Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

KEDRec-LM: A Knowledge-distilled Explainable Drug Recommendation Large Language Model

Kai Zhang, Rui Zhu, Shutian Ma, Jingwei Xiong, Yejin Kim, Fabricio Murai, Xiaozhong Liu

TL;DR

KEDRec-LM tackles explainable drug discovery by marrying structured biomedical knowledge from the Drug Repurposing Knowledge Graph with unstructured evidence from PubMed Central and Clinical Trials using a Retrieval-Augmented Generation pipeline. A three-stage methodology samples disease–drug pairs from DRKG, retrieves contextual background, and distills teacher guidance into an instruction-tuned LLaMA-based student model for drug selection and rationale generation. The approach yields state-of-the-art performance on expRxRec and MIMIC-III in drug selection and rationale quality, with dual-source background information providing the strongest gains. By publicly releasing the expRxRec dataset and KEDRec-LM, the work advances explainable AI in drug discovery and offers a scalable framework for integrating graphs, literature, and distillation. The results highlight the value of grounding LLM reasoning in combined structured and unstructured biomedical knowledge for practical therapeutic hypothesis generation and decision support.

Abstract

Drug discovery is a critical task in biomedical natural language processing (NLP), yet explainable drug discovery remains underexplored. Meanwhile, large language models (LLMs) have shown remarkable abilities in natural language understanding and generation. Leveraging LLMs for explainable drug discovery has the potential to improve downstream tasks and real-world applications. In this study, we utilize open-source drug knowledge graphs, clinical trial data, and PubMed publications to construct a comprehensive dataset for the explainable drug discovery task, named \textbf{expRxRec}. Furthermore, we introduce \textbf{KEDRec-LM}, an instruction-tuned LLM which distills knowledge from rich medical knowledge corpus for drug recommendation and rationale generation. To encourage further research in this area, we will publicly release\footnote{A copy is attached with this submission} both the dataset and KEDRec-LM.

KEDRec-LM: A Knowledge-distilled Explainable Drug Recommendation Large Language Model

TL;DR

KEDRec-LM tackles explainable drug discovery by marrying structured biomedical knowledge from the Drug Repurposing Knowledge Graph with unstructured evidence from PubMed Central and Clinical Trials using a Retrieval-Augmented Generation pipeline. A three-stage methodology samples disease–drug pairs from DRKG, retrieves contextual background, and distills teacher guidance into an instruction-tuned LLaMA-based student model for drug selection and rationale generation. The approach yields state-of-the-art performance on expRxRec and MIMIC-III in drug selection and rationale quality, with dual-source background information providing the strongest gains. By publicly releasing the expRxRec dataset and KEDRec-LM, the work advances explainable AI in drug discovery and offers a scalable framework for integrating graphs, literature, and distillation. The results highlight the value of grounding LLM reasoning in combined structured and unstructured biomedical knowledge for practical therapeutic hypothesis generation and decision support.

Abstract

Drug discovery is a critical task in biomedical natural language processing (NLP), yet explainable drug discovery remains underexplored. Meanwhile, large language models (LLMs) have shown remarkable abilities in natural language understanding and generation. Leveraging LLMs for explainable drug discovery has the potential to improve downstream tasks and real-world applications. In this study, we utilize open-source drug knowledge graphs, clinical trial data, and PubMed publications to construct a comprehensive dataset for the explainable drug discovery task, named \textbf{expRxRec}. Furthermore, we introduce \textbf{KEDRec-LM}, an instruction-tuned LLM which distills knowledge from rich medical knowledge corpus for drug recommendation and rationale generation. To encourage further research in this area, we will publicly release\footnote{A copy is attached with this submission} both the dataset and KEDRec-LM.

Paper Structure

This paper contains 12 sections, 8 equations, 1 figure, 2 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Methodology
  4. Stage 1: Sampling from DRKG
  5. Stage 2: Background Retrieval
  6. Stage 3: Knowledge Distillation
  7. Experiments
  8. Dataset & Experimental Setup
  9. Results on Drug Selection
  10. Results on Reason Generation
  11. Related Work
  12. Conclusion

Figures (1)

  • Figure 1: Overview of KRxELM. Stage 1 samples a disease-drug set which consists of a disease and two drug candidates from the open-sourced rich DRKG; Stage 2 leverages a Language model to embed the drug-disease pair from the sampled disease-drug set, the embeddings are further used for obtaining background information from Clinical Trials and PubMed Central corpora; Stage 3 fits the disease-drug set and the retrieved background information into an instructional prompt template as input to a large language model (student) and a teacher model. Instruction tuning was used to enable the LLM to learn from the teacher model.