Marcelo Der Torossian Torres lifted the clear plastic cover off of a petri dish one morning last June. The dish, still warm from its sleepover in the incubator, smelled of rancid broth. Inside it sat a rubbery bed of amber-colored agar, and on that bed lay neat rows of pinpricks—dozens of colonies of drug-resistant bacteria sampled from the skin of a lab mouse.
Torres counted each pinprick softly to himself, then did some quick calculations. Untreated for the infection, the samples taken from an abscess on the mouse had yielded billions of superbugs, or antibiotic-resistant bacteria. But to his surprise, some of the other rows on the petri dish seemed empty. These were the ones corresponding to samples from mice that received an experimental treatment—a novel antibiotic.
Torres dug up other dishes cultured from more concentrated samples, taken from the same mice who had gotten the antibiotic. These didn’t look empty. When he counted them up, he found that the antibiotic had nuked the bacterial load so that it was up to a million times sparser than the sample from the untreated mouse. “I got very excited,” says Torres, a postdoc specializing in chemistry at the University of Pennsylvania. But this custom antibiotic wasn’t entirely his own recipe. It took an artificial intelligence algorithm scouring a database of human proteins to help Torres and his team find it.
Torres and his colleagues were looking for peptides that are naturally produced by people and that can fight microbes. To do it, they used an AI that scrutinized the chemical makeup of each and every one in the human proteome—the complete set of proteins our bodies can produce. Peptides are small proteins, or fragments of them. They may not resemble classical antibiotics like penicillin. And they don't all originate in the immune system. But they can contain the right chemistry to be lethal to pathogens, because they can dismantle bacterial cell membranes.
This month, Torres’ team reported in Nature Biomedical Engineering that their search turned up 2,603 antibiotic candidates, a feat they accomplished because of AI’s strength in digesting huge data sets. “I think it speaks to the power of AI,” says César de la Fuente, a bioengineer at the University of Pennsylvania and senior author of the study.
The team tested 55 of those candidates in tiny vials, and a majority of them eliminated bacteria. Then, Torres tested two of them in lab mice and found that they stopped infections from growing. “The results are compelling,” says Daria Van Tyne, an expert in bacterial evolution at the University of Pittsburgh School of Medicine, who was not involved in the work. “It's certainly opening a new class of antimicrobial peptides, and finding them in an unexpected place.”
This is the first time anyone has so thoroughly explored the human body for antibiotic candidates. But in using AI to guide the search, the team stumbled upon a mind-bending discovery of something more basic: Many of our proteins that are seemingly unrelated to immunity may have evolved to live double lives as protection against invaders. “The fact that they found so many of them,” Van Tyne says of the peptides, “suggests very strongly that it’s not just coincidence—that they exist for a purpose.”
The global fight against antibiotic resistance could use some new weapons. Antibiotics have gotten less effective as bacteria have evolved tolerance to drugs, in part due to misuse and overuse. The World Health Organization estimates that by the year 2050, 10 million people could die annually from drug-resistant infections as the efficacy of current antibiotics wanes.
Along with vaccines and clean water, antibiotics are one of three “pillars” that let humans double our life span since the 1800s, according to de la Fuente. “Imagine if that disappeared from the equation,” he says.
If antibiotics stop working, surgery and organ transplants would flirt with disaster. Chemotherapy would become more dangerous. Antibiotics are sometimes even crucial for childbirth. “All these other interventions in modern medicine would not be possible, or would be a lot harder without effective antibiotics,” says de la Fuente. And in a worst-case scenario, he says, “we're going to be facing a pre-antibiotic era where just with a minor scratch could be lethal.”
Governments, philanthropies, and pharma companies have pledged billions of dollars to get new drugs approved by 2030. And the natural world has already inspired new ways to kill drug-resistant germs. In 2019, one genetically-altered virus helped save a teen from a deadly infection. But Torres and de la Fuente turned their attention to somewhere even more natural to us: our own bodies
We contain tens of thousands of different proteins. Each is made from amino acid molecules that snap into sequences—referred to as peptides—like Legos. They form big clumps, contort into puzzling shapes, and wiggle microscopically. Each protein usually serves some purpose. Some deliver messages. Others help repair injured tissue. Some, like proteases, chop up other proteins. This specific action typically boils down to a small, evolutionarily preserved sequence of amino acids that are especially eager to lend a proton or electron to molecules around them.
Some peptides contain chemistry that kills microbes. Those found in snake and scorpion venoms attack bacterial cell membranes. Their trick boils down to a couple of things: The sequences are relatively short, positively charged, and amphipathic (not too water-repelling or oil-repelling). Other organisms, including people, have cells that churn out proteins that employ similar tricks. Antimicrobial peptides with these traits are key weapons for the immune function of all living organisms.
The team had this particular brand of chemical defense in mind when they began their search for antimicrobial peptides. De la Fuente’s lab specializes in using AI to discover and design new drugs. Rather than making some all-new peptide molecules that fit the bill, they hypothesized that an algorithm could use machine learning to winnow down the huge repository of natural peptide sequences in the human proteome into a select few candidates.
“We know those patterns—the multiple patterns—that we're looking for,” says de la Fuente. “So that allows us to use the algorithm as a search function.”
The team’s algorithm was based on pattern recognition software that’s used for analyzing images. First, it learned what kills microbes by ingesting a list of peptides that are known to be antimicrobial. Then, it used that knowledge to comb through peptide databases and pick out likely candidates with the right chemical traits—that they should be short (8 to 25 amino acids long), positive, and amphipathic.
Their algorithm gobbled up the entire human proteome and spat out a preliminary list of about 43,000 peptides. Torres narrowed it down to the 2,603 that come from proteins known to be secreted from cells. Some were complete small proteins and hormones. Others were just fragments, encrypted chains within a much larger complex. None of them had ever before been described as antibiotics.
To check that their AI was on the right track, Torres synthesized 55 of the most promising candidates. He tested each one in liquid samples against a “who’s who” of drug-resistant microbes: Pseudomonas aeruginosa, a notoriously rugged infector of lungs; Acinetobacter baumannii, known to spread rampantly in hospitals; Staphylococcus aureus, the germ behind dangerous staph infections—plus others, eight in total. Of the 55, the majority were able to prevent bacteria from replicating.
A few peptides stood out, including SCUB1-SKE25 and SCUB3-MLP22. These peptides live along regions called “CUB domains” that exist in proteins involved in a long list of functions like fertilization, making new blood vessels, and suppressing tumors. The SCUBs are only pieces of the whole. But on their own, they seemed shockingly adept at killing germs. So Torres promoted these two SCUBs to trials in mice.
Torres tested whether either SCUB, or a combination of the two, could eliminate infections in mice with infections under their skin, or in their thigh muscle (a model for more systemic disease). In all cases, bacteria populations sampled from these tissues stopped growing. And in some cases, as Torres noticed on his warm agar, bacterial counts plummeted.
Torres also tested how easily bacteria could evolve resistance to the peptides, in comparison to an existing antibiotic called polymyxin B. After 30 days of exposure, the bacteria could tolerate doses of polymyxin B that were 256 times higher than the original amount, but the SCUBs remained effective at the same dose. (It takes a lot of genetic change for bacteria to adapt to membrane damage.) Of course, that doesn’t mean they’ll never adapt, especially over longer intervals. "Nothing is ever going to be resistance-proof," says de la Fuente. "Because bacteria are the greatest evolvers that we know."
As systematic as the team’s plan was, Torres was still left a bit dumbfounded. “We thought we would have a lot of hits,” he says of the peptides revealed by the AI. But to his surprise, the peptides came from throughout the body. They were from proteins in the eyes, nervous system, and cardiovascular system, not just the immune system. “They’re literally everywhere,” says Torres.
The team thinks life evolved this way to pack as much of a punch as possible into the genome. “One gene codes for one protein, but that protein has multiple functions,” de la Fuente says. “This is a really, I think, clever way for evolution to just keep the genomic information at a minimum.”
It’s the first time scientists have found antibiotic peptides within proteins unrelated to immune response. The idea was “really creative,” says Jon Stokes, a biochemist at McMaster University, Canada, who was not involved in the study, but has been prepping his lab to incorporate AI in the search for small molecule antibiotics. “The take-home for me is: Start looking in unobvious places for antibiotics.”
Researchers look for antimicrobials among organisms that live in the soil and the sea, “but this general idea of identifying what I'll call ‘cryptic’ antibiotics that are within us, I think is really cool,” Stokes continues. “Then the question becomes: Well, if this is true in humans, should we also be looking at other mammals? Should we be looking at reptiles, amphibians, crustaceans?”
AI algorithms can help discover antibiotics in this manner by feeding them known examples of what to look for, then databases of molecules they can search. They can also help invent molecules or optimize existing ones to avoid unwanted side effects. Within the next decade, will we see a drug in clinical use that was discovered, designed, or optimized with machine learning? “Yeah,” says Stokes, “I’d put my money on that.”
But still, there is a lot of work left to turn this discovery into medicine anyone can use clinically, especially when poking around peptides for answers. Peptides don’t have a great track record as antibiotics, says Van Tyne. These molecules often fail because they’re toxic, or they don't move around the body as easily as other drug molecules do. That has made it hard to use them to treat systemic infections. “I don't know that any of these peptides are actually going to become new antibiotics," says Van Tyne.
Torres and de la Fuente both appreciate this uphill battle; when they were designing the study, they chose to use peptides that occur naturally in the human body because they are less likely to be toxic. So far, Torres’ results with the thigh muscle infection in the mice suggests the SCUBs were able to attack a systemic infection. “It's certainly encouraging,” says Van Tyne. “It opens a door that potentially these could be better antimicrobial peptides than the ones that have been tried to be developed and failed.”
That novelty bodes well for the team’s mission. And these early candidates won’t be the only peptide antibiotics that they try. “Our main goal is to have a computer design an antibiotic with very minimal human intervention that will be able to enter clinical trials,” says de la Fuente. “That's our ultimate mission here.”
- 📩 The latest on tech, science, and more: Get our newsletters!
- The 10,000 faces that launched an NFT revolution
- Cars are going electric. What happens to used batteries?
- Finally, a practical use for nuclear fusion
- The metaverse is simply Big Tech, but bigger
- Analog gifts for people who need a digital detox
- 👁️ Explore AI like never before with our new database
- 💻 Upgrade your work game with our Gear team’s favorite laptops, keyboards, typing alternatives, and noise-canceling headphones
