Abstract
The automatic analysis of patent publications has potential to accelerate research across various domains, including drug discovery and material science. Within patent documents, crucial information often resides in visual depictions of molecule structures. PatCID (Patent-extracted Chemical-structure Images database for Discovery) allows to access such information at scale. It enables users to search which molecules are displayed in which documents. PatCID contains 81M chemical-structure images and 14M unique chemical structures. Here, we compare PatCID with state-of-the-art chemical patent-databases. On a random set, PatCID retrieves 56.0% of molecules, which is higher than automatically-created databases, Google Patents (41.5%) and SureChEMBL (23.5%), as well as manually-created databases, Reaxys (53.5%) and SciFinder (49.5%). Leveraging state-of-the-art methods of document understanding, PatCID high-quality data outperforms currently available automatically-generated patent-databases. PatCID even competes with proprietary manually-created patent-databases. This enables promising applications for automatic literature review and learning-based molecular generation methods. The dataset is freely accessible for download.
Introduction
In the fields of drug discovery and materials science, patents are a primary source of chemical knowledge. However, a significant portion of this information is "trapped" inside visual molecular diagrams rather than searchable text. Historically, researchers have relied on expensive, proprietary manual databases like Reaxys and SciFinder, which are limited by human bandwidth, or automated tools like Google Patents and SureChEMBL, which often suffer from lower image-to-structure conversion quality and poor coverage of non-Western patent offices.
PatCID (Patent-extracted Chemical-structure Images database for Discovery) bridges this gap by providing a high-quality, open-access alternative containing 81 million chemical structure images and 14 million unique molecules extracted from over 40 years of global patent documents.
The PatCID Ingestion Pipeline
To process millions of patent PDFs into a searchable database, the researchers developed a three-stage automated pipeline:
- Segmentation (DECIMER-Segmentation): This module scans patent pages to locate and crop out potential chemical images.
- Classification (MolClassifier): A Mask R-CNN model categorizes images into Molecular Structure, Markush Structure, or Background. This allows the system to filter out "dark data" that current tools cannot process reliably.
- Recognition (MolGrapher): This engine converts 2D images into machine-readable molecular graphs (SMILES strings). Using a graph neural network, MolGrapher is robust to low-resolution or unconventional drawing styles often found in older or international patents.
Benchmarking and Performance
PatCID was evaluated against established giants using two new benchmarks: D2C-RND (Random), reflecting recent patent quality, and D2C-UNI (Uniform), which samples across different years and offices to test system robustness.
On the random benchmark, PatCID achieved a 56.0% molecule recall, significantly higher than Google Patents (41.5%) and SureChEMBL (23.5%). It also outperformed manual databases like Reaxys (53.5%) and SciFinder (49.5%). Furthermore, PatCID achieved 100% document recall, successfully identifying at least one chemical annotation for every document in the benchmark set. Comparison against PubChem suggests PatCID contains 6.8 million "exclusive" molecules previously unavailable in major open repositories.
Key Insights and Data Utility
A major finding is that PatCID captures molecules mentioned in the "Description" sections of patents, which manual curators often skip in favor of the "Examples" section. In one test case, PatCID retrieved 78% of these "pre-example" molecules, while manual databases found virtually zero.
By maintaining the exact page coordinates of every extracted image, PatCID enables high-provenance document exploration. Additionally, its focus on major patent offices in the US, Europe, Japan, Korea, and China ensures a more inclusive view of global innovation.
Limitations and Future Work
Current limitations include the lack of stereochemistry (3D spatial arrangement) capture and the inability to fully enumerate complex Markush structures. Future developments aim to integrate text-extracted chemical data and handle polymers.
Conclusion
PatCID represents a significant advancement in open-access chemical informatics. By outperforming established manual and automated databases, it provides a foundational resource for the next generation of AI-driven drug discovery and automated literature review. The dataset and pipeline code are freely available for the research community.
Summary updated