PatCID: an open-access dataset of chemical structures in patent documents

Area of Science:

• Chemistry
• Biotechnology
• Material Science
• Intellectual Property

Background:

• Patent publications contain valuable chemical structure information crucial for drug discovery and material science.
• Accessing and analyzing these chemical structures within patents at scale is challenging.
• Existing chemical patent databases have limitations in comprehensive molecule retrieval.

Purpose of the Study:

• To introduce PatCID (Patent-extracted Chemical-structure Images database for Discovery), a large-scale database of chemical structures from patent documents.
• To evaluate the performance of PatCID in retrieving chemical structures compared to existing state-of-the-art databases.
• To highlight the utility of PatCID for accelerating research and development.

Main Methods:

• Development of PatCID, a database containing 81 million chemical-structure images and 14 million unique chemical structures extracted from patents.
• Comparison of PatCID's retrieval performance against automatically-created (Google Patents, SureChEMBL) and manually-created (Reaxys, SciFinder) chemical patent databases.
• Utilizing state-of-the-art document understanding methods for high-quality data extraction.

Main Results:

• PatCID achieved a molecule retrieval rate of 56.0% on a random dataset.
• PatCID outperformed Google Patents (41.5%) and SureChEMBL (23.5%) in automatic retrieval.
• PatCID's performance was competitive with manually-curated databases like Reaxys (53.5%) and SciFinder (49.5%).

Conclusions:

• PatCID provides high-quality chemical structure data from patents, surpassing existing automatically-generated databases.
• PatCID demonstrates competitive performance against proprietary, manually-curated chemical patent databases.
• The freely accessible PatCID dataset facilitates applications in automatic literature review and molecular generation.