AI’s next pharma challenge: tackling costly and risky clinical trials

The Digital Medicine department at the University of Bern among other universities are increasingly using AI to analyse patient tissue samples. Keystone / Gaetan Bally

AI has transformed the search for new medicines. Now drug companies are betting it can fix the slow, expensive process of testing them in humans.

Ask a pharmaceutical company about the current approach to clinical trials and you’re likely to hear a chorus of complaints about a broken and dysfunctional system. It takes on average at least a decade and roughlyExternal link $1-2 billion (CHF790 million-1.6 billion) to develop a new drug. Around 60-70% of that time is spent in the three phases of human testing. The longer a clinical trial takes, the longer it takes for that medicine to reach patients.

To make matters worse, companies often waste time and money on trials for drugs that aren’t ultimately approved by regulators. For every 100 drugs that begin human trials, about 90 aren’t given the green light because the tests fail to prove the medicines are safe or effective.

“The industry has just accepted that risk and failure are a part of drug development because there has been no other option,” said Kevin Buyens, co-founder and Chief Business Officer at AI biotech TwinEdge Bioscience in western Switzerland. “But this is changing. Digital technologies can help fill the gap.”

Artificial intelligence, especially machine-learning systems, have already helped scientists identify promising drug candidates by sifting through huge volumes of chemical and biological data far faster than humans.

The situation is different in clinical trials. Regulators like the US Food and Drug Administration (FDA) and Swissmedic use information collected in trials to decide whether a drug is safe or effective enough to be given to patients. Regulators have been wary of any shortcuts that could jeopardise patient safety.

But in the last few years, AI technologies have dramatically improved alongside better access to so-called real-world data, such as electronic health records. This has increased confidence in AI’s ability to speed up trials and improve their success rate.

In April, the FDA, the world’s largest medicines regulator, announced plans for a pilot programmeExternal link to assess how AI-enabled technologies can improve efficiency, speed and the quality of decision-making in early-phase clinical trials.

While limited in scope, the initiative is one of the clearest signs yet that regulators are becoming more open to the use of AI in clinical trials to unblock what the FDA describes as a “critical bottleneck” in drug development. It also comes as China is emerging as an important force in early-stage clinical trials because of its ability to conduct them faster and cheaper.

Why trials take so long

Clinical trials have traditionally been a long, drawn-out process for a host of reasons. One major factor is simply the large amount of paperwork, much of which is done manually. The situation has gotten worse as trials are more global, drugs are more complex, and authorities are demandingExternal link more information to assess a therapy’s risks and benefits.

Illustration: Kai Reusser, Swissinfo

Almost half of the time between a Phase 1 trial and an FDA application for approval is “dead time” consumed by paperwork and data transfers, accordingExternal link to former FDA Commissioner Marty Makary. Finding eligible patients and sites is also a problem – studies findExternal link that 80% of trials experience significant delays due to challenges enrolling patients.

Scientists also struggle to predict how patients will respond to a drug. Many trials recruit patients who end up not benefiting from a new drug, which leads to higher failure rates and millions of dollars wasted on unsuccessful testing.

A stark example is Alzheimer’s disease. A studyExternal link found that from 1995 to 2021, some $42.5 billion in private money was spent on over 1,000 clinical trials for Alzheimer’s drug candidates, with 57% of that money spent in Phase 3 trials. Some 184,000 participants took part in the trials for the 235 drug candidates, 95% of which failed to be approved.

AI to the rescue?

Over the past decade, a growing number of companies and start-ups have been trying to figure out how AI can tackle some of these problems. One key area that has seen widespread uptake is the use of AI to write protocols – the detailed study plans researchers use to design and run clinical trials. Writing one typically takes 6-12 months.

Swiss startup RisklickExternal link created Protocol AI, a software that uses generative AI to develop multiple scenarios for a study design. The tool has been tested by Swiss biotech firm Debiopharm, which is now integrating it across the company. Risklick says the tool can reduce the time and cost involved in developing clinical trial protocols by up to 35%.

AI is also starting to be used to improve patient recruitment. US start-up Paradigm Health has developed an AI-powered platform to help pharmaceutical companies identify eligible trial participants. It relies on electronic health records from a vast network of cancer clinics.

Other AI-powered platforms such as Switzerland-based Ancora.aiExternal link and TrialGPT developed by the US National Institutes of Health are empowering patients to find trials for their specific condition and city. StudiesExternal link have found TrialGPT matches patients to trials almost as accurately as human experts in 40% less time. Since 2022, over 23,000 people from 101 countries have used Ancora to search for cancer trials.

Biology AI company Owkin, which has offices in Geneva, has built deep-learning models that are being used to predict patient outcomes but also identify what factors influence those outcomes. It has now signed partnerships with several big pharma companies to customise these AI models for clinical trial decisions.

Swiss pharma giant Novartis, which runs studies that involve thousands of people, has built what it calls the Intelligent Decision System, which is like a computational twin of the entire clinical trial process. The system can simulate parts of a clinical trial to compare likely outcomes at different site locations and troubleshoot potential problems, such as a site dropping out of a trial.

“Every day you can gain in a clinical trial is one less day the patient has to wait for a drug, so you want to speed up trials as much as possible,” Robert McGregor, who leads the company’s Intelligent Decision System, told Swissinfo. “If I change an assumption in my study design, it can completely change the timeline.”

Pharmaceutical companies have been more cautious about using AI in ways that could directly influence regulatory decisions. This includes computational models that simulate how virtual groups of patients might respond to a drug, which could help predict trial outcomes before human testing begins, or even reduce the number of patients needed in some trials.

At an even more personalised level, AI can create digital twins – virtual, dynamic replicas of humans using a mathematical model and a combination of genomic, clinical, and imaging data from real patients. Researchers hope these models can predict how individual patients respond to a drug or even simulate what happens if they receive a placebo instead.

Some companies have started to develop these for research purposes.

TwinEdge Bioscience, founded in 2025, is developing digital twins of individual patient tumours based on molecular data from cancer patients. The company is partnering with pharma companies to test hypotheses about drug targets and truly addressable patients.

SOPHiA GENETICS, which was founded as a start-up out of the Swiss Federal Institute of Technology Lausanne (EPFL) in 2011 and went public in 2021, launched digital twin technology last October. The platform uses each patient’s unique clinical, biological, imaging, and genomic data collected from hundreds of hospitals to simulate treatment responses.

“As an industry, we need to embrace real-world data,” said Jurgi Camblong, the company’s founder. “With this, we can learn things about how to better prescribe or combine drugs, identify sub-populations that would benefit from a drug, and potentially validate that in a clinical trial.”

US-based firm Unlearn has partnered with pharma companies including Basel-based RocheExternal link to generate digital twins that model real patients’ responses to Alzheimer drug candidates based on real-world data. Results suggest digital twins can reduce control arm sizes by around 35%.

The hurdles to AI adoption in trials

Despite the potential benefits, AI’s ability to transform clinical trials still faces constraints. One is gaps in scientists’ understanding of human biology.

“If you understand biology, you can run better clinical trials because you know which population and disease you are targeting,” said Thomas Clozel, founder and CEO of Owkin. But, he added, “we still have little idea why patients develop Alzheimer’s or why some cancers return.”

Data bottlenecks are also an issue. More data is being generated but isn’t necessarily available for clinical research due to privacy laws, and much of it isn’t “AI-ready” – formatted in a way that an AI model can use.

Even if AI models perform well, regulators still need to understand how they reach their conclusions. That is difficult when many AI systems operate as a “black box,” generating information in a way that is opaque or hard to understand.

Nevertheless, regulators are increasingly open to testing new ideas, as reflected in the FDA’s April announcement. The European Medicines Agency has also signalled opennessExternal link to using digital twins for rare disease trials. Swissmedic is also actively engagedExternal link in the development of international AI principles for clinical trials.

“I’m optimistic that in the next few years the way we do clinical trials will change,” said Buyens. “When we see real validation studies of these technologies, it will be impossible for regulators to ignore them.”

Edited by Nerys Avery/vm/gw