September 11, 2024

By: Toni Shears

 

Duxin Sun, PhD, has dedicated his career to developing anticancer drugs, cancer vaccines and nanomedicines for cancer therapy. Yet despite significant advances in his work and the field over the last 30 years, a word that comes up frequently when Dr. Sun discusses his work is “frustration.” 

 

The cancer mortality rate has dropped by one-third, and the FDA has approved 250 new cancer drugs since 2000. Still, 95% of new drugs in development fail — and even then, only 30% of approved drugs extend survival beyond  2.5 months.

 

“Everybody works so hard and we know there are so many rich areas to exploit to make a drug successful,” says Dr. Sun, Charles R. Walgreen Jr. Professor of Pharmacy and Professor of Pharmaceutical Sciences. Yet it takes up to 15 years and costs $1 to $2 billion for a drug to win approval and reach the marketplace. 

 

“The frustration is that developing a drug takes so long, it’s so expensive, and the failure rate is so high. This is a problem everyone wants to solve,” says Dr. Sun, the Associate Dean for  Research at the College of Pharmacy.” 

 

He is convinced there should be a better way. For ten years, he has researched, tested, and advocated for a more focused approach. He is determined that University of Michigan researchers will lead the way to a more efficient, effective drug development process. 

 

The solution, he believes, uses artificial intelligence and machine learning (ML) to speed the process — as many researchers at the University of Michigan are already doing. His solution, called the STAR-guided ML system, uses these technologies in a targeted way. STAR (structure-tissue/cell selectivity-activity relationship) focuses on three overlooked but interdependent factors in the drug development process where the roots of drug failure lurk. 

 

It’s a big, multifaceted problem; to solve it, he has put together a broad multidisciplinary collaboration of experts from pharmaceutical sciences, medicinal chemistry, clinical pharmacy, biomedical and mechanical engineering, computer sciences, computational medicine and bioinformatics,  data science, robotics, clinical oncology, and pharmaceutical companies to work on solving this together.

 

 

Homing in on Root Causes

Developing a successful drug focuses on balancing three aspects: adequate efficacy and manageable toxicity at relevant clinical doses. Most drugs in development fail to win approval for two primary reasons: they lack the strength to produce the desired clinical effect (40% to 50%), or they cause unmanageable or extreme side effects (30%). 

 

Many criteria have been added to improve each step of the drug development process, but the 95% failure rate has not improved in the past 30 years, Dr. Sun notes. Given the current time and cost of the drug development process, he questions whether it is practical and productive to add more steps and criteria without streamlining unnecessary ones in the process. 

 

Instead, his proposed STAR-ML approach focuses discovery on three overlooked key points:

  • Potency and specificity of a selected drug molecule to the on-targets or off-targets, which determines efficacy in tumors at clinical doses
  • Tissue or cell selectivity through on-targets and off-targets, which determines adverse effects in the normal organs at clinical doses
  • Optimal clinical dose balancing of efficacy and safety as determined by potency/specificity and tissue/cell selectivity. 

Sun and his 16 coauthors, including colleagues from  the College of Pharmacy, Medicine, Engineering, Michigan Institute of Data Sciences, Rogel Cancer Center, Lancaster Life Science Group, Aurinia Pharmaceuticals, and Bristol-Meyer Squib Company outline the approach perspective article in the Journal of Medicinal Chemistry. 

 

 

Avoiding “Survivor Bias” 

Dr. Sun says that failure to focus on balancing these points is the root cause of candidate drug failures. “We think we’re picking the targets that are to blame in cancer, but there may be many targets,” he says. Likewise, researchers pick compounds that seem likely to have a clinical effect on those targets, but there may be many others that might work as well.

 

Researchers pick those targets or compounds based on the best available data, but the data may suffer from what he calls “survivorship bias.” He explains this with a metaphor from World War II. 

 

Airplanes returned from bombing runs with damaged wings and tails, so military experts reinforced these areas. However, hits to the engines and cockpit areas were more lethal, but those planes did not survive and never returned, so military leadership was slow to see that these were the more critical areas to protect since people would miss the data of lost planes of the fatal damage.

 

Likewise, in cancer drug development, everyone may have focused on many unimportant things to fix in the process but may have missed key aspects of drug development, he explains. That leads to suboptimal or wrong drug candidate selection, “with insufficient drug target validation and incorrect drug-like properties,” he writes in the perspectives piece. 

ML processes using five key features can help find the relevant targets and the most promising candidate drugs more efficiently and correctly — if they are homed in the right goals, he says. The key is to aim for high potency and specificity in disease tissue, with high selectivity that spares healthy tissue and allows a low dose, as shown in the quadrants below.  

 

“We use AI in every single step of drug development, but we do not address these root causes,” he says. “We need to use AI to do something very different.” 

 

 

 

 

Engaging All Expertise

To develop an effective STAR-guided ML system, “we need pharmaceutical sciences, biology, medicinal chemistry, clinical pharmacy, data sciences, engineering, robotics, computer science and machine learning, and the clinical oncology perspective. “It’s not possible for one department to do this,” Dr. Sun says. “Our goal is to bring everybody together, speaking a common language, with everyone contributing their necessary expertise to a common goal.” 

 

In the perspectives article, Sun and his 16 coauthors conclude that this approach may “improve success rates sevenfold from 5% to 35% by achieving 70% success in each of Phase I/II/III trials and enhance the efficiency of cancer drug development,” and also improve efficiency by reducing the time and cost involved.

 

That is a goal almost everyone can get behind; he says his frustrations with the failures in drug development are widely shared across these disciplines. “When I talk to people about STAR-ML, everybody says, ‘I completely agree there is a problem with too many drug failures, and I want to solve this problem,’” Dr. Sun says. About half of those he has talked with see strong value in his approach; the others ask why his approach will be the best one. 

 

There may be room to improve STAR-ML, he says, “But let’s start with debate. We need to raise awareness of the failings in the current system, not to repeat the same failures over and over, and find ways to solve this problem.”