A real-world randomized clinical trial (RCT) is grounded in sampling theory and follows a rigorous methodology. In practice, this method involves recruiting eligible patients from participating sites based on specific criteria to form the trial sample (eligibility criteria). These patients are then randomly assigned to different treatment groups, or “arms,” whose outcomes are compared. Upon completion, the trial’s results are analyzed using statistical methods, assuming the sample comes from a larger theoretical population defined by the eligibility criteria. However, in reality, the trial group is often defined by the observed characteristics of the sample, as investigators may not perfectly adhere to the eligibility criteria.
Deviations from the ideal RCT can occur due to imperfections in the processes and tools used, such as breaches in randomization or the inability to create perfectly identical treatment arms. According to sampling theory, repeating the trial with the same protocol would likely yield different efficacy estimates due to variability in the patient population. The “true” efficacy is a theoretical concept, and any potential discrepancies between trial estimates and actual efficacy (which remains unknown, even if it could be defined) are in this case attributed to the sampling process.
The trial sample is assumed to be randomly drawn from an infinite “parent population,” allowing inferences about the broader population of interest, with, for instance, extrapolation to future patients. However, the actual process is inverse: the parent population’s structure is inferred from the observed sample, which can vary across repeated trials. Moreover, the parent population may not accurately reflect the real population of interest due to limitations in participating sites representativity of medical practices.
The efficacy estimated from a real-world RCT can be termed “gross efficacy,” as it includes residual factors of variability that are difficult to identify and quantify. These factors are linked to patient and context-dependent characteristics that are challenging to measure. The patients’ characteristics that matter are far to be all measured.
In contrast, an in silico clinical trial (ISCT) is a virtual experiment conducted on simulated patients. Like real-world RCTs, ISCTs aim for unbiased assessments of therapy efficacy but are based on physiological paradigms and mathematical representations of disease and treatment mechanisms. This approach, known as quantitative systems pharmacology (QSP), uses a simulated population that mirrors the real population of interest. Depending on the type of design, ISCTs can potentially provide more accurate predictions of “true efficacy” if the models and virtual populations are carefully validated.
There are two types of ISCT design. One mimics real world RCT with a random sampling in the virtual population, a random allocation of sampled virtual patients in the trial arms with one treatment scenario per arm. It allows to compute a gross efficacy estimate. In the other type, each virtual patient is his or her own control, whatever are the follow-up duration, the endpoints, the number of treatments to be compared. Thus there is no random allocation bias by design. Further, in this type, the group of patients on which the trial relies is not a sample but is the whole virtual population of interest taking into account all the diversity of eligible patients.
The former type predicts the “gross efficacy”. A number of trials can be run. Simulation of drawbacks of real world RCTs can be introduced in the process. The latter leads to a prediction of efficacy closer to the real efficacy. Below, we limit the paper to this type which makes the difference.
ISCTs address several issues that are difficult to manage in real-world RCTs, such as subgroup size, comparison power, and the multiplicity of statistical tests. They are run on the virtual population representative of the entire eligible patients, requiring no statistical inference or extrapolation. This approach also avoids ethical issues and practical limitations associated with patient eligibility and follow-up in real-world trials.
Moreover, ISCTs allow for the computation of “net efficacy,” which more closely approximates real efficacy by eliminating biases and variability factors present in gross efficacy estimates. Each virtual patient serves as their own control, and the protocol is fully mastered, considering all eligible patients’ diversity. Variables impacting efficacy are identifiable, whether they are measurable or not, and accounted for, providing individual benefit distributions rather than average figures.
Beyond “net efficacy” prediction, this type of ISCT enables to profile responders according to the true definition of a responder: “a patient who will suffer the event on control and will not suffer it on treatment”. THis is possible because each patient is his/her own control, whatever is the event.
While RCTs rely on statistical tests for decision-making, ISCTs make these tests irrelevant by basing predictions on a comprehensive representation of real patients at risk. However, methods to estimate the uncertainty of model predictions still need to be defined.
ISCTs are not a replacement for RCTs but rather a complementary approach.
