C-MAGE: Automated Extraction and Organization Engine for Molecular Representations in Chemical Documents

Abstract

To exploit the data present in chemical literature for AI applications in chemistry, an expanded and reliable extraction pipeline is needed. AI-ready chemical databases broaden chemical space coverage and provide benchmarking and training sets for machine learning model development to drive innovation within the field. To curate data from chemical documents, a multimodal approach is needed. C-MAGE is an open-source molecule extraction engine that processes figures and tables in PDF files to provide machine-readable molecule data sets of the structures as CXSMILES strings, an encompassing representation that evades heuristic translation of superatoms to SMILES subunits. This led to a 7.1% increase in accurate molecular representations generated. All data extracted undergo confidence-based sorting to ensure high precision (77.6%) in data set collection and remove the need for manually sorting data point validity. Both the CXSMILES and their confidence values are reproducible and are generated efficiently (1.7 s per molecule). C-MAGE’s use is demonstrated by the generation of an annotated data set containing 1454 CXSMILES strings from 38 chemical documents. C-MAGE thus enables large-scale information extraction of chemical structures in a consistent and reproducible pipeline, complementing the more common text extractions.