How can digital twins be applied to drug evaluation and repurposing?

A dynamic model, not a static dataset

A digital twin in medicine is a dynamic computational representation of a biological system — an organ, a physiological process, or an individual patient. It usually starts from a mechanistic model built from equations describing disease or therapy biology, whose parameters correspond to measurable features such as metabolic rates or cell population dynamics. Sampling those parameters across plausible ranges produces a virtual population in which each parameter set stands for one virtual subject.

That construction matters for drug evaluation because it allows questions to be asked in a specific individual rather than only in a population average. The paper adopts the NASEM framework, which emphasizes three elements: personalization to a specific patient system, dynamic updating between the physical and virtual versions, and the ability to inform decisions that realize value. The authors state plainly that current implementations achieve only part of this vision, limited by data sparsity, model uncertainty, model simplification, and a lack of prospective validation.

Regulatory shifts make in silico evidence usable

The review places digital twins inside a broader regulatory move toward non-animal and computational methods. The FDA Modernization Act 2.0, enacted in 2022, addressed the legal acceptability of non-animal testing methods; in 2024 the FDA created a CDER AI Council, and in January 2025 it issued a draft guidance on using AI to support regulatory decision-making for drugs and biological products. The FDA then announced a phased elimination of animal testing mandates for monoclonal antibodies and other drug classes, and in December the Senate passed the FDA Modernization Act 3.0, which replaces “animal test” with the more inclusive “nonclinical testing method.”

Europe takes a different posture. The EMA advocates computer algorithms for virtual clinical trials and emphasizes a risk-based approach, concentrating oversight on applications with high patient risk or major impact on regulatory decisions. Its guidance requires traceable documentation of data collection and processing, evaluation of data representativeness, strategies to address bias, and a stated preference for interpretable models; it also prohibits incremental learning during a clinical trial. In March 2025 the EMA issued a qualification opinion for an AI tool determining histological activity in NASH biopsies, described in the paper as the first global regulatory qualification of an AI-based clinical trial tool.

China's NMPA issued a 2025–2030 digital and intelligent transformation plan promoting “scene-based application.” Among 41 typical application scenarios, four concern drug evaluation: virtual animal experiments; data mining and decision-model research using real-world traditional Chinese medicine experience; risk-based clinical trial management; and intelligent digital monitoring of adverse drug reactions. These are policy frameworks and identified scenarios, not evidence that digital twins are already validated for those uses.

Cardiac electrophysiology leads on validation

Cardiac electrophysiology is described as the most extensively developed application domain for digital twins in this field, with substantial work on pathological modeling, drug effect evaluation, and treatment planning. The reason is a gap in conventional methods: animal species differ from humans, and in vitro models are oversimplified, so predicting drug-induced cardiotoxicity is unreliable.

One concrete example is a virtual heart built by Camps and colleagues from conventional magnetic resonance imaging and 12-lead electrocardiogram data, using a pseudo-diffusion reaction-Eikonal model combined with approximate Bayesian computation inference. The model captures both an individual's cardiac electrophysiological characteristics and their range of variation. The study then simulated dose-dependent effects of candidate drugs on ventricular repolarization to predict the likelihood of clinically relevant ECG changes such as QTc prolongation. The paper classifies this as a DT-like methodology evaluated retrospectively.

Other work simulates changes in action potential propagation to predict structural consequences of diffuse myocardial fibrosis, a known adverse effect of drugs such as doxorubicin; those simulation results were validated in swine models. That validation supports translational relevance but remains animal validation, not prospective human outcome evidence.

Multi-organ toxicity and safety signal detection

Digital twins are also proposed for liver and kidney toxicity by connecting molecular alterations to organ-level physiological responses. The TransTox platform was developed from transcriptomic profiles of liver and kidney tissue after drug exposure, across a range of doses and treatment durations. It can identify hepatotoxic and nephrotoxic compounds even when they were not in its training set, and it offers mechanistic explanations, such as why thioacetamide selectively damages the liver while sparing the kidneys.

In pharmacovigilance, the paper argues that twins can model drug-related toxicity even with relatively sparse datasets. One cited application predicts neutropenia induced by venetoclax and azacitidine, classified as a canonical digital twin evaluated retrospectively. Twins built from real-world data such as electronic health records and observational studies can model individual patient health trajectories, which could support automated adverse-event monitoring and integration of predictions into case management and signal detection. The authors note that practical deployment faces unresolved data privacy and regulatory compliance issues.

Dosing insight where trials are hard to run

In pharmacokinetics and pharmacodynamics, digital twins integrate high-dimensional patient data — age, sex, genetic profiles, disease status — to calculate and dynamically simulate parameters, letting clinicians estimate drug exposure, metabolic rates, and potential drug-drug interactions for a specific patient before or early in treatment. The paper lists three utilities: individualized PK/PD modeling to inform dosing; estimates of systemic or tissue-specific parameters that are otherwise difficult to measure; and simulation of drug-drug interactions for combination regimens.

Pregnant women are the clearest example of a population where ethical and practical constraints have limited study of drug metabolism and safety, even for commonly used medicines such as prednisone. A pregnant-women digital twin can model maternal-fetal physiology by simulating drug dynamics in vivo, filling a knowledge gap; the paper classifies this as DT-like and retrospective. For narrow-therapeutic-index drugs such as amikacin, where nephrotoxicity concerns push clinicians toward possibly less effective alternatives, twins could support more confident use in critically ill oncologic patients through precise, virtually monitored dose adjustment. Some systems are developed alongside a specific drug or delivery system as drug-DT combination products, using real-time simulation and feedback to inform dosing in complex physiological or local microenvironments.

Most evidence is demonstration, not deployment

The review searched PubMed from database inception to December 31, 2025 and identified 4,370 publications using the term “digital twin,” a number the authors contrast with over 330,000 records for “artificial intelligence” in the same period. The authors also report a marked concentration in implementation trends: both identified patient-intervention digital twins targeted glycemic control. That narrow distribution underscores how much remains to be shown before twins are deployed for broader drug evaluation.

The limitations explain why the synthesis cannot be stronger. The boundary of “digital twin” is unstable, and classification partly depended on authors' own use of the term, so some included studies may not fully satisfy formal criteria. A single database limits coverage and skews toward peer-reviewed academic work over industry applications; English-language restriction underrepresents non-English-speaking regions. Heterogeneity across target populations, disease areas, data sources, outcome measures, implementation and validation approaches, and drug-development stages prevented meta-analysis, and no universally accepted quality assessment tool exists for digital twin studies. Most included literature consists of demonstration studies rather than deployable systems with lifecycle governance.

Continuous data and reimbursement gate deployment

The practical bottleneck is measurement. A twin's dynamic, predictive capacity depends on high-frequency, accurate measurement of critical physiological parameters. Glycemic twins benefit from minimally invasive continuous glucose monitoring, which supplies the reliable, rapid data streams the twin needs. By contrast, PK/PD or toxicological markers for many other diseases rely on traditional non-continuous assays such as high-performance liquid chromatography, which cannot provide that continuous flow. The paper's own framing is that effective twins must be data-responsive and tightly coupled to the temporal evolution of the patient-specific system.

Payment is the second gate. The authors describe an emerging trend to co-develop monitoring protocols during clinical trials, forming a combined product of the drug and its twin. For twins intended for post-market evaluation and monitoring, appropriate payment models will need to be established, and how those are addressed will largely determine pricing and whether a sustainable market exists. They suggest payers could structure reimbursement to reward twins that improve patient-centered outcomes such as quality of life or process metrics such as care efficiency, and that given the near-zero marginal cost of deploying software, alternatives to traditional fee-for-service arrangements may better incentivize high-quality twins without encouraging overuse.

Personalized output is not the same as personal values

The paper raises an ethical problem specific to highly personalized twins. When patients consent to provide data for a specific clinical trial, it is uncertain whether that authorizes future use of their data inside a twin designed to simulate the effects of diverse drugs — a purpose likely beyond the reasonable expectations patients held at the outset.

A second problem is that drug evaluation involves value-based trade-offs, not only calculation. The paper's example is finasteride: initially used for benign prostatic hyperplasia, it later extended to androgenetic alopecia. As the target population shifted from older to younger men, the risk of sexual dysfunction became less acceptable, and the dose was reduced from 5 mg to 1 mg. How much risk a person will tolerate for therapeutic benefit varies across populations and even more between individuals. A twin may not capture the personal rationale and values behind a patient's decision, reflecting only aggregate patterns in the data, so a recommendation can appear personalized while contradicting that patient's true preference.

The authors' proposed response is risk-proportionate governance: the closer a twin's output is to supporting direct clinical or regulatory decisions about individual patients, the more rigorously it must be evaluated for bias and uncertainty. They suggest fairness-by-design toolkits, diverse training datasets, and debiasing algorithms. This is a governance proposal, not a demonstrated result.

The roadmap coordinates stages rather than promising readiness

The paper's central deliverable is a staged development roadmap intended as a coordinating tool to bridge academia, industry, and regulatory agencies, foster consensus, align goals, and guide resource allocation. It has two stated aims: cultivate patient-centered development and evaluation of emerging twins within drug evaluation, and facilitate collaboration among all stakeholders. The roadmap uses a dashboard with a movable pointer: blue marks processes already in place and functioning, yellow marks ongoing development, and red marks processes not yet initiated, with the pointer moving as processes solidify or require further review.

The most decision-relevant reading is therefore a split by evidence maturity rather than by disease area. Cardiac electrophysiology, multi-organ toxicity screening, and PK/PD dosing support have the most concrete demonstrations, but the cardiac and toxicity examples are retrospective or animal-validated, and the pharmacovigilance and pregnancy cases are classified as retrospective or DT-like. The paper's own conclusion is that progress will not come from a single breakthrough but from shared accumulation of data and knowledge across the industry — a transformation of the pharmaceutical ecosystem, not merely a technology advance. For readers weighing a twin for a specific decision, the paper's implicit threshold is proximity to that decision: the closer the output sits to a clinical or regulatory choice about an individual patient, the more prospective validation, bias assessment, and continuous sensor data the application must supply.

Sources used in this answer