An intermediate dataset for US molecular patent grants

Retrieves patent grant data from USPTO weekly releases bulkdata.uspto.gov/data/patent/grant/redbook/{year} and keeps only patents with .mol files for downstream data mining use cases.

Compared to existing work such as [https://github.com/learningmatter-mit/PatentChem/tree/main]. We keep all files of a patent grant, especially .til image files. We also share the data as a git lfs repo instead of a single .zip file to allow easier access.

The data directory

data/{year}/{weekly}/{sub_dir}/{filename}.ZIP

Each file is a full patent grant with original redbook directory structure preserved. weekly is the weekly release date. Expect inconsistencies for the formate of weekly and depth in sub_dir over time.

See [https://www.uspto.gov/learning-and-resources/xml-resources] for DTD and Red Book Viewer.

Data quality

Ideally, USPTO provides the ground truth on chemical structure annotations for the molecular drawings. But the raw data quality appears poor. Some outlandish annotations such as chemical structure annotations on non-chemistry related images suggest an automated annotation process with little human QA is used to annotate the drawings.

A report on improving data quality is left as a future work. (seeking calibrators)

Related resources

(2023) Automated patent extraction powers generative modeling in focused chemical spaces [https://github.com/learningmatter-mit/PatentChem]
(2014) Processing USPTO Patent Data [https://github.com/funginstitute/patentprocessor/] [https://funginstitute.berkeley.edu/wp-content/uploads/2014/06/patentprocessor.pdf]: How to extract information from the .xml file
(2011) SCRIPDB: a portal for easy access to syntheses, chemicals and reactions in patents [https://pmc.ncbi.nlm.nih.gov/articles/PMC3245107/]: extensive data analytics, but database and website no longer exist.

The DATA-CHANGELOG.md file

This file is written by update_data.py to log statistics from each processed weekly.

How to update the dataset

python update_data.py --from_year 2024 will search for new weeklies from 2024 to current year, add them to the data directory and push to upstream.

Please only use this project's root directory as the working directory. You can also copy pre-downloaded weeklies here to avoid re-downloading them.

$ python update_data.py -h

usage: update_data.py [-h] [--quick] [--keep] [--stream] [--verbose] [--from_year FROM_YEAR] [--to_year TO_YEAR]

Downloads and filters patent files for mol files from USPTO website, then adds them to a dataset

options:
  -h, --help            show this help message and exit
  --quick               only processed at most one new weekly file, you can also create a file named 'quick' (touch quick) to enable this flag during runtime to exit after finishing  
                        processing the current weekly file
  --keep                keep the original downloaded files
  --stream              remove all pushed lfs files and objects locally. DO NOT USE WHEN RUNNING PARALLEL SCRIPTS
  --verbose             print verbose logs
  --from_year FROM_YEAR
                        start from this year (default 2001 is the earliest available)
  --to_year TO_YEAR     end at this year (default current year)