Back in the 1990s, the idea of running computer-simulated clinical trials would have seemed far-fetched. The internet had only just emerged, personal computers were the size of suitcases, and...
No clinical trial is perfect. The gold-standard randomized controlled trial does a good job of trying to be, but we know its flaws: selection isn’t truly random, conditions often don’t reflect patients’ real lives, and such studies capture only a fraction of the individual variation across a population.
How about simulating clinical trials on computers, using models of disease and a virtual population that encompasses all our glorious differences? In silico modeling uses mathematical equations to represent our biological pathways and investigate how a new drug affects them.
It may seem a little arrogant – or crazy, even – to assume we can reduce our complex selves to a series of differential equations. Yet for decades, drug developers have used mathematical pharmaco-kinetic and pharmaco-dynamic (PK/PD) models to represent the passage of a drug through the body (PK) and the drug’s effect on our biochemistry and physiology (PD). This modeling helps identify effective and safe dosages, among other things. It is a key part of regulatory submissions.
As more of our lives are quantified and digitalized – from our health records to our genome and all the physiological readings in between – we are ever-better equipped to extend PK/PD modeling to cover a wider range of biological systems. Scientists understand better than ever our genetic and molecular workings. Technology and processing power has made it easier to capture more data from more people, helping fuel and refine models of health and disease.
In real trials, the filter is the fallible human investigator, whose trial design, patient selection and/or data collection may be more or less accurate. In an in silico clinical trial, the filter is methodological . Is this differential equation sufficiently refined to capture what’s going on (and predict what will happen) in a particular organ or system? Has enough of the right data been used to calibrate and validate the model?
Validation is a particularly thorny challenge facing the wider adoption of in silico clinical trials or trial simulation. How closely does a model need to replicate real-life results for it to be considered robust? How wide is the acceptable margin of error between real and simulated data, given the imperfections of both approaches?
These are all questions that drug regulators are grappling with right now. FDA in particular understands the potential of predictive clinical modeling, is eager to welcome it into the approval process and is evaluating Model Informed Drug Development programs.
As more in silico disease models are developed and tested, and as pilot schemes like the FDA’s Model Informed Drug Development (MIDD) evolve, answers will emerge to some of the challenges listed above. One important strength of some predictive modeling platforms (NOVA’s included ) is transparency: you can trace back the source data or assumptions for any given reading. Such sourcing is much harder for a human filter. Errors and biases are difficult to pinpoint.
Simulated trials won’t replace real ones. But by modeling the effect of a drug on a broader range of physiology (purporting, in some cases, to represent a human), they can help design more targeted clinical studies that generate more predictive data. They offer a complementary approach to the in-person trial – one that allows more hypotheses to be tested on an unlimited number of virtual patients, enabling enriched decision-making and reducing timelines and costs.
So many questions about drug efficacy and safety are not – and cannot be – fully addressed by standard clinical trials. Using both real and in silico trials, each with their strengths and weaknesses, provides a powerful way forward for drug R&D.