OREVID-bench: an honest benchmark for drug repurposing on PrimeKG

Which existing drugs could treat this disease? OREVID-bench checks how well a model ranks the ~8,000 drugs of PrimeKG for each disease, on the same knowledge graph and the same splits as TxGNN (Huang et al., Nature Medicine 2024), with three things published evaluations usually miss:

1. A popularity baseline. Ranking drugs by how many diseases they already treat, ignoring the disease, reaches Recall@100 0.36 on TxGNN's zero-shot split (TxGNN: 0.41). Every model should be compared with it.
2. Ranking over all drugs, not pairs. Recall@100 and Precision@K over all 7,957 candidate drugs, which is what a researcher looks at, besides TxGNN's AUPRC.
3. Hard diseases and novelty. ~60 % of the zero-shot test diseases have a near-identical sibling in the ontology with the same drugs; a hard subset removes them. Novelty subsets drop blockbuster drugs or keep only drugs with no prior use, and an external task uses approvals that are not in PrimeKG at all (Open Targets 26.06).

Code, documentation and the full discussion: https://github.com/orevid-org/orevid (benchmark/). Preprint: doi:10.20944/preprints202610.1918.v2. Website built on the same model: orevid.org.

Results (Recall@100, mean over diseases)

Model (config)	Zero-shot	Zero-shot, hard	Disease areas	External (not in PrimeKG)
Popularity, ignores the disease (popularity)	0.358	0.203	0.185	0.130
TxGNN, official code, trained on the same splits (txgnn)	0.412	0.258	0.237	0.121
RotatE, knowledge-graph embedding with PyKEEN (kge_rotate)	0.289	0.162	0.219	0.156
OREVID, graph only: logistic regression, no popularity (orevid_no_popularity)	0.781	0.501	0.591	0.364
OREVID, LambdaRank over meta-paths (orevid_lambdarank)	0.811	0.539	0.617	0.359
OREVID, training labels only, as TxGNN (orevid_lambdarank_train_only)	0.803	0.530	0.619	—

orevid_lambdarank is the production model of the website. The best value of each column is in bold.

TxGNN was retrained with its official code and hyper-parameters; on the zero-shot split it reproduces its published numbers (Recall@100 0.412 vs 0.405; AUPRC 0.910 vs 0.904). Full tables, with Precision@K, AUPRC, the novelty subsets and disease-level confidence intervals: LEADERBOARD.md and the GitHub repository.

What is in this repository

Path	Content
tasks/<task>.json	Candidate drugs (PrimeKG node_index), test diseases with their positive drugs, training popularity
tasks/<task>.split.npz	Row numbers of the validation and test edges in PrimeKG's kg.csv (everything else is training)
disease_areas/	Disease lists of the nine areas (from the TxGNN repository, MIT)
predictions/<model>/<task>.parquet	Scores of every candidate drug for every test disease: columns disease, drug (PrimeKG node_index) and score (higher = more likely to treat)
predictions/<model>/results.json	The metrics bench.py evaluate computes from those files
bench.py	Builds the tasks from PrimeKG and evaluates prediction files (MIT licence)

Tasks. zero_shot_1 … zero_shot_5: TxGNN's complex_disease split reproduced exactly (5 % of the diseases lose all their drug–disease edges). area_*: all drug–disease edges of one branch of the disease ontology are hidden (nine areas, seed 0). external_ot2606: drug–disease pairs approved according to Open Targets 26.06 (ChEMBL) that are not in PrimeKG, trained on zero_shot_1. Each task is a split of every model config in the dataset viewer.

How to use it

from datasets import load_dataset

# scores of TxGNN on the first zero-shot split
preds = load_dataset("orevid/orevid-bench", "txgnn", split="zero_shot_1")

To evaluate a new model: download PrimeKG as used by TxGNN (Harvard Dataverse doi:10.7910/DVN/IXA7BM, files kg.csv and node.csv), train on each task's training edges, write one parquet per task with the columns above, and run

python bench.py evaluate --pred predictions/my_model

The training edges of a task are everything except the rows in tasks/<task>.split.npz (bench.train_edges(load_kg("kg.csv"), "tasks", "zero_shot_1")). Do not use anything that encodes drug–disease therapeutic relations for the test diseases (clinical-trial registries, drug labels, literature saying «drug treats disease»). Python ≥ 3.10 with numpy, pandas, scipy, scikit-learn and pyarrow.

Caveats

• Labels come from PrimeKG/DrugBank and Open Targets: they reward rediscovering known treatments, and a candidate counted as wrong may be right but not yet approved.
• external_ot2606 is out-of-sample but not prospective (median first trial 2008).
• TxGNN's own disease-area splits cannot be reproduced (no seed); these are an approximation with a fixed seed, and diabetes (2 diseases) and adrenal gland (6) are noisy.
• TxGNN outputs a constant score for test diseases with no edge left in the training graph (8 of 335 zero-shot diseases); its ranking there is random. Both protocol differences with TxGNN are measured in benchmark/FAIR_COMPARISON.md of the GitHub repository.

Research use only. These are hypotheses for researchers, not medical advice.

Licence

Task files and predictions: CC BY-NC-SA 4.0, because they are adaptations of sources with non-commercial and share-alike terms (DrugBank, SIDER and DisGeNET through PrimeKG); see LICENSE-DATA.md in the GitHub repository. bench.py: MIT. Please cite OREVID-bench and the original sources, in particular PrimeKG and TxGNN.

Citation

@article{quadrada2026orevid,
  author  = {Quadrada Jan{\'e}, Joaquim},
  title   = {{OREVID}: Open, Evidence-Based Drug Repurposing with a Benchmark Against {TxGNN}},
  journal = {Preprints.org},
  year    = {2026},
  doi     = {10.20944/preprints202610.1918.v2}
}

Code and data archive: doi:10.5281/zenodo.23323226. Contact: contact@orevid.org · issues: https://github.com/orevid-org/orevid/issues