Harnessing GPT-4 for Automated Curation of E3-Substrate Relationships in the Ubiquitin-Proteasome System

The ubiquitin-proteasome system (UPS) is a complex regulatory network involving around 600 E3 ligases that collectively govern the stability of the human proteome by targeting thousands of proteins for degradation. Understanding this network requires integrating vast amounts of information on gene and protein interactions scattered across unstructured literature. Historically, manual curation has been the gold standard for transforming such data into structured databases, but this process is time-consuming, prone to error, and unable to keep up with the rapid growth of scientific publications. To address these limitations, we developed a scalable, cost-effective workflow using GPT-4, a large language model (LLM), to automate the curation of degradative E3-substrate relationships from the literature. By mining approximately two million PubMed papers, we identified 7,829 degradation-related abstracts and curated a structured database of 3,294 unique E3-substrate pairs using GPT-4, achieving an annotation accuracy rate approaching that of human experts. The resulting database of E3-substrate pairs offers valuable insights into the ubiquitin-proteasome system by highlighting understudied E3s and previously unknown UPS substrates in proteome-wide stability experiments. This automated approach represents substantial increase in productivity compared to manual curation and stands as the largest effort to date utilizing LLMs for the automated curation of protein-protein regulatory relationships. We further showed that our approach is generalizable to other enzyme-substrate families, such as deubiquitinases, kinases, and phosphatases. Overall, our study demonstrates the potential of LLMs as a scalable technology for large-scale curation of signalling relationships, substituting and complementing manual curation to accelerate biological research.