Antibiotic resistance is one of the biggest threats to human health worldwide. In 2019, it directly caused 1.27 million deaths and led to nearly 5 million deaths. No new classes of antibiotics have been developed for decades.
Now, the researchers report that they have used artificial intelligence to discover a new class of antibiotic candidates. A team in the James Collins lab at MIT and Harvard's Broad Institute uses an artificial intelligence called deep Xi to screen the antibiotic activity of millions of compounds.
They then tested 283 promising compounds in mice and found several that were effective against methicillin-resistant Staphylococcus aureus (MRSA) and vancomycin-resistant enterococci, some of the most stubborn and hard-to-kill pathogens. Unlike a typical AI model that operates as an incomprehensible "black box", the reasoning of the model can be followed and the biochemistry behind it can be understood.
This progress builds on previous research by the group and others on the importance of this new research in harnessing artificial intelligence to help guide the development of new antibiotics.
I'm very excited about this new work from Collins Lab – I think it's a great next breakthrough. This is a field of study that was not a field of study until five years ago. This is a very exciting and very emerging area of work whose main goal is to use artificial intelligence for antibiotic discovery and antibiotic design. My own lab has been working on this for the past five years.
In this study, researchers used deep Xi to try to discover a new type of antibiotic. They also implement the concept of "explainable AI", which is interesting because when we think about machine Xi and deep Xi, we see them as black boxes. So I think it's interesting to start incorporating explainability into some of the models that we're building that apply AI to biology and chemistry. The authors were able to find several compounds that appear to reduce infections in mouse models, so this is always exciting.
In general, artificial intelligence and machines can systematically and very quickly mine structures or any type of dataset that you provide them. If you consider the traditional antibiotic discovery process, it takes about 12 years to discover a new antibiotic, and three to six years to discover any clinical candidate. Then you have to transition them to Phase I, Phase II, and Phase III clinical trials.
Now, with machines, we've been able to speed up this process. For example, in my own work with my colleagues and I, we can discover thousands or hundreds of thousands of preclinical candidates in a matter of hours instead of having to wait three to six years. I think AI in general has achieved this. I think another example is this work in the Collins lab, where by using deep Xi in this case, the team has been able to classify millions of compounds to identify some that look promising. This is difficult to do manually.
There is still a gap there. You will need a systematic toxicity study followed by a pre-IND [investigational new drug] study. The Food and Drug Administration is asking you to conduct these studies to evaluate whether your potentially exciting drug can be transitioned to a phase I clinical trial, which is the first phase of any clinical trial. Therefore, these different steps still need to be taken. But again, I think it's another very exciting advance in the use of AI in the field of microbiology and antibiotics, which is really an emerging field. Our dream is that one day AI will be able to create antibiotics that can save lives.
Yes, they demonstrated this in two mouse models, which is interesting. It's always more exciting whenever you have data on mouse infections – it shows that these compounds are actually able to reduce infections in real-world mouse models.
As another example of the use of artificial intelligence, we recently mined the genome and proteome of extinct organisms in our own lab, and we were able to identify a number of clinical antibiotic candidates.
I think it's important if one day we're going to think of AI as an engineering discipline. In engineering, you're always able to break down the different parts that make up a structure and understand what each part does. But as far as AI is concerned, especially deep Xi, because it's a black box, we don't know what's going to happen in between.
In order to give us a composite x or y or a solution x or y, it is very difficult to recreate what happened. Therefore, starting to dig into the black box to understand what is actually happening in each step is a critical step for us to achieve our goals. Ability to transform AI into an engineering discipline. The first step in the right direction is to use explainable AI to try and understand what the machine is actually doing. It's no longer a black box – maybe a gray box.
List of high-quality authors