ABC-Net: a divide-and-conquer based deep learning architecture for SMILES recognition from molecular images.

Area of Science:

• Computational chemistry
• Cheminformatics
• Artificial intelligence in chemistry

Background:

• Chemical structures in scientific documents are primarily pictorial, hindering computer-based analysis and knowledge extraction.
• Existing optical chemical structure recognition (OCSR) methods struggle with accuracy for real-world applications.
• Efficiently mining molecular information from vast scientific literature remains a significant challenge.

Purpose of the Study:

• To develop an advanced deep learning model for accurate and efficient optical chemical structure recognition (OCSR).
• To enable direct prediction of molecular graph structures from chemical images, overcoming limitations of current OCSR techniques.
• To facilitate automated knowledge acquisition from scientific literature by improving the interpretation of chemical structures.

Main Methods:

• Developed ABC-Net (Atom and Bond Center Network), a deep neural network model utilizing a fully convolutional neural network (CNN).
• Employed a divide-and-conquer approach, representing atoms and bonds as central points for detection.
• Integrated tasks of point identification and property prediction (atom types, charges, bond types) into a single CNN framework.

Main Results:

• ABC-Net directly predicts graph structures from chemical images, achieving high recovery accuracy.
• The model efficiently processes molecular images by integrating detection and property prediction within a single CNN.
• Experimental results show significant performance improvements over existing publicly available OCSR tools.

Conclusions:

• ABC-Net offers a promising and efficient solution for optical chemical structure recognition (OCSR) problems.
• The proposed method enhances the automated acquisition of molecular information from scientific literature.
• This approach represents a significant step forward in enabling computers to understand and process chemical structures from images.