Virtual patients in clinical trials: what drug modelling can already do, and what regulators still need to see

In 2018, the regional pharmacovigilance centres in France ran a prospective study of hospital admissions caused by adverse drug reactions. Extrapolated to the whole country, their figures pointed to about 212,500 hospitalisations a year, a share of them linked to drugs taken outside recommended use. Part of the problem is arithmetic. The concentration of a drug in the blood depends on the dose, the timing, the patient's weight, kidney function and the other medicines in the body, and very few people can compute that in their head.

Frédéric Dayan, a pharmacist, physicist and former R&D team lead at Dassault Systèmes, co-founded ExactCure in Nice in 2018 to do that computation for each patient. The company's technology, which he describes as a digital twin of the treatment, is now part of the data science unit of Biotrial, a French contract research organisation (CRO), where he is scientific director. In an interview for Tech Anatomy in September 2026, he described how the same modelling moved from one patient taking a pill to whole cohorts in clinical trials, and what regulators accept so far.

From one patient to a virtual cohort

The core of the work is PK/PD modelling. Pharmacokinetics (PK) describes what the body does to a drug: absorption, distribution, metabolism, elimination. Pharmacodynamics (PD) describes what the drug does to the body. ExactCure built these models for thousands of marketed drugs and delivered them through a mobile app for patients, certified as a medical device, with a mirror version for health professionals and integrations sold to companies such as Vidal.

Dayan's argument for modelling is about real-world behaviour. "In a pharmacology lecture, the drug is taken morning, noon and evening. 8 a.m., noon, 8 p.m.," he says (translated from French). "In real life, 8 a.m. does not exist." Doses are late, skipped or doubled, and a model can integrate that.

Around 2024, the team began applying the same logic to clinical trials, simulating groups of patients rather than one. He gives the example of tofacitinib, a Pfizer JAK inhibitor already marketed for inflammatory diseases, being repositioned for a new indication. By his account, the team simulated the step from phase 2 to phase 3 and identified subgroups likely to develop haematological side effects, information a sponsor could use to adjust recruitment. Changes in blood counts, such as low lymphocyte or neutrophil levels, are known effects of the drug. He described the work as a public collaboration with Pfizer; we did not find the corresponding publication. "We try to make the trial live before it takes place."

He distinguishes two families of methods. Mechanistic models try to represent the biology step by step, from the target to the concentration curve, and are calibrated with data. Non-mechanistic models learn patterns from large datasets without describing the mechanism, which is where machine learning comes in. Which one to use depends on the data available and on how well the biology is understood, and the two can be combined.

External controls: statistics more than AI

Asked what is mature today, Dayan points to external control arms. In a classic randomised trial, patients are assigned by chance to the new treatment or a control. An externally controlled trial treats everyone who enrols and builds the comparison group from patients outside the trial, often from registries or health records, matched as closely as possible on their characteristics. "There is not necessarily any AI in it," he says. "It is really statistics."

The approach is used mostly in rare diseases, where randomising a few dozen patients is unrealistic. His practical advice to sponsors is to plan the external arm from the start rather than "arrive as firefighters" once recruitment has stalled, because the treated group has to resemble patients who actually exist in the available databases.

The US Food and Drug Administration published draft guidance on externally controlled trials in February 2023. The European Medicines Agency's reflection paper on single-arm trials, finalised in 2024, states that randomised evidence remains the expected standard and that any deviation must be justified. It explicitly leaves external controls to a separate paper, still at concept stage. In both regions, acceptance is decided trial by trial.

AI-generated patients, retrospectively tested

The newer and more contested idea is to generate virtual patients with AI to complete a trial. Dayan cites a study he finds interesting, though "not perfect". Published in the Journal of Medical Internet Research in 2025 by a Canadian-led team with Austrian and US trial groups, it took ten datasets from nine completed cancer trials, removed the last 10 to 50 percent of recruited patients, and asked four types of generative model to simulate replacements. With the best-performing method, the published conclusions were reproduced in 88 to 100 percent of cases when up to 40 percent of patients were removed. The authors conclude that a trial stopped at 60 percent of its target could, in this setting, be completed by simulation, and they call for confirmation in other diseases.

Europe has already qualified one machine learning method of this family, in a narrower form. In 2022 the EMA issued a qualification opinion on PROCOVA, from the company Unlearn, which uses AI-generated prognostic scores to adjust the analysis of a randomised trial and reduce the size of its control group. The patients remain real; the model sharpens the comparison.

Competing readings

One reading sees a gradual path similar to aeronautics, Dayan's own analogy. Aircraft design moved from wind tunnels to mostly computer simulation over several decades, under a regulator as strict as any drug agency. Seen this way, trials "without patients" are a fantasy today, as he puts it, but trials partly informed by simulation are already here. The FDA's April 2025 plan to reduce animal testing for monoclonal antibodies using computational models and organ-on-chip systems, and the adoption in January 2026 of the international ICH M15 guideline on model-informed drug development, point in that direction.

A second reading is more cautious. Simulated patients can only reproduce the variation present in the data they were trained on, and a retrospective test checks whether a model could recover a result that was already known. Dayan shares part of this concern. "Sometimes it seems magical, and I don't like magic," he says. "I have 100 patients, I can make 100 more, 1,000 more, why not a million more. It no longer means anything."

What the evidence does not show

Most published demonstrations, including the oncology study above, are retrospective. They show that simulation would have reached the same decision, not that it can be trusted before the answer is known. The Biotrial and ExactCure examples described in the interview come from the guest's account and are not independently assessed here. And the regulatory texts cited cover different things: external controls, covariate adjustment, modelling in general. None of them yet sets out how to accept AI-generated patients as evidence of efficacy.

What to watch

Dayan's own bet is the shift from retrospective to prospective validation, with a metric that says how much real information a set of virtual patients carries. Three indicators would show whether that is happening: a pivotal trial in which a simulated arm is specified in advance and accepted by a regulator, the EMA's forthcoming reflection paper on external controls, and how agencies apply ICH M15 to AI-based models rather than classical pharmacometrics. Whether the bottleneck is the technology or, as Dayan argues, its implementation in a heavily regulated industry, those decisions will show how far regulators are prepared to go.

Listen to the episode

MODÉLISER LA VIE DU MÉDICAMENT DANS LE CORPS - Frédéric Dayan, Biotrial

Listen · 35 min