NicheAgent Artifacts: Prompts, Schemas, Alias Maps, and Ontology Rules for Training-Free Spatial Transcriptomics Annotation

Description

This record contains the full set of artifacts used by NicheAgent, a training-free framework for spatial transcriptomics annotation that combines prototype-based inference with ontology-constrained LLM verification.

The archive includes:

• LLM prompt templates used across all stages of the pipeline, including:

  • system prompts,

  • cell-level label selection prompts,

  • strict JSON-output prompts,

  • and role-specific prefixes for Analyst, Consensus, and Reviewer agents.

• JSON schemas defining the exact input–output contracts expected from LLMs, ensuring deterministic parsing and reproducibility.

• Alias maps specifying the mapping between canonical region labels and LLM-facing aliases, including documentation of how aliases are constructed and applied during inference.

• Ontology rule files encoding biological constraints (required and forbidden marker genes) used to validate or reject LLM-proposed labels.

These artifacts correspond exactly to the logic implemented in the NicheAgent codebase and are intended to support transparent reuse, replication, and extension of training-free, LLM-assisted spatial transcriptomics annotation pipelines. The files are provided in plain-text and JSON formats for long-term accessibility.

For full reproducibility, users are encouraged to pair these artifacts with the corresponding dataset-specific alias maps and runtime configuration used in each experiment.