Siouxsie Wiles: People plus machines find new treatment for superbug

New superbug-killing antibiotic discovered using AI! That was how news outlets around the world summarised a recent study published in the journal Nature Chemical Biology.

I’m curious what people imagine when they read or hear a headline like that, so I asked around. One fantastic suggestion was a swarm of tiny robots furiously rummaging through boxes before triumphantly presenting a vial to a lab of scientists. Another suggestion was scientists huddled around a computer furiously typing prompts into ChatGPT!

Those are both great suggestions, but what’s the reality? Let’s dive into that study and find out!

Our story starts with a group of researchers – led by Jonathan Stokes from McMaster University in Canada and James Collins from the Massachusetts Institute of Technology and Harvard in the US – keen to find new antibiotics that kill the bacterium Acinetobacter baumannii.

A. baumannii causes all sorts of different infections, from bloodstream and wound infections to pneumonia and meningitis. It’s becoming more and more of a problem in hospitals where it makes already sick people even more unwell. It’s also becoming more and more resistant to antibiotics making treatment very difficult. That’s why the World Health Organisation calls it one of its “priority pathogens”.

The first question Stokes and colleagues asked was what kinds of compounds do we already have that kill A. baumannii? This bit did involve actual robots. Or at least robotic machines able to quickly and accurately dispense tiny amounts of thousands of compounds – 7684 to be precise – and then add a squirt of bacteria to each of them. I’m very jealous. In my lab we have to do this sort of thing by hand. It’s as tedious and tiring as it sounds.

The bacteria and compounds were incubated together and after a few hours another machine was used to measure whether the bacteria had grown or not. The researchers found that 480 of the 7684 compounds were able to stop A. baumannii growing, at least to some degree. These they classified as “active”.

Next, the researchers fed all that data – whether each compound was active or not – along with a summary of each compound’s chemical structure, to what’s called a message-passing deep neural network. This is a type of artificial intelligence called machine learning. What the researchers wanted the neural network to learn was what sort of chemical structures were able to stop A. baumannii growing.

Then they fed the neural network another set of data – summaries of the chemical structures of another 6680 compounds they had never tested. Now they asked the neural network to predict which compounds could be active against A. baumannii. Then they asked it to tell them which of those compounds had a different structure to any known antibiotics.

In just a few hours the neural network whittled those 6680 compounds to the 240 it classified as potentially active and structurally unique.

Then it was back into the lab to see if they really were able to stop A. baumannii from growing. And from those experiments the researchers identified RS102895 as the most promising anti-A. baumannii compound. They renamed it abautin.

Now the question was, how does abautin work? Back into the lab they went, making resistant mutants and testing it against lots of other bacteria, to figure that out. Finally, they formulated abautin into a cream and showed that it could suppress the growth of A. baumannii in infected wounds in mice. The results look really promising, though the researchers didn’t present any data to confirm whether abautin could treat bloodstream infections or pneumonia.

So, there you have it. Artificial intelligence played an important part in discovering a potential new antibiotic. But it couldn’t have done it without the data generated from the enormous amounts of work put in by the researchers.

And it’s more of those sorts of lab experiments – like testing compounds against bacteria growing in conditions more like they encounter when infecting a person – that will improve the data the neural network can learn from.