Elsevier Adds LG AI Research Vision Model to Reaxys Search

One model, three functions

Precision matters: a single misread bond line can make software identify a different compound. Accurate extraction is especially important in inorganic and organometallic chemistry, where structures are complex.

LG AI Research combined molecule detection, reaction scheme analysis and optical chemical structure recognition (OCSR) in a single model. The AI locates structures in a document, analyzes atoms and the bonds between them, and outputs structured data. Researchers can then check whether a compound has appeared in earlier work or search for related substances and reactions.

Elsevier applies the model in the data extraction and review workflow of Reaxys, which covers substances, reactions, bioactivity, biological targets and properties. The company says it can now pull image-based chemical information faster and more broadly, helping users find new compounds, avoid duplicating existing research and plan synthesis routes.

Testing and next steps

In an evaluation published by LG AI Research, the model outperformed alternative technologies when extracting chemical information from an entire document page. Elsevier checked extraction accuracy against Reaxys's existing benchmarks before deployment and validated its systems and procedures to keep data quality consistent when processing large volumes of documents.

The two companies plan to extend their collaboration to chemical reaction extraction, capturing reactants, products and related details to expand the Reaxys dataset.

Mirit Eldor, head of Elsevier's life sciences division, said the more time chemists spend interpreting diagrams to see whether a substance has been synthesized, the less they have for new discoveries. Lee Hwa-young of LG AI Research said understanding scientific images requires AI designed specifically for chemistry.