AI’s Drug Revolution, Part 1: Faster Trials and Approvals

This is the first in a three-part series from Medscape Medical News on the impact of artificial intelligence (AI) on drug discovery and development. Part 2 is about the use of AI to find new applications for existing drugs. Part 3 reports on the AI’s ability to create new proteins from scratch, streamlining the creation of protein-based therapeutics.

When patients with cancer aren’t responding to treatment, doctors often sequence their tumors, hunting for biomarkers that may point to more effective therapies.

This molecular profiling is rarely done at diagnosis — it’s costly and lacks evidence to support its influence on treatment, according to Christina Curtis, MSc, PhD, director of AI and cancer genomics at the Stanford Cancer Institute.

But for some patients, waiting to use this tool until cancer has already spread could be “a missed opportunity,” she said.

That’s the motivation for a new biomarker-directed clinical trial in patients recently diagnosed with certain subtypes of estrogen receptor–positive (ER+)/human epidermal growth factor receptor 2–negative (HER2−) breast cancer.

Over the past decade, Curtis and her colleagues have used machine learning, a type of AI that identifies patterns in large datasets, to do comprehensive molecular profiling of patients with ER+/HER2− breast cancer. It identified four subgroups of patients with a roughly 50% greater risk for recurrence for up to two decades after diagnosis.

The trial’s goal: To investigate different therapeutic strategies targeting genetic changes in patients at the earliest stage of their disease to see which might improve their outcomes compared with standard care.

“These patients are not recurring from their disease for decades. We have a huge window to do better and optimize their therapy and monitoring,” Curtis said.

How to do that? Typically, patients with ER+/HER2− breast cancer undergo some molecular profiling at diagnosis; gene expression–based assays assess their 5-year risk for recurrence. But by “training” computers to analyze tumor cells’ genomes and transcriptomes, the researchers uncovered telling molecular features about the subgroups — including distinct copy number–altered drivers. Some of these genomic alterations were present as early as stage zero, before cancer starts to spread.

Here’s the catch: None of this could have happened without AI combing through vast amounts of data to pinpoint those biomarkers and predict which existing drugs they might respond to.

“Increasingly, we’re learning that trials that have a biomarker component tend to be more successful,” Curtis said. “A lot of trials, unfortunately, don’t meet their endpoints, so the utility of biomarkers, which rely on some AI or machine learning component, is really critical.”

Incorporating biomarkers is just one example of how researchers are reimagining clinical trials with the use of AI. Other machine learning techniques are enabling more effective preclinical testing or helping researchers capitalize on early-phase clinical trial success to propel treatments through the development pipeline.

That may prove to be AI’s greatest advantage: Inhuman speed. Together, these advancements could deliver the most impactful therapeutics to patients faster than ever before.

Closing the Translational Gap With AI

A lot can go wrong during clinical trials. Many drugs don’t work, cause nasty side effects, or aren’t well absorbed by or excreted from the body despite years of effort and millions or billions of dollars invested in each drug before clinical trials even begin.

It’s a crucial and costly problem known as the translational gap, the discrepancy between effective preclinical lab research and real-world results.

“When you test the animal model, it’s a coin flip to whether it works in humans,” said Kim Branson, PhD, a senior vice president and global head of AI and machine learning at GSK (formerly GlaxoSmithKline).

About 30% of drugs fail during phase 1 clinical trials, while roughly one third of phase 2 drugs make it to phase 3, and only about 25%-30% of phase 3 drugs reach phase 4. Overall, just 1 out of 10 drug candidates that gets through animal testing succeeds all the way through clinical trials and gains regulatory approval.

AI is helping close the translational gap. During the early stages of drug development, AI algorithms are helping scientists whittle down thousands of potential drugs to the most promising candidates.

Take VeriSIM Life’s AI platform, BIOiSIM. The tool simulates drug interactions with biological processes and makes rapid predictions about efficacy and toxicity. It relies on “data lakes” filled with millions of compounds, biological data, physiologic features, and negative data points that explain why compounds may have failed in previous clinical trials.

And oh, the speed: Predictions emerge within minutes to hours, depending on their complexity. The drug candidates are scored based on factors like toxicity, potential off-target effects, and the impact of dosage changes on efficacy. Clients are also given recommendations, such as the types of animal and lab experiments they should use, according to Jo Varshney, PhD, founder, and CEO of VeriSIM Life.

Recently, a client wanted to assess thousands of oncology and neurology compounds to find the five least toxic and most effective options. One of the five molecules identified is now in phase 1 clinical trials.

“We saved them several years and multimillion dollars of investment within each program because we could run all of these virtual experiments in parallel,” said Varshney.

Similar efforts are underway around the world, inside pharmaceutical giants and start-ups alike. Researchers are using machine learning to sift through mounds of healthcare data, creating models that estimate the molecular properties of potential drug candidates. The models “are far from perfect,” according to Branson, but could help researchers decide which compounds they should abandon.

Fueling Clinical Trial Success With Digital Twins

Researchers are also growing organoids, models of organs or tumors, to more accurately represent both individual patients and patient populations. Made with cell and tissue samples from real patients, organoids grow three-dimensionally rather than flat in a dish like human cells.

Tests reveal how these organoids respond to different drugs, drug combinations, and dosages over time. As they treat the organoid, researchers collect omics data, illuminating the tumor’s genetic and molecular traits and developing an immune system profile. The organoids also undergo molecular proximity imaging, which assesses how treatment affects protein dynamics. These findings can be linked back to clinical (whole body) imaging of the tumor in the patient and its morphology and metabolism during treatment.

These data are used to create a “digital biological twin” or virtual representation of each patient’s disease. With these data-rich models, AI algorithms can potentially predict the effects of treatments on patients or populations, resulting in “much greater translational confidence” than animal models provide, Branson said.

“We’re starting to represent the heterogeneity of patients,” said Branson. “Every patient is different, and yet, usually when we’re testing and developing, we don’t have that difference encapsulated.”

GSK and researchers at King’s College London have been using digital biological twins and AI to home in on personalized cancer treatments. If all goes well, their work could help improve the dismal 20% response rate to immunotherapy treatments.

In a recent phase 2 trial led by Memorial Sloan Kettering Cancer Center of a GSK immunotherapy drug, all 23 patients with rectal cancer with a specific genetic mutation completely responded — all of their tumors disappeared, no chemotherapy or surgery required. Some organoid work had informed the trial. Now, researchers are testing new organoids of those patients, with the same and other drugs, in hopes of better understanding which patients would respond similarly well.

“Any time we explore other indications or broadening the patient population, it opens up the potential to positively impact the lives of more patients. And we’re quite sure we can expand it out further,” Branson said.

Improving Recruitment With Machine Learning

Patient recruitment can be a “big bottleneck” for clinical trials, said Danica Xiao, PhD, VP of AI for GE HealthCare. “If you don’t have enough patients, or the patients are not diverse enough to cover different groups, then it’s possible that the trial either gets delayed or canceled.”

To improve recruitment, researchers have been drawing on large language models, such as ChatGPT, to predict eligible patients for clinical trials from large online datasets — including previous clinical trials. But while these models can help, there is much room for improvement. The models lack expertise, and the data are not well curated.

“There’s a gap between the tool and the insights we need,” said Xiao. “Right now, those data are available, but it’s very hard to derive insights.”

Eventually, GSK researchers also hope to use digital biological twins to more effectively match patients to clinical trials, boosting their potential impact. And as large language models evolve, perhaps within the next few years, Xiao can envision a machine learning tool capable of automating clinical trial design, from recruiting participants to spotting potential safety issues.

AI could also help better predict treatment responses in marginalized populations, a problem due to clinical trials’ overreliance on White male participants and failure to report race or ethnicity data.

“As we learn those molecular signatures across diverse populations, we can understand their unique biology,” Curtis said.

Continue on to part 2 and part 3 of AI’s Drug Revolution.