LLMs Match Domain Experts in Microbial Oncogenesis Evidence Extraction

A new study from arXiv (2608.07250) demonstrates that large language models (LLMs) can match domain experts in extracting and appraising evidence from microbial oncogenesis research. The research, which focuses on the link between MMTV-LV and breast cancer, involved creating a benchmark dataset from 24 research papers. Domain experts developed a structured template with 77 items, including multiple-choice, Likert-scale, multi-select, and free-text questions. The performance of four LLMs—Gemini 2.5 Pro, Gemini 2.5 Flash, GPT-5, and GPT-5 Nano—was compared against human experts using novel agreement metrics. The findings suggest that LLMs could enable scalable, expert-level systematic evidence synthesis, potentially accelerating the identification of novel microbe-cancer pairs. This capability is crucial because relevant evidence is currently dispersed and too vast for humans to synthesize comprehensively. The study highlights the potential of AI to reduce cancer burden by uncovering new oncogenic microbes.