Janssen sited COVID vaccine trials with MIT's DELPHI model and finished 6 weeks early

In the spring of 2020, Janssen, the pharmaceutical arm of Johnson & Johnson, faced a scheduling problem no pharma company had faced at this scale. Its COVID-19 vaccine trial was 'event-driven': efficacy could only be measured by counting infections among volunteers, so the trial would finish faster if it ran where transmission was about to spike. But site choices had to be locked in four months before the trials began in September 2020 — exactly the horizon where guessing is worst.

Najat Khan, Janssen's chief data science officer, recruited Dimitris Bertsimas of MIT, whose DELPHI model — a compartmental SEIR model extended with nonlinear, location-specific infection dynamics — was issuing daily infection and mortality forecasts for 120 countries and every US state. The two teams met daily through the summer, feeding Janssen's global surveillance dashboard into the model and refining it with local policy and mask-compliance signals.

The joint team picked trial sites against the forecasts rather than the current case maps. The bet paid off visibly: sites predicted to be hot spots — including in South Africa and Brazil — saw extremely high case counts once the trial began, right as the Beta variant emerged there. That gave Janssen something almost no competing trial had: clinical performance data on the variants that were about to drive the next wave, and one of the most diverse trial populations on record.

The trial concluded roughly six weeks ahead of its original timeline, and INFORMS's review of the project found the model had placed sites in future high-incidence areas with 90 percent accuracy. The vaccine received the US FDA's emergency use authorization on February 27, 2021 — the first single-dose COVID-19 vaccine authorized in the US.

Place trial sites where the epidemic will be, not is

Event-driven studies run on the clock of the phenomenon they measure: the trial's endpoint counted infections, so location choice directly set its speed.

Site selection is irreversible months in advance — leases, staff and ethics approvals lock in long before the epidemic arrives, so the decision is a forecast, whether or not you make one.

DELPHI was deliberately simple (an extended SEIR model) but tuned on real-world data streams, which made it fast to update daily and hard to break.

Forecasts were public on the COVIDAnalytics platform, so the team could benchmark accuracy continuously instead of trusting their own narrative.

Hunting diversity explicitly turned a speed hack into a scientific advantage: South African and Brazilian data made the vaccine legible for the variant era.

Waiting on an uncertain outcome? Forecast it and put your resources where it will occur, not where it last did. Publishing the forecast lets anyone check it and hold you to it.

Janssen's ENSEMBLE trial reported topline results in late January 2021, and the FDA granted emergency use authorization on February 27, 2021 — the first single-dose COVID-19 vaccine in the US. The trial's variant data from South Africa and Brazil shaped how the field read emerging variants. INFORMS chronicled the collaboration and made it a finalist for the 2022 Franz Edelman Award, operations research's leading prize. Bertsimas's group kept DELPHI and its daily forecasts public on the COVIDAnalytics platform through the pandemic.