GaiaLab logs each repurposing candidate with a prediction timestamp and matches it against ClinicalTrials.gov registration records. Matches where the prediction precedes the trial registration date are classified as prospective (forward prediction). All others are retrospective — the system independently surfaced a drug that investigators had already registered for study. Retrospective concordance is not forward prediction. Only entries with a confirmed NCT ID are shown in the top-match panel. Completed trial status indicates the trial finished enrollment; it does not indicate proven efficacy or regulatory approval.
—
Total concordance matches
…
—
Unique drug–disease pairs
—
Matched completed trials
—
Currently recruiting
— flagged·— advisory
Annotation flags
Note: Matched completed trials and Currently recruiting are trial-level counts — a multi-arm match can map to more than one trial/status, so these may sum to more than the total match count.
Semantic Forward Signals · exploratoryAI semantic match · drug present in ClinicalTrials.gov intervention list⛓ How is this verified? →
— semantic prospective signals
Each prediction below was generated by GaiaLab before a clinical trial testing that exact drug for that exact disease was registered. Discovered via pgvector semantic similarity (not exact name matching), then checked against the live ClinicalTrials.gov record to confirm the predicted drug appears in the trial's intervention list. Lead time = days the prediction preceded trial registration.
These are a looser, exploratory signal — not the strict exact-name test. They do not count toward the exact-name Forward-Skill Scorecard (currently 0 confirmed forward hits), and most (13 of 15) are on-label, where a new trial for an approved drug is expected rather than foresight.
Verification provenance: reproducible registry check (drug present in trial intervention list; trial registered after prediction date) performed by an AI-assisted review with operator sign-off — not independent expert peer review. Click any trial to inspect the source record; independent domain-expert re-review is recommended before citation.
Calibration Performance
Forward calibration: not yet demonstrated
The fair measure of forecast calibration is prospective ECE, which is pending until enough predictions precede trial registrations. Until then, confidence scores are not shown to be well-calibrated on forward predictions — read the reliability diagram below.
Brier —(vs 0.25 no-skill · N=— resolved). A low Brier under high ECE reflects class imbalance, not calibrated confidence — read them together, never the Brier alone.Full methodology →
Prediction Calibration — Reliability Diagram
Each bar shows the observed validation rate for predictions grouped by confidence decile. A well-calibrated system tracks the diagonal (green line). Bars above = overconfident; bars below = underconfident.
Epistemic Transparency: Prospective classification uses prediction timestamp vs. trial first-posted date. AI training data predates predictions; some 'prospective' associations may have appeared in preprints before trial registration. Prospective concordance is necessary but not sufficient for independent prediction.
◈
Weekly Co-Occurrence Audit
Prospective: prediction precedes trial registration · Retrospective: trial pre-existed the prediction · Completed ≠ proven efficacy · Featured matches require confirmed NCT ID
Analyzing predictions…
The Prospective Ledger — Building in Real Time
— predictions in the 30–90 day window.
A prospective match fires the first time a new trial is registered after we predicted the drug — and only when no earlier trial for that drug–disease pair already existed. Every GaiaLab prediction is timestamped at analysis time.
Exact name-matching: 0 prospective matches so far — and at this stage that is the expected result, not a failure.
The classifier is deliberately strict: if any trial for that exact drug–disease pair pre-dates the prediction, the match is retrospective. Most predictions here target heavily-studied oncology combinations (KRAS, BRAF, EGFR), where a trial almost always already exists — so they can only ever be retrospective. Genuine exact-name prospective hits depend on predictions in less-trodden spaces (off-label repurposing, neglected diseases) and are expected to be rare and irregular — a slow trickle over months, not a scheduled date. We will not loosen the matcher to make the number look better; a 0 we can trust is the point.
Separately, semantic matching has surfaced 15 AI-verified prospective matches (AI-assisted review + operator sign-off, not expert peer review) — but 13 are on-label (a new trial for an already-approved drug is expected), leaving only 2 genuinely off-label hypotheses. See the “Semantic Forward Signals” panel above — these are exploratory and do not count as exact-name forward hits. Check back weekly — or get notified below.
—
unmatched predictions
—
ripening (30–90 days)
—
no trial yet (90+ days)
—
oldest prediction (days)
🔒 Pre-registered & sealedthe forward-skill claim is falsifiable and un-gameable
On — we sealed a tamper-evident cohort of
— unresolved predictions plus an equal
—-pair shuffled placebo control
(sha256 —) — the set cannot be edited after the fact.
Committed readouts (pass or fail): —.
The claim is rejected at the 180-day readout if the off-label hit rate ≤ the placebo set's rate, or there are 0 off-label hits, or calibration fails to improve (ECE stays > 0.20). A rejected claim will be stated plainly — the tool's value as transparent triage does not depend on it succeeding.
Validation →
Forward-Skill Scorecard · the fair denominator
counted as distinct drug–disease pairs you predicted · last 12 months · retrospective = a trial already existed when the prediction was made
—
prospective-eligible predictions (no trial existed when predicted)
—
confirmed forward hits (exact-name, prospective)
—
retrospective (ineligible) (trial pre-existed — not a foresight test)
Open-Space Watchlist
top open-space predictions by GaiaLab repurposing score (≥50) — where a genuine forward hit can still land · unvetted computational hypotheses, not endorsements
I want to know when GaiaLab's first genuine forward prediction is confirmed.
Get one email when the first prospective match fires. No newsletter — just the signal.
Checked every 6 hours against ClinicalTrials.gov.
Top Verified Co-Occurrence Matches
NCT-CONFIRMED
Filter by Disease Category
⬡ High Confidence Co-occurrences
TARGET IN PANEL
Biological rationale — the drug's target or a key downstream mediator appears in the submitted gene panel. These are not coincidental co-occurrences (note: some links are indirect, e.g. a drug that modulates a cytokine rather than binding it).
🔭
Predictions to Watch
AWAITING A FIRST TRIAL
Unmatched predictions where a genuine forward hit can still land — ranked by GaiaLab's repurposing score (≥50), which filters out low-plausibility noise. A prospective match fires only when a trial for that exact drug–disease pair is registered after we predicted it and none existed beforehand; if one does, subscribers are notified. Unvetted computational hypotheses, not endorsements.
Prediction Age Distribution
Each bar = one day. ■ Ripening (30–90 days)
■ Overdue (>90 days, no match)
■ Young (<30 days). —