Bacteria and antibiotics have been playing a game of cat and mouse for about a century. Unfortunately, the bacteria are gaining the upper hand.
According to the World Health Organization, antibiotic resistance is one of the greatest public health risks and was responsible for 1.27 million deaths worldwide in 2019. With repeated exposure to antibiotics, bacteria quickly learn to adapt their genes to counteract the drugs—and pass the genetic adaptations on to their peers—making the drugs ineffective.
Bacteria with superpowers are also torpedoing medical procedures — surgeries, chemotherapy, Caesarean sections — and jeopardizing life-saving therapies. As antibiotic resistance increases, very few new drugs are being developed. Although studies in petri dishes have targeted potent candidates, some of them also damage the body’s cells and lead to serious side effects.
What if there was a way to retain their bacteria-killing power, but with fewer side effects? This month, researchers used artificial intelligence to redesign a toxic antibiotic, creating thousands of variants and looking for those that retained their germ-killing power without harming human cells.
The AI used in the study is a large language model similar to those behind famous chatbots from Google, OpenAI and Anthropic. The algorithm searched through 5.7 million variants of the original antibiotic and found one that retained its effectiveness but was far less toxic.
In laboratory tests, the new variant rapidly destroyed the bacteria’s “shields” – a fatty bubble that keeps cells intact – but left host cells undamaged. Compared to the original antibiotic, the newer version was far less toxic to human kidney cells in petri dishes. It also quickly eliminated deadly bacteria in infected mice with minimal side effects. The platform can also be easily adapted to test other drugs in development, including those for various cancers.
“We found that large language models represent a major advance for machine learning applications in protein and peptide engineering,” said Dr. Claus Wilke, a biologist and data scientist at the University of Austin and author of the study, in a press release.
Insane in the membrane
Antibiotics work in different ways. Some interfere with the bacteria’s ability to produce proteins. Others prevent their genetic material from replicating, thus stopping them from multiplying. Even more antibiotics specifically destroy their metabolism.
Each strategy took years to research, and safe and effective antibiotics took even longer to develop. But bacteria evolve quickly and can no longer resist these drugs.
The overuse of antibiotics in medicine and agriculture is leading to the emergence of “superbugs” that are resistant to even the most effective drugs currently available. Once a strain of bacteria has learned to bypass one mechanism – such as inhibiting protein production – it effortlessly blocks other drugs that target the same strategy.
Resistance can also spread quickly within a bacterial population. Unlike our genetic material, which is locked in a nut-like structure, bacterial DNA floats freely within its cells. Genetic changes – for example, those that allow bacteria to evade antibiotics – can be transmitted to other similar bacteria through temporary biological “tunnels” that literally connect the two cells. In other words, antibiotic resistance spreads quickly.
That is, if you give them the chance.
For antibiotic resistance to develop, bacteria must survive the initial attack. Extremely lethal treatments, including a class called antimicrobial peptides, kill bacteria before they can adapt. These drugs rapidly destroy the fatty protective barrier that surrounds all bacterial cells. Over decades of work, scientists have created many of these molecules.
The problem? They also damage the membranes that protect our own cells, causing toxicity that renders most of them useless in humans. Although a library of these highly effective antibiotics already exists, they have largely been benched like underperforming ballplayers.
Safe and healthy
The goal of the new study was to rehabilitate antimicrobial peptides by modifying a peptide called protegrin-1. Although it is extremely effective at killing bacteria, it is too toxic for human use. The researchers wanted to see if they could reduce the side effects while maintaining the bacteria-killing effect.
Led by Dr. Bryan Davies, the team had previously developed a system to rapidly screen hundreds of thousands of peptides to determine whether they could kill harmful bacteria.
The system is called SLAY (for Surface Localized Antimicrobial Display) and looks like a bunch of tetherballs, each attached at one end to a biological surface while the other – the antimicrobial peptide – floats around to capture bacteria.
The researchers then developed over 5.7 million protegrin-1 variants. “This is a huge increase in diversity compared to the 18 single mutants” in previous studies, the authors write.
Next, they turned to large language models of AI. Known for their ability to generate text, audio, and videos, these types of algorithms learn by ingesting terabytes of data and can spit out responses to a given prompt. While they’re primarily used to generate text, scientists are increasingly using their ability to “think up” new proteins or other drugs.
The study used several clues to guide the AI’s search: things like that the drug must attack bacterial membranes and break them down without harming human cells. The AI searched the available pool of variants and found one that hit the spot – a new version called bacterially selective protegrin-1.2 – that met all the guidelines.
In tests in Petri dishes, the variant quickly decomposed the membranes in Escherichia colia common type of bacteria often used for research purposes, within half an hour. Human red blood cells, on the other hand, thrived under the same circumstances, even when exposed to concentrations 100 times higher than the bacteria. Instead of indiscriminately killing both bacteria and human cells, the AI-approved antibiotic targeted the pathogen.
Protegrin-1 is known to cause kidney damage. The team pitted protegrin-1.2 against the original and colistin, an antibiotic used as a last resort, in cultured human kidney cells. The variant was superior to the others in terms of safety and showed less cell membrane damage.
The team also treated mice infected with a type of multidrug-resistant bacteria – found in hospitals – with the antibiotic selected by the artificial intelligence. Six days later, the animals treated with the new version had lower levels of bacteria in several organs than the untreated mice. Some showed no signs of infection at all. Compared to Protegrin-1, the new version is “significantly less toxic to mice,” the authors write.
Although the study focused on antibiotics, the team plans to use a similar strategy to develop other drugs that were previously considered too toxic for humans. Recently, another team used AI to determine the structure of small chemicals that are useful for antibiotic and cancer therapies but were previously discarded by chemists as useless for safe and effective drugs.
“Many use cases that were not feasible with previous approaches are now starting to work. I expect that these and similar approaches will be widely used in the development of therapeutics or drugs in the future,” said Wilke.
Photo credit: x / x