30,000 chemical structure images and corresponding graph-based labels curated from United States Patent and Trademark Office (USPTO) documents. The dataset serves as a benchmark for Optical Chemical Structure Recognition (OCSR) and was introduced alongside the MolGrapher architecture to address data redundancy in existing chemical vision tasks.
Use Cases
- Train graph-based neural networks to perform Optical Chemical Structure Recognition (OCSR) by mapping images to molecular graphs
- Evaluate the generalization of image-to-SMILES translation models across diverse molecular layouts
- Benchmark the accuracy of chemical structure extraction algorithms against patent-sourced visual data
Strengths
- Contains 30,000 distinct chemical structures extracted from USPTO patent literature
- Designed specifically to benchmark graph-based visual recognition of chemical structures
- Mitigates the 'batch similarity' issue where molecules in existing datasets share identical substructures
- Serves as the primary evaluation set for the MolGrapher OCSR framework