Towards credible digital twins for basic and preclinical research

Abstract

Digital twins are well established in industrial settings, but there has not been wide adoption in biomedical settings. Digital twins for biomedical applications are now possible with the inclusion of artificial intelligence and the potential to combine mechanistic and clinical models that learn and adjust for human variability.

Digital twins are virtual counterparts of physical systems that learn and evolve through continuous data exchange¹. They are redefining how biomedical scientists connect experiment computation with clinical translation. While industrial digital twins have long optimized aircraft engines and manufacturing processes, the biological equivalent faces unique challenges, including human variability, measurement uncertainty and intrinsic ethical dilemmas inherent to human synthetic replication.

The National Heart, Lung, and Blood Institute Workshop on Digital Twins in Heart Lung and Sleep Research held in the autumn of 2025 convened leaders from academia, government, industry and medicine to consider how digital twins can evolve from theoretical constructs to validated tools for discovery, preclinical innovation and clinical care. A central theme that emerged from the workshop was that digital twins are more than a modelling paradigm and comprise extraordinary potential as a research method. Digital twins can be developed to represent living human systems to unite mechanistic understanding, adaptive learning and continuous data validation.

Digital twins in a biomedical context can derive credibility from mechanistic foundations such as electrophysiological, haemodynamic or biochemical models that represent physical laws¹. Yet physical models alone cannot capture the complexity of living systems where parameters vary across individuals, time and space scales, environments and time. By contrast, purely data-driven models may be used to explain observations but often suffer from poor interpretability and the ability to generalize. The most powerful method in digital twinning is likely to combine both in a meaningful way². Hybrid modelling, possibly through activation of new artificial intelligence (AI)-derived methods, has the potential to integrate mechanistic and data-driven frameworks so that each informs the other. For example, physics-informed neural networks embed conservation laws directly into learning algorithms, constraining their behaviour by the same equations that govern physiology. In cardiovascular and pulmonary applications, these methods can infer hidden parameters such as arterial stiffness or airway resistance from signals that traditional sensors cannot currently measure directly³. With increasing digital twin complexity, hidden parameters tend to become data driven or empirically determined. Hence, digital twins can grow to integrate mechanistic understanding with AI and/or machine learning. A key feature of digital twins is their potential to transform modelling into a dynamic hypothesis testing loop that creates potential to clarify black-box aspects of hidden parameters and AI.

Simulations that start with modelled internal states to predict behaviour (known as forward models) have long been central to computational physiology, predicting behaviour from known structure and function, such as for motor control⁴, for example. The next frontier is inverse modelling, which infers internal states or structural parameters from measured outputs. Inverse lung mechanics models, for example, can deduce regional airway closure patterns from oscillometer data to generate patient-specific insight without invasive procedures⁵. Noninvasive electrocardiographic mapping on the body surface can now be subject to inverse techniques to reconstruct epicardial electrical maps on the heart surface to identify origins of abnormal rhythms⁶. For basic researchers, inverse approaches allow the digital twin to be implemented to test whether a proposed mechanism is physically plausible. Rather than simply fitting data, digital twins can evaluate the feasibility of competing hypotheses under realistic constraints, reconcile disparate data sets and create meaningful feedback loops between experimentation and computation.

Credible digital twins must also follow rigorous verification, validation and uncertainty quantification⁷. Verification ensures that a model is implemented correctly, while validation confirms that the model reproduces the relevant behaviour of a quantity of interest within its defined context of use. For preclinical applications, that context may be narrow, such as predicting clot formation in a microfluidic chip or simulating drug transport in a tissue construct. Clear boundaries of validity are essential. Uncertainty quantification provides a numerical measure of confidence defining when predictions remain reliable and when they begin to degrade. Emerging mathematical tools such as topological data analysis extend these ideas to multiscale data by capturing the geometric and relational structure that persists across resolutions. Such descriptors may become crucial for comparing outputs across models and modalities.

A central limitation in biomedical modelling is the sensor–model mismatch: the data that sensors provide are often not the data that models require. Wearable or laboratory sensors capture proxy measurements such as optical or impedance signals rather than direct physiological quantities⁸. Machine learning offers a way to bridge this mismatch by training models to map sensor signals onto underlying variables while preserving physical constraints. An even more transformative step is to design sensors to serve the model rather than adapting models to existing sensors. Model-driven sensor development guided by simulation and sensitivity analysis would allow focused measurement of the most informative variables and could reduce redundancy, improve accuracy and accelerate discovery by uniting experimental design with computational modelling.

The concept of the digital twin as an experimental collaborator may prove the most transformative of all. A twin can help to determine which experiment will be most informative by identifying the perturbation that best discriminates among competing mechanisms⁹. In systems biology, this adaptive design principle could minimize animal use and focus bench experiments where they yield maximal insight. Microphysiological organ-on-chip systems illustrate this synergy. These miniature physical systems can be coupled with in silico twins that simulate unmeasurable states such as molecular flux or tissue stress. Continuous calibration between the physical and digital domains creates a closed experimental loop in which data refine models and models guide new experiments.

For digital twins to become broadly adopted practical methods rather than bespoke demonstrations, shared methodological standards are needed. Every model should specify its domain of validity, quantities of interest and uncertainty bounds, similar to how experimental assays describe their performance characteristics. Open benchmarking datasets can be developed with reference measurements and standardized metrics for credibility assessment. Provenance metadata following FAIR principles will ensure that models, data and validation protocols remain findable, accessible and reusable across institutions¹⁰. These steps will transform digital twins into reproducible and shareable tools for experimental science. This does not mean that all data and models must exist in a single static form. Instead, they should be developed within a common set of guiding principles and designed for maximal compatibility. New methods can enable models to learn about emerging data types and formats, develop automated ingestion pipelines and communicate across systems, allowing digital twins to evolve dynamically while remaining interoperable and credible. Building credible, transparent and interpretable digital twins will require close collaboration between modellers, experimentalists and clinicians. With shared standards and open science, these systems can evolve from aspirational ideas into core methods for the current and next generation of biomedical research.