Digital Twins in Rare Diseases: From Virtual Trials to Personalized Care

Digital twins are moving from aerospace, automotive, and manufacturing into medicine and drug development, with the potential to transform how we study diseases, conduct clinical trials, and personalize treatment and care. The technology is especially promising for rare diseases, where small patient populations complicate disease understanding, drug development, and care planning. At the same time, rare diseases pose unique challenges for building sophisticated digital twin models, which depend on high-quality, longitudinal data for training. This article provides an overview of the current state of digital twins in the rare disease space and offers an outlook on their future applications.

#advertisement

From Space to Clinic: What Digital Twins Bring to Biomedicine and Healthcare

Digital twins are dynamic, data-driven virtual counterparts of real-world objects or systems. While the term was only coined in 2005 for product life cycle management, the digital twin concept was first adopted by NASA in the 1960s as a living model of its Apollo mission. After decades of successful application in industries such as aerospace, automotive, and manufacturing, digital twins are slowly entering the fields of healthcare, biotech and pharma.

Figure 1

In the biomedical and healthcare contexts, digital twins can represent patients, organs, tissues, diseases, or even single cells as dynamic, individualized models (Figure 1). Researchers and clinicians can use digital twins in various applications (Figure 2):

modeling disease biology for discovering novel drug targets and compounds, and designing medical devices
designing and conducting safer, more efficient, more ethical, and less costly clinical trials
designing more personalized treatment and care models

Figure 2

The complexity of biological systems far exceeds that of engineered devices (even spaceships), which has long hindered the development of digital twins in human biology. Today, however, the surge in data availability and technical progress is driving the creation of more sophisticated models. On the one hand, the catalog of human biological data, from electronic health records, imaging, lab parameters, genomics, multi-omics, and wearable devices, is becoming more comprehensive and accessible. On the other hand, advances in computational and technological infrastructure are boosting scalability and enabling real-time decision-making. Particularly notable is the role of generative AI models, which show promise in synthesizing heterogeneous, unstructured data and providing the scalability and contextual understanding needed to build multi-organ models and even patient-specific twins.

From Understanding to Insight: Why Digital Twins Matter in Rare Disease

Rare diseases can be challenging for doctors to diagnose and treat, and for scientists to investigate. Each rare disease affects a small number of people, sometimes only a few people in one country, resulting in patchy and dispersed data and leaving many open questions when it comes to disease mechanisms, manifestations, trajectories, and treatment success predictors. Creating digital twins of rare disease patients could help researchers gain a greater understanding of the disease to inform diagnosis, treatment, and drug discovery.

An emerging project at the EMBL-EBI is applying digital twin technology to rare diseases, with the aim of building tissue-specific digital twins that simulate healthy and diseased states of tissues affected in rare diseases. The team around Rahuman Sheriff, Senior Project Leader at EMBL-EBI, and Ellie McDonagh, Translational Informatics Director at Open Targets, aims to collect and curate multi-omics data from various sources such as public databases, controlled-access data (e.g. European Genome-phenome Archive) and collaborations with rare disease consortia and patient advocacy groups. Based on this compiled data, the researchers aim to model healthy and diseased tissue twins.

Other initiatives aim to develop organ-specific digital twins, with different players developing cardiac digital twins such as Dassault's Living Heart Project. In a recently published study, researchers from Imperial College London created over 3,800 personalized cardiac digital twins using UK Biobank data. These digital twins enabled the researchers to uncover links among anatomical and physiological variations, an individual’s lifestyle, age, and clinical outcomes. When extrapolated to specific disease contexts, such initiatives hold significant potential for improving our understanding of organ- and system-level manifestations of rare diseases.

Disease understanding that can be gathered using digital twins can also be translated into drug discovery and development efforts. For example, Boston-based Aitia uses causal AI and multi-omic data to develop “Gemini digital twins”. These computational models of human disease are designed to replicate genetic and molecular interactions driving disease biology. Aitia’s researchers use their digital twins to conduct billions of virtual experiments, including patient-level gene and protein knockdowns. They have applied their digital twins to two fatal neurodegenerative rare diseases, Huntington’s disease and amyotrophic lateral sclerosis (ALS). For Huntington’s, Aitia’s digital twin approach resulted in a model with approximately 23,000 nodes and 5.3 million interactions, allowing researchers to uncover a novel target related to cognition and motor function in the disease.

From Trial Design to Augmentation: Digital Twins in Clinical Trials

Clinical trials are generally costly and failure-prone. Patient recruitment, in particular, is a major cost driver, and insufficient enrollment can leave studies underpowered or cause them to fail. In the rare disease space, small patient populations and limited disease understanding further hinder efficient trial design and recruitment. A number of academic groups, biotech companies, and pharma players are pursuing digital twins for clinical trials. Beyond external controls that draw on historical studies and real-world data, digital twins can augment trials by predicting each patient twin’s control outcome and can even be used to suggest adaptive trial decisions to investigators.

While examples in rare disease trials remain limited, the technology can offer financial incentives for sponsors and tangible benefits for participants, including virtual control arms that can reduce costs and timelines and spare patients from being randomized to placebo or to a standard of care presumed less effective than the investigational treatment.

A digital twin pioneer, the San Francisco-based startup Unlearn, uses digital twins to help augment clinical trials, which can reduce patient numbers in control arms, with the potential to accelerate timelines and cut costs. In its partnership with Johnson & Johnson, Unlearn showed that its generative AI-based digital twins could reduce control arm sizes by up to 33% in Phase 3 Alzheimer’s trials. Going beyond augmentation, digital twins could potentially result in fully virtual control arms. For example, in a proof-of-concept study, clinical trial data analytics company Phesi demonstrated that AI-powered digital twins could replace standard-of-care control arms in trials for chronic graft-versus-host disease (cGVHD). cGVHD is a serious complication that can occur as a result of stem cell transplants. Phesi employed its Trial Accelerator™ platform, which contains data from over 100 million patients, and selected 2,042 patients (32 cohorts) to construct the cGVHD digital twin cohort and 438 patients from 8 cohorts to construct the cGVHD digital twin standard-of-care cohort.

AI-based digital twins can also help investigators simulate trials before conducting them, to inform inclusion criteria, endpoints, timelines, and doses. For example, French AI-based trial simulation company Nova employs its jinkō platform to simulate trials, which helps its partners such as AstraZeneca to optimize trial design. Such simulations have also been employed in rare disease trials. Sanofi scientists developed digital twins of patients with Pompe disease, a progressive and often fatal lysosomal storage disorder that affects fewer than 50,000 people worldwide. Using data from multiple clinical trials, they created and validated quantitative systems pharmacology (QSP)-based digital twin models of individuals with infantile-onset Pompe disease, for which patient numbers are very small. Their QSP digital twins captured variations in disease history and severity and enabled the simulation of a head-to-head trial of Sanofi’s next-generation enzyme replacement therapy, Nexviazyme vs. Lumizyme (standard of care), while adding an additional layer of disease understanding for this understudied patient population.

San Francisco-based startup Trace Neuroscience plans to employ Unlearn’s Digital Twin Generator for ALS (ALS DTG), which has been trained on more than 13,000 clinical records obtained from various sources, to predict the behavior of patients’ digital twins for a control group and a group treated with an antisense oligonucleotide that targets the UNC13A protein.
Notably, at the pre-trial stage, models don’t have to be fully formed digital twins. Phesi, for example, uses so-called Digital Patient Profiles, precursors to digital twins, which are based on a statistical view of patient attributes. Digital Patient Profiles or similar models could be used, even when full digital twin implementation is not feasible, to support trial design, decision-making, and to enable sponsors to engage regulators around digital twin strategies early on.
However, digital twins for clinical trials still have a number of limitations due to the complexity of human biology and a lack of understanding of novel disease contexts and drug behaviors. For example, digital twins currently can’t reliably predict novel adverse events because they’re trained on past datasets. Since these models depend on existing knowledge, they generally work best for well-understood monogenic diseases, while their utility is far more limited in complex and less well-studied conditions.

From Decision to Personalization: Digital Twins for Treatment and Care

Beyond clinical trials, digital twins can support treatment personalization and care decisions by integrating diverse data sources into predictive models to optimize treatment outcomes. This can be especially valuable for rare diseases, where evidence to guide treatment is often sparse.

For example, a recent proof-of-concept study explored using large language models (LLMs) to build digital twins for rare gynecologic tumors. The model integrated clinical data from 21 patient cases with insights from 655 publications to generate personalized treatment plans for metastatic uterine carcinosarcoma—revealing options traditional analyses might miss.

Digital twins can also address questions on dosing and drug-drug interactions. French digital health startup ExactCure, in partnership with Dassault Systèmes’ 3DEXPERIENCE Lab, is developing a full-body digital twin model to improve medication safety and personalization. Their model integrates characteristics such as age, gender, kidney function, genotype (e.g., CYP2D6 status), and lifestyle factors to simulate individual drug metabolism and anticipate underdoses, overdoses, and harmful interactions. Beyond broader therapeutic areas, ExactCure also aims to focus on rare diseases, and has announced a collaboration with the Fondation Adolphe de Rothschild Rare Disease Reference Centre for Wilson's disease and with Orphalan with the goal of personalizing treatment for Wilson's disease.

Other companies employ digital twins to aid more targeted interventions and lifestyle modifications. For example, digital twin startup Twin Health uses wearables (e.g., CGMs, smartwatches) to build a virtual representation of a person's metabolic state and recommend diet and activity changes; in a randomized trial of 150 people with type 2 diabetes, 71% on Twin's program achieved an A1c < 6.5% at one year, versus 2.4% with standard care.

Such digital twin–aided lifestyle interventions and adherence guides could also benefit rare disease patients, especially pediatric populations, by enabling more personalized and potentially more engaging interventions. In one study, children with asthma showed high acceptance of digital twins for day-to-day disease management.

From Data Scarcity to Regulatory Frameworks: Unique Challenges and Opportunities for Digital Twins in Rare Disease

A key challenge for generating AI-based digital twins in the rare disease area lies in the availability of data for training deep learning models. Even for more common diseases, training datasets rarely exceed 5,000 patients, which is small for deep learning. This problem is exacerbated in rare diseases, where limited case numbers make it hard to build large, balanced datasets of multidimensional, multimodal, and longitudinal data needed for reliable training. Moreover, data access can be challenging, as sources are often fragmented and lack interoperability. Sparse data can also introduce bias and raise privacy concerns for rare disease patients. To extend toward rare disease digital twins, researchers aim to aggregate and curate relevant data and, where necessary, generate disease-specific datasets. An alternative route to address data sparsity in the rare disease space is the use of synthetic data. In a 2023 study, researchers applied a conditional generative adversarial network to create synthetic cohorts for the rare hematologic malignancies myelodysplastic syndromes (MDS) and acute myeloid leukemia (AML). Their synthetic data mimicked the clinical features of real patients while preserving privacy.

Another way to overcome data challenges is to employ models better able to work with smaller, noisier datasets. For example, a recent publication introduced a large language model dubbed TWIN-GPT to create patient-specific digital twins. The authors reported that the LLM with its embedded clinical knowledge could establish cross-dataset associations from limited data, which enabled simulation of individualized disease trajectories and improved outcome prediction while preserving privacy.

The development of successful digital twin strategies in the rare disease space will require supportive policy, shared infrastructure, and privacy-preserving data sharing initiatives, for example through federated learning, as well as clear regulatory frameworks that outline how to validate and qualify digital twin technologies.

In Europe, several initiatives strive to build a coordinated digital-twin ecosystem for the healthcare space. These include the EMBL-EBI/Open Targets project for building rare disease digital twins; the European Commission’s Virtual Human Twins (VHT) Initiative launched in 2023 with over €125 million in funding and over 76 stakeholders and the EDITH project, which will provide a repository, and future simulation platform needed for credible, multiscale twins. While the VHT and EDITH initiatives do not solely focus on rare diseases, the infrastructure they create will open new opportunities for rare disease research by providing multi-scale, data-rich simulation platforms that enable researchers to model complex disease dynamics.

Regulators recognize the potential of digital twin technology; as of August 2025, there is no FDA or EMA regulation dedicated solely to digital twins in drug trials. However, the EMA’s CHMP has provided a Qualification Opinion for Unlearn’s PROCOVA™. PROCOVA™ is a statistical method that allows Phase 2/3 trials with continuous endpoints to adjust the primary analysis using prognostic scores derived from models of patients’ outcomes, which are often implemented via digital twins. Since then, the EMA has outlined their AI Action Plan, which reinforces expectations on transparency, robustness, data governance and human oversight when AI models inform clinical evidence. Additional regulations on AI for medical use are captured under the EU AI Act.

In the U.S., the FDA has not yet qualified digital twins as a tool, but in a Center for Drug Evaluation and Research (CDER) discussion paper from 2023, which was updated in 2025, the agency detailed current and future uses of digital twins, with favorable views on the potential of the technology to speed up drug development and use in placebo arms. According to FDA case discussions, the agency may allow the use of prognostic-score–based digital twins in clinical trials—initially in exploratory analyses during Phase 2, and potentially as the primary analysis in Phase 3—provided that the model demonstrates strong credibility, validation, and representativeness of the target patient population.

While these initial signals from regulatory agencies appear favorable for implementation of digital twins, as they will become more widely used, clear regulatory guidance will be invaluable to assure patient safety and trust by both investigators and patients. This might also include mandates on model explainability, which are discussed in the context of clinical AI use.

References

Topic: Next-Gen Tools