Based on Transformer's artificial neural network, the image of organic structure is transformed into molecular structure

Mankind is entering the age of artificial intelligence . Chemistry will also be changed by modern deep learning methods , This requires a large amount of qualitative data for neural network training .

The good news is , Chemical data 「 Well preserved 」. Even if a compound was originally in 100 Synthesized years ago , About its structure 、 Information on the nature and mode of synthesis is still relevant today . 

The bad news is , There is no accepted standard method for presenting chemical formulas . Chemists usually use many techniques to express familiar chemical groups in the form of abbreviated symbols . But chemists have different personal habits , The Convention will also change . For computer algorithms , This task seems insurmountable .

come from Skoltech 's start-ups Syntelly And researchers at Lomonosov State University in Moscow have developed a new type of computer-based software Transformer The artificial neural network , It can transform the image of organic structure into molecular structure . To train this network , A comprehensive data generator is developed , It can simulate various drawing styles randomly 、 functional group 、 Placeholder for functional groups （R The base ） And visual pollution .

The research 「 Image2SMILES: Transformer-Based Molecular Optical Recognition Engine」 entitled , On 2022 year 1 month 11 Published in Japan 《 Chemistry–Methods》 On .

 

Data generator GitHub Address ：https://github.com/syntelly/img2smiles

Thesis link ：https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cmtd.202100069