Every drug repurposing prediction GaiaLab generates is timestamped, assigned a confidence score, and automatically cross-referenced against ClinicalTrials.gov. This page is the unfiltered record.
Claims without calibration data are marketing. Here's ours. · Full calibration methodology →
Each analysis run adds to this dataset. As predictions accumulate and ClinicalTrials.gov checks complete, calibration curves populate here automatically. No other biological intelligence platform publishes this data continuously — that is the standard we hold ourselves to.
daaed957…); integrity independently re-verifiable.docs/PREREGISTRATION_drug-repurposing-2026-07-17.md · power & blinded-adjudication amendment (2026-07-25): docs/PROSPECTIVE_ADJUDICATION_AND_POWER.md (to be published as a new DOI version).| Drug | Disease Context | Confidence | Outcome | Trial NCT IDs | Recorded |
|---|---|---|---|---|---|
Every headline metric, with its scope and known limitation. Retrospective figures test correspondence with already-existing evidence; prospective figures — the only fair test of forward predictive ability — are pending. We publish these whether or not they flatter us.
| Benchmark | Dataset | Type | Metric | Known limitation | Readout |
|---|---|---|---|---|---|
| Drug repurposing — retrospective AUROC | N=529, 22 disease areas | Retrospective | 0.545 AUROC | Modest signal just above the 0.50 random baseline; not a clinical predictor. | Complete |
| Drug repurposing — temporal holdout AUROC | n=22 curated approvals, 8/8 neg. controls | Retrospective | 0.90 AUROC | Small curated set; a different, easier test than the retrospective AUROC above. Not a clinical predictor and not a prospective result. | Complete |
| Drug repurposing — prospective forward skill | Sealed cohort; live at /new-trials | Prospective | Pending | The genuine test. Pre-registered; 0 confirmed forward hits so far — reported honestly. | 2028 (drug lockbox) |
| Calibration — Brier score | N=608 resolved predictions | Retrospective | 0.08 (vs 0.25 no-skill) | Below 0.25 partly reflects class imbalance, not proven calibration — read WITH ECE (~0.40). | Live |
| LassaAI — national early-warning AUROC | 313 weeks / 6,456 cases (2020–2025), 2024 hold-out | Retrospective | 0.880 (vs 0.849 naive) | Beats a naive baseline by only ~0.03 — not statistically significant (DeLong p=0.61). Seasonality-driven. Preprint: Zenodo. | Prospective 2026–27 season |
Trial correspondence is not efficacy. A registered or completed ClinicalTrials.gov trial for a drug–disease direction indicates that direction is being pursued in research — it does not establish that the drug is an effective treatment, nor that GaiaLab predicted it before the trial existed.
How predictions are recorded: At the end of every analysis, up to 10 drug candidates are saved with their confidence score, disease context, target genes, and a timestamp. Records are immutable — no retroactive changes.
Trial-correspondence check: Each prediction is queried against ClinicalTrials.gov API v2 (clinicaltrials.gov/api/v2/studies) using the drug name + disease context. We use precise labels: completed-trial correspondence (a completed disease-matched trial exists — the internal status code is validated, but this means a matching trial was found, not that efficacy is proven); active-trial correspondence (an active/recruiting trial exists — direction pursued, outcome pending); and no trial correspondence identified (none found — counts against the correspondence rate). None of these establish treatment efficacy or forward-prediction skill.
What this is NOT: This does not measure whether GaiaLab identified the drug before the trial started (we don't have that date information). It measures whether the drug+disease direction is being/was pursued in a clinical setting — a proxy for research relevance, not therapeutic efficacy.
Calibration curve: A well-calibrated system shows higher-confidence predictions matching trials at higher rates than lower-confidence ones. This is research direction correspondence, not therapeutic outcome accuracy. A true efficacy calibration curve requires prospective trial completion data — that data will be added as it matures. We publish what we have, not what looks best.
Data source: GET /api/predictions · GET /api/predictions/calibration — public, no auth required.