关键词: 科学文献, 光学结构识别, 数据挖掘, 化学知识库

Abstract: Objective： To create a protocol that could be used to construct chemical information database from scientific literature quickly and automatically. Methods: Scientific literature, patents and technical reports from different chemical disciplines were collected and stored in PDF format as fundamental datasets. Chemical structures were transformed from published documents and images to machine-readable data by using the name conversion technology and optical structure recognition tool CLiDE. In the process of molecular structure information extraction, Markush structures were enumerated into well-defined monomer molecules by means of QueryTools in molecule editor ChemDraw. Document management software EndNote X8 was applied to acquire bibliographical references involving title, author, journal and year of publication. Text mining toolkit ChemDataExtractor was adopted to retrieve information that could be used to populate structured chemical database from figures, tables, and textual paragraphs. After this step, detailed manual revision and annotation were conducted in order to ensure the accuracy and completeness of the data. In addition to the literature data, computing simulation platform Pipeline Pilot 7.5 was utilized to calculate the physical and chemical properties and predict molecular attributes. Furthermore, open database ChEMBL was linked to fetch known bioactivities, such as indications and targets. After information extraction and data expansion, five separate metadata files were generated, including molecular structure data file, molecular information, bibliographical references, predictable attributes and known bioactivities. Canonical simplified molecular input line entry specification as primary key, metadata files were associated through common key nodes including molecular number and PDF number to construct an integrated chemical information database. Results: A reasonable construction protocol of chemical information database was created successfully. A total of 174 research articles and 25 reviews published in Marine Drugs from January 2015 to June 2016 collected as essential data source, and an elementary marine natural product database named PKU-MNPD was built in accordance with this protocol, which contained 3 262 molecules and 19 821 records. Conclusion: This data aggregation protocol is of great help for the chemical information database construction in accuracy, comprehensiveness and efficiency based on original documents. The structured chemical information database can facilitate the access to medical intelligence and accelerate the transformation of scientific research achievements.

Key words: Scientific literature, Optical structure recognition, Data mining, Chemical information database