QuantHealth has raised $45 million to expand an AI platform that simulates clinical trials before a single real patient is enrolled — a bid to cut the enormous waste in drug development.
The Series B was led by Qumra Capital, with Sanofi Ventures, Pitango HealthTech and others participating, bringing total funding to $70 million since the company was founded in 2020.
What it does
The system builds virtual simulations of trials, predicting how patients would respond to a treatment and flagging design changes — different endpoints, patient populations or protocols — that could raise the odds of success.
It pairs AI with large biomedical knowledge graphs and uses transfer learning to extend predictions to rare diseases and small patient groups where data are scarce.
The company says it has simulated more than 600 trials across 30 disease areas, with up to 90% accuracy in predicting patient outcomes — figures it reports itself.
Which failures this could actually prevent
Roughly 90% of clinical trials fail, and separating the causes clarifies what a simulation platform can address.
Some trials fail because the drug does not work — the target does not drive the disease, or the molecule does not engage it adequately in humans. No simulation trained on prior clinical data can reliably foresee that, because the information does not exist before the trial.
Others fail despite an effective drug, through design error: an endpoint too insensitive to detect the effect, a population too heterogeneous, a comparator too strong, a trial underpowered, a duration too short.
The second category is where simulation has a genuine claim. Those are questions about study design given assumed drug behaviour, and they can be modelled from historical data on how similar populations behaved in similar trials.
What transfer learning contributes
The rare-disease application addresses the field’s hardest data problem.
Machine learning generally requires substantial training data, and rare diseases have very little — few patients, few prior trials, sometimes none at all in the specific condition.
Transfer learning takes a model trained where data are abundant and adapts it to a related setting where they are not, on the premise that much of what was learned still applies. For clinical trials the assumption is that disease progression, placebo response and dropout behave in partly generalisable ways.
Whether that holds is genuinely uncertain, and rare diseases frequently differ from common ones in exactly the respects that matter.
How to read the 90% figure
“The clinical stage of drug development has remained largely untouched by AI innovation,” said CEO Orr Inbar.
The accuracy claim needs care for reasons beyond it being self-reported. Predictive accuracy depends entirely on how the prediction was framed and what counted as correct, and a model predicting the more likely outcome in a field where most trials fail achieves high accuracy by predicting failure.
What matters is discrimination — distinguishing the trials that will succeed from those that will not — and whether the predictions were made before outcomes were known or fitted retrospectively. Retrospective accuracy on completed trials is a much weaker claim than prospective performance.
The verification problem
Catching flawed designs early could spare patients from ineffective treatments and speed useful drugs along. Establishing that it does is unusually difficult.
The counterfactual is unobservable. If a company changes its trial design on the platform’s advice and the trial succeeds, nobody knows whether the original design would have failed — and if the platform advises against running a trial that is then not run, no outcome exists to check against.
That structural problem is why such platforms are difficult to evaluate from outside, and why participation from a pharmaceutical company’s venture arm is informative: those investors see internal validation that outsiders cannot.
What comes next
QuantHealth plans to use the money to improve its models, cover more diseases, and extend into drug positioning and commercialisation.
Where synthetic patients are already accepted
Simulated trial participants are not solely a startup proposition — a related idea has regulatory precedent worth distinguishing from this one.
External control arms use real patient data from prior trials or registries in place of a concurrent placebo group. Regulators have accepted them in specific circumstances: rare diseases where randomising patients to placebo is impractical or ethically difficult, and conditions with predictable natural history where historical outcomes are a defensible comparison.
That acceptance is narrow and hard-won. The objection is confounding — patients treated at a different time received different supportive care, were diagnosed by different criteria and were selected differently — so an apparent treatment effect may reflect the era rather than the drug.
Simulation of the kind described here is a further step. An external control arm uses real patients who really existed; a simulated arm uses modelled patients whose behaviour is inferred. The first is contested and sometimes accepted; the second is currently a planning tool rather than evidence, and the distinction matters when reading claims about what such platforms do.
That extension moves toward questions — which patients to target, how to differentiate against competitors — that are commercially valuable and further from verifiable prediction. As with any predictive platform, the real test is whether forecasts hold up against actual trial outcomes. Business news, not investment advice.