By Natalie Grover and Martin Coulter
LONDON (Reuters) - Major drugmakers are using artificial intelligence to find patients for clinical trials quickly, or to reduce the number of people needed to test medicines, both accelerating drug development and potentially saving millions of dollars.
Human studies are the most expensive and time-consuming part of drug development as it can take years to recruit patients and trial new medicines in a process that can cost over a billion dollars from the discovery of a drug to the finishing line.
Pharmaceutical companies have been experimenting with AI for years, hoping machines can discover the next blockbuster drug. A few compounds picked by AI are now in development, but those bets will take years to play out.
Reuters interviews with more than a dozen pharmaceutical company executives, drug regulators, public health experts and AI firms show, however, that the technology is playing a sizeable and growing role in human drug trials.
Companies such as Amgen (NASDAQ:AMGN), Bayer (OTC:BAYRY) and Novartis (SIX:NOVN) are training AI to scan billions of public health records, prescription data, medical insurance claims and their internal data to find trial patients - in some cases halving the time it takes to sign them up.
"I don't think it's pervasive yet," said Jeffrey Morgan, managing director at Deloitte, which advises the life sciences industry. "But I think we're past the experimentation stage."
The U.S. Food and Drug Administration (FDA) said it had received about 300 applications that incorporate AI or machine learning in drug development from 2016 through 2022. Over 90% of those applications came in the past two years and most were for the use of AI at some point in the clinical development stage.
ATOMIC AI
Before AI, Amgen would spend months sending surveys to doctors from Johannesburg to Texas to ask whether a clinic or hospital had patients with relevant clinical and demographic characteristics to participate in a trial.
Existing relationships with facilities or doctors would often sway the decision on which trial sites are selected.
However, Deloitte estimates about 80% of studies miss their recruitment targets because clinics and hospitals overestimate the number of available patients, there are high dropout rates or patients don't adhere to trial protocols.
Amgen's AI tool, ATOMIC, scans troves of internal and public data to identify and rank clinics and doctors based on past performance in recruiting patients for trials.
Enrolling patients for a mid-stage trial could take up to 18 months, depending on the disease, but ATOMIC can cut that in half in the best-case scenario, Amgen told Reuters.
Amgen has used ATOMIC in a handful of trials testing drugs for conditions including cardiovascular disease and cancer, and aims to use it for most studies by 2024.
The company said by 2030, it expects AI will have helped it shave two years off the decade or more it typically takes to develop a drug.
The AI tool Novartis uses has also made enrolling patients in trials faster, cheaper and more efficient, said Badhri Srinivasan, its head of global development operations. But he said AI in this context is only as good as the data it gets.
In general, less than 25% of health data is publicly available for research, according to Sameer Pujari, an AI expert at the World Health Organization.
EXTERNAL CONTROL ARMS
German drugmaker Bayer said it used AI to cut the number of participants needed by several thousand for a late-stage trial for asundexian, an experimental drug designed to reduce the long-term risk of strokes in adults.
It used AI to link the mid-stage trial results to real-world data from millions of patients in Finland and the United States to predict the long-term risks in a population similar to the trial.
Armed with the data, Bayer started the late-stage trial with fewer participants. Without AI, Bayer said it would have spent millions more, and taken up to nine months longer to recruit volunteers.
Now the company wants to take it a step further.
For a study to test asundexian in children with the same condition, Bayer said it plans to use real-world patient data to generate a so-called external control arm, potentially eliminating the need for patients taking a placebo.
That's because the condition is so rare in the age group it would be difficult to recruit patients, and could raise concerns about whether it was ethical to give trial participants a placebo when there are no proven treatments available.
Instead, Bayer aims to mine anonymised real-world data of children with similar vulnerabilities.
Bayer said it hoped that would be enough to help discern how effective the drug is. Finding real-world patients by mining electronic patient data can be done manually, but using AI speeds up the process dramatically.
While unusual, external control arms have been used in the past instead of traditional randomised control arms where half the participants take a placebo - mainly for rare diseases where there are few patients or no existing treatments.
Amgen's drug Blincyto, designed to treat a rare form of leukaemia, received U.S. approval after adopting this approach, although the company had to conduct a follow-up study to confirm the drug's benefit once it was on sale.
Blythe Adamson, senior principal scientist at Roche subsidiary Flatiron Health, said the advantage of AI was that it let scientists examine real-world patient data quickly, and at scale.
She said it could take months to trawl through data from 5,000 patients using traditional methods: "Now we can learn those same things for millions of patients in days."
OVERESTIMATION RISK
Drugmakers typically seek prior approval from regulators to test a drug using an external control arm.
Bayer said it was in discussions with regulators, such as the FDA, about now relying on AI to create an external arm for its paediatric trial. The company did not offer additional detail.
The European Medicines Agency (EMA) said it had not received any applications from companies seeking to use AI in this way.
Some scientists, including the FDA's oncology chief, are worried drug companies will try to use AI to come up with external arms for a broader range of diseases.
"When you're comparing one arm without randomization to another arm, you are assuming that you have the same populations in both. That doesn't account for the unknown," said Richard Pazdur, director of the FDA's Oncology Center of Excellence.
Patients in trials tend to feel better than people in the real world because they believe they are getting an effective treatment and also get more medical attention, which could in turn overestimate the success of a drug.
This risk is one of the reasons regulators tend to insist on randomised trials as all patients believe they are getting the drug, even though half are on a placebo.
Gen Li, founder of clinical data analytics firm Phesi, said many companies were exploring AI's potential to reduce the need for control groups.
Regulators, however, say that although AI has the potential to augment the clinical trial process, evidentiary standards for a drug's safety and effectiveness will not change.
"The main risks with AI are that we want to make sure we don't get the wrong answer to the question of whether a drug works," said John Concato, associate director for real-world evidence analytics in the Office of Medical Policy in the FDA's Center for Drug Evaluation and Research.