AI finds molecules that kill bacteria, but would they make good antibiotics?

Antibiotic drugs have been around for less than a century, but the rise of drug-resistant bacterial strains and a dearth of new antibiotics reaching the clinic in recent decades threaten to undermine our ability to beat deadly infections. Using a combination of machine learning and experimentation, researchers say they’ve identified several molecules that may be effective antibiotics (Cell 2020, DOI: 10.1016/j.cell.2020.01.021). Experts applaud the computational approach, but say these compounds are not likely to be the drugs we need.

Because bacteria have evolved resistance to some of the most common antibiotics, scientists want to find molecules or structural features that can kill microbes in new ways. Computational chemist Regina Barzilay and bioengineer James J. Collins, both of the Massachusetts Institute of Technology, and colleagues put their heads together and designed a machine-learning approach to find new antibiotics. They trained their algorithms to recognize structural features of different molecules—not just antibiotics—and to predict whether a given structure will inhibit Escherichia coli growth.

The researchers first set their model loose on the Drug Repurposing Hub, a database of about 6,000 molecules known to be useful against various diseases. One compound had predicted antimicrobial activity against E. coli, low predicted toxicity, was dissimilar from known antibiotics, and was verified in the lab to inhibit E. coli growth. This molecule, called SU3327, is an inhibitor of the enzyme c-Jun N-terminal kinase, which is a target for cancer and other diseases. The group renamed the molecule halicin in a nod to the computer HAL from the movie 2001: A Space Odyssey. An unpublished 2017 study also identified this molecule's antibiotic activity, but those researchers chose not to pursue it because of its similarity to a compound that the US Food and Drug Administration was already evaluating. Jonathan M. Stokes of Broad Institute of MIT and Harvard who is a coauthor of the new study says the group was not aware of the 2017 research until March 2.

In the new study, subsequent lab tests showed that the molecule can kill a number of resistant bacterial strains. The researchers think it kills bacteria by disrupting the movement of protons across the cell membrane, disrupting the microbe’s ability to move or store energy. In mice, halicin treated Acinetobacter baumannii-infected skin wounds and Clostridium difficile gut infections.

The group also used their machine-learning algorithm to search 100 million molecules from the ZINC15 database and found 8 potential antibiotics, including ZINC000100032716, but did not follow up their search with lab tests.

The results show “how much can be achieved when skilled practitioners and machine-learning teams work together,” says Günter Klambauer, who leads an AI drug-discovery lab at Johannes Kepler University. But he criticized the group for training their algorithms on only a couple thousand molecules and looking for just a few biological effects, saying the model could have been stronger with broader training that took into account multiple effects..

Antibiotics experts also praised the group’s methods but were unimpressed by halicin. Several said the group’s results did not suggest the compound is the kind of antibiotic doctors need. The nitroaromatic group in the molecule resembles structures in known broad-spectrum antibiotics, suggesting that how the researchers trained the algorithm was too limited and didn’t allow the program to find truly novel structures: “We need new antibacterial chemotypes that may be hard to find through this approach,” says Richard E. Lee, an antibiotics researcher at St. Jude Children’s Research Hospital. And nitroaromatic groups can be toxic to human patients, as can the membrane-associated mechanism of action the researchers describe, says Shahriar Mobashery, a biochemist at the University of Notre Dame. Mobashery nonetheless applauded the approach, saying the paper’s strength was in its methodology.

Collins says it’s fair to point out their molecules’ similarities to existing antibiotics. But he stresses that one of the major values in using machine learning is its speed in searching for antibiotic-like molecules. It took their model about 4 days to evaluate more than 100 million molecules.

Lee also said halicin was probably not a viable drug. He says we need antibiotics that are effective against systemic infections, not those on the skin or in the gut. Still, Collins says, there is a need for effective topical treatments of A. baumannii. John H. Rex, chief medical officer at the drug discovery firm F2G, echoed that critique. Rex, a former editor of the journal Antimicrobial Agents and Chemotherapy, says the work did not meet his standards for announcing a new antibiotic drug lead. In addition to his concerns about its efficacy against the most dangerous bacteria, Rex said the group’s tests of halicin’s toxicity fell short. He says he’d like to have seen the researchers test its toxicity against mammalian cells, which could provide a clearer picture of whether halicin would be toxic in the human bloodstream.

Still Rex is glad the paper got published. This work “could open up new chemical space” where effective antibiotics might be found, he says, but the group hasn’t found a drug candidate.

“It is quite easy to kill bacteria, even the tough ones, with toxic agents—and quite easy to find those,” says antibiotic expert Lynn Silver, who worked at Merck & Co. for 2 decades. But she says finding drugs is much harder: “Even well studied antibacterials fail in clinical trials due to toxicity.”

Collins says the group is looking to establish partnerships to continue the pre-clinical evaluation of halicin and the other molecules. Barzilay says they also plan to improve their model in the hopes of predicting molecules’ antibacterial mechanism in addition to their activity, which could help them find more narrowly-targeted drugs. And they’re adding a component to the model that should allow it to design wholly new molecules, rather than finding possible antibiotics from databases of existing compounds.