Published September 15, 2026
| Version 0.1
Dataset
Open
Parkinson's Disease Ontology (sciai-lab)
Authors/Creators
Description
# An Ontology That Grows From Evidence
*SciAI is building a Parkinson's-disease ontology with an unusual property:
its instance layer is machine-produced at corpus scale — 942,000 graded,
provenanced assertions from 100,000 papers — instead of hand-curated.*
## The problem with ontologies
Most published biomedical ontologies are, in practice, glorified taxonomies.
What distinguishes a true ontology is the instance layer: individuals linked
to classes and relations, so that inference over existing statements becomes
possible. Building that layer by hand is enormous work that was almost never
financed — the Gene Ontology is the rare counterexample, sustained by decades
of funded curation.
## Our particularity: instantiation is industrialized
Every paper passing the SciAI extraction pipeline produces ontology instances
as a side effect: 81 structured field values per paper, plus typed
relationship assertions (`X —induces→ oxidative damage`), each carrying
confidence, effect direction, provenance (PMID, text section), and a context
snippet. The core of the ontology **is** the extraction schema: its typed
properties are the ontology's property vocabulary, and schema versions *are*
ontology versions.
The power sits at the bottom. The domain layer on top is deliberately thin —
eight rough domains derived from the structure the extraction schema actually
grew into — with vocabulary anchors (genes → HGNC, proteins → UniProt,
chemicals → MeSH/ChEBI) for cross-topic linkage.
## Uncertainty is data, not noise
Biomedical knowledge has wobble effects; that cannot be out-ontologized. Our
assertion model is graded by design. Class-level knowledge is *induced* from
instance volume: an edge type with 60% positive and 40% negative instances is
a **contested claim** — surfaced as such — not a fact. A claim's value *is*
its instance distribution:
- **Settled** — large instance count, direction ratio far from 50/50
- **Contested** — large count, split near 50/50 (a first-class object)
- **Sparse** — small count, flagged as under-evidenced, not wrong
## PD-first, transferable spine
Depth in one disease over breadth across biomedicine. The type spine
(document & evidence, clinical, genetic, preclinical, mechanism,
computational, synthesis) is generic enough to transfer to other diseases;
the branches grow where the PD data demands them. One domain — the PD
Mechanism Core — is the disease-specific hypothesis space: standing questions
about named mechanisms (alpha-synuclein, mitochondria, inflammation, …),
answered per paper.
## The refinement loop
extract → normalize → graph → curation findings → ontology refinement → new
schema version → re-extract → better graph → deeper findings.
Curation is expert-adjudicated and machine-pre-digested. The first empirical
map produced by this loop is already visible in the field fill rates: they
approximate where PD research effort actually sits — and where the controlled
vocabularies of the field still fail, which tells us what to fix next.
---
*Internal working form is JSON/JSON-LD; a community-facing export is planned
when v1.0 stabilizes. The instance layer (942,333 assertions, 100,835 papers,
state of 2026-09-07) grows with every extraction cycle.*
Files
README.md
Files
(64.6 MB)
Additional details
Dates
- Available
-
2026-09-15v0.1
Software
- Development Status
- Active