An exploration of using LLMs (deepseek-v4-pro) to assess patent applications and grants for antibody-drug conjugate (ADC) components. Identified components are joined to public data sources: ChEMBL, Uniprot, and AACT/ctgov. Joined data is published as Parquet files for browsing and reporting (see link above).

Still squarely on the 80 side of the 80/20 data engineering rule. Lots of data cleanup and normalization. Next steps are to improve filtering to reduce false positives then on to joining other data sources.

AI Used:

  • DeepSeek for patent assessments. 500 patent assessment prompts was about 10.00 USD.
  • Claude Chat for design and troubleshooting. Basically a Google and StackOverflow replacement.
  • Qwen 3/3.5 and DeepSeek via local llama.cpp for code review and spec adherence.

Miscellaneous:

  • Pipeline could be adaptable to other subsets of patent data by using different LLM prompts.
  • Started with Airflow mostly due to resume pressure. Tried prefect and stayed.
  • Stayed local and lo-fi: 300G of USPTO data on external drive, local Forgejo, local pipeline runs.

Prefect flow: ADC pipeline