BIONEWS
Sponsors use medical digital twins — computational models that mirror a patient, disease, or trial population — to shrink traditional placebo or standard treatment comparison groups in their trials.
A regulatory affairs director at a mid-size biotech opens the FDA's response to her pre-submission package and reads a question her industry cannot yet answer in any standardized way: "Please describe the credibility assessment methodology used to qualify this model for its stated context of use." The question is not idiosyncratic. It is the new normal for any sponsor whose trial math now leans on a computational model, and the window for getting the answer right is open right now.
The technology in question is a medical digital twin: a computational model that mirrors a specific biological system, a patient, a disease trajectory, or a trial population with enough fidelity to predict outcomes that have not yet happened. In clinical research, the most common application is the synthetic control arm, a stand-in for the patients a trial would otherwise have to enroll, treat, and follow in the traditional control group. Used well, synthetic controls shrink the control arm, compress timelines, and improve statistical power without adding sites or spend. Used badly, they substitute a model's assumptions for human evidence and quietly inflate apparent effect sizes.
The credibility question cuts at the mechanism. A mechanistic digital twin for disease forecasting is built in four sequential steps, and each one compounds the uncertainty of the one before it. First, a model of the underlying biology is constructed, often with differential equations for drug pharmacokinetics, receptor binding, or cellular signaling. Second, the model is calibrated against historical patient-level data. Third, the calibrated model is validated in a held-out dataset drawn from a different population or trial context. Fourth, the sponsor defines the context of use explicitly: which subgroup, which indication, which outcome, which timeframe. The load-bearing risk sits at the third and fourth steps, the validation-and-context-of-use boundary, because that is where a model calibrated to one population gets pressed into service for another.
Regulators have noticed. The FDA has now stacked three relevant guidances in under three years. The November 2023 final guidance on Assessing the Credibility of Computational Modeling and Simulation in Medical Device Submissions set the floor for model credibility in device work. A January 2025 draft guidance extended the conversation to AI used in regulatory decisions for drugs and biologics. Then in June 2026 came M15, General Principles for Model-Informed Drug Development, which reaches directly into the kind of trial math that synthetic controls are now underpinning. A separate January 6, 2025 press release proposed a credibility framework for AI models used in drug and biological product submissions. The scaffolding is being assembled while sponsors are already building on the foundation, and that is the gap the regulatory affairs director is now trying to navigate.
Europe is moving on a parallel track. In September 2022, the European Medicines Agency issued a Qualification Opinion for PROCOVA (Prognostic Covariate Adjustment), an early regulatory signal that covariate-adjustment and synthetic-control methodologies are advancing toward formal qualification. PROCOVA is not a digital twin, but it is the regulatory logic a synthetic control arm is built on, and the EMA opinion is the closest thing the field has to a qualified methodology for the use case. The fact that the FDA and EMA are converging on the same problem from different angles is itself a signal: the credibility standards will be written, and they will be written soon.
The honest read is that the technology is real, the promise is real, and the trust framework is underbuilt. Sponsors are filing protocols on epistemologies that are not yet standardized, and the regulatory affairs director's question is the system catching up. What to watch over the next 12 months: finalization of the January 2025 AI draft guidance, the first public Type B meeting summary in which a credibility-assessment question produces a concrete sponsor response, the next EMA qualification opinion that names a digital-twin use case explicitly, and whether M15 settles into routine review practice or stays advisory. Each of these will narrow the gap between a model that runs and a model that a regulator is willing to accept as evidence.