Linked Parquet
Stable entity IDs, accepted cross-references, source record IDs and evidence remain joinable.
Open biomedical infrastructure for agents + ML
A 100M+ edge biomedical graph, designed for ML and agents.
Jouvence-Graph is built for agents and predictive graph ML. It combines an Open Targets-derived core with ontologies, molecular interactions, expression, regulatory genomics, pharmacology and clinical evidence — while keeping source identifiers and provenance attached.
Open the viewer preview
Full database: local install
Explore the notebooks
Access the full database locally
View the code
- 100M+ audited edge rows
- 112 documented Parquet datasets
- 15 node families
- Open code
A source-aware path
01 Assertions are compact. Evidence stays rich.
The data product
One graph.
Three machine interfaces.
The canonical graph is stored as linked Parquet datasets, exported relation-by-relation for PyG, and paired with a fail-closed inference layer that composes only approved biological relations. Agents can inspect the same contracts, manifests and evidence used by predictive models.
PyG sidecars
Relation-wise edge indices, node maps and manifests support bounded PyG loading and training smokes. Full-scale multi-relation training and biological performance are not yet claimed.
Explicit path rules
Versioned, approved relation chains generate candidates while retaining support paths and leakage constraints.
Agent-readable
Documentation, schemas and safety contracts tell agents what to query, train and never assume.
What you can do
A graph to read,
audit and learn from.
Explore an entity and its neighborhood
Start from a disease, gene, protein or molecule and follow typed biomedical relationships.
Trace assertions back to evidence
Inspect the provenance, scores, studies, assays and conflicts behind a graph assertion.
Build machine-learning representations
Export heterogeneous subgraphs to PyG and attach text, molecule, protein or sequence embeddings.
Test therapeutic hypotheses
Train graph models and test zero-shot links with TxGNN-compatible APIs, PyG exports and split-aware leakage controls.
What is inside
A broad biomedical core — with status made explicit.
Jouvence-Graph absorbs the useful graph-shaped layers of Open Targets and extends them with ontology, interaction, expression, regulatory, pharmacology and clinical sources. Files that are metadata-only, leakage-prone, semantically invalid as topology or not yet independently accepted stay out of the default training graph. The code is MIT-licensed; upstream data licenses still apply, and requester-pays cloud access may incur infrastructure charges.
Canonical now
Validated graph + sidecars
Open Targets-derived variants, genetics, interactions, pharmacogenomics and drug–target/disease layers; GO, HPO and UBERON; BioGRID protein interactions; HPA tissue protein expression; Cellosaurus tissue/species provenance; ClinicalTrials.gov trial text and treatment-link metadata.
Reviewed or staged
Built, but not all default topology
ReMap TF–enhancer support, Cell Ontology hierarchy and tissue placement, Reactome protein-native membership, ChEMBL protein-native targets, protein-native disease associations, DepMap essentiality, GDSC response, cell-line proteomics and additional Cellosaurus mappings.
Not canonical yet
Requires a source-native contract
Direct miRBase/miRTarBase regulation, Complex Portal membership and structured UniProt/BioGRID PTM events. These are not advertised as finished until endpoint, license, evidence and leakage gates pass.
What is deliberately excluded from default topology?Open Targets payloads that are metadata rather than biological adjacency; paper/dataset inventory edges; unsafe gene→protein projections; generic phenotype/pathway Cartesian joins; trial outcomes that leak treatment labels; and any inferred candidate without an approved composition rule.
Architecture choices
Provenance is not a footnote.
SOURCE → ASSERTION → EVIDENCE
Source-native semantics
Relationships follow the assertion and endpoint types actually provided by each source. Source-specific details remain in evidence instead of exploding the graph taxonomy.
COMPATIBILITY
TxGNN, without breaking compatibility
The txgnn namespace and the TxData, TxGNN and TxEval classes remain the public compatibility boundaries.
REPRODUCIBILITY
Executable notebooks
A fixture-backed suite lets you explore the concepts without reading the full graph or depending on private infrastructure.
INTEROPERABILITY
Linked Parquet · PyG
Every logical dataset is cataloged; reference identifiers and source records remain in nodes, evidence or metadata sidecars instead of being discarded during export.
GUARDED SCALE
Heavy workloads are intentionally explicit
Local reads remain bounded. Full scans, synchronization and production exports are routed to suitable infrastructure with validation and independent promotion gates.
Where to go
Choose your entry point.
01↗
02↗
03→
04↗
05↗
06↗
07↗
Install
Clone the repository, synchronize the environment with uv and load the API.
Explore
Work through the user-facing notebook suite.
Why this architecture?
Why Open Targets is a foundation, biology constrains the graph and Neo4j remains an optional view.
Understand
KG architecture, provenance, exports, embeddings and runbooks.
Reproduce
Use the scientific reproduction notebooks and construction pipelines.
Contribute with an agent
Start with context, safety and reviewability rules.
Read the upstream project
Find the original TxGNN implementation and publication.
Agent-ready
Give your agent the right context.
Not a vague “set up the project.”
These prompts establish the right entry points and prevent an agent from accidentally launching heavy workloads on your machine.
Clone https://github.com/jkobject/jouvence-graph.git and enter the repository. Read AGENTS.md before taking any action. Use uv, verify Python 3.11+, then install the local environment with the dev and notebooks groups. Run only the bounded, fixture-backed checks described in the README. Do not launch a full KG scan, bulk LaminDB sync, production export or heavy training job on this machine. Finish with a summary of the commands you ran, their actual results and any optional access still required.In the Jouvence-Graph repository, read AGENTS.md, README.md and notebooks/README.md. Prepare and run the public notebook suite in fixture mode, without accessing the live KG. Then explain the node / assertion / evidence / feature model, show me where to start when exploring a disease or molecule, and clearly distinguish local demonstration, bounded live access and production processing.You are contributing to Jouvence-Graph. Start with AGENTS.md, then read only the documentation specific to the task. Before changing anything, inspect Git state and isolate pre-existing changes. Respect source-native relationships and the assertion/evidence separation. Use uv for tests. Canonical writes and heavy workloads require explicit authorization and the designated infrastructure. Provide a reviewable diff, the tests actually executed and the remaining validation limits.Predict, but keep the derivation