Hi all,
I would like to share an open project that relies heavily on Open Targets data, and to thank the team for it.
OREVID (https://orevid.org) ranks existing drugs for each disease on PrimeKG extended with Open Targets 26.06 (associations, mechanisms of action, clinical records) and STRING, and shows the evidence behind every candidate. It is non-profit and open source (MIT code; data CC BY-NC-SA 4.0 because of some PrimeKG sources).
Open Targets plays three roles:
- Features: gene–disease associations by evidence type, and ChEMBL targets and mechanisms of action (including their direction).
- A temporal test: dated clinical records let me train only on what was known before 2015 and check the pairs that later reached phase 3 or approval (pre-registered): Recall@100 0.32 vs 0.17 for a popularity baseline.
- An external exam: drug–disease approvals in Open Targets 26.06 that are not in PrimeKG, which no model saw during training (OREVID 0.36, TxGNN 0.12, popularity 0.13).
«Known drug» evidence and literature «treats» relations are never used as features, since they are the answer.
While building it I reported two data issues, Approved indications derived from multi-drug labels and abbreviations (26.06) · Issue #4579 · opentargets/issues · GitHub (approvals derived from multi-drug labels and abbreviations) and Multiple sclerosis merged into "multiple sclerosis symptom measurement" in 26.09 · Issue #4573 · opentargets/issues · GitHub (disease merges in 26.09); thanks to David Ochoa for the quick answers.
- Preprint: https://doi.org/10.20944/preprints202610.1918.v2
- Code: GitHub - orevid-org/orevid: Open, evidence-based drug repurposing: a biomedical knowledge graph that ranks existing drugs for each disease and shows the evidence behind every candidate. Research use only, not medical advice. · GitHub
- Benchmark (OREVID-bench, with TxGNN retrained on the same splits): orevid/orevid-bench · Datasets at Hugging Face
Feedback is very welcome, especially on how the Open Targets data are used.
Joaquim