A clinical trial traditionally answers its most important question by comparing real people. One group receives an experimental medicine, another receives placebo or standard treatment, and investigators measure whether outcomes differ enough to demonstrate that the new intervention works. Digital twins challenge that model by asking whether detailed mathematical and artificial-intelligence models of individual patients could simulate some treatment trajectories that would otherwise need to be observed directly in a conventional control group.
That future moved another step closer to regulatory relevance on September 1, 2026, when the U.S. Food and Drug Administration and Duke-Margolis Institute convened an Innovation in Quantitative Medicine Summit to help establish a cross-sector Quantitative Medicine Innovation Network. The initiative is intended to identify scientific and regulatory gaps and accelerate broader adoption of modeling and simulation across drug development and clinical care. It follows final FDA adoption in June of the ICH M15 guidance establishing principles for planning, evaluating and documenting model-informed drug-development evidence.
What is a digital twin in medicine?
A medical digital twin is a computational representation of an individual patient or biological system that attempts to predict how that system may evolve under different conditions. Depending on the application, a twin could combine demographics, laboratory results, imaging, genetics, physiology, disease history, drug exposure and population-level data within mechanistic or machine-learning models. The model can then simulate what might happen if the patient receives one treatment, another treatment or no treatment at all.
The important word is “predict.” A digital twin is not a perfect electronic copy of a human body and cannot reproduce every biological process operating inside a person. Its usefulness depends on whether it accurately captures the mechanisms and variables relevant to the clinical question being asked, which means a model can be highly useful for one decision and completely inappropriate for another.
How could digital twins actually change a clinical trial?
One application is dose selection. Pharmaceutical companies frequently test several doses because they need to understand how drug exposure relates to biological response, efficacy and toxicity. Model-informed drug development can combine pharmacokinetic, pharmacodynamic and disease-progression information to predict which doses deserve further testing, potentially reducing unnecessary trial arms and helping sponsors move more efficiently into confirmatory studies.
A more ambitious application involves synthetic or digital control patients. Instead of enrolling the full conventional number of participants into a placebo or standard-care arm, developers could potentially use validated models to predict outcomes for some patients under the control condition. Real participants would still be needed to anchor and verify the model, but the simulated evidence might reduce how many people must receive placebo when reliable historical and biological data are available. Researchers writing in Nature Medicine have described digital twins and in-silico trials as an increasingly realistic component of the evolving regulatory landscape while emphasizing the need for careful oversight and public-sector involvement.
Why would patients want fewer people assigned to placebo?
The answer becomes most compelling in serious diseases.
Imagine an oncology trial in which participants know they have an aggressive cancer but face a chance of receiving a control therapy that investigators already suspect is inferior. Randomization remains scientifically valuable because it protects against bias, but patients may understandably prefer access to the experimental therapy. If validated digital controls could safely reduce the proportion randomized away from the investigational treatment, recruitment could become easier and fewer people might need to receive an unattractive control regimen.
Rare diseases create an even stronger case. When only several hundred eligible patients exist worldwide, assigning half of them to placebo can consume a substantial portion of the total population available for research. Digital twins could theoretically preserve statistical comparison while directing more patients toward active treatment, although regulators would need extremely high confidence that the model accurately represents untreated disease.
Could digital twins make drug development faster and cheaper?
Potentially, because several expensive development decisions arise from uncertainty rather than from the physical cost of manufacturing trial tablets.
Drug companies routinely conduct studies to determine dose, interaction effects, special-population exposure and expected disease progression. FDA’s model-informed drug-development program already recognizes that quantitative approaches can improve trial efficiency, raise the probability of regulatory success and help optimize dosing in circumstances where dedicated studies may not always be necessary.
A 2026 review involving AstraZeneca researchers described potential digital-twin applications across target discovery, preclinical development, clinical trials, regulatory review, manufacturing and post-market care. The authors argued that twins could reduce uncertainty, timelines and failure rates but highlighted major unresolved challenges involving data integration, model reliability, regulatory acceptance and privacy.
Is the FDA already accepting simulations instead of real clinical evidence?
Modeling already influences regulatory decisions, but that is not the same as replacing pivotal randomized trials with virtual patients.
FDA uses and reviews quantitative models for pharmacokinetics, exposure-response relationships, dose selection, drug interactions and other development questions. The agency’s MIDD Paired Meeting Program gives selected sponsors opportunities to discuss modeling strategies directly with FDA scientists, while the new ICH M15 framework creates common expectations for documenting and evaluating model-informed evidence.
The regulatory direction is therefore evolutionary rather than revolutionary. Models are being incorporated into more decisions as evidence of reliability accumulates, but the evidentiary threshold rises sharply when a model is asked to replace information that would otherwise come from a real patient. A simulation used to select a Phase 2 dose carries a different consequence from a virtual control used to help establish whether a new cancer therapy should be approved.
How would regulators know whether a digital twin is trustworthy?
Validation must be tied to the exact context in which the model will be used.
A model predicting drug concentration from age, body size and kidney function can be tested against thousands of real pharmacokinetic measurements. A twin intended to predict survival over three years in metastatic cancer is far more difficult because tumor biology, subsequent treatment and individual disease trajectories introduce enormous variability.
Researchers working on causal inference and digital twins argue that future trial applications will need careful definitions of treatment effects, transportability and patient heterogeneity rather than simply producing a high headline accuracy score. A model might predict average outcomes well while performing poorly in an important demographic or molecular subgroup, which could introduce systematic bias into a clinical trial rather than removing it.
Could AI make digital twins much more realistic?
Artificial intelligence can help integrate data types too complex for conventional statistical models, including imaging, longitudinal electronic-health-record data, continuous wearable measurements and molecular profiles. Generative and machine-learning methods can also discover relationships between variables that researchers did not explicitly encode. That makes AI a powerful component of digital-twin development, but it also creates explainability and stability problems when the system’s predictions depend on patterns that scientists do not fully understand.
The FDA is already exploring AI-powered virtual models in regulatory science. Its AI4TOX program includes AnimalGAN, which aims to develop virtual-animal digital twins for predicting toxicological outcomes of untested compounds as part of efforts to reduce, refine or replace some animal studies. This does not mean virtual humans are ready to substitute for clinical participants, but it demonstrates that the agency itself views simulated biological systems as a serious regulatory-science tool rather than merely a technology-industry slogan.
Could wearables make a patient’s digital twin update continuously?
This is where digital twins begin moving from static models toward something closer to their engineering namesake.
A conventional model might use information collected at baseline and generate one prediction. A more advanced twin could continuously ingest heart rate, activity, sleep, glucose, respiratory patterns or other measurements from digital-health technologies and update its estimate as the patient’s physiology changes. FDA has been actively supporting development of digital-health measures for clinical trials and recently convened a workshop specifically addressing statistical considerations for digitally derived endpoints.
Continuous data could allow future trials to detect subtle treatment effects that clinic visits miss. A neurological therapy, for example, might influence balance, movement or reaction time before a traditional questionnaire shows deterioration, while a heart-failure medicine might improve everyday activity long before hospitalization rates diverge. The FDA has explicitly identified capturing early manifestations of chronic disease outside conventional healthcare settings as one potential use of digital-health technologies in drug development.
What could go wrong if trials rely too heavily on virtual patients?
The most dangerous failure would be a model that looks scientifically sophisticated but systematically predicts the wrong outcome.
Digital twins learn from existing datasets, and existing healthcare data reflect historical treatment patterns, demographic imbalances and unequal access to care. A model trained on patients from major academic hospitals may not represent people treated in rural settings, while a model developed largely in one ethnic population could perform differently elsewhere.
There is also a deeper scientific risk. If drug developers repeatedly use models derived from the same historical datasets, incorrect assumptions can become self-reinforcing. A real randomized control group has one enormous advantage over a simulated patient: reality does not care whether the model developer’s assumptions were elegant.
Could digital twins eventually reduce animal testing too?
This may happen sooner than widespread replacement of human trial controls.
Toxicology studies often use animals because researchers need to understand how a compound affects multiple organs before exposing people to an experimental drug. Computational systems combining chemical structure, biological pathways, organ models and historical toxicology data could eventually identify compounds with predictable toxicity or demonstrate when additional animal studies are unlikely to provide meaningful information.
FDA’s AnimalGAN program is explicitly exploring generative-AI virtual animal models to predict toxicological outcomes while supporting the principles of replacement, reduction and refinement of animal studies. If such models become sufficiently validated, pharmaceutical companies could eliminate some low-information experiments while reserving animal studies for questions where computational prediction remains inadequate.
Will digital twins ever completely replace clinical trials?
For most medicines, probably not.
Human biology is too variable, and clinical outcomes are influenced by behavior, adherence, environment and unmeasured factors that no model captures perfectly. Regulators will continue demanding direct human evidence whenever uncertainty is high and consequences are serious.
The more realistic future is hybrid trials. Real patients will provide biological ground truth, while digital twins help select doses, simulate scenarios, enrich enrollment, reduce selected control-group requirements and interpret variation between patients. Quantitative models could therefore make trials smaller or smarter without eliminating the people whose outcomes ultimately determine whether treatment works.
What could a clinical trial look like by the 2030s?
A patient might enter a study with genomic data, medical history, imaging and months of wearable measurements already incorporated into a computational model. The digital twin could predict likely disease progression under standard therapy, while the patient receives the investigational medicine and continuously streams real-world physiological measurements back into the model. Investigators could identify early deviations from the predicted trajectory and adapt selected aspects of treatment or monitoring under a prospectively defined protocol.
The control arm might contain fewer people because simulated trajectories supplement rather than replace randomized patients. Dose adjustments could be informed by individualized pharmacological models rather than broad population averages, while regulators could review a transparent digital record showing how the model was developed, validated and changed throughout the trial.
The technology is not ready to make that the default today, and the ethical stakes are too high to rush. Yet the regulatory infrastructure is clearly moving toward greater use of quantitative evidence. With FDA establishing a Quantitative Medicine Innovation Network, implementing international model-informed development standards and experimenting with virtual biological models of its own, the question is gradually shifting from whether simulations belong in drug development to how much responsibility regulators will eventually allow them to carry.
