For many years, the industry has streamlined the maintenance of technology products, such as jet engines, by creating a simulated version of the object on the computer, known as a digital twin.
By running simulated actions on the digital twin, engineers can achieve feats that would not be possible in the real world, such as testing what conditions might make an engine fail, and how long it might take for failure to occur. The idea is to predict these conditions and avert them.
Efforts have been underway for years to bring the same idea of digital twins to life sciences.
A digital twin can simulate on the computer the structure of an organ, such as a lung, heart, or brain; the areas of a cell where drugs can bind to treat disease; or the entire state of health of a patient in a clinical trial.
The development of those simulated objects can be tracked over time to anticipate organ failure, drug viability, or patient outcomes.
Also: Hooking up generative AI to medical data improved usefulness for doctors
Part of a broader phenomenon known as "digital twins for health," the use of digital twins to advance drug development is one of the most promising uses of the technology. While still early, data from startups and drug companies suggest digital twin simulation can speed up drug development and shave costs, thus providing a clear life-saving and economic incentive.
"Digital twins are the future of medicine," Charles Fisher, founder and CEO of seven-year-old AI startup Unlearn AI said in an interview with ZDNET. "They basically replace our entire medical system; you basically turn medicine into a predictive science, you run experiments, and you optimize patients' treatments."
Sean McClain, founder and CEO of Absci, which is finding novel therapies using digital twins combined with generative AI, is similarly bullish about decreasing the approval time for new drugs, increasing success rates, and lowering costs. "You do all that, and you start to enable personalized medicine."
By any measure, drug development needs an overhaul. Creating new drugs, or even repurposing old ones comes with an enormous cost. A new drug takes, on average, 10 years to develop, from fundamental chemistry through clinical trials to regulatory approval. It can cost almost $3 billion, and the failure rate of most new drug candidates is 96%.
While the need is great, the digital twin efforts are gaining steam from the increasing opportunity afforded by artificial intelligence, especially generative AI.
Also: How GenAI got much better at medical questions - thanks to RAG
Digital twins have massive data requirements for their highly detailed simulations. Generative AI's proven ability to consume and model such mountains of information has provided a new tool in the quest to build ever more sophisticated simulations.
The concept of digital twins was outlined almost 20 years ago in the realm of product development, or, more specifically, "product lifecycle management." The impetus was making better products.
Originally, it was just a digital replica of the asset in question, a simulation on the computer akin to an avatar. Over the years, the digital twin has evolved from something static to something dynamic, being updated and improved with new data in a kind of feedback loop.
Such a replica can be applied in numerous ways to health. There is no single digital twin, but there are different forms a twin can take based on what is being replicated or modeled.
A digital twin, for example, can simulate hospital processes to predict resource utilization. In 2020, scientists at Oregon Health & Science University explored such an application to predict the usage of ventilators during the COVID-19 pandemic.
Digital twins can also simulate individual organs of the body to predict organ failure or disease progression, as is the case with The Living Brain project at the Icahn School of Medicine at New York's Mount Sinai Hospital.
However, it is in the exploration of new drugs that the digital twin has shown the most promise in recent years. The market for tools to speed up drug discovery and development via digital twins is estimated by market research firm Visiongain at a little over $300 million this year and is expected to increase by 31.3% annually, compounded, over the next decade.
Digital twins are being amplified by the use of artificial intelligence, which can aid the construction of simulations. A white paper released last year by the US Food and Drug Administration framed digital twins as an application of AI.
"Digital or computerized 'twins' of patients can be used to model a medical intervention and provide biofeedback before patients receive the intervention," Patrizia Cavazzoni, director of the FDA's Center for Drug Evaluation and Research wrote in the paper.
The application of AI to digital twins, said Cavazzoni, "can help bring safe, effective, and high-quality treatments to patients faster."
McClain's Absci is trying to revamp the front end of drug development, where the discovery of drugs happens. The company's AI-driven software tools, combined with its own wet lab, are a virtual reinvention of laboratory procedures.
The digital twin, in this case, involves simulating the structure of a protein to test what happens if one selectively changes the amino-acid bases.
In a sense, Absci is making specific digital twins of real-world molecules. However, the vision is much broader: personalized medicine.
"You could take a patient sample, have an AI be able to predict what target you should be going after for that particular patient, that particular disease, and then be able to design an antibody to treat it," he said.
Today, the company uses generative AI "to predict antibodies from scratch that can bind to a target of interest," he explained.
In March, Absci reported the development of novel antibodies against what's called "human epidermal growth factor receptor 2," or HER2, a human oncogene that has been linked to some forms of breast cancer.
The AI model had been fed no data on existing, successful antibodies against HER2, and no explicit information about how to successfully attach to -- or "bind" to -- HER2.
Also: Generative AI could lower drug prices. Here's how
McClain called it a breakthrough. Traditionally, scientists in a wet lab would use an animal's immune system to generate an antibody. With generative AI, the antibody can be created as a computer model.
"We've gone from this paradigm of searching for a needle in a haystack to now being able to create the needle," he said.
Absci's lead drug candidate in its pipeline, ABS-101, is a treatment for irritable bowel disease. The novel antibody, developed using generative AI, binds to the TL1A protein in immune cells whose over-expression has been linked to a variety of inflammatory autoimmune diseases. The antibody was developed from scratch in just 14 months and at a cost of less than $5 million.
"You are starting to see these AI-generated antibodies and small molecules make it into the clinic," said McClain.
Using another metaphor, if the antibody being designed is the key to a lock, the lock itself -- the part on the antigen where the antibody should attach itself -- remains something of a mystery.
In the next three to five years, said McClain, the goal is to "actually start to predict where an antibody should bind to a target to give us the biological response that we want -- to start to predict the biology."
That is going to take more biological data and require feeding that data into the digital twin to revise and improve the picture of both lock and key.
In contrast to the front end of the process with Absci, digital twins are supplementing the clinical trial process by simulating human participants.
As the FDA white paper frames the matter, "AI/ML can be utilized to build in silico representations or replicas of an individual that can dynamically reflect molecular and physiological status over time.
"In comparison to a participant in a clinical trial that received an investigational treatment, the digital twin could potentially provide a comprehensive, longitudinal, and computationally generated clinical record that describes what may have happened to that specific participant if they had received a placebo."
Toward that end, drug giant Bayer has used simulations to "predict the level of blood glucose that successfully informed insulin dosing." Bayer has developed digital twins to simulate control groups, or those taking a placebo, which it says can replace the use of humans as controls.
These fledgling efforts are leading to new tools for trials.
Bayer released an open-source software project to advance those kinds of simulations, called Open Systems Pharmacology. The company has also partnered with venture-backed startup Altis Labs of Toronto, Canada, which develops software specifically tailored to employ AI trial simulations. Such creations are limited, Bayer noted, but a lack of "robust" data on which to base the formation of the digital twins. Another limitation is the lack of a standard by which regulators can accept the results from such simulations.
Starutp Unlearn AI uses forms of deep learning AI to extrapolate from patient data to predict outcomes that will happen over time, and how they may change with interventions. It's a simulation of the drug trial, in other words.
A 2019 study by Unlearn AI into Alzheimer's disease modeled likely outcomes for 1,909 patients for whom cognitive data were gathered for 18 months from an existing database called the Coalition Against Major Diseases database.
Unlearn AI's neural network, trained on the data, was able to "forecast a patient's future state," in the sense that it can take part of the 18-month data and predict how multiple factors, such as changes in a patient's ability to recall words, will shift.
In a position paper published this month in the journal Clinical Translational Science, Unlearn researchers said that "AI-generated DTs [digital twins] can help to ameliorate many of the troubles that plague clinical trials, such as their exorbitant cost, lengthy process of experimentation, and high failure rate."
One approach is to use past clinical trial data to "create a probabilistic forecast" of how a participant in a new trial may respond to treatment. That, in turn, "can be used in a variety of ways to optimize clinical trial design." A digital twin may, for example, convince regulators to allow for a smaller control group in a trial, which reduces the time and cost needed to gather participants.
The company has received $130 million in venture capital backing to take such models of patients into clinical settings. Unlearn is currently involved in clinical trials with parties Fisher could not name. There are no results yet to report from those trials.
"If we could predict how a patient could respond, you would be able to run CT without control groups at all," he said. Having zero human controls, thereby reducing the number of people that need to be recruited, would "save an enormous amount of time and money, and dramatically speed up medical innovation."
It would also benefit patients because they would all receive the therapy being tested, rather than some being given a placebo. "That's why patients participate in the first place: to get access," he said.
Fisher foresees clinical trials having progressively fewer control participants over time.
"Today, we have the ability to decrease from 25% to 50%" of a control group's normal size," which can shave four to six months off of the typical enrollment time for a trial. A lot depends on the particular indication and therapy being tested in trials, he said.
The models Unlearn creates give the company mathematical certainty, said Fisher, that it can predict what sort of control a trial needs.
Similarly, for rare diseases or pediatric cases, digital twins can help when it's hard to gather enough participants. Unlearn AI wants to use historical control data to augment the sample size of smaller trials.
The company is working on ways to develop more sophisticated digital twins using generative AI. In May, it described in a pre-print research report what it calls Digital Twin Generators, a kind of factory for making digital twins of trial participants.
The AI model is trained on data for 13 different disease indications from three distinct therapeutic areas, including neurodegenerative conditions such as Alzheimer's, immunological afflictions such as rheumatoid arthritis, and general medical issues such as hypertension.
The model can ingest and learn from a vast set of data, including completed clinical trials, observational studies, and registries gathered from 156 clinical trials and 33 observational studies, totaling 400,000 patients and more than 3.8 million patient visits.
As with Absci, the ultimate goal is to use more fine-grained patient data to arrive at personalized medicine. That would, as Fisher put it, turn medicine into a mostly mathematical optimization problem.
"A lot of our focus in the long term is on the R&D aspects for unlocking the ability to go from clinical trials into these other larger areas," said Fisher. "That requires us to move to more general models of disease, a more general digital twin."
The "more general" digital twin will depend upon broad "foundation models" for health, he said, that can handle many diseases, not just one. And that, in turn, will require future advances in AI that he predicted will go beyond the limitations of today's transformer-based generative AI.
Unlearn AI's patient model is part of a bigger trend in using AI to predict clinical trial outcomes.
Startup Insilico Medicine offers a software program called InClinico, which uses generative AI to predict outcomes of Phase II clinical trials, which is the part where initial efficacy is determined. The company says the technology can predict with 79% accuracy how a Phase II trial will play out.
Such efforts have nascent support from regulators. The US Food and Drug Administration in 2022 granted an emergency use authorization for a drug called anakinra, known by its commercial name of Kineret, to treat COVID-19-induced pneumonia and prevent severe respiratory failure.
Because the clinical trial that was originally used to develop anakinra was not available in the US, the FDA based its approval in part on neural network predictions of how US patients would be likely to respond if they had been given the test.
ZDNET independently reached out to the FDA for further comment on how digital twin technology in drugs can more broadly impact consumers. We will update the story after speaking with them.
The quest for digital twins that can design antibodies or simulate clinical trials is a decade-plus endeavor, said Absci's McClain.
As the Kineret example at the FDA shows, events such as COVID-19 can inject a measure of urgency that can speed the use of new technology. Even so, there's a limit to how fast data can be gathered and digital twin simulations can be put through the learning loop.
Just as generative AI models like OpenAI's GPT-4 require massive data gathering, the biological data, and the clinical trials data will take years yet to gather, process, and feed into models.
As Unlearn AI pointed out in its position paper, "Electronic health records (EHRs) contain an abundance of information acquired at the point of care, but are collected in a wide variety of formats and often include poorly labeled diseases and conditions."
That means that "existing AI models are constrained by both the availability of these data and by the algorithm's ability to learn from these records due to the inherent heterogeneity of disease processes."
Data gathering of such scale will raise ethical and regulatory issues concerning patient privacy that have yet to be articulated much less addressed. On top of all the ways that people have fretted for decades about how medical records should be managed, patient biological data will likely be an even thornier issue, both because of the scale of the data, and because of the more complex ways it may now be used to train AI models.
If using Wikipedia was controversial to book authors, how would humans feel about being digital twins' "ground truth?"
Also: Dana-Farber Cancer Institute finds main GPT-4 concerns include falsehoods, high costs
Blockchain technologies are one possible answer to patient autonomy and privacy. The blockchain could conceivably offer managed, authenticated access to all the medical data being gathered in a Pap smear, X-ray, or sonogram. Such implementations have yet to be designed, however.
Generative AI in life sciences also requires greater scrutiny, not just in the data but in the design.
That fact is obvious in the initial work by medical professionals testing generative AI. Researchers at the renowned Dana-Farber Cancer Institute in Boston described in April how several months of implementing OpenAI's GPT-4 for operations produced inaccurate responses from the chatbot.
The many obstacles including data gathering, learning from the output of digital twin simulations, and refining the AI itself, mean that it's going to take an entire ecosystem to make digital twins reach that goal of personalized medicine, said Absci's McClain.
"At the end of the day, it's a team sport," he said. "Not one company is going to solve this, it's going to be all of us as a collective."
