In silico clinical trials use computer simulations of human physiology to test drugs, medical devices, and treatments on virtual patients rather than, or alongside, real people. The approach draws on computational models that mimic how the body absorbs a drug, how bone responds to an implant, or how a tumor interacts with the immune system. These simulated trials are already influencing regulatory decisions and, in at least two documented cases, have helped accelerate drug approvals by replacing traditional control groups with synthetic ones built entirely from data.
How Virtual Patients Are Built
A virtual patient is a computer-generated representation of a real human body, or a specific organ or tissue, that responds to simulated interventions according to biological and physical rules coded into a mathematical model. There is no single recipe. Virtual patients can be generated as digital twins through statistical inference or by randomly assigning physiological parameters and then checking whether the result looks biologically plausible. The choice of method depends on study goals and available data.1PubMed Central. Virtual Patients in Clinical Trials for Drug Development: A Narrative Review At one end of the spectrum, researchers build a single high-fidelity digital twin matched to a specific person’s medical records. At the other, they generate entire populations of thousands of virtual subjects whose traits span the full range of human variation.
Generating a complete population is more than just randomizing numbers. One computational approach uses global search algorithms driven by statistical model checking: starting from a quantitative model of human physiology plus drug behavior, and incorporating biological and clinical knowledge from experts, the software produces a population of virtual patients whose responses cover the entire spectrum of possible phenotypes the model can produce, with each virtual patient distinguishable from the others by user-defined criteria.2Bioinformatics. Complete populations of virtual patients for in silico clinical trials The goal is to avoid blind spots: if a real population would include someone with an unusual combination of liver function and body weight, the virtual population should include that person too.
Increasingly, these populations are generated with the help of machine learning trained on historical clinical trial data and real-world health records. Rich datasets from past trials allow AI models to produce holistic forecasts of how an individual patient’s health might unfold, creating what some researchers call AI-generated digital twins.3PubMed Central. Increasing acceptance of AI-generated digital twins through clinical trial applications The richer the historical data feeding the model, the more realistic and clinically useful the resulting virtual cohort becomes.
Predicting Drug Behavior Before Anyone Takes a Pill
One of the most mature applications of in silico trials is in pharmacology, where physiologically based pharmacokinetic (PBPK) models simulate how a drug moves through the body: absorption, distribution, metabolism, and elimination. These models can incorporate differences across age groups, disease states, and genetic variants that affect drug-processing enzymes.4PubMed Central. Leveraging In Silico and Artificial Intelligence Models to Advance Drug Disposition and Response Predictions Across the Lifespan
A concrete example comes from the breast cancer drug tamoxifen. Tamoxifen’s effectiveness depends heavily on a liver enzyme called CYP2D6, and patients carry different genetic variants of the gene coding for that enzyme. Researchers built a PBPK model of tamoxifen and its metabolites and ran virtual versions of an ongoing clinical trial (called TARGET-1) that was testing whether patients with certain CYP2D6 variants would benefit from a higher dose. The virtual trial predicted that the probability of the real trial achieving its primary endpoint was only about 47% on average. But as the simulated population grew larger, that probability rose to roughly 67% at 260 virtual patients, and the analysis also revealed that wide variability in levels of the active metabolite endoxifen was dragging success rates down.5PubMed Central. Application of PBPK Modeling and Virtual Clinical Study Approaches to Predict the Outcomes of CYP2D6 Genotype-Guided Dosing of Tamoxifen That kind of insight, delivered before the real trial is even done, lets investigators adjust their strategy rather than discover a problem after years of enrollment.
Testing Medical Devices Without a Single Surgery
Drug trials get most of the headlines, but in silico methods have found an equally important home in medical device evaluation. For devices that interact mechanically with tissue or bone, finite element analysis (FEA) can simulate what happens under realistic loads and stresses. The approach replaces or supplements some of the bench testing and animal studies traditionally required before a device reaches patients.
One team used FEA to compare two nasogastric tube securement devices, modeling the strains each device places on facial tissue. The device held in place by a strap-style mechanism produced maximum tissue strains of 52% to 79%, while a hook-style device generated strains of 404% to 434%, several times higher across all tested loading conditions.6PubMed Central. Assessment by finite element analysis modelling of tissue strains associated with the use of two different nasogastric tube securement devices That gap would be difficult to detect in a short bedside trial but becomes obvious in simulation, where you can apply forces from every direction and measure the result instantly.
Orthopedic implants offer another case. Researchers developed and validated a finite element model of humeral stem stability in shoulder replacements. They first confirmed the model’s physics against bench-top micromotion testing, then expanded it to a population-based clinical model that could replicate a known clinical finding: grit-blasted stems carry a higher risk of loosening than porous-coated ones.7PubMed. Developing and Validating a Model of Humeral Stem Primary Stability, Intended for In Silico Clinical Trials Once validated, a model like this can test new stem designs against a virtual population of bones varying in density, shape, and loading pattern.
For high tibial osteotomy plates used in knee surgery, an in silico clinical trial compared a personalized plate design against a generic one. The trial simulated physiological activities at three healing stages and used a logistic model to calculate the odds ratio of fatigue failure between the two designs.8Communications Medicine. Personalised high tibial osteotomy has mechanical safety equivalent to generic device in a case–control in silico clinical trial Running this comparison in silico avoided the need for a randomized surgical trial to answer a question that was ultimately about engineering performance.
Hip replacements are getting the same treatment. A recent study simulated the wear of metal-on-plastic hip implants across five patient profiles ranging from sedentary elderly individuals to active young people, factoring in six daily activities including walking, stair climbing, and sit-to-stand transitions. The results showed that both the type and frequency of motor tasks significantly influence implant wear, with stair climbing and highly active patients producing the most wear regardless of age.9PubMed. How Patients’ Lifestyle Affects the Wear of Hip Implants: An In-Silico Study Standard lab wear tests use repetitive machine-driven cycles that look nothing like how people actually move through their day, so this kind of simulation fills a gap that bench testing leaves wide open.
Cardiac Safety and Cancer Immunotherapy
Drug-induced heart rhythm problems have historically been one of the leading reasons drugs get pulled from the market after approval. Catching those problems earlier is a big deal, and in silico cardiac models have made real progress here. The Virtual Assay software, for example, runs simulated drug trials across populations of human cardiac cell models. It has demonstrated accurate, mechanistic predictions of drug-induced pro-arrhythmic toxicity and has been adopted in both industry and regulatory workflows.10Journal of Computational Science. The virtual assay software for human in silico drug trials to augment drug cardiac testing Rather than relying solely on animal heart tissue, which beats differently from a human heart, the software tests compounds against virtual human cells that carry the ion channels responsible for rhythm.
In oncology, researchers have used in silico trials to study something that has puzzled clinicians about immunotherapy: its distinctive survival curve shapes. Unlike chemotherapy, where patients who respond tend to do so quickly and uniformly, immunotherapy produces delayed responses, durable long-term survivors, and a subgroup that progresses despite treatment. Three different mathematical models of cancer-immune dynamics, each built on different assumptions, all reproduced these distinctive patterns when used to assemble virtual patient cohorts undergoing immunotherapy, chemotherapy, or combination therapies.11PubMed Central. In silico cancer immunotherapy trials uncover the consequences of therapy-specific response patterns for clinical trial design and outcome The practical payoff is that researchers can now use these simulations to test which trial designs, endpoint definitions, and sample sizes are best suited for immunotherapy, rather than discovering mid-trial that a design optimized for chemotherapy is giving misleading results.
Why Pediatric and Rare Diseases Are the Strongest Use Case
If there is one area where in silico trials could have their biggest impact, it is pediatric rare diseases. These conditions affect tiny patient populations, many of them children, creating a painful collision between the need for rigorous evidence and the practical impossibility of enrolling enough patients in a traditional trial. Virtual patients offer a way forward. Machine learning and mechanistic computational approaches can generate cohorts that reflect the heterogeneity of these small populations, potentially maximizing the impact of trials that might otherwise be underpowered.12PubMed Central. Transforming Pediatric Rare Disease Drug Development: Enhancing Clinical Trials and Regulatory Evidence With Virtual Patients
One published example involved congenital pseudarthrosis of the tibia, a rare pediatric bone condition. Researchers ran an in silico trial on 200 virtual subjects, each simulated to receive either no treatment or bone morphogenetic protein (BMP) therapy. The trial showed that BMP significantly reduced the severity of the condition, though the effect was highly variable from one virtual subject to the next. Using machine learning, the team stratified the virtual population into adverse responders, non-responders, responders, and asymptomatic individuals.13Scientific Reports. In silico clinical trials for pediatric orphan diseases That kind of stratification is exactly what a real pediatric trial would struggle to achieve because there simply are not enough patients to fill each subgroup. The ethical dimension matters too: in silico approaches raise the possibility of refining, reducing, and ultimately partially replacing the need to expose vulnerable children to experimental treatments before the treatment’s behavior is better understood.14PubMed Central. Digital twins, synthetic patient data, and in-silico trials: can they empower paediatric clinical trials?
Synthetic Control Arms That Have Already Changed Approvals
One of the most tangible regulatory successes of in silico methods so far involves synthetic control arms. In a traditional randomized trial, some patients receive a placebo or the current standard of care, forming the control group. A synthetic control arm replaces that group with virtual patients whose outcomes are simulated based on historical and real-world data, meaning every real participant in the trial receives the experimental treatment. Beyond the efficiency gain, this eliminates ethical concerns about assigning seriously ill patients to a placebo.
Two real cases stand out. In 2015, a synthetic control arm of 68 virtual patients was created for alectinib, a drug for non-small-cell lung cancer. The approach accelerated its FDA approval and advanced insurance coverage in European countries by about 18 months. Separately, a synthetic control arm of 694 patients was used to speed up the approval of blinatumomab for a rare form of acute lymphoblastic leukemia.15PubMed Central. The Case for AI-Driven Cancer Clinical Trials – The Efficacy Arm In Silico These are not theoretical demonstrations. Real regulatory agencies accepted these synthetic arms and made approval decisions partly on their basis.
Mixing Virtual and Real Patients in Hybrid Designs
Most researchers do not see in silico trials as a wholesale replacement for human studies. The more realistic near-term picture is hybrid trial designs that blend virtual and physical patients. The central idea borrows from Bayesian statistics: the probability of a given clinical event is treated as a combination of the observational likelihood from real patients and a prior probability supplied by virtual patient predictions. A loss function monitors how many virtual patients can be added per real patient enrolled before the virtual data starts biasing the real-world signal.16Briefings in Bioinformatics. In silico clinical trials: concepts and early adoptions
This framework is especially appealing for the early phases of trials, where safety signals are the priority and patient numbers are small. If a virtual cohort can reliably predict which dose ranges are dangerous, fewer real patients need to be exposed to those ranges. Later-phase confirmatory trials would still rely heavily on real patients, but the hope is that the virtual component sharpens the design enough to make the real trial shorter, smaller, or both.
How Regulators Decide Whether to Trust a Model
None of this works if regulators cannot evaluate whether a computational model is good enough for the decision it is being asked to support. The field has converged on the ASME V&V 40 standard as the main framework for credibility assessment. The standard asks teams to define a specific “context of use” for the model, perform a risk analysis to set acceptability thresholds, and then work through verification (does the math do what we think it does?), validation (do the outputs match reality?), and uncertainty quantification (how wrong could the predictions be?).
Applying this framework to a biomechanical model used to predict fracture risk, one team found that the risk was judged medium and the credibility levels acceptable, though the single biggest source of uncertainty was in how material properties were assigned to the bone.17PubMed. Credibility assessment of computational models according to ASME V&V40: Application to the Bologna Biomechanical Computed Tomography solution In a different application, the same framework was applied to a computational model of hemolysis in centrifugal blood pumps, illustrating how the same model can require different levels of validation activity depending on its intended use and the associated risk.18PubMed Central. Assessing Computational Model Credibility Using a Risk-Based Framework: Application to Hemolysis in Centrifugal Blood Pumps
The key insight from the credibility framework is that there is no universal bar for “good enough.” A model used to screen out obviously toxic compounds early in development faces a lower credibility threshold than a model submitted as the primary evidence for a market approval. Risk drives the requirements.
Major regulatory agencies have taken steps to formalize their acceptance. The FDA has developed model credibility guidelines and AI-specific guidance. The European Medicines Agency promotes what it calls 3R guidelines (refine, reduce, replace) with explicit room for computational evidence. Japan’s PMDA supports computational validation through dedicated subcommittees.19PubMed. Regulatory Adoption of AI, ML, Computational Modeling & Simulation in In-Silico Clinical Trials for Medical Devices: A Systematic Review Regulatory credibility has also grown alongside advances in machine learning, multi-omics data integration, and predictive simulation, particularly through initiatives like model-informed drug development.20PubMed Central. In Silico Research Is Rewriting the Rules of Drug Development: Is It the End of Human Trials?
What the Models Still Get Wrong
For all the progress, the limitations are real and worth understanding. The most fundamental problem is that any simulation is only as good as the biology coded into it. Human physiology is not fully understood at the resolution needed for every disease, and simplifications that work for one organ system can fail spectacularly in another. A cardiac electrophysiology model might faithfully reproduce ion channel behavior but miss a metabolic feedback loop that matters at the whole-heart level. A PBPK model might nail drug clearance in a healthy liver but struggle with the unpredictable pharmacokinetics of a cirrhotic one.
Data bias is another concern. Virtual populations are generated from historical datasets, and if those datasets underrepresent certain ethnic groups, age ranges, or comorbidity profiles, the virtual population inherits those blind spots. Ethical and regulatory concerns around replacing humans with digital data, data privacy, and security remain unresolved and need to be addressed before virtual patient data can be adopted widely.21PubMed Central. Digital twins, synthetic patient data, and in-silico trials: can they empower paediatric clinical trials? A virtual population that looks statistically complete may still miss a subgroup that experiences a rare but serious adverse event, precisely the kind of signal that real-world surveillance is designed to catch.
There is also the question of whether these models can predict what they have never seen. Most validated models are tested against known clinical outcomes. Predicting the behavior of a genuinely novel drug class, one that acts through a mechanism not yet represented in the model’s equations, is a fundamentally harder problem. One early assessment did suggest impressive capability: a 2009 in silico method prospectively predicted the clinical potential of 156 drug targets that had no approved drug at the time. Eight years later, the predictions lined up well with actual clinical outcomes.22Trends in Pharmacological Sciences. Trends in the exploitation of novel drug targets But that success involved target-level predictions, not full trial simulations, and the gap between predicting whether a target is worth pursuing and predicting exactly how a trial will turn out remains large.
Open-Source Tools and Reproducibility
A practical barrier to wider adoption has been the lack of standardized, freely available software for running and validating in silico trials. Most early work used proprietary or custom-built tools that could not be easily inspected or reproduced by other teams. That is starting to change. An open-source web application called SIMCor, built in R using the Shiny framework, was recently released to support proof-of-validation for virtual cohorts and computer-based simulations. It offers a suite of standard analytic techniques for comparing virtual cohorts with real datasets, including one-variable, two-variable, and multivariate comparisons, and provides options for applying validated virtual cohorts in in silico trials.23Scientific Reports. An open source statistical web application for validation and analysis of virtual cohorts
Tools like this matter because reproducibility is the currency of scientific credibility. If only the group that built a model can run it, regulators and independent researchers have to take the results on trust. Openly available validation software lets anyone with the right dataset check whether a virtual cohort actually matches the real population it claims to represent. As the field matures, the expectation will likely shift from “show us your model” to “let us run your model ourselves,” and the infrastructure to support that is only now being built.