Grounding QA Generation in Knowledge Graphs and Literature: A Scalable LLM Framework for Scientific Discovery

Therapeutic biomarkers are crucial in biomedical research and clinical decision-making, yet the field lacks standardized datasets and evaluation methods for complex, context-dependent questions. To address this, we integrate large language models (LLMs) with knowledge graphs (KGs) to filter PubMed abstracts, summarize biomarker contexts, and generate a high-quality synthetic Q/A dataset. Our approach mirrors biomarker scientists' workflows, decomposing question generation into classification, named entity recognition (NER), and summarization. We release a 24k high quality Q/A dataset and show through ablation studies that incorporating NER and summarization improves performance over using abstracts alone. Evaluating multiple LLMs, we find that while models achieve 95.45\% accuracy on multiple-choice questions, performance drops to 59.67\% on open-ended Q/A, highlighting the need for synthetic data to address the issue of novel discovery. By addressing a critical resource gap, this work provides a scalable tool for biomarker research and demonstrates AI’s broader potential in scientific discovery.