Foundational Research: Ontology-Native AI
An open, conformant language model for biomedical knowledge graphs
To our knowledge the first openly published language model fine-tuned for the Biolink Model and GO-CAM, the open standards that make biomedical knowledge interoperable. It turns a grounded biomedical statement into a schema-conformant knowledge graph, and it is released as a research prototype with its dataset and evaluation harness so the field can check the work.
Term conformance
0 → 100%
vs the untuned base model
Hallucinated terms
0.42 → 0.00
eliminated
Unseen graph shapes
0 → 89%
holds out of distribution
Why biology ontology conformance
Biomedical knowledge is scattered across hundreds of incompatible databases. The open ontologies — the Biolink Model (the schema behind the Monarch Initiative and the NCATS Biomedical Data Translator), the Gene Ontology and GO-CAM, and the OBO Foundry — are the interoperability backbone that makes that knowledge FAIR: findable, accessible, interoperable, reusable.
Language models are becoming the default interface to that knowledge. Off the shelf, they hallucinate ontology terms: asked to write a Biolink graph, a strong 30B code model invents roughly 42% of the terms it uses and produces a fully conformant graph 0% of the time. In a knowledge graph that is worse than no help at all. A single fabricated predicate or non-resolvable identifier silently breaks interoperability, poisons every downstream query, and is indistinguishable from a real edge until something fails.
What we built
The core idea is to never trust the model to invent structure. Instead we generate the structure correctly from real ontology releases, then teach the model to reproduce it. Every training target was built programmatically with the Biolink Model Toolkit and the GO-CAM data model, and validated twice before training.
| Component | What it is |
|---|---|
| Three conformant tasks | Biolink knowledge-graph construction, GO-CAM causal activity models, and OBO term grounding (GO, MONDO, HP, ChEBI, Cell Ontology) — the structuring, mechanism and normalisation work every biomedical data programme needs. |
| Correct-by-construction data | Every training target is generated programmatically from real ontology releases with the Biolink Model Toolkit and the GO-CAM data model, never hand-authored, so the gold standard is valid by construction. |
| Real, verified entities | Identifiers drawn from real OBO releases — GO (38,245), MONDO (32,095), HP (19,836), ChEBI (218,253), Cell Ontology (3,335) — and gene/protein IDs verified against mygene.info. No invented CURIEs enter the data. |
| Double validation | Each example passes both the official toolkit (bmt category/predicate membership and subject/object range closure; the GO-CAM schema) and an independent membership check before training. |
| Generalisation holdout | 90 multi-association graphs held out entirely — a composition the model never trains on — to test conformance on genuinely unseen structure. |
| Open release | Model, LoRA adapter, the full dataset and the generators and validators that produced every number above, published for independent reproduction. |
The results
Measured on held-out prompts against the untuned base model. The point of the fine-tune is not fluency, which the base model already has, but conformance: using real ontology terms, in structurally valid ways, and doing so on graph shapes it never saw in training.
| Metric | Base model | Fine-tuned | Meaning |
|---|---|---|---|
| Biolink term conformance | 0% | 100% | every term used exists in the Biolink Model |
| Hallucinated-term rate | 0.42 | 0.00 | fraction of biolink: terms that are invented |
| Out-of-distribution conformance | 0% | 89% | multi-association graphs never seen in training |
| Structural conformance | n/a | 1.000 | subject and object categories satisfy the association ranges |
| GO-CAM schema validity | 0% | 100% | output validates against the GO-CAM causal-model schema |
| OBO grounding accuracy | 0% | 100% | correct ontology identifier chosen from candidates |
The out-of-distribution number is the honest one to watch: 89% term conformance on multi-association graphs the model never trained on, a composition strictly harder than anything in its training set. We report it, and the model's weaker properties, openly in the model card rather than hide them.
The conformance gate is the validator, not the model
The right architecture for standards work is neuro-symbolic: the model drafts, and a symbolic layer verifies. Conformance here is not hoped for, it is enforced by a validator in the loop that rejects 100% of out-of-ontology terms — invalid predicates, non-existent CURIEs, categories that violate the schema. The model proposes; the validator disposes.
This division of labour is also where we are honest about the model's limits. On its own it does not reliably refuse a malformed request — asked to encode a nonsensical relationship, it will still try to structure it. That does not compromise the guarantee, because safety lives in the symbolic validator that catches every invalid term, not in the model's self-restraint. A language model should never be the last line of defence in front of a scientific knowledge graph; the validator is.
Where it fits
This is the biology counterpart of our open language model for IES4 defence data: the same correct-by-construction method, pointed at a new family of standards. Both pair a fine-tuned drafter with the open-ontologies validation engine that checks conformance deterministically. The language model removes the blank-page problem; the validators guarantee correctness. Neither is asked to do the other's job.
The practical uses are the unglamorous, load-bearing ones: harmonising heterogeneous multi-omic and clinical data into one FAIR knowledge graph, normalising free-text mentions to the correct ontology identifier, and building the mechanism graphs that large biomedical programmes query at scale. It is a structuring and interoperability tool, not a source of biomedical truth — every output is meant to be validated before use.
"A general model asked to write a niche standard will confidently make it up. The fix is not a bigger model, it is correct-by-construction data and a validator in the loop. Publish both, with the weak spots named, and the field gets a tool it can actually build on."
Fabio Rovai, Tesseract Academy
This is an independent, self-initiated demonstration built entirely on published open standards and open tools. It is not affiliated with, or endorsed by, the Monarch Initiative, the Gene Ontology Consortium, the OBO Foundry or the Biolink Model project. Model weights and dataset are Apache-2.0; the ontology terms remain under their source licences. Associations in the training data are schema-correct but synthetic, not curated biological assertions.
Model, dataset and evaluation harness
Weights and LoRA adapter, the full training data, and the code that produced every number above.
