Elsevier has revealed intentions to implement AI technology designed for chemistry to upgrade its chemical database and search platform, Reaxys. This innovative technology will facilitate the extraction and curation of data such as images, illustrations, and reaction diagrams from scientific articles and patents more effectively and accurately than previously possible.
This AI, referred to as MolMole, was created by LG AI Research and combines molecule detection, reaction-diagram interpretation, and optical chemical structure recognition (OCSR) into a single model. Its functionalities comprise extracting machine-readable structures from visuals found in scientific literature.
Mirit Eldor, managing director of life sciences at Elsevier, emphasizes how this technology can save chemists significant time that would have otherwise been spent interpreting images, thus propelling chemistry discovery forward. The collaboration with LG AI Research is anticipated to transition more chemistry from visuals into an easily accessible format on Reaxys.
As per LG AI Research, its technology outperforms rivals in comprehensive document page chemistry extraction. Each extraction’s accuracy is validated against existing benchmarks in Reaxys, with the system undergoing thorough testing within Elsevier’s data infrastructure.
Lutz Weber, co-founder of MolGenie, recognizes the notable advancements in AI models like MolMole, which have boosted productivity for Reaxys curators when compared to traditional methods. However, challenges persist, such as the inability to extract metal–organic complexes or frameworks.
In spite of the encouraging progress, Weber highlights constraints in evaluating LG’s technology due to the absence of available source code and independent validation possibilities. This situation has prompted criticism from industry professionals like Christoph Steinbeck, an analytical chemist working on Decimer, who points out the difficulty in reproducing results or permitting broader industry benchmarking because of restrictive licensing.