Penn Engineers is building AI models to create new antibiotics, and early testing shows several studies, as well as existing approved drugs.
The University of Pennsylvania has developed a generative artificial intelligence (AI) model that can design new antibiotics, marking what researchers say as a major step forward in the fight against drug-resistant bacteria.
In research published in Cell BiomaterialsPen Engineering engineers and university colleagues present AMP diffusion, a generative AI tool for designing antimicrobial peptides (AMPs). This is a short amino acid chain that kills bacteria. In animal studies, some of the AI-designed molecules functioned as effectively as existing FDA-approved antibiotics and showed no detectable side effects.
For decades, scientists have warned of the looming crisis of antibiotic resistance. New drugs have proven difficult and slow to develop, and health systems around the world are struggling to keep up. Generation AI, best known for creating images and text, can provide a radically faster approach.
Nature's dataset is finite. With AI, we cannot design the evolution of antibiotics.
“Nature's dataset is finite. With AI, we never tried to evolve antibiotics,” says César de La Fuente, presidential associate professor at the University of Pennsylvania.
Pranam Chatterjee, an assistant professor of Penn Engineering who started the project at Duke University, adds that “it leverages the same AI algorithms that generate images, but is augmenting them to design powerful new molecules.”
From mammoths to microorganisms
De La Fuente's lab has been using AI to scrutinize unconventional biological sources of antimicrobial properties, ranging from wool mammoth proteins to animal venom and ancient microorganisms. However, the rate at which resistance emerges exceeds these findings. “Unfortunately, antibiotic resistance continues to increase faster than new antibiotic candidates can be discovered,” he says.
Meanwhile, Chatterjee's group focuses on using AI to design peptides for difficult-to-treat diseases. Their collaboration was a natural match. “It seemed like a natural fit,” Chattersee said. “Our labs know how to use AI to design new molecules, and De La Fuente Lab knows how to use AI to identify powerful antibiotic candidates.”
Drug-resistant bacteria are on the rise worldwide, spreading faster than new antibiotics can be developed. The World Health Organization (WHO) calls it one of the biggest threats to global health. Credit: Saiful52/ShutterStock
Medicine's diffusion model
Large-scale language models such as ChatGPT predict the next word in sequence, while diffusion models start with random noise and gradually purify it all over the coherent to generate content. This is the principle behind creative AI systems such as Dall-E and stable diffusion.
Instead of teaching the model the ABC in biology, I started with a fluent speaker.
Amp-Diffusion applies the same concept to biology. Instead of forming pixels, the amino acid sequence is shaped into plausible peptides. “It's almost like tuning the radio,” says de la Fuente. “You start with static and then the melody finally pops up.”
Unlike other teams who have tried antibiotic diffusion models, Penn's approach leaning towards ESM-2, a protein language model for Meta trained in hundreds of millions of sequences. By building on existing “psychic maps” of how proteins fit, AMP-Diffusion is more likely to generate candidates more rapidly and become biologically effective.
“Instead of teaching models about ABC in biology, we started with fluent speakers,” says Chatterjee. “That shortcut allows you to focus on peptide design with real shots to become a drug.”
From 50,000 ideas to two winners
AMP diffusion generated approximately 50,000 peptide sequences. However, even some of the labs are not possible to test it. To narrow down the field, researchers relied on another AI tool, Apex 1.1, which was developed in De La Fuente's lab. We ranked candidates based on predicted germ-killing power, novelty and diversity.
It's exciting to see how the molecules generated in our AI actually worked. This shows that the produced AI helps combat antibiotic resistance.
This led the team to synthesize 46 peptides for laboratory and animal testing. Two molecules stood out in mouse models of skin infection. Antibiotics performed on par with levofloxacin and polymyxin B and used against resistant bacteria have been established. Importantly, no adverse side effects were observed.
“It's exciting to see how the molecules produced in AI actually worked,” says Chatterjee. “This shows that the production AI helps combat antibiotic resistance.”
What's next?
Researchers see this as just the beginning. AMP spread may ultimately be improved to design antibiotics for specific infections or tailored to prioritize molecules with particularly desirable drug-like properties. “We've shown that the model is working, but now if we can guide it to enhance its beneficial drug-like properties, we can create a treatment that we can tackle right away,” says Chatterjee.
For De La Fuente, long-term ambitions are radical speed. “Ultimately, our goal is to compress the antibiotic discovery timeline over the years,” he says.
Such breakthroughs could prove transformative as drug resistance is currently listed among the global health threats by the World Health Organization.
