Not only was the drug able to inhibit all A. baumanni strains with known resistance, but it was specific to that bacteria alone.
Developing a new antibiotic can take as long as 15 years. But in as few as two, the target bug can develop resistance to it. Jon Stokes, Ph.D., a professor from Canada’s McMaster University, sees a world where antibiotics get to patients more quickly—with a little help from artificial intelligence.
“I think that [AI models], when appropriately trained and leveraged for various prediction tasks, can help increase the rate at which we discover new candidates for drug development,” Stokes told Fierce Biotech Research.
Alongside Massachusetts Institute of Technology Professor James Collins, Ph.D., Stokes has developed a machine learning algorithm that can screen thousands of compounds at once to see whether any of them have activity against a particular bacteria. The tech has yielded one discovery already, and, now, it’s made another: an agent that, at least in mice, appears to fend off antibiotic-resistant strains of the deadly bacteria Acinetobacter baumanni. The findings were published May 25 in Nature Chemical Biology.
“We believe that this is a tremendously powerful, efficient and accurate way to move drug development into a new era,” said Akhila Kosaraju, M.D., CEO at Phare Bio, a nonprofit biotech linked to Stokes’ and Collins’ labs. “It’s particularly well-suited for antibiotics, where you can quickly test what’s generated out of the AI in a petri dish and determine how efficacious those predictions are.”
Public health officials have been sounding the alarm about antibiotic resistance for more than a decade, but the problem is growing faster than scientists can come up with solutions. Annual deaths from
antibiotic-resistant infections in the U.S. have grown from 23,000 to 35,000 in the past six years, data from the Centers for Disease Control and Prevention show. Globally, the figure is closer to 5 million a year, according to the World Health Organization (WHO). The organization expects that to double to 10 million by 2050.
But developing novel antibiotics is an arduous process. Bacteria evolve quickly, and it’s tough to find new ways to target them that haven’t already been tried. Twenty novel
cancer drugs were approved by the U.S. FDA in 2022 alone, while roughly the same number of new antibiotics have been approved in the past 13 years combined.
Given the onerous task at hand, AI’s integration into antibiotic R&D feels inevitable. Stokes has been working on that since he was a postdoc in Collins’ lab at MIT, where he helped train and employ the lab’s machine learning algorithm for antibiotic discovery. The AI's first breakthrough in 2020 was the identification of an old
diabetes drug called halicin that was able to kill several different strains of multidrug-resistant bacteria in mice.
“That really launched our entire effort,” Kosaraju said. The finding led to an investment in Collins’ lab and the launch of Phare with funding from the TED Audacious Project. Now, researchers in Stokes’ and Collins’ labs perform the computational work using their AI platform then hand their hits off to a team at Phare to translate them into therapies. The translational arm helps refine the AI, too, by providing feedback Stokes can incorporate into the model to improve its predictive power.
“I can say to Jon, ‘Hey, I notice we’re getting a lot of this particular kind of compound coming out of the AI, and for industry standards, it would be more helpful to have X, Y or Z,’” Kosaraju said. “Then he can iterate on the platform, incorporating more drug-like characteristics into the training dataset.”
Big screen, narrow spectrum
In their latest study, the MIT and McMaster teams chose to focus specifically on A. baumanni, enemy No. 1 on the WHO’s list of priority pathogens and a deadly foe in
hospital-acquired infections. They screened around 7,500 molecules against the bacteria in a dish to see which ones could inhibit its growth. They then used those data to train their machine learning model to identify chemical structures with anti-A. baumanni effects.
When the model was ready, they used it to analyze the structures of 6,680 compounds from the Drug Repurposing Hub, an open-access library of molecules with potential therapeutic benefits. The model turned up several hundred candidates. The team moved forward with 240 of them, selecting chemicals with very different structures from existing antibiotics—an important consideration, as bacteria might have cross-resistance to medicines that look like ones they’ve encountered before.
Additional screening rounds revealed a winner, which the researchers dubbed “abaucin” for its ability to kill the bacteria. Further in vitro studies revealed something surprising: Not only was abaucin able to inhibit all A. baumanni strains with known resistance, but it was more potent relative to other bacteria species.
“What’s interesting is that the model wasn’t trained to predict narrow-spectrum compounds—it was only trained to predict compounds with activity against Acinetobacter,” Stokes said. “The fact that abaucin displayed enhanced activity against Acinetobacter was simply a unique laboratory observation.”
In vivo studies confirmed their findings. Using mice with wounds that had been infected with A. baumanni, the researchers repeatedly applied either an unmedicated cream or one containing abaucin over the course of 24 hours. When they analyzed the tissue, they found that the wounds that had been treated with the cream looked virtually the same as they did pre-
infection, while the untreated wounds were inflamed and colonized by the bacteria.
The researchers noted that conventional antibiotics might outperform abaucin against A. baumanni in a clinical setting. Though the antibiotic killed the bacteria in a petri dish, it didn’t sterilize cultures completely, unlike fluoroquinolones, beta-lactams and other older antibiotics. Still, in the face of multi-drug resistance, less-than-perfect performance is a reasonable trade-off.
“So many isolates of A. baumanni are resistant to conventional antibiotics that we need something structurally and functionally novel,” Stokes said.
Traversing the 'valley of death'
Next comes the task of taking abaucin from the bench to clinical trials—the phase between exciting preclinical data and human testing that’s been called the “valley of death” in drug development.
“That’s where most drugs fail—right during the handoff from academia to industry, where you’re trying to do some of the highest risk, low probability of success preclinical work,” Kosaraju explained. “This is kind of the less glorious part of drug development and incredibly important.”
That valley is so steep in the world of antibiotics that it’s practically a chasm, she added. The financial reward for developing a new antibiotic is much smaller than for, say, a novel oncology drug or a blockbuster weight loss medication, and it takes much longer to recoup an investment—30 years, on average. Antibiotics are prescribed for shorter durations than other medicines, and, while treatments for deadly,
hospital-acquired infections are badly needed by susceptible populations, those populations are relatively small. They’re even smaller for any one pathogen.
Few major pharmas are still in the game, GSK being the most active, while others like Johnson & Johnson or Pfizer commit to initiatives like the AMR Action Fund to aid the effort instead.
All that adds up to make developing narrow-spectrum antibiotics with potentially better resistance profiles, like abaucin, an even riskier bet than usual for investors, despite a growing need for them.
“It’s harder to make that choice, instead of going after something broad-spectrum where you can treat many different kinds of
infections,” Kosaraju said.
Phare’s nonprofit status acts as a bridge over the chasm of death. Funding for the next phase of development—finding analogues of abaucin with better medicinal profiles and conducting studies to support future human testing on them—will come from philanthropy and grants in addition to what’s provided by the Audacious Project. Then, once Phare emerges with a viable product, it plans to partner with a company in Big Pharma for further development or spin out its own venture-funded company, according to Kosaraju.
“This helps us bridge to the point that our drugs are de-risked, where the chance of success is much higher,” she said. On top of that, she added, “the nonprofit model enables us to do truly innovative work—to go after the most difficult-to-treat pathogens on the WHO list and stay laser-focused without the pressures of commercial funding.”
Let the robots in (but keep the humans, too)
Like other AI tools in drug discovery, Stokes’ and Collins’ machine learning platform is already mitigating expenses by streamlining the process of mining for candidates. While the model is ultimately a “suggestion generation machine,” Stokes said, it cuts down the number of experiments humans have to do—and the price. On average, it takes drug developers 4.5 years and $2.5 million to get from discovery through the pre-investigational new drug application stage, Kosaraju said. Thanks to the AI platform, Phare estimates that time can be cut down to 2.5 years at a third of the cost.
Those figures may have room to fall further as the technology’s predictive power improves with the addition of more data sets. It will get better at being able to tell what compounds are most likely to work not only against microbes, but in animals and humans, too.
“I think that in the future we will be able to more robustly apply machine learning models to help us not only predict which compounds might be antibacterial in a dish, but also those that perhaps are most likely to successfully traverse preclinical and clinical development,” Stokes said.
The platform can’t do all the work alone, of course. Just as important as training the models on new data is optimizing the way researchers interact with them, Stokes explained. Like any other laboratory tool, scientists must learn to use AI effectively.
“Maybe this is a misconception on my part, but I feel that oftentimes, people think of AI models as these static entities,” Stokes said. “I don’t see them as that. I see these models as entities that work with us, and that we have to learn to work with them appropriately in order to maximize the net benefit to drug discovery.”