Self-Refining Biological Agents: Mitigating Hallucinations in Single-Cell Annotation

ABSTRACT

Automating cell-type annotation in single-cell RNA-seq data using Large Language Models often leads to confident but biologically inaccurate classifications. We present a 'Self-Refining Agent' framework that mitigates these hallucinations through an internal reasoning loop. The agent proposes initial cell-type labels based on marker genes, then autonomously queries external knowledge bases (like Cell Ontology) and cross-references spatial transcriptomic contexts to verify its own predictions. This multi-step, self-correcting workflow improves annotation accuracy in complex tumor microenvironments by 42% compared to zero-shot LLM prompts, providing a highly reliable tool for downstream immunological analysis.