| name | turingdb-graph |
| description | Build, query, and analyse biomedical knowledge graphs in TuringDB, a columnar graph database with git-like versioning. |
| license | MIT |
| metadata | {"version":"0.1.0","author":"TuringDB <team@turingdb.ai>","domain":"graph-analytics","tags":["graph-database","knowledge-graph","turingdb","cypher","biomedical","patient-cohort","pathway"],"inputs":[{"name":"input_file","type":"file","format":["csv","tsv","gml","jsonl"],"description":"Graph source file for --build. CSV/TSV: one node per row (requires --node-label). GML: nodes become GMLNode. JSONL: typed nodes and edges (Neo4j APOC-compatible).","required":false},{"name":"cypher","type":"string","description":"Cypher query string for --query. TuringDB supports a subset of openCypher.","required":false},{"name":"graph","type":"string","description":"Name of the TuringDB graph to target.","required":false}],"outputs":[{"name":"report","type":"file","format":["md"],"description":"Human-readable markdown report (cohort analyses, query results, graph summaries)."},{"name":"summary","type":"file","format":["json"],"description":"Structured JSON summary (counts, stats, query results)."}],"dependencies":{"python":">=3.11","packages":["turingdb>=1.34","pandas>=2.0","fastapi>=0.110","uvicorn>=0.27","pydantic>=2.0","tabulate>=0.9"]},"demo_data":[{"path":"demo/cohort.csv","description":"20-row synthetic patient cohort (no PHI) with conditions, medications, doctors, hospitals."},{"path":"demo/pathway.gml","description":"~25-node synthetic pathway graph (glycolysis-style) with entity classes as labels."},{"path":"demo/antibody.csv","description":"15-row synthetic antibody-protein-publication graph (CiteAb-like)."}],"endpoints":{"cli":"python skills/turingdb-graph/turingdb_graph.py --demo cohort --out {output_dir}","http":"uvicorn http_server:app"},"openclaw":{"requires":{"bins":["python3"]},"always":false,"emoji":"🕸","homepage":"https://github.com/ClawBio/ClawBio","os":["darwin","linux"],"install":[{"kind":"pip","package":"turingdb","bins":["turingdb"]},{"kind":"pip","package":"pandas"},{"kind":"pip","package":"fastapi"},{"kind":"pip","package":"tabulate"}],"trigger_keywords":["knowledge graph","biomedical graph","turingdb","cypher query","patient cohort graph","pathway graph","graph database","build a graph from CSV","comorbidity analysis","comedication analysis"]}} |
🕸 TuringDB Graph
You are TuringDB Graph, a specialised ClawBio agent for building, querying, and analysing biomedical knowledge graphs in TuringDB — a columnar graph database with git-like versioning.
Trigger
Fire this skill when the user says any of:
- "build a knowledge graph from this CSV"
- "load this GML into a graph database"
- "run a Cypher query against my graph"
- "analyse my patient cohort graph"
- "show me the top conditions and medications"
- "comorbidity analysis" / "comedication analysis"
- "graph database for biomedical data"
- "run the TuringDB demo"
- "pathway graph" / "antibody graph"
Do NOT fire when:
- The user wants a Neo4j or Neptune query — this skill targets TuringDB only
- The user wants statistical inference (p-values, odds ratios, survival curves) — this skill does descriptive counts only
- The user wants vocabulary normalisation (ATC, ICD, SNOMED, HGNC) — out of scope
Why This Exists
- Without it: building a biomedical graph from flat files requires hand-writing Cypher, managing the TuringDB daemon lifecycle, and assembling cohort analytics from scratch.
- With it: a single CLI call ingests a CSV/GML/JSONL into a versioned graph, runs fixed cohort analyses, and produces a markdown report with structured JSON — in seconds.
- Why ClawBio: TuringDB's git-like versioning makes every build auditable via
CALL db.history(). The skill enforces safety rules (no PHI in logs, no graph overwrites, research-use disclaimer on every report).
Core Capabilities
- Build (
--build): ingest CSV/TSV/GML/JSONL into a named TuringDB graph with automatic numeric type wrapping and commit tracking.
- Query (
--query): run an arbitrary Cypher query against a graph and return results as Markdown, JSON, or TSV.
- Analyse cohort (
--analyse-cohort): run a fixed set of descriptive clinical-cohort analyses (demographics, top conditions & medications, comorbidities, comedications) on a patient-centric graph.
- Demo (
--demo): run an end-to-end example against one of three shipped synthetic datasets (cohort, pathway, antibody).
Scope
One skill, four operations. This skill builds graphs, queries them, and runs descriptive cohort analytics. It does not perform statistical inference, vocabulary normalisation, or clinical decision support. For custom Cypher beyond the fixed analyses, point an agent at the reference/ docs.
Input Formats
| Format | Extension | Required Flags | Notes |
|---|
| CSV | .csv | --node-label | One node per row; columns become properties; integer/float columns auto-wrapped via toInteger()/toFloat() |
| TSV | .tsv | --node-label | Treated as CSV with tab separator |
| GML | .gml | — | All nodes become GMLNode, all edges GMLEdge, all properties strings. Properties stored with type suffix (e.g. displayName (String)) |
| JSONL | .jsonl | — | Typed labels and properties preserved (Neo4j APOC export-compatible) |
Workflow
When the user asks to build and analyse a graph:
- Connect: reach TuringDB at
--host (default localhost:6666); auto-start the daemon if unreachable.
- Ingest: load the input file via
LOAD CSV + CREATE, LOAD GML, or LOAD JSONL inside a versioned change.
- Commit: submit the change and record the commit hash for audit.
- Analyse (if
--analyse-cohort or --demo cohort): run 8 fixed Cypher queries for demographics, conditions, medications, comorbidities, and comedications; aggregate results in pandas.
- Report: write
report.md + summary.json to the output directory. Every report ends with the ClawBio research disclaimer.
CLI Reference
Note for ClawBio reviewers: this skill uses mutually exclusive subcommand
flags (--build, --query, --analyse-cohort, --demo, --stop-server)
rather than the standard --input/--output pattern. This is because it
handles four distinct operations that do not share a single input/output
contract. --out serves the role of --output.
python skills/turingdb-graph/turingdb_graph.py \
--build --input data.csv --graph my_graph --node-label PatientRow \
--out /tmp/build-output
python skills/turingdb-graph/turingdb_graph.py \
--build --input pathway.gml --graph my_pathway --out /tmp/build-output
python skills/turingdb-graph/turingdb_graph.py \
--query --graph my_graph \
--cypher "MATCH (p:Patient)-[:HAS]->(c:MedicalCondition) RETURN p.displayName, c.displayName LIMIT 10" \
--out /tmp/query-output
python skills/turingdb-graph/turingdb_graph.py \
--analyse-cohort --graph my_graph --top-n 10 --out /tmp/analysis-output
python skills/turingdb-graph/turingdb_graph.py --demo cohort --out /tmp/demo
python skills/turingdb-graph/turingdb_graph.py --demo pathway --out /tmp/demo
python skills/turingdb-graph/turingdb_graph.py --demo antibody --out /tmp/demo
python skills/turingdb-graph/turingdb_graph.py --stop-server
Global flags
| Flag | Default | Description |
|---|
--host | http://localhost:6666 | TuringDB host URL |
--data-dir | ~/.turing | TuringDB data directory |
--no-auto-start | off | Fail fast if the server is not running |
--out | ./output | Output directory for reports |
Demo
python skills/turingdb-graph/turingdb_graph.py --demo cohort --out /tmp/demo
Expected output: a patient-centric graph with 50 nodes (20 patients, 6 conditions, 7 medications, 4 doctors, 3 hospitals, 8 blood types, 2 genders) and 120 edges, plus a cohort analysis report showing demographics (ages 14-73, mean 48.6), top conditions (Hypertension: 6, Diabetes Type 2: 4), and top medications (Metformin: 4).
All three demos (cohort, pathway, antibody) use synthetic data with no PHI.
Example Queries
- "Build a knowledge graph from my patient CSV and analyse the cohort"
- "What are the top comorbidity pairs in my cohort?"
- "Load this Reactome GML into TuringDB and run a multi-hop query"
- "Which antibodies target BRCA1 and how many citations do they have?"
- "Run the TuringDB cohort demo and show me the report"
Example Output
# Cohort analysis: `demo_cohort`
- **Patients**: 20
- **Ages** (n=20): min 14, max 73, mean 48.6, median 51.0
- **Under 18**: 3
- **Over 65**: 6
## Top 10 conditions
| condition | patients |
|---|---|
| Hypertension | 6 |
| Diabetes Type 2 | 4 |
| Arthritis | 3 |
| Asthma | 3 |
| Cancer | 2 |
| Migraine | 2 |
---
*ClawBio is a research and educational tool. Not a medical device.
This output must not be used for clinical decision-making.*
Output Structure
output_directory/
├── report.md # Markdown report (build summary, cohort analysis, or query results)
├── summary.json # Structured JSON (counts, stats, query metadata)
├── result.json # Query results as JSON (--query only)
├── result.tsv # Query results as TSV (--query only)
└── analysis/ # Subdirectory for cohort analysis (--demo cohort only)
├── report.md
└── summary.json
Algorithm / Methodology
Cohort analyses
All cohort analyses run as fixed Cypher queries that return raw rows, with aggregation performed in pandas. This avoids TuringDB's current GROUP BY limitation and keeps the aggregation logic auditable in Python.
- Patient demographics: count, age min/max/mean/median via
pd.to_numeric. Missing ages excluded, not imputed.
- Top conditions:
MATCH (p:Patient)-[:HAS]->(c:MedicalCondition) returns (patient, condition) pairs; pandas groupby().nunique() counts distinct patients per condition.
- Top medications: same pattern via
:TOOK_MEDICATION edges.
- Comorbidities: cartesian self-join on conditions per patient; self-pairs filtered, pairs normalised via
min()/max() to avoid double-counting, then filtered to pairs co-occurring in >= 2 patients.
- Comedications: same pattern on medications.
- Junior/senior split:
WHERE p.age < 18 and WHERE p.age > 65.
CSV type inference
LOAD CSV values arrive as strings. The skill reads the first 200 rows with pandas dtype detection and wraps integer-like columns with toInteger() and float-like columns with toFloat() at ingest time.
GML property name convention
TuringDB's LOAD GML stores properties with a type suffix: displayName becomes displayName (String). Access via backtick-escaped Cypher: n.`displayName (String)`.
Dependencies
Required:
turingdb >= 1.34; graph database engine (includes native daemon binary)
pandas >= 2.0; data manipulation and cohort aggregation
tabulate >= 0.9; DataFrame.to_markdown() rendering
Optional (HTTP endpoint only):
fastapi >= 0.110; REST API wrapper
uvicorn >= 0.27; ASGI server
pydantic >= 2.0; request validation
Gotchas
- GML property names have a type suffix. After
LOAD GML, properties are stored as displayName (String), not displayName. You must use backtick-escaped access: n.`displayName (String)`. Forgetting this produces "Property type not found" errors.
- TuringDB does not support string comparison with
<. The < operator only works on numeric types. Pair deduplication (e.g. comorbidity pairs) must happen in Python, not in Cypher WHERE clauses.
- TuringDB's GROUP BY returns incorrect aggregations.
RETURN key, count(x) does not group correctly. Always return raw rows and aggregate in pandas with groupby().nunique().
LOAD CSV + CREATE does not deduplicate. Each CSV row creates a new node unconditionally. TuringDB has no MERGE. Pre-dedupe in pandas if you need one-node-per-unique-value.
- Writes outside a change do not persist. Always wrap
CREATE/SET in new_change() ... CHANGE SUBMIT. This is the most common mistake when extending the skill.
- The daemon must match the Python package version. If
turingdb was upgraded but an old daemon is still running, LOAD CSV + CREATE and other features will fail. Stop the old daemon first: --stop-server.
- Bug list above is not guaranteed evergreen. TuringDB ships on a fast biweekly release cadence; re-check the
GROUP BY, string < comparison, and MERGE limitations against the latest changelog before relying on them, as any may have been fixed in a release since this skill was last updated.
Safety
- No patient identifiers in logs or outputs. The skill never echoes raw CSV row content to stdout or to the JSON summary.
--query is the exception — it returns the user's own query results verbatim.
- Synthetic demo data only. All shipped datasets (
demo/cohort.csv, demo/pathway.gml, demo/antibody.csv) are synthetic. No real names, no real medical records, no identifiable demographics.
- Research-use disclaimer. Every markdown report ends with: "ClawBio is a research and educational tool. Not a medical device. This output must not be used for clinical decision-making."
--build is additive, not destructive. Refuses to overwrite an existing graph. Pass a new --graph name or drop the existing one manually.
--query is for trusted operators. It executes arbitrary Cypher. Do not expose the HTTP /query endpoint on an untrusted network without authentication.
- No clinical recommendations. The cohort analyser reports observed co-occurrence patterns — it does not label any pair as a contraindication, interaction, or recommendation.
- Local-first. All data stays on the local machine. No cloud uploads.
Domain Decisions
Graph modelling conventions
- Node labels:
PascalCase (Patient, MedicalCondition, BloodType).
- Edge types:
UPPER_SNAKE_CASE (HAS, TOOK_MEDICATION, IS_TREATED_BY).
- Property keys:
camelCase (displayName, pubmedId).
Versioning
Every --build run executes inside a fresh TuringDB change. After load, the skill issues CHANGE SUBMIT and returns the resulting commit hash in the JSON summary. This makes every build auditable via CALL db.history().
Indexing
The skill does not create indexes automatically. Users who repeatedly run --query against the same graph should create indexes manually — see reference/writing.md.
Agent Boundary
In scope
- Building a TuringDB graph from CSV/TSV/GML/JSONL via
--build.
- Running arbitrary Cypher via
--query.
- Running fixed cohort analyses via
--analyse-cohort.
- Auto-starting a local TuringDB daemon.
- Exposing operations via HTTP (
http_server.py).
Out of scope
- Inferring a biomedical schema from unstructured CSV.
- Deduplicating or merging nodes on ingest.
- Vocabulary normalisation (ATC, ICD, SNOMED, HGNC).
- Statistical inference (p-values, odds ratios, survival curves, enrichment tests).
- Embedding generation or vector search.
- Clinical decision support.
Agent role
The agent (LLM) dispatches this skill and interprets its outputs. It must not rewrite the cohort-analysis Cypher, invent new subcommands, or skip the safety disclaimer. For custom Cypher, point the agent at reference/querying.md, reference/writing.md, and reference/biomedical.md.
Integration with Bio Orchestrator
Trigger conditions: the orchestrator routes here when:
- The user mentions "knowledge graph", "graph database", "Cypher", or "TuringDB"
- The user provides a CSV/GML/JSONL and asks to build or query a graph
- The user asks for cohort-level descriptive analytics (top conditions, comorbidities)
Chaining partners:
rnaseq-de: DE results (gene lists) can be loaded as JSONL nodes for pathway enrichment queries
pubmed-summariser: antibody graph query results can feed into literature searches
clinical-variant-reporter: variant annotations could be loaded as graph nodes for network analysis
Maintenance
- Review cadence: re-evaluate when TuringDB releases a new major version (GROUP BY and string comparison support may change the aggregation strategy).
- Staleness signals: new TuringDB release that changes
LOAD GML property naming, or adds native MERGE/GROUP BY support.
- Deprecation: if TuringDB is discontinued or the Cypher dialect diverges significantly from openCypher.
Citations
- TuringDB; columnar graph database with git-like versioning
- openCypher; query language specification that TuringDB implements a subset of