Can artificial intelligence (AI) help beat superbugs? AI isn’t simply confined to chatbots that generate text and images. There’s a lot of interest in using AI to discover new drugs. But how much of this is actual hope and how much of it is hype?

An interdisciplinary team of researchers from Massachusetts Institute of Technology, and McMaster University, Ontario, is using deep learning — a kind of AI that attempts to simulate the way the brain works by learning from large data sets — to identify promising chemical compounds. Reporting their research in scientific journals, the team was able to identify two such compounds that could potentially be used as superbug-neutralising antibiotics if they make it through the gamut of preclinical tests and human trials which is where most potential drugs fail. The researchers named them halicin and abaucin, and, in lab settings, these were able to fight specific superbugs.

What’s more, the approach that the team used could also be helpful in finding promising drug candidates for other diseases.

Finding a drug is like solving a puzzle. Sometimes you get lucky — as no doubt, Alexander Fleming did with the serendipitous discovery of penicillin. More often than not, these days, finding a drug involves painstakingly testing large libraries of chemical compounds to find pieces that fit a disease puzzle and then optimising these pieces, so they work better. Computers are used in this process by large teams of biologists, engineers, and chemists.

So far, so good. How can AI help speed up this process? Since the chemical universe is incredibly large, AI could help to filter out promising compounds for specific diseases faster and cheaper. In principle, AI could also come up with drugs that humans wouldn’t think of rationally because the path to discovery isn’t charted specifically by people to the AI model. It’s a case in which the black-box intelligence of AI might come in handy.

In a paper published in Nature Chemical Biology last month, the researchers focused on a problematic superbug known as Acinetobacter baumannii. This is a nasty bacterial species, often found in hospitals, which can lead to serious ailments and even death. It can survive on hospital doorknobs and equipment and pick up antibiotic resistance genes from its surroundings that make it virtually undefeatable by common antibiotics.

So, how did AI help?

The research team trained an AI model on a library of thousands of potential drug compounds. The model was told which drugs stopped bacterial growth and which didn’t, and it was allowed to learn to build its own criteria for identifying superbug-killing properties. After this training phase, the model was then exposed to compounds it hadn’t seen before and asked which ones it thought might stop growth. It filtered out compounds of which 240 were tested in a lab. One of these compounds — abaucin — was singled out for further study.

Abaucin is highly potent against A. baumannii and has limited effect on other bacteria. This is a desirable trait in an antibiotic since broad-spectrum antibiotics can indiscriminately kill both good and bad bacteria. Good gut bacteria form part of the microbiome and are necessary for good health.

Previously, the researchers had shown the effectiveness of the AI approach in identifying halicin, a potential antibiotic effective against multiple superbugs. In that case, the AI model had been trained to find drugs that stopped the growth of Escherichia coli (E. Coli), another kind of superbug.

Armed with these successes, the team plans to extend their AI-based approach to find other promising antibiotics that fight infection caused by other superbugs, including Staphylococcus aureus and Pseudomonas aeruginosa. These two bacteria are two of the most common causes of multidrug-resistant infections globally.

Other frontiers

Finding new drugs isn’t the only way AI might help us defeat superbugs. AI could help us optimise existing antibiotics and help doctors to diagnose and treat infections with our existing arsenal of antibiotics.

With some modifications, the general approach of allowing an AI model to discover pieces that fit a disease puzzle could be applied to other diseases like cancer.

But I think it is also important to take a balanced perspective. According to the World Health Organization, there are only 77 new antibacterial treatments in clinical development. Most of these are modifications of existing antibiotics.

Depressingly, even fewer antibiotics will make it to market. According to a research article in the journal Biostatistics, around 86% of all drug candidates that were developed between 2000 and 2015 failed to meet their stated clinical outcomes.

Finding drug leads is only the first step of a long and arduous process. In the safety and efficacy tests, for instance, halicin and abaucin might fail to meet the standards we expect from approved drugs. In that case, the AI method might be the most lasting contribution from these studies.

To my knowledge, AI has not led to the discovery or creation of a single approved drug yet. Of course, the counterargument is that the field of AI-assisted drug discovery is still very new and advances in computing have only recently made the use of AI practical. It can take decades for identified compounds to make it. We’ve not had that much time yet to see how AI can help.

In sum, it’s hard to say with any certainty that AI models in use now will lead to the next life-saving antibiotics. But I think these early steps are promising. AI will get faster and better at finding antibiotics. And given the scale of the problem of superbugs, we will need all the help we can get.

Anirban Mahapatra is a scientist by training and the author of a book on COVID-19. The views expressed are personal.