Extracting structure–activity relationships from patents with large language models

Abstract Background Persistent clinical failures in drug development arise from multiple factors, including imbalanced molecular properties, underscoring the need for systematic structure–activity relationship analysis to link structural modifications with optimized efficacy, selectivity, and safety of small-molecule drugs. Compared with academic literature, drug-related patent documents contain earlier and richer structure–activity relationship information that can support preclinical decision making. However, extracting structure–activity relationship information from patents is technically challenging due to ambiguous terminology, fragmented tables, dispersed compound descriptions, and the need to correctly link compound identities with activity endpoints. To address these issues, we developed PatSAR, a prompt-engineering-centered LLM workflow implemented through the ChatGPT web interface, in which finalized task-specific prompts guide document-by-document extraction of SAR information from patent PDF documents. Results When evaluated on 214 PARP-related patents containing 8226 expert-annotated SAR entities, PatSAR achieved a recall of 79.05%, a precision of 99.65%, and an F1 score of 88.16%. In a multi-target generalization test set of 200 recent patents across ten non-PARP small-molecule targets, PatSAR achieved a recall of 85.25%, a precision of 98.97%, and an F1 score of 91.60%. In an external comparison on DocSAR-Patent-96, PatSAR achieved a Table Recall of 81.86%, compared with the reported Doc2SAR result of 74.09% on the patent subset. Conclusions These results support the potential of large language models for patent information extraction and highlight the role of prompt engineering in improving task-specific extraction performance. The proposed PatSAR framework illustrates the promise of hybrid human–AI workflows for SAR curation and patent analysis.