MolMole: Molecule Mining from Scientific Literature

TL;DR

MolMole addresses the challenge of extracting molecular structures and reaction data from unstructured scientific literature by unifying detection, reaction parsing, and OCSR into a single page-level pipeline. It introduces ViDetect, ViReact, and ViMore, and couples them with a novel 550-page benchmark and evaluation metric for end-to-end document-level chemical data extraction. Empirically, MolMole achieves state-of-the-art performance on page-level tasks and competitive results on public OCSR benchmarks, with robust handling of complex layouts such as two-column diagrams and wavy-bond representations. This work enables direct, end-to-end extraction of chemical data from documents, facilitating integration into public databases and accelerating chemoinformatics research and discovery.

Abstract

The extraction of molecular structures and reaction data from scientific documents is challenging due to their varied, unstructured chemical formats and complex document layouts. To address this, we introduce MolMole, a vision-based deep learning framework that unifies molecule detection, reaction diagram parsing, and optical chemical structure recognition (OCSR) into a single pipeline for automating the extraction of chemical data directly from page-level documents. Recognizing the lack of a standard page-level benchmark and evaluation metric, we also present a testset of 550 pages annotated with molecule bounding boxes, reaction labels, and MOLfiles, along with a novel evaluation metric. Experimental results demonstrate that MolMole outperforms existing toolkits on both our benchmark and public datasets. The benchmark testset will be publicly available, and the MolMole toolkit will be accessible soon through an interactive demo on the LG AI Research website. For commercial inquiries, please contact us at \href{mailto:contact_ddu@lgresearch.ai}{contact\_ddu@lgresearch.ai}.

Figures (13)

• Figure 1: MolMole pipeline. ViDetect detects molecular regions in document images, while ViReact extracts reactants, products, and conditions from reaction diagrams. ViMore processes the identified molecular structures, converting them into SMILES or MOLfiles.
• Figure 2: Hallucination effects from generative models: repeated SMILES generation (top) and incorrect chemical bias (bottom).
• Figure 3: ViMore results of Polymer (top) and Wavy line (middle). ViMore preserves the molecule coordinates from the image (bottom).
• Figure 4: Examples of reaction extraction in two-column documents. (a) and (b) show MolMole successfully extracting reaction information that spans from the first to the second column. In contrast, (c) and (d) show results from RxnScribe, which fails to capture the full reaction due to its reliance on cropped, isolated diagrams.
• Figure 5: Sample ViDetect results from our testset.