OPEN TARGETS
Evidence platform
- Integrates target–disease evidence
- Optimizes discovery and evidence access
- Exposes rich source payloads
- Provides a strong biomedical foundation
Architecture decision · July 2026
Why not just
Open Targets?
Because a large biomedical evidence platform is not automatically a biologically faithful, causal or machine-learning-ready graph. Jouvence-Graph starts from Open Targets, then changes the scientific contract.
THE THESIS
The graph is not a pile of links. It is a typed claim about what biology permits a model to learn.
01
The foundation
Open Targets is excellent.
It is not the whole model.
Open Targets gives Jouvence-Graph a high-value core: genetics, target–disease evidence, variants, drugs, pharmacogenomics, interactions and harmonized identifiers. Reusing that work is the rational choice.
But Open Targets is built to aggregate and serve evidence for target identification. Jouvence-Graph has a different downstream objective: construct training adjacency, features and supervision for heterogeneous graph models, while preserving exactly what each source observed. That requires saying no to some available links and adding biological layers that Open Targets was never designed to make ML-ready as one graph.
JOUVENCE-GRAPH
Predictive graph contract
- Selects biologically valid message-passing adjacency
- Separates observations, context and labels
- Preserves endpoint level and causal qualifiers
- Exports auditable Parquet and bounded PyG surfaces
02
Biological relevance
Model what was measured.
Not what an ID mapping makes imaginable.
The most consequential decision in Jouvence-Graph is source-native modeling. Normalization may change an identifier; it must not silently change the biological level of an assertion.
Genes are not proteins
A gene-level association stays gene-level. A protein edge requires a direct protein or isoform endpoint. We do not turn every gene xref into protein evidence.
Prediction is not observation
A motif hit is not observed TF binding. Coordinate overlap is not regulation. Population frequency is not pathogenicity.
Context is not adjacency
Papers, datasets, trial text and dense metadata remain evidence, features or catalog objects unless they express a defensible biological relation.
Direction matters
RNA expression cannot prove that a protein or isoform exists. Direct protein observation can support a carefully derived gene-expression claim, with the derivation retained.
03
Causal relevance
Causality is carried by evidence and typed edge features.
A relation name alone cannot tell a model whether inhibiting a target should help or harm a disease. Jouvence-Graph keeps stable biological relations, then attaches source-backed mechanism, sign, action and context.
drug action×disease mechanism×effect direction→candidate benefit or harm
Causally disciplined, not a fully causal KG.Broad disease associations can still be statistical, biomarker, expression or model-organism evidence. Signed treatment inference remains fail-closed until the required direction and mechanism features are materially present and independently reviewed.
One assertion, many evidence rows
The edge is deduplicated. Assays, studies, populations, scores, releases and conflicts remain separately auditable.
Conflict is data
Mechanism and direction aggregate to single, consensus, conflicting or unknown. Required signs fail closed on the last two.
Inference stays inferred
Approved path rules generate candidates in separate inferred surfaces. They never overwrite or masquerade as observed source assertions.
Leakage beats convenience
Trial outcomes, circular disease evidence and downstream consequences are excluded from training inputs when they could reveal the target label.
04
Why not Neo4j today?
A graph database stores a graph.
It does not make the graph scientifically correct.
For Jouvence-Graph today, Neo4j would duplicate the canonical data without solving the hardest problems: endpoint semantics, evidence multiplicity, leakage, reproducible promotion and graph-ML export.
Our dominant workloads are columnar
Versioned bulk reads, anti-joins, row-level audits, hashes, relation exports and feature assembly map naturally to Parquet, DuckDB and PyArrow.
Evidence can dwarf topology
Materializing every source record as a graph object would multiply storage and operational cost across more than 100M audited edge rows.
ML still needs another export
TxGNN and PyG consume tensors, node maps, features and split contracts. A Neo4j database would still need a deterministic export back into those representations.
A second truth would drift
A live editable database introduces schema migrations, indexes, backups, access control and synchronization. Canonical Parquet artifacts remain cheaper to version, review and reproduce.
The precise position
Neo4j is a useful view.
Not the vault.
We would use a generated, read-only Neo4j projection for scientist-friendly neighborhood browsing, Cypher queries, demos and bounded subgraph extraction. Every projection should be reproducible from canonical Parquet, carry a manifest and never become an independently edited source of truth.
Canonical Parquet→validated export→optional Neo4j view
The resulting architecture
Open Targets as a foundation.
Biology as the constraint.
Evidence as the memory.
Parquet as the truth.
That combination makes Jouvence-Graph less convenient to build than a generic property graph. It also makes every training edge easier to defend.