By MARIN WOLF, The Dallas Morning News
Between major academic medical institutions, a slew of pharmaceutical companies and a new federal biotechnology hub, there’s no shortage of research studies taking place in North Texas.
But this latest trial is different: It didn’t use any animals or humans for testing.
In a Farmers Branch office over three months, North Texas biotech firm Praedicare completed a clinical trial for a drug candidate that could treat a nontuberculosis bacterial lung infection faster and more effectively than the current standard regimen.
Praedicare used an artificial intelligence-based program to analyze a drug in what the company is calling the first-ever virtual clinical trial. It’s a system that could save lives, money and time in a drug development process that takes around 10 to 15 years on average, said Praedicare’s CEO, Dr. Tawanda Gumbo.
The move toward AI has gripped every industry in recent years, although it is often more heavily associated with tech and business than with medicine.
Farmers Branch-based Praedicare sees AI models as a way to revolutionize clinical research by weeding out ineffective drugs before they make it to the expensive and lengthy regulatory approval process. Only 10% of drug candidates make it from phase 1 trials to federal approval.
“We can now utilize our vast libraries of patient profiles in tandem with wet lab models, mathematical models and AI to accurately predict whether molecules will achieve disease clearance, making the process of drug development faster and more effective, with less risk to patients,” Gumbo said.
In search of speed
All drug candidates that gain approval of the Food and Drug Administration follow more or less the same path. First, a drug is developed. Then, it goes to preclinical research, where it’s tested in animal or test tube models.
If the drug is shown not to be toxic during the preclinical stage, it can move to the four stages of clinical trials: phase 1, to evaluate safety and dosage; phase 2, to evaluate efficacy and side effects; phase 3, to monitor adverse reactions in a larger participant pool; and phase 4, to evaluate safety and effectiveness using several thousand volunteers.
Praedicare’s study, published in The Journal of Infectious Diseases, takes care of both the preclinical research and the first two phases of clinical trials. The research compared two treatments for Mycobacterium avium complex, a sometimes deadly lung infection caused by bacteria that live in soil and water.
Federal regulators designated the lung disease as an orphan, meaning it’s a condition for which researching new treatments isn’t considered profitable. The designation is meant to incentivize drug developers to target these diseases, an often difficult task given the financial and time stakes.
The current standard treatment requires three antibiotics over 18 months and sees a 43% cure rate. The new treatment is only one drug, an antibiotic called ceftriaxone, that Praedicare predicted to have a cure rate of 80% after six months of use.
Gumbo has spent years trying to speed up the drug testing process. A decade ago, the FDA and European Medicines Agency approved his hollow fiber system model, which tests drugs on human cells rather than in animals. It’s used to satisfy the preclinical requirements, and it can be more accurate than tests on mice, rats or chimpanzees.
“The biology of animals, especially small animals, is quite different from humans,” said Olivier Elemento, director of the Englander Institute for Precision Medicine at Weill Cornell Medicine. “These models use human data and capture human biology as opposed to animal biology.”
Praedicare evaluated the standard treatment on five MAC strains using the model. The effectiveness rates were the same in the hollow fiber system model as they were in previous human studies.
“That’s what gives us the confidence in this system,” said Gumbo, who previously worked at Baylor University Medical Center, UT Southwestern Medical Center and the Cleveland Clinic.
A promising start
The AI team, led by Gesham Magombedze, then used mathematical modeling of how a patient would respond to the new drug, looking at 1,000 simulated patients to predict its effectiveness and 10,000 simulated patients to find the best dosage.
Praedicare uses predictive rather than generative AI, meaning that the AI system is not creating new information, but is instead predicting trends using existing data.
The mathematical modeling section of the study replaces the first two clinical trial phases of the FDA regulation process, which can take anywhere from months to two years.
Praedicare has used the same predictive AI models in conjunction with traditional phase 1 and 2 trials and has found that its computer-generated results are essentially identical to those of lab researchers.
In total, Praedicare spent three months building and running the cell testing system and AI models for ceftriaxone, a fraction of the time traditional clinical trials require, Gumbo said.
Praedicare and some pharmaceutical partners are asking the National Institutes of Health for funding to move forward with a phase 3 trial.
Ceftriaxone is also an established antibiotic that’s safely used to treat other infections.
MAC is just one disease for which Praedicare has created a virtual model for drug candidates. It has similar models for obesity, Type II diabetes and nonalcoholic steatohepatitis fatty liver disease.
This isn’t the first time that AI has been used in drug development. Elemento at Weill Cornell Medicine said AI models are increasingly used alongside genomics to create treatments personalized to individual patients. What Praedicare did that was unique was applying a predictive model to a large number of simulated patients.
“But I think we’ll be seeing more of that, the combination of models and virtual patients, as the models become more and more reliable,” Elemento said.
“It doesn’t mean that they won’t have to do an actual clinical trial, but it means that you can essentially explore other parameters, the landscape of possible combinations of treatments, and make a statement about how often it’s likely to work in a patient population,” he said.
