LLM-scCurator: Data-centric feature distillation for zero-shot cell-type annotation

Zero-shot cell-type annotation with language models degrades when marker lists are dominated by biological-noise programs (e.g., ribosomal/cell-cycle/stress). We present LLM-scCurator, a data-centric, backend-agnostic framework using pre-prompt noise masking and Gini-informed distillation to recover identity markers. Across benchmarks, LLM-scCurator improves ontology-aware hierarchical accuracy from 78.4% to 86.1% (52 clusters) over naive prompting, approaching supervised/reference-transfer performance without training data. LLM-scCurator extends to spatial transcriptomics (Visium, Xenium) for reference-free discovery of fine-grained niches.