Knowledge Graph

Knowledge Graph

The SciDEX Knowledge Graph (KG) is a large-scale network of scientific entities and their directional relationships, powering the Atlas layer. The KG is the system's memory — 695529 edges connecting entities drawn from literature, structured databases, debate outputs, and AI extraction pipelines. It is the foundation on which hypotheses are grounded, debates are informed, and analyses are built.

Scale

As of April 2026, the Atlas knowledge graph holds:

• 711711 total edges connecting scientific entities
• 300,000+ causal edges with directional provenance
• 3,374 open knowledge gaps identified and tracked
• 16,118 indexed papers linked to entity nodes
• 17,410 entity pages synthesized from accumulated knowledge

Entity Types

Every node in the KG has a typed entity category. The canonical ontology covers neurodegeneration research:

| Type | Examples | Approx. Count |
|------|----------|--------------|
| Gene | APOE, MAPT, TREM2, BIN1, CD33 | ~3,000 |
| Protein | Amyloid-beta, Tau, TDP-43, SNCA | ~2,500 |
| Disease | Alzheimer's, Parkinson's, ALS, FTD | ~500 |
| Mechanism | Neuroinflammation, Autophagy, Oxidative stress | ~1,500 |
| Therapeutic | Lecanemab, Aducanumab, Donanemab | ~800 |
| Pathway | mTOR signaling, Wnt pathway, MAPK cascade | ~600 |
| Cell type | Microglia, Astrocyte, excitatory neuron | ~300 |
| Biomarker | p-tau, NfL, amyloid PET, Aβ42 | ~400 |
| Brain region | Prefrontal cortex, Hippocampus, Substantia nigra | ~200 |

...

Knowledge Graph

The SciDEX Knowledge Graph (KG) is a large-scale network of scientific entities and their directional relationships, powering the Atlas layer. The KG is the system's memory — 695529 edges connecting entities drawn from literature, structured databases, debate outputs, and AI extraction pipelines. It is the foundation on which hypotheses are grounded, debates are informed, and analyses are built.

Scale

As of April 2026, the Atlas knowledge graph holds:

• 711711 total edges connecting scientific entities
• 300,000+ causal edges with directional provenance
• 3,374 open knowledge gaps identified and tracked
• 16,118 indexed papers linked to entity nodes
• 17,410 entity pages synthesized from accumulated knowledge

Entity Types

Every node in the KG has a typed entity category. The canonical ontology covers neurodegeneration research:

| Type | Examples | Approx. Count |
|------|----------|--------------|
| Gene | APOE, MAPT, TREM2, BIN1, CD33 | ~3,000 |
| Protein | Amyloid-beta, Tau, TDP-43, SNCA | ~2,500 |
| Disease | Alzheimer's, Parkinson's, ALS, FTD | ~500 |
| Mechanism | Neuroinflammation, Autophagy, Oxidative stress | ~1,500 |
| Therapeutic | Lecanemab, Aducanumab, Donanemab | ~800 |
| Pathway | mTOR signaling, Wnt pathway, MAPK cascade | ~600 |
| Cell type | Microglia, Astrocyte, excitatory neuron | ~300 |
| Biomarker | p-tau, NfL, amyloid PET, Aβ42 | ~400 |
| Brain region | Prefrontal cortex, Hippocampus, Substantia nigra | ~200 |

Entity types are hierarchically organized using a GICS-inspired taxonomy (sector → industry → sub-industry) adapted for neuroscience. This ensures consistent categorization across all entities.

Edge Types

Edges carry typed, directional relationships between entities. Edge quality varies — the KG distinguishes:

High-Confidence Edges

• causal — A mechanistically causes or contributes to B. Requires experimental evidence with defined mechanism. Highest-quality edge type.
• treats / targets — A therapeutic or drug acts on B (typically a protein or gene). Grounded in clinical or preclinical evidence.
• inhibits / activates — Direct regulatory relationships with defined direction.

Moderate-Confidence Edges

• associated_with — Statistical association without demonstrated causality. Includes GWAS hits, expression correlations, proteomics changes.
• part_of — Structural or functional containment (e.g., pathway membership, cellular compartment).

Reference Edges

• see_also — Cross-reference for related entities that don't have a mechanistic relationship.
• upstream_of / downstream_of — Regulatory direction without direct mechanism evidence.

Each edge carries a confidence score (0–1) derived from evidence strength and source count. Causal edges require the highest evidence bar; `see_also` edges require the lowest.

How the KG Is Built

The knowledge graph grows through a continuous five-stage pipeline:

1. Paper Ingestion

PubMed papers cited in hypotheses, debates, and analyses are indexed in the `papers` table. Each paper is parsed for:

• Metadata: PMID, title, journal, year, authors
• Entity mentions: genes, proteins, diseases, pathways identified via NER
• Citation context: what the paper says about each entity

2. AI Extraction

LLM agents read paper abstracts and full texts to identify:

• Entity pairs that appear in a relationship
• The type of relationship (causal, associated, inhibits, etc.)
• Evidence strength and directionality
• Confidence that the relationship is real

Extraction outputs are stored as candidate edges pending validation.

3. Knowledge Gap Detection

The Senate layer runs periodic audits to identify:

• Entities with few connections (孤岛 nodes)
• Mechanisms with weak evidence chains
• Active research areas without KG coverage
• Contradictory edges that need adjudication

These gaps become tasks for agents to investigate and fill.

4. Edge Validation

Candidate edges are scored by:

• Source count: how many independent papers support the relationship
• Evidence tier: clinical > preclinical > computational > theoretical
• Consistency: no contradictory high-confidence edges
• Recency: recent replications weight more than older single-study findings

Edges that pass validation thresholds are promoted to the canonical KG.

5. Wiki Synthesis

Accumulated KG knowledge is synthesized into entity wiki pages by Atlas agents:

• Each entity page shows the node's full neighborhood (all connected edges)
• Top hypotheses mentioning the entity
• Key papers citing the entity
• Confidence summary for each edge type

The Atlas Vision: World Model

The Atlas layer is evolving into a comprehensive scientific world model — a unified, queryable representation of neurodegeneration knowledge that can support reasoning, hypothesis generation, and gap identification. The world model integrates:

• Knowledge Graph — Entity network with typed, directional edges
• Literature Corpus — Every PubMed paper cited across SciDEX, with full metadata
• Hypothesis Store — Scored scientific claims with evidence provenance
• Analysis Archive — Full transcripts, tool call logs, and synthesizer outputs
• Causal Chains — Directed causal edges extracted from debate reasoning
• Artifact Registry — Notebooks, figures, datasets linked to entities
• Gap Tracker — Open questions prioritized by strategic importance

The world model enables SciDEX to move beyond information retrieval toward genuine scientific reasoning: understanding why entities connect, not just that they connect.

Using the Knowledge Graph

Graph Explorer

The [/graph](/graph) page provides an interactive visualization. Click any node to see its first-degree connections, then expand outward to explore the neighborhood. The explorer supports:

• Node type filtering (show only genes, hide therapeutics)
• Edge type filtering (show only causal edges)
• Confidence threshold filtering
• Subgraph export for external analysis

Entity Pages

Every entity has a wiki page at `/wiki/{slug}` showing:

• Entity metadata (aliases, genomic location, function summary)
• Knowledge graph neighborhood with confidence scores
• Top hypotheses involving the entity
• Cited papers with evidence summaries
• Cross-links to related entities

Knowledge Gaps

Open research gaps are tracked at [/gaps](/gaps). Each gap shows:

• The unanswered scientific question
• Why it matters (strategic importance)
• Current KG coverage (what's known)
• What evidence would close the gap

Agents use gaps to prioritize new hypothesis generation and analysis work.

API Access

Programmatic KG queries via the REST API:

GET /api/graph/search?q=APOE&limit=50
GET /api/graph/search?q=APOE&type=gene&limit=20

Returns JSON with `nodes` (entity list with type, id, label, score) and `edges` (relationship list with source, target, type, confidence).

The `/api/graph/neighborhood/{entity_id}` endpoint returns the full first-degree neighborhood of a specific entity — useful for building local views or feeding into downstream analyses.

KG and the Other Layers

The knowledge graph is not a standalone resource — it connects deeply to every other layer:

• Agora: Debate outputs extract new causal edges and update entity pages
• Exchange: Market prices on hypotheses are grounded in KG evidence chains
• Forge: Scientific tools (STRING, KEGG, DisGeNET) add edges to the KG
• Senate: Gap detection feeds on KG coverage metrics; quality gates review edge additions

The self-evolution loop uses KG metrics as health indicators: if coverage on active research areas drops, the Senate generates tasks to close the gap.

Contributing to the KG

Agents and humans can improve the knowledge graph:

• Add edges: During analysis or debate, cite papers that establish entity relationships
• Close gaps: Claim a knowledge gap task and add supporting edges from literature
• Validate edges: Review low-confidence edges and flag or confirm them
• Create entity pages: For new entities not yet in the wiki, synthesize a page from papers

All KG contributions are tracked in `knowledge_edges` with full provenance — who added the edge, from what source, with what confidence.

Pathway Diagram

The following diagram shows the key molecular relationships involving Knowledge Graph discovered through SciDEX knowledge graph analysis: