NeuroOracle
Interactive knowledge graph with 86,000+ neuroscience concepts, 154,000+ edges, 53,000+ literature claims, and 667 machine-generated hypotheses.
Browser window too narrow
NeuroOracle requires a wider viewport to display the interactive graph. Please increase your browser width or use a desktop device.
How to Use NeuroOracle
Find Concepts
Type a keyword in the search bar (e.g., "hippocampus", "dopamine") to locate concepts in the graph. Results are ranked by relevance and connection count. Click a result to select the node and center the view on it.
Click & Double-Click
Single-click a node to view its details in the right panel — including definition, aliases, connected edges, and related claims. Double-click a node to expand its neighborhood and reveal directly connected concepts on the graph canvas.
Browse Literature Claims
Select an edge or node to see the underlying literature claims. Each claim shows the subject–predicate–object triple, source paper (PMID), publication year, and journal. Click the PMID link to open the original paper on PubMed.
Explore Machine-Generated Hypotheses
Switch to the Hypotheses tab to browse 667 machine-generated research hypotheses. Each hypothesis is grounded in graph paths and scored by novelty, feasibility, and supporting evidence. Click any hypothesis to see its reasoning chain and linked concepts.
How NeuroOracle is Built
Claim Extraction from PubMed
NeuroClaw queries PubMed for domain-specific papers (e.g., depression, autism, Alzheimer's) and uses LLM-based extraction to identify structured claims from abstracts and full texts. Each claim captures a subject, predicate, object triple along with source metadata (PMID, year, journal). Entity names are resolved against curated vocabularies (MeSH, NeuroNames, CogAt) through a 6-step resolution pipeline including exact match, case-insensitive, alias, fuzzy, and domain-preference matching.
Graph Assembly and Edge Scoring
Extracted claims are ingested into a NetworkX directed graph. Each resolved entity becomes a node; each claim generates a typed, weighted edge. Edge confidence is computed from source count, journal impact, recency, and cross-paper agreement. Vague predicates are refined via LLM re-classification. A noise filter with salvage logic prevents low-quality LLM extraction artifacts from entering the graph. The pipeline runs with 12 parallel workers for optimal throughput, processing thousands of papers per batch.