Pathways Knowledge Graph

119K nodes. 835K edges. Every human biological pathway, protein interaction, and GO annotation in one graph.

Part of the Samyama ecosystem — loaded into and queried via the graph engine at samyama-ai/samyama-graph. This repo holds the loader and source-data specifics for the KG.

---

We loaded Reactome pathways, STRING protein interactions, Gene Ontology, WikiPathways, and UniProt, then asked:

"Which protein has the most interaction partners?"

MATCH (p:Protein)-[:INTERACTS_WITH]-(other:Protein)
RETURN p.name AS protein, count(DISTINCT other) AS partners
ORDER BY partners DESC LIMIT 5

Protein	Partners
TP53	571
UBC	524
EGFR	441
APP	395
ESR1	384

TP53 -- the most connected hub in the human interactome. Powered by Samyama Graph.

See all 100 benchmark queries →

---

Documentation

New here? Start with the guides:

Guide	What it covers
GETTING_STARTED.md	prerequisites (Python ≥ 3.10) · install · run the engine (Docker) · load the graph · first query
docs/QUERYING.md	ask questions via MCP (Claude), the HTTP API, or the Samyama CLI
Biomedical Benchmark	100 example queries

---

Demo

A narrated terminal walkthrough — loads a real Reactome subset and answers four pathway-biology questions in plain Cypher (hierarchy, protein hub-centrality, protein-dense pathways).

The demo loads a fast, representative subset of REAL Reactome data: all 2,848 human pathways and their CHILD_OF hierarchy, plus the first 40,000 human protein-participation rows from UniProt2Reactome (2,745 proteins, 35,160 PARTICIPATES_IN edges). Edges are created one-by-one, so this caps the ~900K-row participation file to keep the demo under a minute. Load the full KG with python -m etl.loader --data-dir data.

# Run
source ~/projects/venv/bin/activate
PYTHONUNBUFFERED=1 python -m demo.demo

# Re-record
asciinema rec --overwrite --cols 92 --rows 32 --idle-time-limit 2.0 \
  -c "bash -c 'source ~/projects/venv/bin/activate && PYTHONUNBUFFERED=1 python -m demo.demo'" \
  demo/pathways.cast
agg demo/pathways.cast demo/pathways.gif
aws s3 cp demo/pathways.gif s3://samyama-data/demos/pathways.gif

---

Schema

5 node labels -- Protein (37,990), GOTerm (51,897), Complex (15,963), Reaction (9,988), Pathway (2,848)

9 edge types -- ANNOTATED_WITH (265K), INTERACTS_WITH (228K), PARTICIPATES_IN (140K), CATALYZES (121K), IS_A (59K), COMPONENT_OF (8K), PART_OF (7K), REGULATES (3K), CHILD_OF (3K)

5 data sources -- Reactome, STRING v12.0, Gene Ontology, WikiPathways, UniProt (all human, organism 9606)

Quick Start

Full walkthrough → GETTING_STARTED.md (prerequisites, Docker, loading, querying).

Load from snapshot (recommended)

Needs Python ≥ 3.10 for the tooling and Docker for the engine:

pip install -r requirements.txt
docker run --rm -p 8080:8080 -p 6379:6379 public.ecr.aws/f9f6l5u4/samyama-graph:1.1.0

curl -LO https://github.com/samyama-ai/samyama-graph/releases/download/kg-snapshots-v3/pathways.sgsnap  # ~9.6 MB
curl -X POST http://localhost:8080/api/tenants -H 'Content-Type: application/json' -d '{"id":"pathways","name":"Pathways KG"}'
curl -X POST http://localhost:8080/api/tenants/pathways/snapshot/import -F "file=@pathways.sgsnap"

Build from source

git clone https://github.com/samyama-ai/pathways-kg.git && cd pathways-kg
pip install -r requirements.txt          # or: pip install -e ".[dev]" for tests
python -m etl.download_data --data-dir data        # ~1.9 GB
python -m etl.loader --data-dir data --url http://localhost:8080     # → pathways tenant

Example Queries

-- Pathway crosstalk: shared proteins between pathways
MATCH (p1:Pathway)<-[:PARTICIPATES_IN]-(prot:Protein)-[:PARTICIPATES_IN]->(p2:Pathway)
WHERE p1.name < p2.name
WITH p1, p2, count(prot) AS shared WHERE shared >= 100
RETURN p1.name, p2.name, shared ORDER BY shared DESC LIMIT 10

-- TP53 two-hop neighborhood
MATCH (tp53:Protein {name: 'TP53'})-[:INTERACTS_WITH]-(hop1:Protein)-[:INTERACTS_WITH]-(hop2:Protein)
WHERE hop2 <> tp53
RETURN DISTINCT hop2.name AS protein LIMIT 15

Use with Claude (MCP)

python -m mcp_server.server --url http://localhost:8080 --graph pathways   # against a running engine
python -m mcp_server.server --data-dir data                               # embedded, loads on startup
python -m mcp_server.server --url http://localhost:8080 --list-tools       # see all tools

Register it with Claude and ask in natural language — full steps in docs/QUERYING.md.

Part of the Biomedical Trifecta

This KG is one of three biomedical knowledge graphs that together form Samyama's billion-edge benchmark: Clinical Trials (27M edges) + Pathways (835K edges) + Drug Interactions (388K edges), merged on load with PubMed (1.04B edges).

These four graphs are also loaded as part of a larger merged corpus — fifteen KGs plus two synthetic patient cohorts in a single store. BiomedQA has the node, edge and merge counts, and states how and when they were measured.

Links

Samyama Graph	github.com/samyama-ai/samyama-graph
The Book	samyama-ai.github.io/samyama-graph-book
Benchmark (100 queries)	Biomedical Benchmark
Contact	samyama.dev/contact

License

Apache 2.0