UPenn researchers use AI to create next-generation antibiotics

Antibiotic resistance has become one of the most pressing threats to global health. Infections once treatable with a simple prescription are becoming harder, sometimes impossible, to cure. Doctors warn that without new solutions, even routine surgeries could become dangerous. Scientists are racing to outpace the rise of resistant bacteria, and a team from the University of Pennsylvania believes artificial intelligence may be the breakthrough we’ve been waiting for.

Their new system, which they've named AMP-Diffusion, is designed to generate antimicrobial peptides—small proteins that organisms naturally employ to fight infection. The peptides eliminate bacteria in mechanisms that are harder to resist than traditional antibiotics.

The research shows that AI can be used to design totally new peptides, screen them for safety, and even pick out variants that are just as potent as FDA-approved drugs in animal experiments.

Drug discovery is a normally slow and expensive process. Scientists must sift through enormous chemical libraries, searching for a few molecules of interest. It can take years to identify one candidate worth testing in animals. AMP-Diffusion changes that process entirely.

Instead of trial and error, the system uses a type of AI known as a diffusion model. These models start with random "noise" and progressively convert it into something meaningful—akin to programs that generate realistic photos from text prompts. In this case, instead of tweaking pixels, the AI reshapes amino acid sequences, the building blocks of proteins.

Marcelo D.T. Torres, who led the research along with colleagues César de la Fuente and Pranam Chatterjee, explains that the researchers trained the model on a large database of natural proteins. This gave the system a "map" of how real proteins come together. From that data, it generated 50,000 potential antimicrobial peptides in silico, an overwhelming number compared to what could be manually designed by humans.

Of course, no lab could test tens of thousands of sequences, so the researchers used another AI tool, called APEX, to rank them. APEX selected the most hopeful candidates, filtering out those that were too similar to existing peptides and keeping a varied assortment of new designs.

From there, the scientists synthesized 46 peptides and tested them in the lab. The results were surprising: more than three-quarters showed an ability to kill bacteria, including some of the most stubborn drug-resistant types that make the rounds in hospitals.

"These peptides basically poke holes in bacterial walls," Chatterjee says. "They make the membrane leaky or disrupt its electrical balance, which kills the cell very quickly."

Even more encouraging, the peptides left human cells largely alone. Tests on red blood cells and mammalian tissue revealed little toxicity, suggesting that the peptides are selective in their attack.

The most persuasive data came from animal experiments. The scientists infected mice with bacterial skin abscesses and administered the AI-designed peptides to them. The outcome was comparable to what is typically seen with drugs like polymyxin B and levofloxacin, two antibiotics that doctors only resort to when everything else fails.

Importantly, the mice did not suffer from any side effects of treatment. "It's exciting to see that our AI-generated molecules actually worked in living organisms," says Chatterjee. "This shows that generative AI can help combat antibiotic resistance."

Traditional antibiotics usually target single bacterial proteins or enzymes. Bacteria often develop mutations in those proteins, making the drugs useless. Peptides attack differently. They destroy membranes and provide bacteria with fewer escape routes. That should also make resistance more challenging to develop as easily.

De la Fuente, who has spent years searching for antibiotic candidates in the unconventional—from mammoth proteins to animal venom—says nature only offers so many options. "Nature's dataset is finite; with AI, we can design antibiotics evolution never tried," he says.

This approach could lead to entirely new classes of drugs, ones not limited by what evolution has already developed.

The work was a collaboration between two labs with complementary expertise. De la Fuente's laboratory had been combing natural proteins for antibiotics, while Chatterjee's laboratory was working to design therapeutic peptides for diseases that are recalcitrant to traditional drug design.

"It was a natural fit," Chatterjee says. "Our lab is skilled at creating new molecules using AI, and the de la Fuente Lab is adept at identifying promising antibiotic candidates."

By combining these strengths, they created a pipeline that can go from random noise on a computer screen to validated drug candidates in the lab.

One of the biggest advantages of this method is speed. Traditional drug discovery takes years, but AMP-Diffusion can generate thousands of candidates in a few days and narrow them down in short order. That more than 75% of the peptides tested in the lab showed activity illustrates how effective the system is. Chemical libraries typically provide far lower hit rates.

This efficiency would conserve millions of dollars and significantly cut down the time it would take to bring a novel antibiotic to clinical trials. Theoretically, the system could also be adapted to design peptides for other disorders, including cancer, autoimmune disorders, or even viral infections.

Despite the positive results, the researchers are quick to point out the difficulties. The model employed is based on early diffusion models that are less efficient than more modern ones. A change to more contemporary frameworks might make the process even faster and more precise.

And then there's the issue of biased training data. Public peptide databases are replete with examples for targeting bacteria but less so for viruses, fungi, or parasites. That renders AMP-Diffusion more skilled at designing antibacterial peptides than anything else—for now. Growing and balancing the training sets will be key to more general uses.

However, the proof of principle is strong: generative AI can create peptides that not only appear good on a computer but also function in living organisms.

If refined and scaled up, AMP-Diffusion could reduce the process of antibiotic discovery from years to a matter of days. Such speed would be critical at a time when antibiotic resistance is increasing around the world. Hospitals could one day have access to entirely new classes of medications created by AI, giving doctors more alternatives when traditional antibiotics fail.

Aside from fighting infections, the approach could have a wide range of medical fields. The same system could be used to design peptides to treat cancer, metabolic diseases, or even psychiatric diseases. By extending beyond what nature has offered, scientists might open the doorway to drugs evolution didn't create.

For now, AMP-Diffusion proves that AI can move beyond creating images and text to creating medicines that might save lives. That leap might be the beginning of a new era where humans and algorithms join forces to fight diseases that were once unbeatable.

Research findings are available online in the journal Cell Biomaterials.

Like these kind of feel good stories? Get The Brighter Side of News' newsletter.