Clinical trials
Drug programs unfold over years. A world model can learn that path from the clinical and regulatory record, then forecast what comes next for a program it has never seen.
It can ask:
- What is the next regulatory gate?
- When is the next material event likely to happen?
- Which programs look unusual compared with similar histories?
- Where is withdrawal pressure building before it becomes obvious?
Clinical Trial Oversight
Clinical Trial Oversight learns from the dated ClinicalTrials.gov registry history. The current AACT 2.0 serving family combines trial state, sponsor history and current facts for six- and twelve-month outcome forecasts. The serving tools turn those forecasts into per-trial probabilities, attention lists and expected counts.
Research note: Clinical Trial Oversight, with parts on risk ranking and conditional scenarios
Results: In the September 27 development read, six-month AUROC was 0.819 for enrollment shortfall, 0.832 for termination and 0.806 for completion, with calibration evaluated separately. Conditional scenarios let a trial team supply a possible move and compare the forecast with the same future date with nothing filed. A scenario shows how the model's forecast responds to a supplied change; it is not a measured effect of making that change.
Watch it work: Explore a trial registry's event stream
In the Factory: Clinical trial registry streams · members only
Earlier work: Project Confirm Oncology
Project Confirm turns approvals, trials, submissions and labels into a changing model of the FDA oncology world. Whole drug programs were held out of training, so the model had to work zero-shot on assets it had never seen.
Across 17,703 events and 3,275 held-out windows, it reached 0.857 AUROC on the next regulatory gate. The withdrawal readout reached 0.754 AUROC on 587 windows, including 20 positives. Those positive windows preceded withdrawal by a median of about 24 months; this measures window timing, not warning performance. In the Imbruvica example, the January 2020 cutoff preceded the later withdrawal by 40 months.
Earlier scored results · The detailed report · review access
Where this goes
The FDA record is one world. Clinical development, trial operations, safety, manufacturing and commercial uptake create others. The aim is a model that can follow a program across all of them: not a static risk score, but a living view of what changed, what is likely to happen next, and which questions are worth asking now.