AI discovers new antibiotic

While the world races to find drugs and a vaccine to combat the Covid-19 coronavirus, we also face a significant threat from antibiotic resistant bacteria. But this week researchers at  the Massachusetts Institute of Technology announced they’ve been using Artificial Intelligence, in other words computers that can kind of “think” for themselves, to dream up new infection-fighting drugs. They’ve used the new system to discover an antibiotic called halicin and it’s proving to be very good at stopping bacterial superbug infections in experimental mice, and could soon be tested on people. The other piece of good news is that this AI system might also help to develop drugs to fight the new coronavirus too. Chris Smith spoke to Jim Collins...

Jim - The issue we set out to solve was antibiotic resistance. We are facing a global crisis: bacterial pathogens, those that cause nasty infections, are growing increasingly resistant to antibiotics, which means they're not responsive to antibiotics, due to the overuse of these drugs. The number of new antibiotics being developed and approved is dramatically dropping, as pharmaceutical companies and biotech companies are getting out of the antibiotics industry. So here at MIT we decided to see if we could harness the power of artificial intelligence, also called AI, in order to address the antibiotic resistance crisis.

Chris - Well, just before we consider how you've done that, Jim, you said something quite interesting, which is that we need a lot of antibiotics, but pharmaceutical companies are deserting this industry like rats off a sinking ship. Why?

Jim - The economics for antibiotics is broken and there are many reasons for it. One is that it costs just as much to develop an antibiotic as it does another drug, for example, say one to treat cancer or to treat blood pressure. Antibiotics, however, are only prescribed for very short periods of time. You take an antibiotic for a few days, maybe a week, whereas you take a blood pressure medicine for the rest of your life. And so the economics for delivery and use are quite low. Second, is that as one comes up with new antibiotics, they're actually not being prescribed out of concern for resistance to arise. So as new antibiotics were being developed and approved, they were being shelved and so companies couldn't even sell their product.

Chris - So how can AI help to address what is an economic disincentive for the pharmaceutical industry to go down this path?

Jim - We wanted to see if we could harness AI to quickly and inexpensively expand our antibiotic arsenal and thereby significantly de-risking what's called the preclinical stage. So the stage of development before you move to human trials, coming up with whole new molecules that could overcome existing resistance and thereby dramatically dropping the economic barriers. So in our project we trained an AI model, so this is a computer based model, with information on existing drugs against E. coli. So E. coli is a bug found in our guts, often harmless to us, but in many cases can be harmful. We've all heard of scares around contaminated meats. In this case we look to see which of these FDA drugs would have some antibacterial activity, so could it inhibit the growth of E. coli or kill it? We took those data, along with the information about each of those drugs, to train this computer-based model to learn molecular features, so features of those molecules that are associated with an antibiotic or being antibacterial.

Chris - So the AI is interrogating the actual physical shape of the molecule. It knows what works against these particular microbes?

Jim - We train it so that it can learn which of the features, even at an atom by atom level, of these chemical compounds appear to be associated with antibacterial activity. We then applied the model to a drug repurposing library that consisted of 6,100 molecules that had been developed as drugs, or were explored as possible drugs, to treat various conditions. And we asked the model to do two things. One was to identify molecules that are predicted to be antibacterial or good antibiotics, but two, to then only identify within that subset, molecules that don't look like our existing antibiotics. And in that drug repurposing library, one molecule fit those criteria and that molecule is what we call HALICIN, which is one of the most powerful antibiotics discovered to date. Its ability to kill a wide range of bacterial pathogens included pathogens that are pan-resistant, that is resistant to effectively all antibiotics; TB, tuberculosis, which has the highest number of deaths as a bacterial pathogen around the world. It was able to kill a pathogen called C. diff, that's a nasty gut pathogen and HALICIN was also able to kill a bug called Acinetobacter baumannii, which is a weird sounding bug, but it's also called the Iraqi bug. The US soldiers and UK soldiers are coming back from service in Iraq and Afghanistan with Acinetobacter baumannii infections, basically in skin wounds, and our existing antibiotics don't treat it well and we showed that HALICIN could actually treat those infections very effectively in skin wound models in mice.

Chris - Now given that the world is currently in the grip of a potential pandemic caused by this coronavirus, is it possible that you could take the technique you've developed and then ask, of the enormous repertoire of drug molecules we already have, could we try some of those and see if we can find a drug that we might not have considered but which might give us an opportunity to intervene with this new pandemic?

Jim - In principle, one could consider using this platform or a related platform to address that challenge. The difficulty in applying it to a viral infection, like the coronavirus, is getting after appropriate data to train the AI model. So what we don't have in many cases are good experimental setups in the lab, that would allow us first to explore which molecules are having an effect. And so for example, in the case of coronavirus, I'm not aware of good cell-based models to get those data. If one had those cell-based experiments that you could generate the data, then then I think this type of platform could become really quite valuable.