从有机合成程序中提取结构化数据,使用微调的大型语言模型
Qianxiang Ai1, Fanwang Meng1, Jiale Shi1
1Department of Chemical Engineering, Massachusetts Institute of Technology Cambridge MA USA ccoley@mit.edu.
Digital discovery
|August 19, 2024
概括
这项研究对大型语言模型 (LLM) 进行了微调,以从有机化学文献中提取结构反应数据. 该模型在将非结构化文本转换为下游应用程序的开放反应数据库 (ORD) 记录时实现了高精度.
科学领域:
- 有机化学 有机化学
- 计算化学计算化学
- 数据科学数据科学数据科学
背景情况:
- 在有机化学中越来越多地依赖数据驱动的方法和机器学习 (ML),这凸显了对结构化反应数据的需求.
- 很大一部分化学数据存在于文献中的非结构化文本中,这使得手动数据提取成为反应预测等应用的瓶.
- 自动化工具对于有效地将非结构化文本转换为结构化格式至关重要,以释放化学大数据的潜力.
研究的目的:
- 微调一个大型语言模型 (LLM) 来从有机合成过程中提取结构反应信息.
- 将非结构化文本转换为标准化格式,遵循开放反应数据库 (ORD) 方案.
- 通过促进创建一个全面的,机器可读的有机反应数据库,使下游应用成为可能.
主要方法:
- 在一个有机合成程序文本数据集上微调一个大型语言模型 (LLM).
- 开发一个系统来提取反应细节,并将它们映射到开放反应数据库 (ORD) 模式.
- 评估模型在准确性,语法正确性以及识别复合引用和推断反应角色的能力方面的表现.
主要成果:
- 精心调整的LLM成功地从非结构化文本中生成了语法正确的ORD记录.
- 对ORD"消息" (例如,化合物,处理,条件) 实现了91.25%的平均准确性.
- 对于单个数据字段 (例如,化合物标识符,数量) 和公认的化合物引用令牌,证明了高准确性 (92.25%).
结论:
- 开发的LLM有效地从有机化学文献中自动提取结构反应数据.
- 这种方法显著提高了为反应预测和条件建议等应用程序创建结构化数据集的效率.
- 该模型推断反应作用和处理复杂文本结构的能力为更先进的计算化学工具铺平了道路.
相关概念视频
01:14Polymer Classification: Architecture
2.7K
Polymers are classified as linear or branched on the basis of their chain architecture. The polymer chains in linear polymers have a long chain-like structure with minimal to no branching at all. Even if a polymer features large substituent groups on the monomer, which appear as branches to the skeleton, it is not considered a branched polymer. A branched polymer contains secondary polymer chains that arise from the main polymer chain. The branching occurs when the polymer growth shifts from...
2.7K
01:26Polymer Classification: Stereospecificity
2.4K
Polymerization generates chiral centers along the entire backbone of a polymer chain. Accordingly, the stereochemistry of the substituent group has a significant effect on polymer properties. Polymers formed from monosubstituted alkene monomers feature chiral carbons at every alternate position in the polymer backbone. Relative to the predominant orientation of substituents at the adjacent chiral carbons, the polymer can exist in three different configurations: isotactic, syndiotactic, and...
2.4K
02:43E1 Reaction: Stereochemistry and Regiochemistry
9.3K
One of the critical aspects of the E1 reaction mechanism, as also observed in E2, is the regiochemistry, with multiple regioisomers obtained as products. In the example discussed, the presence of water as a weak base favors elimination over substitution to generate two alkenes. Given that alkenes’ stability increases with the number of alkyl groups across the double bond, typically, E1 reactions lead to the Zaitsev product, for this is more substituted and stable than the Hofmann product.
9.3K
02:43E2 Reaction: Stereochemistry and Regiochemistry
11.4K
Elimination reactions of alkyl halides can yield one or more alkenes depending on the specific regiochemical and stereochemical considerations. While the regiochemistry of the reaction governs the location of the double bond in the product, the stereochemical requirements often influence the geometry.
When a substrate with two different β hydrogens undergoes an E2 elimination, the presence of a strong base can yield two regioisomeric alkenes. The more-substituted alkene is the major...
When a substrate with two different β hydrogens undergoes an E2 elimination, the presence of a strong base can yield two regioisomeric alkenes. The more-substituted alkene is the major...
11.4K
02:17What is Organic Chemistry?
73.9K
Organic chemistry is the study of compounds of carbon called organic compounds. Organic compounds either originate from living organisms or are synthesized by chemists. A defining trait of these compounds is the presence of carbon as the principal element, which is bonded to other carbon atoms and other elements such as hydrogen, oxygen, nitrogen, and sulfur. The existence of a wide array of organic molecules is a consequence of carbon atoms’ ability to form up to four strong bonds to...
73.9K
02:28Electrophiles
10.5K
This lesson explains the definition, classification, and characteristic features of an electrophile that are key features of nucleophilic substitution reactions. An analysis of their charge and orbital picture helps understand their reactivity for seeking electrons. Electrophiles can be classified into positive and neutral species. Other classes include free radicals and polar functional groups.
While a positive electrophile, like a proton, reacts due to its vacant, low-energy 1s orbital, the...
While a positive electrophile, like a proton, reacts due to its vacant, low-energy 1s orbital, the...
10.5K