GraphReco: Probabilistic Structure Recognition for Chemical Molecules

Area of Science:

• Chemistry
• Computer Science
• Bioinformatics

Background:

• Chemical structure information is typically conveyed via images, which are not machine-readable.
• This limitation hinders advancements in chemical data mining and drug discovery.

Purpose of the Study:

• To develop a novel optical chemical structure recognition (OCSR) model.
• To address and explain ambiguity challenges in the graph assembly process for OCSR.
• To improve the machine readability of chemical structure images.

Main Methods:

• A rule-based probabilistic OCSR model was developed.
• A novel line detection algorithm was created for precise bond identification.
• A candidate mechanism with a probabilistic graph was designed to resolve atom/bond ambiguities.

Main Results:

• The proposed OCSR model effectively tackles ambiguity challenges in graph assembly.
• Achieved superior recognition accuracy on large image datasets.
• Outperformed existing state-of-the-art OCSR solutions.

Conclusions:

• The developed OCSR model enhances the conversion of chemical structure images into machine-readable formats.
• This advancement has significant implications for chemical data mining and drug discovery.
• The model's probabilistic approach effectively resolves ambiguities in image-based chemical structure recognition.