Before the First Pill

The Patient Who Never Had to Exist

The Johns Hopkins pilot trial, completed in early 2026 with FDA approval, did not use a new drug. It used a new idea: that before you treat a patient, you can treat their digital replica first. The replica — built from the patient's cardiac imaging, electrophysiology data, and medical history — had already been through the procedure hundreds of times. The surgeon came to the operating room knowing where to aim.

This is the promise of the digital twin in medicine. Not a generic model of a human body, but a patient-specific mathematical simulation — one built from your data, trained to behave the way your biology behaves, and available to absorb the risks of experimentation that would otherwise fall on you. The twin fails first. You learn from its failures. You take the drug, or the surgery, after it has already proved itself — in silico.

The technology has been gathering force for a decade in research labs and pharmaceutical back offices. In 2024, it broke into the regulatory mainstream. In August of that year, the FDA published its first formal guidelines for "in silico clinical trials" — a document acknowledging that mathematical simulations of patients could, under the right conditions, substitute for actual human subjects in portions of the drug development process. In January 2026, the FDA and the European Medicines Agency jointly published ten Guiding Principles for Good AI Practice in drug development, with digital twins listed as a priority application area.

The pharmaceutical industry had been waiting for exactly this kind of regulatory signal. The average drug takes 12 years and $2.6 billion to bring to approval. Roughly ninety percent of drugs that enter Phase I clinical trials — the first time they are tested in humans — fail before they reach patients. A technology that could filter for failure before the first human subject is enrolled does not merely save money. It saves lives — the lives of people enrolled in trials of drugs that were always going to fail.

How You Build a Body in a Computer

A digital twin is not a diagram. It is not a simplified stick-figure approximation of human physiology. It is a computational model that attempts to replicate the dynamics of a specific biological system — an organ, a tissue, or eventually a whole person — with enough fidelity that simulations run on it produce results that predict what would happen inside the real body.

The data inputs are extensive. A cardiac digital twin might draw on three-dimensional imaging from an MRI scan, electrophysiology measurements, genomic data that identifies variants affecting ion channel behaviour, blood pressure and heart rate histories from wearable monitors, and the results of previous drug responses. A cancer twin might incorporate tumour genomics, proteomics, the immune microenvironment, and prior treatment history. The richer the data, the more accurate the twin.

One of the most mature examples of this approach is the Living Heart Project, a collaboration launched in 2014 between Dassault Systèmes, the FDA, and more than a hundred academic and industry partners. The project's goal was to build a validated computational model of the human heart that could be used to simulate the effects of drugs and medical devices. By 2024 that model had been validated against clinical data and used in regulatory submissions for cardiac devices. The FDA accepted the results as partial evidence of safety and efficacy — the first time a computer simulation played a formal role in cardiac device approval.

The Unlearn.AI approach focuses not on organs but on the most expensive part of any clinical trial: the control arm. Standard randomized controlled trials require a group of patients who receive a placebo — not because researchers want to deny them treatment, but because without that comparison group, you cannot know whether a drug's apparent effects are real. Control arms are ethically uncomfortable, statistically necessary, and expensive. Unlearn builds AI-generated digital twins of each patient in the trial — a prediction of what that patient's disease trajectory would have been without the drug — and uses those predictions to augment or partially replace the real control arm.

A 2025 study published in Alzheimer's & Dementia, co-authored with AbbVie, found that using digital twin controls in Alzheimer's disease trials could reduce the required sample size by approximately 14.5 percent while maintaining statistical power. A separate analysis using a Phase 2 trial dataset found that a digital-twin-augmented trial required 1,855 participants to detect a 25 percent disease-slowing effect at 90 percent power — versus 2,170 in a standard trial. The difference is not trivial. Each patient enrolled in a clinical trial is a human being navigating a serious disease. Fewer required subjects means faster trials, lower costs, and patients in control arms who have a better chance of receiving active treatment.

From Population to Person

Modern medicine has always been built on population averages. Clinical trials recruit hundreds or thousands of patients with a shared diagnosis and ask whether a drug works across that population. The result is a drug approved for "patients with Type 2 diabetes" or "patients with metastatic breast cancer" — statistical generalizations that obscure vast individual differences in genetics, metabolism, immune function, and disease progression. The drug that works for sixty percent of the population fails the other forty. Nobody knows in advance which group you are in.

The aspiration that a computer could predict individual drug responses has a long history in pharmacology. Pharmacokinetic/pharmacodynamic (PK/PD) modelling has been used for decades to predict how drugs move through the body and what they do when they get there. These models describe population behaviour well and individual behaviour approximately. The jump from "approximately" to "specifically" is what digital twin technology attempts to make — by feeding in individual patient data, by using machine learning to calibrate the model to the individual, and by running simulations at a scale and speed that population trials cannot approach.

The companies now working in this space span the full range of the drug development pipeline. Unlearn.AI works on clinical trial augmentation for neurological diseases — Alzheimer's, Parkinson's, multiple sclerosis — where the control arm problem is particularly acute, trials are long, and the patient population is vulnerable. Turbine, a Budapest-based company, builds tumour simulation systems that test oncology drug combinations in a virtual cancer cell before anyone decides which patients to enroll. inSilico Medicine uses generative AI to design novel drug molecules and then simulates their behaviour in digital patient models before synthesis. Novadiscovery models individual patient disease progression to predict who will respond to which treatment.

None of these companies are fully replacing human trials. What they are doing is changing where failure happens — moving it earlier in the pipeline, when the costs are lower and no human being has been harmed by a drug that was always going to fail.

The regulatory trajectory matters. The FDA's August 2024 guidelines for in silico trials did not approve digital twins as a wholesale replacement for human subjects. They established the conditions under which computational models could produce evidence acceptable to regulators — requirements for validation, for transparency, for uncertainty quantification. They also signalled that the agency was prepared to accept this evidence in certain contexts, particularly for devices, and increasingly for drugs. The joint FDA-EMA principles of January 2026 extended that signal to the European regulatory space. The international alignment is deliberate: a drug developed with digital twin evidence in the United States needs to work in Europe too.

The End of One-Size Medicine

If digital twin technology matures as its proponents expect, the implications for medicine are not incremental. They are structural. The entire architecture of drug development — the long march through Phase I, II, and III trials, the decade of attrition, the enormous cost that ends up baked into drug prices — rests on a specific epistemic problem: we do not know, before we test a drug in humans, how a human will respond to it. Digital twins do not eliminate that problem. But they move its resolution upstream.

The first shift is in the composition of clinical trials themselves. If digital twin controls can reliably substitute for a portion of the placebo arm, trials get smaller, faster, and cheaper. Smaller trials can reach statistical significance with more targeted patient populations — which means the drug can be designed and tested for a specific genetic subgroup rather than a diagnostic category that lumps together people with very different underlying biology. The promise is not one drug for everyone but the right drug for you specifically.

The second shift is in what counts as a drug candidate in the first place. If you can simulate a drug's behaviour in a virtual patient before synthesising it in a laboratory, you can run thousands of candidate molecules through the model and discard the ones that fail before anyone makes them. This is already happening. inSilico Medicine's AI-designed drug for idiopathic pulmonary fibrosis entered Phase II trials in 2023 having been identified, designed, and pre-screened computationally in ways that would have taken years by conventional methods.

The third shift is harder to quantify but perhaps the most important. Clinical trials, as currently designed, systematically exclude people whose biology is complex or whose circumstances are difficult — elderly patients on multiple medications, pregnant women, people with rare genetic variants, patients in low-income countries without access to trial centres. A virtual patient population has none of these exclusions. You can build a digital twin of a 78-year-old woman on seven concurrent medications and run the drug trial in her model as easily as in a 45-year-old man in perfect health. The drug you approve can actually reflect the diversity of the people who will take it.

The Johns Hopkins cardiac twin trial is a glimpse of what this looks like at the level of an individual patient. A person with a dangerous arrhythmia, lying in a hospital bed, while a computer somewhere runs ten thousand simulations of their heart. The surgeon arrives knowing where the problem is. The procedure takes two hours. The patient goes home.

That is not science fiction. That trial already happened. What is still being built is the infrastructure — regulatory, computational, ethical, and clinical — that will allow it to happen for everyone, for every disease, before the first pill is prescribed.