ReactionDataExtractor 2.0: A Deep Learning Approach for Data Extraction from Chemical Reaction Schemes.

Area of Science:

• Chemistry
• Computer Science
• Data Science

Background:

• Chemical knowledge is often shared through graphical reaction schemes.
• Extracting data from these schemes is challenging for machines.
• Existing tools require manual preprocessing and have limited scope.

Purpose of the Study:

• To present ReactionDataExtractor v2.0, a novel tool for automated chemical data extraction.
• To overcome limitations of current machine-based data extraction from chemical diagrams.
• To enable high-throughput analysis of image-based chemical information.

Main Methods:

• Utilized a combination of neural networks and symbolic artificial intelligence.
• Developed a modular system architecture for balancing speed and accuracy.
• Employed a data-driven approach with synthetically generated data for model tuning.

Main Results:

• Achieved F1 score metrics between 75% and 96% on a test set of reaction schemes from journal articles.
• Demonstrated the tool's effectiveness in extracting data from graphical chemical representations.
• Showcased potential for improved metrics via subdomain-specific model tuning.

Conclusions:

• ReactionDataExtractor v2.0 offers an autonomous, high-throughput solution for image-based chemical data extraction.
• The tool effectively bridges the gap between graphical chemical representations and machine readability.
• Future work can enhance accuracy through targeted model optimization for specific chemical subdomains.