VivaBench 3.0

VivaBench scores drug-repurposing models only on regulatory approvals that happened after the date the model's data stops (its knowledge cutoff). Each system ranks its whole drug list for a disease using only data from before the approval; we then ask whether the drug that was actually approved is among its top 30 picks.

Result

On 48 new FDA/EMA approvals (2025–2026), BioAtlas found the approved drug in its top 30 for 37.5% of cases vs 2.1% for TxGNN; on 85 subtype- and biomarker-defined approvals, 35.3% vs 0%.

These use BioAtlas with trial-registration data. The pre-registered primary comparison, without trial data: 15.3% vs 0% (85 approvals, 95% CI +8.3 to +24.3 points).

BioAtlas is VivaMed's ranker; TxGNN (Huang et al., Nature Medicine 2024) is a published drug-repurposing AI.

Why trust it. Every approval is dated after all the data either system read. The analysis plan was fixed and pushed to a private repository before scoring; the plan and its SHA-256 fingerprint are published here. The test cases were scored exactly once. The cases, BioAtlas's ranks and an offline re-scorer are in this repository.

Results · Why it matters · How it works · Quick start · Limitations · Citation

Results

Two configurations of BioAtlas are scored. BioAtlas reads disease genes, drug mechanisms and earlier approvals; it is the pre-registered primary system. BioAtlas + trial data also reads clinical trials registered before the cutoff, and is always secondary. Clinical-trial registrations are a legitimate early signal that a drug is heading for approval; the primary comparison shows what BioAtlas finds without them.

Comparison	Cases	BioAtlas	TxGNN	Difference, points (95% CI)
Primary	85 subtype- and biomarker-defined approvals (2024–2026), own drug lists	15.3% (13/85)	0.0% (0/85)	+15.3 (8.3 to 24.3)
Secondary, with trial data	the same 85, own drug lists	35.3% (30/85)	0.0% (0/85)	+35.3 (22.9 to 61.8)
Secondary, with trial data	48 new FDA/EMA indication approvals (2025–2026), own drug lists	37.5% (18/48)	2.1% (1/48)	+35.4 (20.0 to 51.0)
Secondary, without trial data	the same 48, own drug lists	12.5% (6/48)	2.1% (1/48)	+10.4 (−2.1 to 23.4): includes zero
Secondary, same drug list	40 of the 85 cases, both systems ranking the same 2,822 drugs	22.5% (9/40)	2.5% (1/40)	+20.0 (7.1 to 33.3)
Secondary, same drug list	the same, without the 23 cases that overlap BioAtlas's tuning data	15.0% (3/20)	5.0% (1/20)	+10.0 (0.0 to 17.2): not a clear difference

The 95% intervals resample whole disease families, so related diseases are not counted twice. Every other comparison, including those that do not favour BioAtlas, is in the white paper and in results.json.

Why it matters

TxGNN's released zero-shot AUROC of 0.91 measures how often it ranks a correct drug above one sampled wrong drug, drawn from drugs that have a training indication. Asked to rank all 7,957 drugs it knows, the same released results place only 40% of correct drugs in its top 100. On TxGNN's own test split, a no-learning baseline that copies the drugs of the 3 most similar diseases beats it (Recall@100 0.69 vs 0.40); BioAtlas uses a related signal (paper §6). On real approvals after its training data, TxGNN put the approved drug in its top 30 in 1 of 48 and 0 of 85 cases. VivaBench measures the task a drug-discovery team actually faces: given what was known on a date, which drug should we look at first?

How it works

1. Dated cases. The 48 new approvals come from FDA approval letters and EMA decisions, each dated by that event. The 85 subtype- and biomarker-defined approvals come from FDA labels, FDA biomarker tables and the CIViC cancer-variant database; 23 of their dates are upper bounds, and the rule below uses the earliest proven date.
2. No look-ahead. For each case, a system uses the latest data snapshot whose cutoff is on or before the approval date, and every input holds only data from before that cutoff. BioAtlas's data ends on 31 December 2023 or 31 December 2024; TxGNN's released checkpoint is dated 19 March 2023, an upper bound on its data. All 133 scored cases were checked against this rule before scoring: 0 violations.
3. Pre-registered. The analysis plan (provenance/VB3_PREREG_2026-10.json, SHA-256 90a6c78425501d7a6a2c0ea5eb39907f9bd952d6059858f67d1ffd02627458bf) was committed and pushed to VivaMed's private GitHub repository at 2026-10-03 17:40:40 UTC, one minute before the single scoring run; the repository is private, so the published hash is what outsiders can check.
4. Scored once. The test cases were fixed by fingerprint (SHA-256 hash) and scored exactly once; nothing was re-run or tuned afterwards.
5. Same rules for both systems. Each system ranks its own full drug list (BioAtlas + trial data 7,047 drugs, BioAtlas 4,901, TxGNN 7,957). Drugs with equal scores are treated the same way for both systems, and a case a system cannot answer counts as a miss.

The systems. BioAtlas, VivaMed's ranker, matches drugs by mechanism to disease genes (curated sources, genes that interact with the disease genes, human genetics) and adds drugs already approved for closely related diseases. BioAtlas + trial data also uses late-stage clinical trials registered before the cutoff. TxGNN (Huang et al., Nature Medicine, 2024) is a published AI model (a graph neural network) for drug repurposing.

Quick start

from datasets import load_dataset

cases = load_dataset("vivamed/Vivabench", "cases", split="train")
print(cases[0])

To re-score BioAtlas from the shipped files (Python standard library only):

huggingface-cli download vivamed/Vivabench --repo-type dataset --local-dir vivabench
python3 vivabench/score_offline.py vivabench   # prints OFFLINE SCORE PASS or FAIL

It recomputes the BioAtlas rates on the own-list and same-list rows; the remaining rows are reported from results.json. TxGNN's per-case predictions are not redistributed, so its rates can be checked only by re-running TxGNN.

Files

Path	Contents
data/cases.csv	All 135 cases (133 scored; 2 subtype/biomarker cases had no usable data snapshot): set (track A = new approvals, B = subtype/biomarker approvals), disease (MONDO id), approval date, drug (DrugBank id, FDA application number where available); a flag marks new-approval rows where an AI model located the indication sentence in the letter
data/bioatlas_ranks.csv	Where BioAtlas ranked the approved drug, per case
data/case_ids/	Which cases enter each comparison
results.json	The complete output of the scoring run
score_offline.py	Re-scores BioAtlas from the files above
paper/	White paper (PDF) and its LaTeX source
figures/	The figures on this page (PNG, PDF and the data behind each)
croissant.json	Machine-readable dataset description (Croissant 1.0)
provenance/	The analysis plan and its timestamp receipt, and the records used to build and check this release

Limitations

• Trial data drive much of the lead on the 48 new approvals. Without them, BioAtlas scores 12.5% (6/48) against TxGNN's 2.1%, a difference whose 95% interval (−2.1 to 23.4 points) includes zero. A drug in late-stage trials for a disease is a strong hint that approval is coming.
• Fresher data. TxGNN's checkpoint is dated 19 March 2023, while BioAtlas reads snapshots that end on 31 December 2023 or 2024, always before the approval. 7 of BioAtlas's 13 primary hits were drugs that its own data already listed for the same disease before the cutoff (the new approval was a new biomarker-defined patient group or a new product). Newer inputs are an advantage of the system, not proof of better ranking.
• Small samples, wide intervals. 48 and 85 cases give the wide intervals shown above. Secondary comparisons are reported without correction for multiple comparisons.
• Labels were checked by AI review, not clinicians. In a sample of 50 of the 87 subtype/biomarker labels (not uniformly random: all 9 exact-subtype cases were included), three automated reviewers, all the same AI model (the two blind reviewers agreed with kappa 0.34), disputed 8 (16.0%, 95% CI 8.3 to 28.5%), all at disease level. Build-time checks found 4 of 40 new-approval labels and 2 of 40 subtype/biomarker labels in error. No label was changed after scoring.
• Different drug lists. Hitting the top 30 is easier in a shorter list. On the same drug list the gap remains (48 new approvals with trial data: 37.2% vs 5.2%, n = 39; 85 subtype/biomarker approvals: 22.5% vs 2.5%, n = 40). In 43 of the 85 cases the approved drug is not in TxGNN's drug list, so TxGNN cannot rank it; of the 42 cases it could rank, it placed none in its top 30. Removing the 23 cases that share a drug–disease pair with BioAtlas's tuning data leaves 11.3% vs 0.0% on each system's own list (n = 62), and 3 of 20 vs 1 of 20 on the same list, where the interval (0.0 to 17.2 points) reaches zero.
• Not every case is a subtype. 2 of the 85 primary cases are approvals for a whole disease rather than a biomarker-defined group. Both systems rank per disease; TxGNN answers a subtype with its parent disease's list.
• Same authors. VivaMed built both the benchmark and BioAtlas. The protection is procedural: the plan was fixed before scoring, the test cases were sealed, and they were scored once.

Citation

@misc{vivamed2026vivabench,
  title        = {VivaBench: drug--indication predictions scored on approvals made after the models' knowledge cutoff},
  author       = {{VivaMed}},
  year         = {2026},
  version      = {3.0},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/vivamed/Vivabench}}
}

See also CITATION.cff. An earlier release (v2.1) was withdrawn after we found an evaluation error; details are in provenance/.

Licence and sources

CC BY 4.0. Case labels come from U.S. FDA records (public domain; FDA does not endorse this dataset), the European Medicines Agency (reused with acknowledgement: EMA is the source of that material), CIViC (CC0), DrugBank Open Data identifiers via UniChem and Wikidata (CC0), HGNC symbols, DailyMed/openFDA labels (U.S. National Library of Medicine; this dataset does not reflect the most current data available from NLM), and the Mondo Disease Ontology (CC BY 4.0). Full attributions are in the white paper.