MolMiner: You Only Look Once for Chemical Structure Recognition

Area of Science:

• Chemistry
• Computer Science
• Bioinformatics

Background:

• 2D molecular structures in scientific literature lack machine readability.
• Decades of printed documents present a significant backlog for data processing.
• Existing Optical Chemical Structure Recognition (OCSR) systems often rely on rule-based methods.

Purpose of the Study:

• To develop a practical software tool for Optical Chemical Structure Recognition (OCSR).
• To convert 2D molecular depictions from scientific documents into machine-readable formats.
• To leverage deep learning for enhanced chemical structure recognition.

Main Methods:

• Developed MolMiner software utilizing deep neural networks for semantic segmentation and object detection.
• Implemented a distance-based algorithm to construct molecular graphs from recognized atoms and bonds.
• Trained and evaluated the system on benchmark datasets and real-world scientific papers.

Main Results:

• MolMiner achieved state-of-the-art performance on four benchmark datasets.
• The software demonstrated high accuracy on a self-collected external dataset from scientific papers.
• MolMiner showed comparable performance in real-world OCSR tasks with a user-friendly interface.

Conclusions:

• MolMiner offers a valuable and practical solution for Optical Chemical Structure Recognition.
• The deep learning-based approach provides a significant improvement over traditional rule-based systems.
• MolMiner is a useful tool for daily applications in scientific research and data management.