Img2Mol - accurate SMILES recognition from molecular graphical depictions

Area of Science:

• Cheminformatics
• Computational Chemistry
• Artificial Intelligence in Chemistry

Background:

• Automatic recognition of molecular structures from images is a long-standing challenge in chemistry.
• Recent neural machine translation advances allow encoding molecular structures into latent representations with low errors.

Purpose of the Study:

• To develop a fast and accurate model for inferring molecular structures from their graphical depictions.
• To translate these depictions into the Standard Molecular Input Line Entry System (SMILES) format.

Main Methods:

• Utilized a deep convolutional neural network (CNN) for learning from molecule images.
• Integrated a pre-trained decoder to translate the CNN's latent representation into SMILES strings.
• Developed the Img2Mol model, combining image-based deep learning with a translation decoder.

Main Results:

• The Img2Mol model achieved high accuracy in translating molecular depictions to SMILES.
• Successfully translated up to 88% of molecular depictions correctly.
• Demonstrated a fast and precise method for chemical structure inference from images.

Conclusions:

• Img2Mol offers a significant advancement in automatically recognizing molecular content from images.
• The model provides a practical solution for converting chemical structure images to machine-readable SMILES format.
• A pre-trained version of Img2Mol is available for non-commercial use to facilitate research.