---

What is NeuroOracle?

NeuroOracle is the knowledge-graph component of NeuroClaw, an autonomous research framework for neuroimaging. It combines two complementary information sources to provide a quality-grounded foundation for hypothesis generation:

1. Curated structured databases — concepts and relations imported from NeuroNames, MeSH, DisGeNET, BrainMap, and Cognitive Atlas, all aligned to UMLS semantic types.
2. PubMed-derived scientific claims — evidence-weighted edges extracted from neuroimaging literature using LLM-based claim extractors, with provenance (PMID, p-value, sample size, study type) preserved on every edge.

Together they form a graph of approximately 89K concept nodes and 174K edges, covering brain anatomy, diseases, genes, neurotransmitters, drugs, cognitive functions, imaging features, connectivity, visual stimuli, and emotion/vigilance labels.

This Space provides an interactive explorer for browsing the graph, inspecting evidence chains behind individual claims, and visualising multi-hop hypothesis paths.

Why dual-source matters

Existing autoresearch systems either rely on free-form LLM ideation (no quality anchor) or on a single curated KG (limited coverage and stale evidence). NeuroOracle's dual-source design is what enables NeuroClaw to:

• Generate hypotheses with traceable evidence chains back to specific PubMed papers
• Filter or re-rank hypotheses using evidence weights (effect size, sample size, replicability)
• Iterate the graph itself in response to new findings, rather than treating the KG as a static asset

NeuroClaw ecosystem

NeuroOracle is one of three modules within the broader NeuroClaw system:

NeuroClaw is the umbrella framework; NeuroOracle is its scientific memory; NeuroBench measures how effectively the agent can use that memory to do real research work.

What you can do here

• Browse concepts across 13 domain tags (neuroanatomy, disease, gene, drug, imaging_feature, connectivity, cognitive_function, visual_stimulus, emotion, vigilance, paradigm, dataset, ml_model)
• Inspect claims — every PubMed-derived edge carries the source paper, predicate (is_biomarker_of, predicts, correlates_with, etc.), and structured evidence fields
• Trace hypothesis paths — multi-hop reasoning examples such as visual stimulus → functional ROI → anatomical region, or imaging feature → gene → disease
• Filter subgraphs by domain, dataset, or relation type for focused exploration

Links

• 🏠 Project homepage: https://cuhk-aim-group.github.io/NeuroClaw/
• 💻 Source code (GitHub): https://github.com/CUHK-AIM-Group/NeuroClaw
• 📄 Technical report (arXiv): https://arxiv.org/abs/2604.24696
• 🧠 NeuroOracle docs page: https://cuhk-aim-group.github.io/NeuroClaw/neuro-oracle.html

Citation

If NeuroOracle or NeuroClaw is useful for your research, please cite the NeuroClaw technical report:

@article{neuroclaw2026,
  title   = {NeuroClaw: Closed-Loop Agentic AI for Executable and Reproducible Neuroimaging Research},
  author  = {NeuroClaw Team},
  journal = {arXiv preprint arXiv:2604.24696},
  year    = {2026},
  url     = {https://arxiv.org/abs/2604.24696}
}

License

MIT — same as the NeuroClaw repository. See https://github.com/CUHK-AIM-Group/NeuroClaw/blob/main/LICENSE for full terms.

Contact

For questions or issues, please open an issue on the NeuroClaw GitHub repository.