Authors: Prof. Jean-Pierre Boissel, Claudio Monteiro, Emmanuel Pham, François-Henri Boissel
In silico clinical trials (or “ISCTs”), as with any new approach, has to battle against persistent myths that delay large-scale adoption. In spite of the accumulation of use cases demonstrating the value of ISCTs, as well as regulatory agencies’ stated interest in this technology – notably the US FDA, there remain misconceptions which this article will attempt to clear up.
ISCTs consist in simulating a trial on a virtual population, following the desired protocol, by using a computer that applies mathematical models of the disease and the treatment of interest to each patient in this population.
Mathematical formalism is used to represent complex biological systems, where differential equations translate the dynamics of the process of interest (disease, treatment), enabling the disease and the treatment effects to progress in silico over time. All levels of physiological phenomena involved can be taken into account, from the genome and its expression to the clinical events affecting the patient. The approach called Quantitative Systems Pharmacology (or “QSP”) is rooted in the careful curation of scientific knowledge compiled in the scientific literature on the phenomena of interest.
While the model is deterministic, the second critical building block, which is the virtual population, introduces biological, clinical, and environmental variability, as well as patient diversity. This diversity determines variability – a statistical property of patient descriptor distributions – but it is also crucial in itself when considering patients and individual response to treatment. It is designed to mimic the real population at risk of the disease; hence. In some cases, when each real patient has his virtual counterpart, it is composed of digital twins.
ISCTs serve to fill gaps in knowledge about a new health product that conventional experimental methods cannot fill with real patients. ISCTs produce quantified predictions of efficacy (and safety) based on the profiles of virtual patients. These predictions are verifiable and have been verified. They are especially asked for in diseases where innovations are frequent and patients are rare (e.g., non-small cell lung cancer drugs targeting patients harboring EGFR exon20 mutations) or outcomes are quite infrequent over a long term of observation yet a burden for public health, making real-life clinical trials difficult to run.
Now let’s debunk the 7 most frequent myths about ISCTs.
It is neither realistic nor ethically desirable to imagine a future in which drug candidates are tested solely on virtual patients. You won’t hear any ISCTs expert make such a wild claim.
Insights derived from simulated trials can significantly improve decision-making as relates to the design of future trials on real subjects, making the whole process more efficient from a time, cost and ethical standpoints. By helping in streamlining the design of clinical trials and, for example proposing the optimal drug regimen, the approach further leads to decrease the number of patients and total duration of exposure. In addition, the modeling and simulation process makes observed treatment effects more understandable. As Singh et al. state in the Journal of Pharmacokinetics and Pharmacodynamics (2023): “The efficient discovery and development of a safe and efficacious drug requires a mechanistic understanding of physiological and biochemical processes that contribute to drug exposure and efficacy.”
What is likely to happen however is a future in which more trials are conducted on smaller patient populations thanks to the more precise characterization of optimal responder profiles. And this is what should make obsolete the dilemma between carrying out trials on a small and very homogeneous sample at the risk of losing generality or on a large sample, with unselected patients, with many patients exposed unnecessarily and leading to an indication that is not very selective. And in the shorter run, we have started to witness the emergence of hybridization strategies, whereby e.g. a conventional control group is augmented and expanded with virtual patients.
Statistical models, epidemiological or pharmacoeconomic models, as well as PK or PKPD compartmental models – all based essentially on data – have been around for a while.
The mathematical modeling of disease mechanisms – leading to what the US FDA calls in silico testing – is relatively new in the field. Publications of the first QSP models representing diseases from molecular phenotypes to clinical events date back to the beginning of the current century (In silico clinical trials: concepts and early adoptions, Francesco Pappalardo, Giulia Russo, Flora Musuamba Tshinanu and Marco Viceconti, Briefings in Bioinformatics, 20(5), 2019, 1699–1708).
Absolutely not! A map of all the biological entities and the functional relationships between them describing the various biological processes known to be implicated in normal and abnormal physiology form the basis of a QSP model we are speaking of here (Bridging Systems Medicine and Patient Needs J-P Boissel, C Auffray, D Noble, L Hood, F-H Boissel, CPT Pharmacometrics Syst. Pharmacol. (2015) 4, e26; doi:10.1002/psp4.26). So these models are based on knowledge, not raw datasets. A first graphical and textual version of the model is produced and discussed with experts in the field, e.g. hematologists. It is then converted into mathematical equations, which capture the dynamics over time, to enable simulations.
A QSP model of a given disease sits on the shoulders of giants, aggregating decades of scientific research.
This statement negates the vast body of literature available describing methods and processes to evaluate a model’s credibility and robustness. On one hand the model represents what is known in science and this knowledge integrates by itself the dynamics of the modelled process thus a part of the simulation output. The other part comes from the inputs that are individual virtual patient profiles. These patients come from a virtual population which is designed to represent a real population of interest. The virtual population represents the real population of interest by calibrating patient descriptor values on the basis of accessible and verifiable databases of real patients. Inputs are thus not decided arbitrarily by modelers. The entirety of the processes, from knowledge selection to data selection is traceable. Fraudsters attempting to achieve a predetermined result would be very easily exposed.
QSP places causal relationships at the center of the process of model design, making it mechanistic rather than correlation-based. Furthermore, by representing mathematically available knowledge on the pathophysiology of interest, a QSP model captures the disease dynamics as the current science has established it and response to treatment over time. This means QSP models are robust when extrapolating beyond existing datasets, which is arguably not the case with correlation-based statistical models.
From the colloquial example of Inuit dialects having a much larger number of adjectives to qualify snow than any other language, to the Standard Model’s failure to explain gravity, it is hard to argue there exists such a thing as a model perfectly representing reality.
The justification for QSP models originates from this trivial observation: the human mind is limited in its ability to assemble vast, interacting, quantitative, redundant, and retroactive pieces of knowledge. So while not perfect, models applying mathematical formalism to represent complex biological systems is better than any other alternative.
But proof is in the pudding: ISCTs have successfully predicted real outcomes, notably in the case of a non-small cell lung cancer drug developed by AstraZeneca (Flaura2 use case).
Quite the contrary. First, ISCTs make it possible to fill gaps in files submitted to Health Technology Assessment (HTA) agencies that real-life clinical trials may have left open.
ISCTs make it possible to predict more accurately long-term efficacy beyond the duration of a conventional trial, the effects on clinical outcomes when development could only be achieved with surrogate endpoints, to delineate the target population more precisely, to compare the treatment to comparators that could not be taken into account during development, and to explore interactions with established treatments (standard of care) when necessary.
Second, and this is still some way in the future, QSP models will eventually be deployed in medical practice to individualize treatment decisions based on a given patient’s profile.
In many aspects of life, human decision-making is supported by the predictive capabilities of mathematical models. While imperfect, these models compensate for our limited ability to account for the vast amount of accumulated knowledge in any given field.
ISCTs, given their ability to achieve superior outcomes when developing innovative therapies, will eventually become a staple of drug development. In a recent article, Bei et al. state that the number of QSP model submissions to the US FDA has more than doubled since 2020 (“Landscape of regulatory quantitative systems pharmacology submissions to the U.S. Food and Drug Administration: An update report”).
