As disease-related data increase, digital twin technology promises to impact precision medicine from clinical trials to clinical care—in silico
As healthcare increasingly becomes informed by the analysis—via machine learning and artificial intelligence—of patient-specific, or disease-specific data, a new tool is steadily making its way onto the scene: patient digital twins.
While this terminology may lead some to imagine a complete, in silico, digital recreation of a human being, a digital twin is rarely that involved. After all, scientists estimate there are more than 37 trillion cells in a human body, so digitally recreating this level of complexity is not a realistic consideration.
Further, the use of digital twins is not restricted to healthcare or even to twins in the human sense, they are simply digital models of real-world objects and are currently used across a broad swath of industries including manufacturing, aerospace, transportation, and food and beverage, among others. The term has been used for nearly 20 years, and some maintain that the first use of digital twinning—though it was years before the term was coined—was by NASA engineers who used flight simulators to mimic the conditions of Apollo 13 in order to devise a plan on how to return the crippled space craft to earth.
But just as the term “precision medicine” has operated under differing definitions depending on who you ask and their frame of reference, the definition for a digital twin, especially in relation to healthcare, drug discovery, and clinical trials is equally hard to pin down.
Eric Stahlberg, director bioinformatics and data science at the Frederick National Laboratory for Cancer Research and one of the authors of a paper in Nature Medicine entitled “Digital twins for predictive oncology will be a paradigm shift for precision cancer care” said he relies on the definition used by the cross-industry Digital Twin Consortium, which defines a digital twin as “a virtual digital model that represents a specific real-world system or object, and is updated in real time to continuously mirror the state of the real-world instance.”
That means something as small as a single cell can have a virtual digital twin that life scientists can use for research purposes, running the twin through multiple scenarios of perturbations to better understand how that cell would react under particular conditions; or a digital twin of a disease like cancer or Alzheimer’s disease which can be used to inform drug discovery, clinical trials, and clinical care; or even a digital twin of a human heart that can help model for an individual patient the specific characteristics of their cardiovascular disease, in silico, to be used to develop an effective treatment regimen.
Stahlberg, who is also a part of the Envisioning Computational Innovations for Cancer Challenges (ECCIC) noted that developing patient digital twins to improve care requires a multidisciplinary team. This team provides the expertise in computing and data science to create AI-driven cancer models to digitize treatment advice beyond that of the more traditional tumor board.
“It’s truly exciting to see the amazing energy that results when cancer scientists, biologists, computer scientists, data scientists, oncologists, software engineers, physicists, and mathematicians work together to imagine the future for cancer research and work together to find truly creative solutions that move cancer research forward and advance computing at the same time,” Stahlberg said.
Built on the backs of data
What is true for cancer care is also true for other health conditions as well including cardiovascular disease, central nervous system (CNS) diseases, and others. These areas are likely to see the most traction for digital patient twins, or digital tumor twins in the coming years, for one very simple reason: an abundance of publicly available data from past academic research and clinical studies that can be tapped to train the artificial intelligence algorithm of these diseases to effectively model how they will respond to potential new drugs or treatment modalities.
According to Charles Fisher, co-founder and CEO of patient digital twin company Unlearn, which works with drug developers to incorporate digital twin technology in their clinical trials, the company’s focus on CNS and autoimmune disorders is precisely because of the drug development efforts that have existed for years around Alzheimer’s disease and Parkinson’s disease in the CNS area, and also autoimmune diseases like rheumatoid arthritis, Crohn’s disease and ulcerative colitis, among others.
“Ahead of the work with clinical trials, we need to build the model that’s going to be used—we call them digital twin generators,” Fisher said.
To do this, Unlearn aggregates large disease-specific datasets—often built from a number of smaller datasets—and then evaluates these data on three parameters: The first is how many patients in the dataset; the second is how rich is the information and does it include diverse data like imaging, blood tests, and/or genomics data; and the third is how much follow up has been done on the patients.
“We’re typically looking for datasets where we’re getting a lot of data from patients that have this disease that is quite rich and with a long follow up, because we’re building time series models of patients,” Fisher said. At Siemens Healthineers, the approach of building deep datasets of patients is similar, but is applied in cardiovascular disease leveraging the company’s strength in imaging systems and technology. According to Peter Aulbach, head of technology and innovation marketing at Siemens Healthineers, advanced imaging, such as its computed tomography (CT) systems are powerful platforms for collecting data that can be analyzed by AI algorithms.
“We build assistive systems and patient twinning helps us with two major things: what is the next logical step either clinically, or what is the next procedure, to deliver a seamless patient experience” Aulbach said. “And we do this with evidence so there is less trial and error.”
Over the years, Siemens has invested heavily in improving the imaging capability of its platforms and as it has produced ever finer and more detailed images, AI has stepped in to see.. and catalog variations that are not discernible to the human eye—variations that have become biomarkers of disease. With each file weighing in at as much as two gigabytes, the company’s imagining platforms are generating data the company uses to train its models of cardiovascular disease.
“There is so much information hidden in the raw data of our machines, so is there a disease we can identify early enough, based on simulations of the DNA molecular information and CT information? Once we train AI to identify early signs of the disease progressing, we can alarm doctors to take precautions with their patients,” Aulbach said.
The use of digital twin technology is perhaps most advanced within the drug discovery and clinical trials arena, where its application promises to shorten timelines and lower the cost of bringing new drugs to market. While regulatory agencies like the FDA and EMA have already gone down the road of the use of algorithms for diagnostic tests and devices—information held in the so-called “black box”—currently, the agencies are still developing their own protocols for how and when the use of patient digital twins within therapeutic clinical trials may need to be regulated.
But according to Fisher, this may be a moot point. “The question that the regulators care about is will this clinical trial that is analyzed, including these data from the digital twins of the patients, will that give them the right answer? That’s really what they care about,” he said.
Leveraging digital twins of patients in this setting relies on two separate factors, Fisher continued. The first is the algorithm developed for the particular disease that will be addressed by the candidate drug; the second is how this algorithm will be used within the study. When used in a randomized controlled trial (RCT), Fisher said his company has mathematically proven theorems to demonstrate that the error rate of trials leveraging digital twins in the control arm is the same compared with standard RCTs. Even if the algorithm used underestimates, or overestimates, the treatment effect, the results can be squared by comparing the placebo group with the treatment group.
What happens is that if there is a bias in the model, that bias is in the treatment group,” Fisher said. “When I look at the difference between the treatment outcome and the predicted outcome, I see a bias, and when I look at the difference between the patient control group their observed outcome and the predictor, I see the bias. So when I estimate the treatment effect, that bias ends up cancelling itself out.”
The net effect of all this is the ability for drug sponsors to effectively run their clinical trials with fewer patients, which can significantly shorten the time for trial enrollment and save the sponsor money, while potentially bringing a drug candidate to market sooner.
The European Medicines Agency (EMA) has also weighed in on this approach. In March, EMA published a draft opinion validating Unlearn’s twin RCT approach, noting it is suitable for both phase II and phase III trials. At the time of his interview with Inside Precision Medicine for this feature, Fisher said he anticipated a final validation of the approach from EMA.
The FDA may not be far behind EMA according to François-Henri Boissel, co-founder and CEO of NOVA Discovery, a company focused on developing disease models and digital patients to streamline drug discovery and clinical trials. Speaking in a mid-September webinar hosted by Endpoints, Boissel noted that the FDA is in the process of working collaboratively with industry stakeholders and companies that create digital twins of patients to create a framework of expectations for their use and how to validate the disease models used for the digital twin creation. Using this information, it can assess how intensely each trial’s dosing will rely on these digital twins versus a traditional trial with all human subjects.
“If it’s just a marginal contribution in terms of information, most likely [the trial] is going to be approved—not the product itself and the drug candidate, but the use of in silico [patients],” he said. “If you’re really going for an aggressive use of an in silico approach, potentially, in certain circumstances, even going as far as replacing at least [some] human subjects, then you will need to face very stringent validation thresholds.”
While data is the backbone of building and training algorithms for the creation of digital twins of patients’ diseases to help drive precision medicine, the pathway is not without its hurdles. Top of mind for many working in the space is recognizing the limits of the technology today to model complex disease, but also ensuring that the data provided from twinning is unbiased and can help foster health care equity.
“When developing algorithms for digital twins in oncology, it is very important to recognize that these models will have limits in their ability to recreate the reality of the complex biology,” noted Stahlberg. “To make things even more challenging, the complexity only grows as one moves up the size scale from the molecular level to where one would envision the digital twin for an individual patient. Uncertainty will be inherent to any of the algorithms, so the challenge is how best to work with the underlying uncertainty while providing meaningful insight from the digital twin when making decisions.”
As the limits of the capabilities of patient digital twins are better understood, there also exists an opportunity to seek out the datasets necessary to allow these models to be more representative of the broader population. Aulbach said that he can see a future where there are multiple patient digital twins representing different ethnicities which will provide more equitable clinical care.
But first, the data—especially genomic data—must catch up.
“Good quality data are essential for developing and validating algorithms,” Stahlberg said. “Where there is an absence of data, there will also be an absence of digital twins, so it will be important for equity to assure that suitable data are available across all demographics.”
Verena Kallhoff, vice president, omics and precision and health director, data integrity and learning network at Equideum Health sees a much broader potential for patient digital twins in the future that can influence not just clinical trials or patient care, but also include methods to understand how patients go about their daily lives.
“We know it’s really hard to change people’s behaviors,” Kallhoff said. “But if we have that digital twin, we can understand how can we alter their behavior to actually make them take the drug, to find a time point that works in their daily schedule where it can be tacked on to something they already do.”
Kallhoff also noted that patient digital twin models could be the impetus for changing care paradigms at a population level and how health and wellness are managed. “There is an opportunity to decrease costs. For some countries, there may be an incentive, for instance, to say, ‘we know for Alzheimer’s disease, changing the eating habits and activity levels for these patients has a pretty big effect.’ This could let us figure out what and how we can implement it in a way that is sustainable for populations,” she said.
Data from wearable devices will also be layered on top of imaging, genomic, and other health data in the future to help create even more accurate digital twins of patients.
“You can imagine getting that data from wearables, and the big tech companies have millions of humans’ wearables data,” said Jorge Nieva, associate professor of clinical medicine at the Keck School of Medicine of USC. “Now, the trick is how do you link them? How do you link that human wearables data, to the medical record, to the tumor information to make really good digital twins? Because I think if digital twin technology is going to be effective, I think you need all three of those dimensions.”
Chris Anderson, a Maine native, has been a B2B editor for more than 25 years. He was the founding editor for Security Systems News, and Drug Discovery News, and led the print launch and expanded coverage as editor in chief of Clinical OMICs, now named Inside Precision Medicine. He is an avid supporter of the Boston Red Sox, the Boston Bruins, Aston Villa FC, and wishes U.S. professional sports would adopt the practice of relegation. In his spare time, he can be found with headphones on playing Beatles and Bob Marley songs on his Fender Jazz bass while dwelling on the hundred degrees of separation between his abilities and those of Pino Paladino. He also mixes the best Manhattan in the world—prove him wrong.