Patient digital twins and in silico trials

The phrase “digital twin” covers two fundamentally different things, and telling them apart matters more than defining the phrase. One is a mechanistic model of physiology derived from conservation equations. The other is a statistical forecast of a patient’s trajectory learned from historical data. They work differently and they fail differently.

The mechanistic branch: physiology as a set of compartments

Physiologically based pharmacokinetics (PBPK) represents the body as compartments corresponding to real organs, connected by blood flow. Its parameters are not fitting coefficients but measurable quantities: organ volumes, perfusion rates, unbound fraction in plasma, tissue-to-plasma partition coefficients, and the abundance of cytochrome P450 isoforms and transporters in liver and gut. That is exactly why such a model extrapolates into regions with no data: to predict exposure in renal impairment or in a three-year-old, you change a physiological parameter rather than refit the model. On that basis regulators accept PBPK to support drug-interaction and paediatric dosing conclusions — the FDA issued guidance on the format and content of these analyses in 2018. The larger version of the same approach is quantitative systems pharmacology, which adds mechanism of action at the level of signalling pathways.

The statistical branch: variance reduction, not a substitute control

The second approach takes historical control data, trains a model that forecasts an individual’s trajectory, and enters that forecast into the analysis of a randomised trial as a covariate. What is gained is specific. The forecast explains part of the between-patient variability in outcome, residual variance falls, and the required sample size drops at constant power. Randomisation stays, the control arm stays, and unbiasedness is preserved because the model is trained without sight of the current trial’s data or its allocation. The European Medicines Agency issued a qualification opinion on this method in 2022. A fully synthetic arm is a different proposition: it imports every bias of an external control, from shifting standard of care to differences in eligibility.

Where the limit runs

The formal limit is a single one and it applies to both branches: a model is credible only within the context of use for which it was calibrated and validated. There is an engineering formalisation of that requirement — ASME V&V 40 sets out how to scale verification and validation effort to the consequence of the decision the model informs, and it is used as the framework for medical devices.

Two consequences follow that product rhetoric usually blurs. First, identifiability: a mechanistic model with hundreds of parameters and a dozen observable quantities admits many parameter sets that fit the data equally well and extrapolate differently. Second, a virtual patient is not a source of evidence about a drug’s effect: it reproduces the distributions embedded in its training data, so a rare toxicity absent from the source population will be absent from the simulation too. A model can shrink a trial; it cannot replace the experimental comparison.