Large language models (LLMs) are increasingly used in biomedical applications, including literature mining (PMID: 40188094), drug discovery (PMID: 38730226; 41362614; https://arxiv.org/abs/2510.27130), clinical decision support (PMID: 40753316), and patient data analysis (PMID: 41034564). Hybrid approaches combining LLMs with structured knowledge bases and retrieval-augmented generation (RAG) improve performance and interpretability (PMID: 38830083; https://www.biorxiv.org/content/10.1101/2025.05.08.652829v2) . However, LLM-based systems remain vulnerable to hallucinations and generate associations that lack explicit evidence and traceability. This limits their reliability in high-stakes biomedical research. There is an urgent need for methods that systematically ground and validate LLM-derived associations using structured biomedical knowledge, such as knowledge graphs, to enable transparent, evidence-based discovery.
Thesis Type |
|
Status |
Open |
Presentation room |
Seminar room I5 6202 |
Supervisor(s) |
Stefan Decker |
Advisor(s) |
Yongli Mou |
Contact |
mou@dbis.rwth-aachen.de |
The goal of the thesis is to conduct a systematic evaluation of hybrid LLM-RAG, LLM-KG, LLM-KG-RAG systems for biomedical association retrieval and evaluate precision on domain-specific literature. In particular, the thesis will compare agentic systems built using LLMs, knowledge-base-driven approaches, and RAG frameworks. The task is to assess their performance, strengths, limitations, and suitability for biomedical tasks with focus on using multi-modal drug-target-disease association discovery. The work involves designing and implementing computational experiments, benchmarking results, and providing insights into how hybrid agentic systems support research and clinical applications in biomedicine.
Tasks and resources:
- PubMedQA (for biomedical QA: https://pubmedqa.github.io/), MedQA (for clinical QA: https://arxiv.org/abs/2009.13081)
- PrimeKG, BiomedKG, DREAMwalk (additional KGs: https://academic.oup.com/bib/article/23/6/bbac404/6712301)
- BioBERT / PubMedBERT as domain-specialized encoder to support retrieval for RAG
You will gain experience in:
- Implement a complete framework for testing LLMs/KGs and RAG for biomedical domain-specific tasks
- Integrating structured biomedical knowledge bases with AI models
- Applying retrieval-augmented generation techniques
- Evaluating AI systems in terms of accuracy, reasoning, and usability
Additional References:
[1] Zhou, J., Li, H., Chen, S. et al. Large language models in biomedicine and healthcare. npj Artif. Intell. 1, 44 (2025). https://doi.org/10.1038/s44387-025-00047-1
[2] Hayat, S.; Joppich, M.; Kramann,R.. NephroAIX: Large Language Model (LLM)-Guided Knowledge Graph for Explainable Target Discovery in Human Kidney Diseases Using Single-Cell Data: FR-OR014. Journal of the American Society of Nephrology 36(10S):10.1681/ASN.20254q7nxd06, October 2025. | DOI: 10.1681/ASN.20254q7nxd06
[3] KG-LLM-Bench: A Scalable Benchmark for Evaluating LLM Reasoning on Textualized Knowledge Graphs, Link https://arxiv.org/abs/2504.07087
Application instructions:
The thesis is a collaboration between Prof. Stefan Decker (RWTH Aachen) and Prof. Sikander Hayat (Icahn School of Medicine, Mt. Sinai, New York & UKA, Aachen). Please send your application to Yongli Mou, Msc. (mou@dbis.rwth-aachen.de) and Sikander Hayat, PhD (shayat@ukaachen.de, sikander.hayat@mssm.edu).
Requirements:
- Knowledge of machine learning, natural language processing, and programming in Python
- Interest in biomedical data and AI applications in life sciences
- Previous knowledge of single-cell transcriptomics data (desirable but not essential, our team will help you with our biomedical/single-cell omics data)