Source-linked AI summary
From Knowledge to Treatment: Large Language Model Assisted Biomedical Concept Representation for Drug Repurposing
Chengrui Xiang, Tengfei Ma, Xiangzheng Fu, Yiping Liu, Bosheng Song, Xiangxiang Zeng
TL;DR
Biomedical KGs support drug repurposing, but existing methods often miss common-sense mechanistic constraints relevant to treatment decisions. LLaDR uses LLM-derived biomedical concept semantics to fine-tune KG embeddings. The paper reports state-of-the-art benchmark performance and robustness under KG noise and semantic perturbations.
Problem
Biomedical KGs and structure-focused methods often overlook common-sense mechanistic constraints, although such priors are important for safe treatment decisions and generalization to complex diseases.
Source: §1 · p. 2Method
LLaDR extracts treatment-related textual representations of biomedical entities with LLMs and uses them to fine-tune KG embeddings.
Source: §1 · p. 2, §3.2 · p. 3Results
LLaDR is reported to achieve state-of-the-art benchmark performance and robustness under KG noise and semantic perturbations.
Source: §1 · p. 2Takeaways & Limitations
LLaDR combines structural and semantic information for drug-disease prediction and is described as applicable to biomedical KGs such as disease-gene networks.
Source: §5 · p. 8Errors or biases in LLM-generated descriptions can affect performance and reliability, while scaling to graphs with millions of entities may substantially raise computational costs.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
Drug repurposing plays a critical role in accelerating treatment discovery, especially for complex and rare diseases. Biomedical knowledge graphs (KGs), which encode rich clinical associations, have been widely adopted to support this task. However, existing methods largely overlook common-sense biomedical concept knowledge in real-world labs, such as mechanistic priors indicating that certain drugs are fundamentally incompatible with specific treatments. To address this gap, we propose LLaDR, a Large Language Model-assisted framework for Drug Repurposing, which improves the representation of biomedical concepts within KGs. Specifically, we extract semantically enriched treatment-related textual representations of biomedical entities from large language models (LLMs) and use them to fine-tune knowledge graph embedding (KGE) models. By injecting treatment-relevant knowledge into KGE, LLaDR largely improves the representation of biomedical concepts, enhancing semantic understanding of under-studied or complex indications. Experiments based on benchmarks demonstrate that LLaDR achieves state-of-the-art performance across different scenarios, with case studies on Alzheimer's disease further confirming its robustness and effectiveness. Code is available at https://github.com/xiaomingaaa/LLaDR.
1 Introduction
Drug repurposing can accelerate drug development, but existing KG methods often overlook common-sense biomedical constraints that matter for treatment decisions. LLaDR integrates LLM-derived concept semantics into KG embeddings and is reported to achieve state-of-the-art benchmark performance and robustness.
1 Introduction
- Drug repurposing identifies new therapeutic uses for existing drugs and can accelerate drug development.The growing complexity of biological systems and availability of heterogeneous biomedical data create a need for computational approaches that integrate and reason over large-scale sources.Source: §1 · p. 1
- Common-sense mechanistic constraints, including treatment incompatibilities, are underrepresented in biomedical KGs and matter for safe and effective treatment decisions.Structure-only embeddings rarely capture these priors, limiting reliability when generalizing to rare or mechanistically complex diseases.Source: §1 · p. 2
- LLaDR integrates LLM-extracted concept-level semantics into KG embeddings to improve biomedical concept representations and treatment-aware drug repurposing.The framework uses LLM-generated textual descriptions to guide KGE fine-tuning.Source: §1 · p. 2
- The authors report state-of-the-art benchmark performance and robustness under KG noise and semantic perturbations.They also describe LLaDR as the first framework integrating common-sense biomedical concept knowledge into KG-based drug repurposing.Source: §1 · p. 2
2 Related Work
Prior drug-repurposing approaches incorporate biomedical graphs and language models, but often neglect high-level biomedical constraints or face computational and scale limitations. LLaDR extracts biomedical concept semantics with LLMs and integrates them into KG embeddings through hierarchical alignment.
2.1 Drug Repurposing
- Existing drug-repurposing methods often neglect contraindications and mechanistic incompatibilities, which are essential for clinically valid repurposing decisions.The proposed concept-aware framework incorporates biomedical priors into reasoning for more trustworthy, mechanistically grounded decisions.Source: §2.1 · p. 2
2.2 LLM incorporated Knowledge Representation
- Earlier PLM-KG approaches use classification, prompting, cross-modal fusion, or fused embeddings, but can be computationally intensive, limited to small-scale models, or costly for specialized tasks.LLaDR instead extracts biomedical concept semantics with LLMs and integrates them into KG embeddings through lightweight hierarchical alignment.Source: §2.2 · p. 2
3 LLaDR Framework
LLaDR enriches biomedical knowledge graph embeddings with LLM-derived entity descriptions, then fine-tunes them to retain textual semantics and graph structure.
3 LLaDR Framework
- LLaDR has two phases: generate text-based entity representations, then use them to fine-tune knowledge graph embeddings.The framework aims to combine graph structure with semantic information from text.Source: §3 · pp. 2–3
3.1 Generating Discription
- For each knowledge-graph entity, LLaDR prompts an LLM for a description, embeds that description, and concatenates it with the entity embedding.The prompt requests a concise, informative description in a consistent format.Source: §3.1 · p. 3
3.2 Knowledge Graph Fine-Tuning
- Fine-tuning uses text embeddings to align structural representations with text while optimizing link prediction.The semantic anchoring constraint is intended to preserve semantic integrity and reduce overfitting in large clusters and sparse graphs.Source: §3.2 · pp. 3–4
- The link-prediction objective contrasts observed triples with sampled negative tails, while the total loss combines anchor and link losses.The anchor loss draws similar entities or relations closer; the link loss preserves graph relationships.Source: §3.2 · pp. 4–5
4 Experimental Setup
LLaDR is evaluated on DRKG drug-repurposing link prediction against language-model, graph-neural-network, and structure-based baselines. The experiments compare overall performance, generation settings, semantic masking, robustness to KG noise, and Alzheimer’s disease candidates.
4 Experimental Setup
- The evaluation uses DRKG tail-entity replacement, sampling 50 disease candidates per test triple and averaging five runs across three therapeutically relevant relation types.MR, MRR, Hits@N, and AUC are the reported metrics.Source: §4 · pp. 5–6
4.1 Experimental Results
- LLaDR outperforms the compared PLM-based, GNN-based, and structure-driven methods across all reported evaluation metrics on DRKG.With GPT-3.5-turbo descriptions, LLaDR-DistMult reports a 24.5% relative improvement in MR, a 6.7% enhancement in H@10, and a 3.1% increase in AUC.Source: §4.1 · p. 6, Table 1 · p. 6
- The prompt experiment compares no prompt, the original prompt, and a refined prompt using GPT-4o-mini.Full results are reported in Appendix A.4.Source: §4.1 · p. 7, Table 2 · p. 7
- The temperature experiment compares text generation at 0.1 with the original setting of 0.7 for downstream drug repositioning.Full results are reported in Appendix A.4.Source: §4.1 · p. 7, Table 3 · p. 7
- LLaDR maintains stable performance under substantial KG noise, while the baseline models’ accuracy declines as noise increases.The perturbations add or delete unknown associations; the authors attribute LLaDR’s resilience to regularization and LLM-generated initial embeddings.Source: §4.1 · p. 7, Figure 3 · p. 8
- Masking increasing proportions of LLM-generated descriptions—20%, 40%, and 60%—significantly lowers metrics across models.The authors interpret the decline as evidence of the descriptions’ role in performance.Source: §4.1 · p. 7, Table 4 · p. 7
4.2 Case Analysis
- The Alzheimer’s case analysis identifies Dasatinib and Quercetin among the top-ranked candidate treatments and examines their potential mechanisms of action.The authors report literature support for their potential in Alzheimer’s therapy and refer to Appendix A.7 for the top 10 predictions.Source: §4.2 · p. 8, Figure 4 · p. 8
5 Conclusion
LLaDR combines structural embeddings with LLM-derived semantics to improve drug-disease prediction. Its representations capture structural and semantic relationships, with potential application to biomedical KGs such as disease-gene networks.
5 Conclusion
- LLaDR combines structural embeddings with LLM-derived semantics and surpasses traditional methods in drug-disease prediction accuracy.Source: §5 · p. 8
- The approach captures structural and semantic relationships relevant to drug discovery and applies to biomedical KGs such as disease-gene networks.Source: §5 · p. 8
- LLaDR enhances DRKG for drug repositioning as a scalable solution for knowledge graph improvements in drug discovery.Source: §5 · p. 8
Limitations
The authors identify dependence on LLM-generated descriptions, scalability, and interpretability as limitations requiring further investigation. These issues may constrain reliability, large-graph deployment, and clinical adoption.
Limitations
- Errors or biases in LLM-generated descriptions can affect performance and reliability.The authors identify descriptions generated by GPT-3.5-turbo and GPT-4o-mini as examples.Source: Limitations · p. 9
- Computational costs may rise substantially on graphs with millions of entities, posing barriers to deployment at scale.Source: Limitations · p. 9
- LLMs' black-box nature complicates clinical adoption and raises transparency concerns that may limit stakeholder trust and acceptance.Source: Limitations · p. 9
A.1 Datasets
The DRKG dataset is divided into training, validation, and test sets; the test set contains meaningful drug-disease triplets for repositioning experiments.
A.1 Datasets
- DRKG is split into 3,889,539 training entries, 429,959 validation entries, and 39,829 test entries.Source: Table 5 · p. 12
- The test set primarily contains meaningful drug-disease triplets used to identify new therapeutic applications.Source: Table 5 · p. 12
A.2 Hardware
Experiments used an RTX 3080, with DRKG training taking two hours per session.
A.2 Hardware
- Training and testing used an RTX 3080, and each 100,000-step DRKG training session took two hours.Source: §A.2 · p. 12
A.3 Drug Repurposing Results
Table 6 reports that LLaDR consistently outperforms PLM-based, GNN-based, and structure-based models across all evaluation metrics for tail-entity prediction.
A.3 Drug Repurposing Results
- LLaDR outperforms PLM-based, GNN-based, and structure-based models across all reported evaluation metrics.The comparison uses averaged results from ten DRKG runs for tail-entity predictions.Source: Table 6 · p. 12
- The comparison evaluates tail-entity predictions in the drug-repurposing task.Source: Table 6 · p. 12
- Table 6 reports results averaged across ten runs for DRKG.Source: Table 6 · p. 12
A.4 prompt,temperature,KL and bge
The appendix tests description-generation choices and embedding variants, alongside robustness, semantic masking, and Alzheimer’s drug predictions. It reports that prompting outperforms no prompts and that progressively masking descriptions degrades performance across three embedding models.
A.4 prompt,temperature,KL and bge
- Using low-temperature descriptions does not significantly decrease the experimental metrics.Source: Table 7 · p. 13
- Descriptions generated without prompts perform significantly worse in drug repurposing than prompted descriptions.The prompt examples include a structured biomedical-entity template and outputs describing HER2.Source: Table 8 · p. 13, noprompt: (gpt-4o-mini) · p. 13, goodprompt: (gpt-4o-mini) · p. 14
- The authors tested replacing the text embedding deviation constraint with Kullback-Leibler divergence.The passage does not report the experiment’s outcome.Source: Table 9 · p. 14
- Embeddings generated by the closed-source large model remain highly effective in downstream drug-repositioning tasks.The test uses bge-v1.5-small for entity and description embeddings.Source: Table 10 · p. 15
A.5 Robustness Experiment Results
- LLaDR’s performance remains consistent under progressive noise injection and triple removal, including at the tested 20%, 40%, and 60% levels.Source: Table 11 · p. 15
A.6 Mask Experiment Results
- Progressively masking GPT-4o-mini descriptions worsens performance across TransE, DistMult, and RotatE.The experiment tests 20%, 40%, and 60% masking.Source: Table 12 · p. 16
A.7 Top 10 Drugs
- Most of the ten predicted Alzheimer’s disease drugs are associated with relevant literature on Alzheimer’s.The authors report this finding from an online survey.Source: Table 13 · p. 16