RexDrug: Reliable Multi-Drug Combination Extraction through Reasoning-Enhanced LLMs

TL;DR

RexDrug is proposed, an end-to-end reasoning-enhanced relation extraction framework for n-ary drug combination extraction based on large language models and established as a scalable and reliable solution for complex biomedical relation extraction from unstructured text.

Abstract

Automated Drug Combination Extraction (DCE) from large-scale biomedical literature is crucial for advancing precision medicine and pharmacological research. However, existing relation extraction methods primarily focus on binary interactions and struggle to model variable-length n-ary drug combinations, where complex compatibility logic and distributed evidence need to be considered. To address these limitations, we propose RexDrug, an end-to-end reasoning-enhanced relation extraction framework for n-ary drug combination extraction based on large language models. RexDrug adopts a two-stage training strategy. First, a multi-agent collaborative mechanism is utilized to automatically generate high-quality expert-like reasoning traces for supervised fine-tuning. Second, reinforcement learning with a multi-dimensional reward function specifically tailored for DCE is applied to further refine reasoning quality and extraction accuracy. Extensive experiments on the DrugComb dataset show that RexDrug consistently outperforms state-of-the-art baselines for n-ary extraction. Additional evaluation on the DDI13 corpus confirms its generalizability to binary drugdrug interaction tasks. Human expert assessment and automatic reasoning metrics further indicates that RexDrug produces coherent medical reasoning while accurately identifying complex therapeutic regimens. These results establish RexDrug as a scalable and reliable solution for complex biomedical relation extraction from unstructured text. The source code and data are available at https://github.com/DUTIR-BioNLP/RexDrug

Figures (7)

• Figure 1: Overall framework of RexDrug
• Figure 2: (a) Task performance under two reasoning strategies. "Full Framework" denotes the complete RexDrug training pipeline. (b) Expert evaluation of reasoning-trace quality with anonymized model identities for RexDrug and GPT-4o. (c) Case study contrasting reasoning traces generated by the single-model and multi-agent settings.
• Figure 3: An instruction template for multi-agent collaborative generation of structured reasoning traces.
• Figure 4: Fine-grained robustness analysis of RexDrug. (a) Higher-order drug combinations ($\geq$3 drugs): performance on Pos-E, Pos-P, Any-E, and Any-P. "Pos" and "Any" indicate metrics for positive combinations and all combinations, respectively; "-E" and "-P" refer to exact match and partial match scores. (b) NO_COMB negatives: discrimination performance measured by micro-F1 on the NO_COMB subset. (c) NO_COMB case study: qualitative comparison of GPT-4o and RexDrug on a representative NO_COMB example.
• Figure 5: Representative penalized example illustrating that the reasoning chain incorrectly attributes anemia to the aspirin–cilostazol combination and lacks sufficient evidence-to-conclusion bridging, resulting in reduced medical and logical accuracy scores.