Elsevier said on September 15 that it is using LG AI Research's chemical-image recognition technology to turn drawings in papers and patents into searchable substance data for Reaxys, its chemistry research platform. The integration brings information embedded in scientific images into the database's extraction and curation workflow.
Chemical drawings carry information that ordinary text search can miss: which atoms are connected, the types of bonds between them, and their spatial arrangement. Reading one bond incorrectly can identify a different compound. Elsevier says better extraction should help researchers check whether a substance has appeared before and find chemistry relevant to planning experiments.
The announcement points to LG's MolMole research, first published in 2025. Its workflow begins by converting complete PDF pages into images. A detector called ViDetect locates molecular drawings; ViMore then identifies atoms and bonds and assembles them into computer-readable structures. Outputs include SMILES, a text notation for molecular structure, and MOLfiles, which can preserve the drawing's layout for inspection. A separate module, ViReact, identifies starting materials, conditions and products in reaction diagrams.
Processing the whole page matters because researchers typically start with papers and patents, rather than a collection of neatly cropped molecule images. LG's development account highlights scanned patents, where noise and crowded layouts complicate recognition. Keeping detection and conversion in the same workflow addresses both finding a drawing and interpreting what it contains.
In the 2025 paper's 300-page patent test, MolMole scored 0.891 against 0.738 for the tested DECIMER extraction pipeline on F1, a measure balancing precision and recall. A successful result required both locating the molecule and matching its encoded structure. This was the authors' selected benchmark and comparison, not a measurement of today's Reaxys service. The project now links to the patent pages and annotations publicly; its separate 250-page article set remains unreleased because of copyright restrictions.
Elsevier says extraction pipelines are checked against existing Reaxys benchmarks before going live. Its announcement gives no numerical live-service accuracy, processing speed or count of newly recovered substances. The current application concerns substance extraction; extending the work to reaction extraction is described as the next stage. The broader research toolkit's capabilities therefore exceed the deployment scope established by this announcement.
For researchers, the immediate significance is the use of image recognition inside a maintained chemistry database. Our assessment is that its practical value will depend on how much previously missed chemistry becomes retrievable without adding incorrect structures. Published coverage and error measurements from the operational pipeline would make that benefit easier to judge.