Developing new antibiotics presents a complicated challenge, but scientists are now using artificial intelligence to design new drugs to address the problem. In May researchers at the Massachusetts Institute of Technology and McMaster University published a study on their use of an AI algorithm to identify an antibiotic that can kill a particularly resistant type of bacterium. The pathogen, called Acinetobacter baumannii, can lead to serious infections, including meningitis and pneumonia, and is often found in hospital settings. It’s also a leading cause of infections in military personnel in the Middle East.
The findings are significant because they show how AI can be used to hasten the development of new antibiotics to fight drug-resistant bacteria. Using AI and machine learning—the subset of AI that involves using algorithms to find patterns in data—dramatically reduces the number of experiments that humans would need to screen a potential drug for efficacy. It also greatly reduces the cost because the computer modeling weeds out compounds that aren’t promising.
Antibiotic resistance, the phenomenon in which bacteria become resistant to the drugs used to treat them, is a growing threat that could cause up to 10 million deaths annually by 2050. The World Health Organization (WHO) estimates that in 2019 alone 1.27 million people died globally from drug-resistant bacterial infections.
“We need new antibiotics because we’re facing a crisis due to the fact that the number of resistant bacterial pathogens is growing while our pipeline of new antibiotics is diminishing,” says James Collins, a professor at M.I.T.’s Institute for Medical Engineering & Science and co-senior author of the new study. “The former is arising due to numerous factors—largely the overuse and misuse of antibiotics, both in health care and in agriculture. And the latter is largely caused by a broken economic market for antibiotics.”
Antibiotic development presents something of a catch-22. The cost of developing a new antibiotic is enormous; it’s on par with the cost of developing a new cancer drug. But unlike cancer drugs, which can be taken for months or even years, antibiotics are usually taken for a relatively short period of time and often just for a single infection. Because of stewardship programs designed to decrease antibiotic resistance, any new antibiotic is likely to be reserved by health care workers until it’s really needed. And long drug approval periods and the widespread availability of generic drugs mean that there is little financial incentive for drug companies to develop new antibiotics.
“There is a real market failure in the antibiotic space, in the antimicrobial space—and I’d include some antifungal medicines in that bucket as well,” says Jocelyn Ulrich, deputy vice president for policy and research at the industry group Pharmaceutical Research and Manufacturers of America (PhRMA). She points out that antibiotic resistance is a naturally occurring phenomenon and that the only ways to control it are to use tools, such as those for infection prevention and control, to slow down the use of new products.
“We have seen the number of companies, especially the larger companies, decline from 20 or so down to just a handful who are still even in the space,” Ulrich says. “And as a result, we have a much smaller pipeline of novel therapeutics coming through.”
Scientists are hoping to change that. In the recent study, Collins and his fellow researchers exposed A. baumannii to thousands of potential drug compounds to see which of them blocked the pathogen’s growth. They used those data to train a computer model to predict a compound’s antibacterial activity based on its structure.
The team used the model to analyze 6,680 compounds in just a few hours—a process that would have taken a few weeks without AI. The analysis narrowed the batch down to a few hundred possibilities—240 of which Collins and his colleagues tested in the lab. Among these, the researchers identified nine antibiotics, including one that was effective at killing A. baumannii. Importantly, the compound is “narrow spectrum,” meaning it doesn’t kill other species of bacteria. This is beneficial because it reduces the chance of other bacteria spreading resistance against the drug and does not compromise the gut’s overall microbiome.
Collins explains that the new drug, named abaucin, works by disrupting the bacterium’s protective outer layer, the cell membrane. In tests on mice, it was effective against wound infections caused by A. baumannii. Abaucin also worked against a number of drug-resistant strains of A. baumannii that had been isolated from human samples and subsequently grown in the lab.
Collins’s team used the computer model to test thousands of compounds. “Now imagine you want to go from [testing] thousands to many billions of molecules,” he says. “It would be effectively impossible [for humans] to curate, purchase and test all of those molecules. And yet for billions of compounds, it still only takes several days to analyze with AI. So we are able to explore much, much larger chemical spaces that really would not be available to us without these computer models.”
“I think these tools have the potential to speed many aspects of the drug development process, but it’s early days,” Ulrich says. “Whenever we have really large data sets and can very efficiently analyze that data, that’s certainly time-saving.” But she notes that the findings are only in animal models. “You still have to do all the work to develop that compound into something that can be metabolized in the human body, and they need to do thorough clinical trials and other things,” she says. “I’d say there’s immense potential and immense excitement around some of these tools.”
Aleksandra Mojsilović is an IBM fellow and head of AI Foundations at IBM Research. She has been instrumental in research showing how AI can be used to develop new therapeutics and has co-authored a paper showing how IBM’s AI system could help accelerate the processes of finding new antibiotics.
Mojsilović agrees that AI can accelerate research by reducing the time it takes to search thousands or millions of compounds. “But it can go way beyond that,” she says. “You can train the models to actually predict properties of existing molecules quickly, which allows you to screen or predict how good the molecule is, or identify the unknown properties, such as toxicity.” Additionally, with generative AI, computer models can be trained on the existing molecules to learn their “representation,” or characteristics, Mojsilović says. Researchers can then design molecules that have never been seen before in nature.
In response to the market challenges in developing new antimicrobial drugs, in 2021 U.S. senators Michael F. Bennet of Colorado and Todd Young of Indiana introduced the PASTEUR Act in Congress. The bipartisan bill would create an incentive program and government investment of $6 billion for the development of new antivirals and antibiotics—and give the government unlimited access to the drugs once they’re approved by the Food and Drug Administration. The bill was just reintroduced in April 2023, after more than 200 organizations, including PhRMA, signed a letter in support of the act in March.
If the bill passes, researchers will still be in a race against time to develop new antibiotics against the most dangerous drug-resistant pathogens—and that’s where AI could be critical. The FDA recently released a paper to facilitate discussion among developers, manufacturers, regulators, academic groups and other stakeholders on the use of AI and machine learning throughout the drug-development process.
AI has been getting a lot of negative attention lately for the ways it could be misused, but it could also be a very powerful tool in helping us solve some of our most pressing challenges.
“I’m very hopeful,” Collins says. “I think our AI tools, our technology platforms to discover, design and develop new antibiotics, are expanding with each year.”