We are used to the fact that almost every innovation in the world of artificial intelligence is accompanied by an enormous amount of unspecific and less tangible things. Promises, promises and promises that, from time to time, come to fruition. Today, luckily, it is one of those cases. A neural network has just identified a compound, halicin, which has shown great promise in attacking difficult-to-kill bacteria.
But that’s not the best. The best part is that with the discovery of halicin, the use of deep learning approaches applied to the world of antibiotics shows its potential. A potential that comes at the best time : in the middle of a deep drug crisis that threatens the near future of modern medicine .
There are no antibiotics for so many superbugs
The last time a human being said “we have discovered a new class of antibiotics” was the 1980s and many of us were not even born. In essence, all the antibiotics that have been released during these three decades are variations of drugs that have been found before . The reason? It is like “looking for a needle in a haystack” if the haystack had several tens of hectares. The Wellcome Trust estimates that the process required to find one would take no less than 15 years and cost no less than € 1 billion. All this without knowing if it will be a successful process or not.
In any other scenario, one that included an alarming growth of multi-resistant bacteria, this would be nothing more than a historical curiosity more typical of a television contest than of a medium like ours. The problem is that this is not the case. The truth is that our ability to find new molecules has slowed almost as fast as concerns about antibiotic resistance have grown . Luckily, we have (what we think is) a secret weapon :
Deep learning vs multi-resistant bacteria
The rumor that machine learning was going to be a good methodology to tackle this problem has been around for years, but that has little merit. I don’t think there is any contemporary problem for which an artificial intelligence-based solution has not been proposed. The difference is that now, thanks to James Collins, a synthetic biologist at MIT, we have a new method to try to discover new antibiotics quickly and efficiently through machine learning.
Using compounds known to suppress the growth of E. coli , Collins’ team trained a neural network using machine learning to identify potential antibiotics. Once ready, the researchers used it to examine thousands of molecules registered in numerous existing chemical libraries and attempted to predict their effectiveness. The researchers found that nearly 50% of the compounds identified by the network were effective in vitro at killing E. coli .
A very hopeful start
This, despite the fact that we are still talking about 50%, is excellent news because, although these huge libraries of molecules have been available for years, teams of researchers do not have an efficient way of selecting those that are most likely to have antibiotic properties. If, as the Collins team at Cell suggests , neural networks can be used to identify good candidates we could save a lot of time and resources .
The Collins network also identified a compound (halicin) that seems to have good results with pathogens such as Clostridium difficile , Koch’s bacillus or Acinetobacter baumannii . We should not fool ourselves, the research on these neural networks is in its infancy and there is still a lot of work to do , but it is comforting to see how this type of approach leaves the world of possibilities and is materialized in interesting projects.
