注于关键化学子结构来改进分子表示,在多个数据集上表现优于现有的方法.
科学领域:
- 计算化学计算化学
- 人工智能的人工智能
- 药物发现 药物发现 药物发现
背景情况:
- 分子性质预测 (MPP) 对人工智能辅助药物发现 (AIDD) 是至关重要的.
- 自主监督学习 (SSL) 对MPP有希望,但往往忽略了关键的分子亚结构.
- 现有的SSL模型很难有效地利用子结构信息来改进分子表示.
研究的目的:
- 开发一种新的自主监督学习框架,CAFE-MPP,它明确结合了化学意识的子结构,用于增强分子表示.
- 通过专注于片段级表示来提高分子性质预测的性能.
- 通过有效的代表性学习来解决AIDD中的数据稀缺挑战.
主要方法:
- 引入了一种基于碎片的新型分子图 (FMG) 来捕捉亚结构之间的拓关系.
- 采用自我监督的对比学习框架,在碎片级别使用硬负对进行预训练.
- 使用 Graphormer 模型,从 MPP 的片段嵌入中生成最终的分子表示.
主要成果:
- 在MPP的11个基准数据集中,CAFE-MPP在7个基准数据集中取得了最先进的表现.
- 与其他六种著名的自我监督方法相比,表现优越.
- 展示了CAFE-MPP学习的表示,隐含地编码关键的碎片信息相关的分子性质.
结论:
- CAFE-MPP有效地学习了化学意识的分子表示,在AIDD中显著推进了MPP.
- 碎片级对比学习方法提高了预测性能,并解决了现有的SSL方法的局限性.
- 在药物发现方面,CAFE-MPP为开发更准确,更有效的数据模型提供了有希望的方向.
相关概念视频
Mass Spectrometry: Molecular Fragmentation Overview
3.2K
The ionization of a molecule into a molecular ion inside the mass spectrometer causes instability in the molecule's structure due to the loss of an electron. This eventually leads to the fragmentation or breaking of some bonds in the molecule. The fragmentation occurs predominantly at specific bonds to yield relatively stable fragments.
One type of fragmentation pattern is the cleavage of a single bond in the molecular ion. The cleavage leads to a radical and a cation. The cleavage can...
One type of fragmentation pattern is the cleavage of a single bond in the molecular ion. The cleavage leads to a radical and a cation. The cleavage can...
3.2K
Predicting Molecular Geometry
34.5K
VSEPR Theory for Determination of Electron Pair Geometries
34.5K
Mass Spectrometry: Aromatic Compound Fragmentation
1.8K
Upon ionization, aromatic compounds generate a molecular ion that is observed as a prominent peak in their mass spectra. For example, the molecular ion peak for benzene appears at a mass-to-charge ratio of 78, while toluene is observed at a mass-to-charge ratio of 92. The molecular ion benzene is highly stable and does not readily undergo further fragmentation due to the significant amount of energy required to disrupt the aromatic stability of the benzene ring. In contrast, the molecular ion...
1.8K
Mass Spectrometry: Alcohol Fragmentation
3.6K
Alcohols (R-OH) ionize to lose one non-bonded electron from the oxygen atom, forming molecular ions. Due to their tendency to fragment rapidly, the intensity of the molecular ion peak in the mass spectrum is weak or sometimes absent. The fragmentation patterns for alcohols occur in two ways, i.e. ⍺-cleavage and dehydration. During ⍺-cleavage, the bond at the ⍺-position adjacent to the hydroxyl group cleaves to give a resonance-stabilized cation and a radical. However,...
3.6K
Mass Spectrometry: Cycloalkane Fragmentation
1.4K
In mass spectrometry, cycloalkanes exhibit distinct fragmentation patterns due to the inherent stability of their molecular ions compared to linear or branched alkanes. The ring structure of cycloalkanes provides additional stability to the molecular ions, often resulting in prominent ion peaks in the mass spectrum.
For example, cyclohexane molecular ions have a mass-to-charge ratio (m/z) of 84, which tends to produce a stronger signal than linear alkanes like hexane. This stability comes from...
For example, cyclohexane molecular ions have a mass-to-charge ratio (m/z) of 84, which tends to produce a stronger signal than linear alkanes like hexane. This stability comes from...
1.4K
Mass Spectrometry: Branched Alkane Fragmentation
1.1K
This lesson delves into the mass spectrometry of branched alkane fragmentation. Branched alkanes possess secondary or tertiary carbon atoms, which generate relatively stable carbocations if the cleavage occurs at the branching point. The high stability of carbocations drives the instant fragmentation of branched alkanes. Accordingly, the branched alkane's molecular ion peak is very weak or invisible in the mass spectra, especially in comparison to a linear alkane.
1.1K


