The December 2022 issue of IEEE Spectrum is here!

Close bar

AI Helps Scientists Discover Powerful New Antibiotic

An AI algorithm surfaces one drug that's already eradicated "superbugs" in the lab—and eight more that show promise in computer simulations

3 min read
MIT researchers used a machine-learning algorithm to identify a drug called halicin that kills many strains of bacteria. Halicin (top row) prevented the development of antibiotic resistance in E. coli, while ciprofloxacin (bottom row) did not.
MIT researchers used a machine-learning algorithm to identify a drug called halicin that kills many strains of bacteria. Halicin (top row) prevented the development of antibiotic resistance in E. coli, while ciprofloxacin (bottom row) did not.
Images: Collins Lab/MIT

Deep learning appears to be a powerful new tool in the war against antibiotic-resistant infections. One new algorithm discovered a drug that, in real-world lab tests, killed off a broad spectrum of deadly bacteria, including some antibiotic-resistant strains. The same algorithm has unearthed another eight candidates that show promise in computer-simulated tests.

How does one make an antibiotics-discovering neural network? The answer, counter-intuitively, is not to hold its hand and teach it the rules of biochemistry. Rather, a little like Google’s successful AlphaZero super-AI chess and Go programs, this group’s deep learning model must figure things out from scratch.

“We don’t have to tell the computer anything—we just give it a molecule and a property label, which in our case is, ‘Is it antibacterial?’” says researcher Jonathan Stokes, postdoctoral fellow at MIT’s Department of Biological Engineering. “Then the model on its own learns what molecular features are important, which molecular features are more strongly or more weakly predictive of antibiotic activity.”

And as the AlphaZero researchers found, once a good deep learning model gets going on a well-defined problem, without humans jumping in to teach it a bunch of rules, new frontiers sometimes open up.

Stokes said he and his co-authors from MIT, Harvard, and McMaster University in Hamilton, Ontario repurposed a deep learning algorithm designed to figure out chemical properties of molecules. The chemical property algorithm in this case outperformed other computer simulation programs in predicting, say, a simulated molecule’s solubility.

Stokes said the new research treated antibiotic efficacy as another chemical property for this same algorithm to predict.

The group trained its neural net on a database of more than 1,000 FDA-approved drugs and another group of natural compounds isolated from sources like plants or dirt. These 2,335 molecules all had well-known chemical structures and well-known antibiotic or non-antibiotic properties.

Once the model had been trained, they pointed it at a drug repurposing database of more than 6,000 compounds that have either been FDA approved as drugs or had at least begun the FDA approval process.

Stokes said the team was focused on two parameters in particular—antibiotic efficacy (which their deep learning algorithm determined) and chemical similarity to other known antibiotics (calculated by a well-known mathematical formula called the Tanimoto Score). They wanted to discover compounds in the Broad Institute’s Drug Repurposing Hub that were highly effective antibiotics. But they also wanted these potential antibiotics to be as chemically distant from any other known antibiotic as possible.

They desired the latter because chemical cousins of known antibiotics can also prove ineffective against antibiotic-resistant strains of infections.

And that is how the group lit on a drug they called halicin. Originally developed as an anti-diabetes drug, halicin seemed to be an antibiotic nothing like, for instance, the tetracycline family of antibiotics or the beta-lactam antibiotic group (of which penicillin is a member).

“It didn’t obviously fit into an existing antibiotic class,” he said. “You could kind of stretch your imagination and say, ‘Maybe it belongs to this class.’ But there was nothing obvious, nothing clean about it. And that was cool.”

So they tested halicin against known dangerous bacteria like E. coli. They also tested halicin as a cream against a skin infection grown on a lab mouse that is generally not treatable by any antibiotic today.

“We took halicin and applied it topically. We applied it periodically over the course of a day,” Stokes said. “Then we looked to see how many live Acinetobacter baumannii existed at the end of a day’s worth of treatment. And we found it had eradicated the infection.”

Bolstered by this success, the group then applied the model to a much broader virtual depository, the so-called ZINC 15 online database of more than 120 million molecules.

Again they searched for the intersection of an effective antibiotic that was also as chemically distinct as possible from known antibiotics. That’s how they turned up another eight candidates. None of these eight have, however, been tested in the lab as halicin has.

The group describes the entire process in a recent paper in the journal Cell.

Stokes said the group is now applying its deep learning model to discovering novel so-called narrow-spectrum antibiotics.

“We’re training new models to find antibiotics that are only active against a specific bacterial pathogen—and that do not have activity against the microbes living in your gut,” he said.

Plus, he said, narrow-spectrum antibiotics will be much less likely to trigger broad-spectrum antibiotic resistance. “Our current antibiotics are active against a ton of different bacteria. And the fact that they’re active against a ton of different bacteria promotes the dissemination of antibiotic resistance. So these narrow-spectrum therapies… will less strongly promote the rampant dissemination of resistance.”

The Conversation (0)

Will AI Steal Submarines’ Stealth?

Better detection will make the oceans transparent—and perhaps doom mutually assured destruction

11 min read
A photo of a submarine in the water under a partly cloudy sky.

The Virginia-class fast attack submarine USS Virginia cruises through the Mediterranean in 2010. Back then, it could effectively disappear just by diving.

U.S. Navy

Submarines are valued primarily for their ability to hide. The assurance that submarines would likely survive the first missile strike in a nuclear war and thus be able to respond by launching missiles in a second strike is key to the strategy of deterrence known as mutually assured destruction. Any new technology that might render the oceans effectively transparent, making it trivial to spot lurking submarines, could thus undermine the peace of the world. For nearly a century, naval engineers have striven to develop ever-faster, ever-quieter submarines. But they have worked just as hard at advancing a wide array of radar, sonar, and other technologies designed to detect, target, and eliminate enemy submarines.

The balance seemed to turn with the emergence of nuclear-powered submarines in the early 1960s. In a 2015 study for the Center for Strategic and Budgetary Assessment, Bryan Clark, a naval specialist now at the Hudson Institute, noted that the ability of these boats to remain submerged for long periods of time made them “nearly impossible to find with radar and active sonar.” But even these stealthy submarines produce subtle, very-low-frequency noises that can be picked up from far away by networks of acoustic hydrophone arrays mounted to the seafloor.

Keep Reading ↓Show less