Clinical trial capacity · 41 diseases · GBD 2021 burden

If AI hands us a thousand drugs, can we test them?

Three things can stop a drug pipeline: the serial clock, the patient pool, and the site network — and which one binds depends almost entirely on the disease. Pick a disease, pick where you can run trials, and set how strong the molecules are. The model runs a year-by-year queue against a pool that regenerates only as fast as people fall ill.

Disease

Ranked by global DALYs. Success rates, trial durations and endpoint type follow the therapeutic area.

Indication

Candidates entering phase I 500 The AI-discovery volume, for this one indication.

Effect size vs. state of the art 1.0×

Where trials can run

Each region contributes patients in proportion to its share of this disease's burden, and trial slots in proportion to its share of global trial starts. Both ceilings bind, and the mismatch between them is the whole story for diseases like tuberculosis and malaria.

Site network vs. today 1×

Participation

Decomposed the way Unger's meta-analyses measure it: a patient enrols only if a trial is available, they are eligible, and they agree. Defaults reproduce the observed rate for this disease.

Trial available to the patient 44% Unger 2019: 55.6% of patients had no trial open at their institution — the single largest barrier.

Meets eligibility criteria 78% 21.5% of those with a trial available were excluded by criteria.

Agrees when offered 55% Unger 2021: patients accept 55.0% of the time. Refusal is not the bottleneck.

Resulting participation rate —

Phase success rates

Four transitions, from BIO / Informa / QLS, 12,728 phase transitions over 2011–2020.

Phase I → II —

Phase II → III —

Phase III → filing —

Filing → approval —

Starting conditions

Year 0 is the moment the programme starts. Trials already recruiting are competing for the same patients, so the new programme does not get the whole flow on day one. Nothing is ever recovered when those trials end — see below.

Flow already claimed by trials now recruiting 70% Share of this year's enrollable patients already going to trials that started before year 0.

Those trials stop recruiting within 4y

Site-network growth 3%/yr New sites, staff and coordinators each year. Grows the site side only — it cannot make more people fall ill, so it does nothing where patients already bind.

Network growth ceiling 2.0× Trained investigators, coordinators and monitors are finite.

Usable pool

—

enrolments / yr

Enrolments needed

—

all phases

Approvals

—

of 500 candidates

First approval

—

years from start

Last approval

—

years from start

DALYs averted

—

per year, at steady state

Deaths averted

—

per year, at steady state

Capacity occupancy over time

Patients under study each year. The ceiling moves: legacy trials release capacity, growth adds it.

Phase load

Phase	Trials	n each	Draws from	Patients	Duration	Peak load

Same settings, every disease

Everything except the disease is held fixed — including candidate count, so this is a like-for-like test of where capacity binds. Each row uses its own area's success rates, endpoint and durations; if you drag a rate slider by hand it applies to the selected disease only.

Table view

Disease	Area	Global DALYs	Usable pool/yr	Phase III n	Years to clear	Approvals	DALYs averted/yr	Binding constraint