disPathCuration

Finding disease names in pathway labels.

Motivation

Open Targets derives target-disease evidence from Reactome. That evidence depends on manual curation of which pathways belong to which disease.

This project tests one hypothesis: disease pathways are undercurated. Many pathways and biological processes carry a disease name in their label, yet produce no corresponding target-disease evidence. If the hypothesis holds, the missing links can be recovered automatically from data that Open Targets already distributes.

The hypothesis is untested. Step 4 below is what decides it.

Approach

1. Diseases. Take disease names and synonyms from the Open Targets disease ontology (EFO/MONDO).
2. Pathways and processes. Take pathway and process labels from the Open Targets Platform facets file.
3. Match. Search pathway labels for chunks of disease names, yielding candidate disease-pathway links. The matching strategy is not yet chosen. Exact substring, normalised n-grams, and fuzzy string methods each need evaluating.
4. Compare. Score the extracted set against current Open Targets Reactome target-disease evidence, separating links that are already curated from links that are absent.
5. Pipeline. Package the extraction so candidates can be regenerated when the ontology or the facets file changes.

Pipeline

flowchart TD
    subgraph inputs["Open Targets Platform, release 26.06"]
        DIS["disease dataset<br/>EFO/MONDO"]
        TGT["target dataset<br/>gene symbols"]
        RXL["reactome dataset<br/>pathway labels"]
        EVR["evidence_reactome<br/>curated links"]
    end

    subgraph index["Disease chunk index (build_disease_index.py)"]
        LD["load_diseases<br/>drop measurement EFO_0001444<br/>drop process GO_0008150"]
        SF["surface_forms<br/>name + exact synonyms, normalised"]
        CH["chunks<br/>specificity, coverage, sharing"]
        LD --> SF --> CH
    end
    DIS --> LD

    subgraph matcher["Matcher (run_matcher.py), precedence template &gt; whole_form &gt; alias"]
        T["template<br/>Defective GENE causes DISEASE"]
        W["whole_form<br/>full name or synonym span"]
        A["alias<br/>distinctive token, gene-symbol filtered"]
    end
    CH --> T
    CH --> W
    CH --> A
    TGT --> A
    RXL --> T
    RXL --> W
    RXL --> A
    T --> M["matches<br/>pathway-disease pairs"]
    W --> M
    A --> M

    subgraph eval["Evaluation (evaluate.py), evidence as recall floor"]
        R["recovery<br/>exact / more-specific / more-general"]
        NR["non-recovery reasons"]
        NV["new candidate links"]
    end
    M --> R
    M --> NR
    M --> NV
    EVR --> R
    EVR --> NR
    EVR --> NV

    M --> J["LLM judge<br/>annotate low-confidence tier"]
    J --> OUT["outputs (CSV):<br/>new candidates + specificity fixes"]
    NV --> OUT
    R --> OUT

Loading

The template extractor and the evidence comparison are Reactome-specific; the chunk index, whole_form, alias and the judge are source-agnostic. Design notes and the release-automation checklist are in docs/pipeline.md.

Verification and outputs

An optional LLM judge removes context false positives that the deterministic matcher cannot see: hypoxia in "Cellular response to hypoxia", dependence in "Dependence Receptors". It annotates the low-confidence tier (alias and low-specificity whole_form) with a keep or reject verdict; it never deletes rows and never sees the structurally certain template matches. On release 26.06 it rejected 13 of 151 candidates.

The judge needs no ANTHROPIC_API_KEY: by default it calls the Claude Code CLI in print mode, using the existing Claude Code login. See docs/pipeline.md.

python scripts/run_matcher.py --source reactome   # matches_reactome.csv, new_candidates_reactome.csv, ...
python scripts/run_judge.py   --source reactome   # matches_judged_reactome.csv with keep/reject verdicts

Sources: Reactome and GO

The pipeline runs over any label source through one abstraction (sources.py). A source declares where its labels come from, whether the Reactome-only template extractor applies, where gene membership lives, and whether a curated disease baseline exists.

	Reactome	GO biological_process
Labels	reactome dataset	go, filtered to non-obsolete biological_process
template extractor	yes	no (GO has no "causes" labels)
Gene membership	target.pathways	target.go (aspect P)
Curated baseline	evidence_reactome	none exists
Matches (26.06)	351	165
Judge rejection rate	6/151 (4%)	33/77 (43%)

GO describes processes, not disease pathways: only 0.3% of its labels carry a disease cue, so it yields fewer and noisier matches, and there is no baseline to compute recovery against. The high judge rejection rate (plant-type hypersensitive response, cardiac muscle hypertrophy) shows the LLM judge is essential for GO, not optional. Run either source:

python scripts/run_matcher.py --source go
python scripts/run_judge.py   --source go

Unified output

After running the matcher and judge for each source, combine.py merges them into one table:

python scripts/combine.py --sources reactome go   # disease_pathways_all.csv

Every row carries a source column (Reactome_26.06, GO_26.06), the judge verdict, and a matchedSynonym column giving the original disease surface form that fired. The last explains otherwise puzzling matches: the GO label type I hypersensitivity matches the disease IgE responsiveness, atopic because type I hypersensitivity is one of that disease's exact synonyms.

Data sources

Release 26.06 is pinned. The latest symlink currently points at a run whose manifest reports a build failure, so it is not used.

Input	Source
Disease names and synonyms	disease dataset (EFO/MONDO)
Pathway and process labels	reactome dataset
Comparison baseline	evidence_reactome dataset

The facet search index is not published to the FTP in any release from 24.09 onwards, so pathway labels are taken from the reactome dataset the index is built from.

Disease chunk index

The first stage parses diseases and turns them into searchable chunks. Whole disease names rarely appear verbatim in a pathway label (EFO names average 5.4 tokens), so matching has to happen on sub-spans, and most sub-spans are worthless: type, 2 and disease occur across thousands of diseases.

Design decisions, each grounded in the data:

• Measurement terms are excluded. 25,049 of the 47,080 disease rows sit under the EFO measurement therapeutic area (GWAS traits such as "Increased circulating ACTH level"). They are not diseases, and their vocabulary would dominate any index. This leaves 22,031 diseases.
• Exact synonyms only. The primary name plus exact synonyms. Narrow, broad and related synonyms are held back as the main source of false matches.
• Abbreviations are matched case sensitively. Acronyms such as PDS are routed to a separate whole-token dictionary so they do not collide with ordinary words.
• Chunks are scored on two axes. specificity is the summed inverse document frequency of a chunk's tokens; coverage is the share of its surface form's IDF mass the chunk carries. Both are needed: hyperargininemia (a whole disease name) and cause (a fragment of "death by undetermined cause") share the same specificity, and only coverage separates them.

Build it with:

python scripts/build_disease_index.py

This writes disease_surface_forms.parquet and disease_chunks.parquet under data/<release>/. Full detail is in docs/chunking.md.

Matching

Three complementary extractors turn chunks into pathway-disease matches, in order of precedence:

• template: the disease slot of a "causes" label (Defective ABCC6 causes PXE), which licenses acronyms the structure proves are diseases.
• whole_form: a complete disease name or exact synonym found as a token span. Fragments are never matched.
• alias: a distinctive single disease token (HIV), hard filtered against gene symbols, generic terms and short acronyms. The lowest-confidence method, flagged for review.

Global abbreviation matching is deliberately excluded: it recovered 10 curated links while adding 61 acronym collisions. Full detail, filters and examples are in docs/matching.md.

python scripts/run_matcher.py

Recovery against curated evidence

The Reactome evidence is treated as a recall floor, not a ceiling: the project premise is that it is incomplete. Of its 653 curated pathway-disease pairs, the matcher recovers 122 (18.7%), of which 28 resolve a more specific disease than the curator recorded. The low figure is a property of the evidence, not the method: 450 of the not-recovered pairs sit on pathways whose label does not name the disease at all, so no label-based method can reach them.

The matcher also produces two actionable outputs: 263 new candidate links absent from the evidence, and the specificity improvements where the label names a subtype the curation recorded only as a parent.

Approach status

Step	State
1. Parse diseases into scored chunks	done
2. Load pathway and process labels	done
3. Match chunks against labels	done
4. Compare against evidence, per reason	done
5. Package as a pipeline	partial

Development

pip install -e '.[dev]'
pytest

Licence

MIT. See LICENSE.