Image2InChI: Automated Molecular Optical Image Recognition.

Area of Science:

• Artificial intelligence in drug discovery
• Computational chemistry
• Machine learning for chemical informatics

Background:

• Accurate identification of chemical structures from molecular images is crucial for AI-driven drug discovery.
• Efficient and automated conversion of molecular images into machine-readable formats is a significant challenge.

Purpose of the Study:

• To propose an automated deep learning model, Image2InChI, for molecular optical image recognition.
• To enhance the integration of image patch and InChI prediction through a novel feature fusion network with attention.

Main Methods:

• Utilized an improved SwinTransformer as an encoder for image feature extraction.
• Employed a Transformer Decoder with patch embedding for InChI prediction.
• Implemented a novel feature fusion network with attention to combine image and InChI data.

Main Results:

• Achieved 99.8% accuracy for InChI prediction (InChI acc).
• Demonstrated high performance in Morgan fingerprint (94.1%), maximum common structures (94.8%), and longest common subsequence (96.2%) accuracy.
• Validated the model's effectiveness in improving accuracy and efficiency of molecular image recognition.

Conclusions:

• The Image2InChI model significantly enhances automated molecular image recognition.
• Provides a valuable reference for optical chemical structure recognition for InChI generation.
• Facilitates AI applications in drug discovery through efficient image-to-structure conversion.