Digital Human Twins: Transforming the Future of Human Health

2 December 2025

We are approaching a paradigm where human-based in vitro systems are not only human-relevant experimental platforms for understanding disease mechanisms and developing new drugs, but also computationally integrated avatars of patient physiology.

What makes digital human twins (DHT) a priority to pursue is the realisation that human diseases are heterogenous in their etiology and that the patients’ response to treatments are shaped by their age, sex, genetics, lifestyle, and health status.

The long and arduous road of developing new drugs is fraught with difficulties, not least of which is the inability of animal models to recapitulate human-specific physiology, disease etiology and patient-specific variations, causing high failures rates in clinical trials.

By integrating the in vitro, clinical and real-world -generated multi-modal data on patient physiology and disease, that is processed and analysed using AI/machine learning, digital human twins can predict the patient response to compounds that is validated against real patient data, enabling dynamic updates based on new inputs.

For example, predictions based on biobank patient/cohort-derived in vitro data can be validated against clinical data from the same patient/cohort. In vitro studies can also be run alongside a clinical trial in a prospective parallel study, in which data collected in patients – plasma, tissues, biomarkers, histology, pharmacokinetics/pharmacodynamics (PK/PD) - are used to directly validate or refine the in vitro model. 

Personalized digital twins can be powerful tools for improving patient outcomes, minimizing side effects, continuously monitoring and adjust treatments, enabling in effect the long-sought personalized medicine.

In the meantime, digital human twins can also be semi-personalized. For instance, a robust digital human twin model trained and validated on large-scale biobank and clinical cohort data can serve as baseline for predicting responses to a drug for a specific patient, by parameterizing the digital human twin model with patient-specific in vitro-derived data on drug efficacy, pharmacokinetics and toxicity.

Crucially for the development of entirely new untested compounds, population-validated digital human twins can also reliably predict outcomes for a new class of drug. Eliminating preclinical animal testing, this massive feat can be achieved by analysing structure-activity relationships (ChEMBL, PubChem database), ligand-based target prediction (machine learning) and pathway enrichment analysis to infer affected biological processes (human-based in vitro multi-omics – genomics, epigenomics, transcriptomics, proteomics, metabolomics). Drug response simulations can be run across virtual digital human twin cohorts, to stratify and identify patient populations that are most likely to benefit from the new drug. 

Once the digital human twin model has demonstrated predictive accuracy across a population, it can be used to predict effectiveness of response and optimal dosage for a drug for vulnerable patient populations, a pregnant female patient with diabetes for instance. This can be done by measuring in vitro the efficacy biomarkers, drug metabolism rates and toxicity thresholds in patient-derived tissues, by identifying outlier responses compared to the population digital human twin model, and by re-parameterizing the digital human twin population model to reflect pregnancy and diabetes physiology (glucose metabolism, hepatic metabolism, renal clearance) using published physiologically based pharmacokinetic (PBPK) models.

And while digital human twin is still in its infancy stages, the latest technologies and initiatives are converging rapidly toward this goal:

• Generating data

Digital human twins will need data, not just any data but high-quality multi-modal human data. Given thousands of species-specific differences between animals and humans from molecular to organism level, the integration of data from animal experiments would have deleterious effects on predictive efficacy of this technology. Human-based in vitro platforms are ideally suited for providing experimental human and patient-specific multi-omics, histology, functional and PKPD data. Biological, microbiome, health status, behavioural, lifestyle, drug response, and Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) data are collected in clinical trials, while portable sensor devices and electronic health records ensure real-time access to a trove of relevant patient data.

• Managing data

In the EU, the European Health Data Space (EHDS) operates under Regulation (EU) 2025/327, enabling access to high-quality, anonymised/pseudonymised datasets via EU member states’ Health Data Access Bodies. In this manner, the EU directly supports research, innovation and building digital twins, while ensuring compliance with data protection regulations. While the US approach to managing patient data does not rely on a single centralized data space like it is the case in the EU, it operates within frameworks of Health and Human Services’ HIPAA protected health information privacy rules, FDA’s guidance for Computational Modeling, Simulation, and Digital Health, and FDA’s Real World Evidence program, that work with de-identified datasets and align with digital twin development. 

• Bridging the bench to the bedside

By recapitulating human-specific physiological responses at molecular, cellular, tissues, organ and multi-organ level, human-based in vitro devices are uniquely equipped to relate experimental to clinical/real-world outcomes. Both DHT and in vitro systems can be customized to represent specific diseases, specific patients and specific circumstances, allowing for personalized therapies. Both methods complement each other, as bioengineered human tissues enable high-throughput testing of compounds that is not feasible in humans and human digital twins provide the organism-level context that completes the picture.

• Processing and analysing data

DHT can leverage a combination of AI supervised, unsupervised, self-supervised and reinforcement machine learning. For example, to predict the patient’s response to a drug starting from patient-derived in vitro tissues, DHT can use deep neural networks with supervised learning, in which the model maps input patient-derived multimodal data to known clinical outcomes from similar patients or tissues. Unsupervised learning can group patients based on similarity in their multi-omics profiles and can infer the underlying co-expression gene modules or patterns of activity of biological pathways. By testing different compound types and dosing strategies on the DHT, the reinforcement learning agent can adapt the therapy to changing patient states as to maximize the health outcome.

• Scaling up

Large-scale projects have constructed resources for various human cells and tissues for biomedical research, allowing the use of reproducible datasets that are comparable across labs.

By building a comprehensive gold-standard reference map of every cell type in the human body across over 9 000 donors of diverse ages, origins, genetic backgrounds and health states, including disease-specific variations in gene expression profiles, the Human Cell Atlas supports inference of biological pathways necessary for prediction of drug effects and personalized medicine.

Unlocking the potential of high resolution analysis of multi-omics, detection of heterogeneity within tissues, and identification of cell-specific drug targets, single-cell technologies are increasingly used. For instance, by advancing multiplex barcodes systems, mass spectrometer functional protocols and data processing software, the non-profit research organization Parallel Squared Technology Institute has succeeded in significantly scaling up proteome profiling of human cells.

The UK Biobank has built the world’s most comprehensive dataset of biological, health and lifestyle information to date from 500 000 individuals, available for health-related research by global public and private organisations.

Operating at the intersection of microfluidics, robotics and multimodal AI, Vivodyne has the capacity to run functional analyses of human drug responses on thousands of lab-grown fully-functional human tissues. Its fully automated robotic workflow ensures reproducibility at AI-scale throughput, generating human-relevant multi-omics data that is processed using supervised and unsupervised learning to identify new disease targets, model their druggability, and evaluate efficacy and toxicity of therapeutic candidates.

As the same time as companies advance on developing customized digital human twins, this concept is already strategically implemented at the EU level through projects like Virtual Human Twins which ambitions to build a standardized, interoperable ecosystem with emphasis on regulatory-grade validation, cross-border data sharing, and integration with EHDS.

Given that digital human twins integrate a variety of distinct technologies, its advancement requires collaboration across disciplines - from bioengineering and medicine to mathematical modelling and AI. This convergence of skills and methods supports a shift from poor animal model proxies of human diseases to patient-specific physiology simulations and from empirical trial-and-error to data-driven predictions, accelerating development of innovative therapies and democratising access to personalized medicine.