Researchers develop AI system to automatically extract complex drug combinations from medical literature

Finding combinations of drugs that work well together in medical literature is a labor-intensive task that could benefit from automation. Researchers have now developed RexDrug, a system that uses advanced reasoning in large language models to automatically extract descriptions of multi-drug combinations from unstructured biomedical texts.

The challenge with current approaches#

Existing methods for identifying drug relationships in medical texts have focused primarily on simple pairs of drugs. Scaling up to capture more complex therapeutic regimens—where multiple drugs interact in ways that require understanding their collective compatibility and the distributed evidence across paragraphs—remains difficult. This limitation has hindered efforts to leverage the vast reservoir of pharmacological knowledge published in academic papers.

How RexDrug works#

The framework employs a two-phase training approach. In the first stage, a multi-agent collaboration generates reasoning traces that resemble expert analysis, which then guides the model's learning. The second stage applies reinforcement learning with a custom reward system designed specifically for drug combination extraction, further improving both the reasoning quality and accuracy of identified drug interactions.

Validation and results#

Testing on the DrugComb dataset demonstrated that RexDrug surpassed competing baselines for extracting n-ary (multi-drug) relationships. The model also generalized effectively to simpler binary drug-drug interactions when evaluated on the DDI13 corpus. Human expert review confirmed that the system produces medically coherent reasoning while correctly identifying complex therapeutic combinations, positioning it as a practical tool for mining pharmacological insights from published research.

The authors have made their code and data publicly available to support further development in biomedical relation extraction.