The deep learning will perhaps help save the lives of millions of people. Indeed, it is by relying on this type of machine learning techniques that researchers from the Massachussetts Institute of Technology (MIT) have succeeded in finding a revolutionary new antibiotic, capable of killing most bacteria harmful to it. humans, including some that are resistant to all antibiotics known to date.
This discovery is all the more important as bacteria tend to become increasingly resistant, due to overconsumption of antibiotics. However, the pharmaceutical industry has great difficulty in overcoming this problem, because finding new molecules by traditional methods is very expensive and takes time. This new method based on artificial intelligence could therefore be a game-changer.
Find molecules by scanning files
To achieve this result, the researchers developed a prediction model capable of saying whether a molecule is capable of killing the E. Coli bacteria or not. This model was trained on a set of 2,500 molecules, whose biochemical properties are well known. It was then applied to a computer library of 6,000 representations of molecules. The software has selected one that was previously rather studied to treat diabetes.
Tests on dozens of bacterial cultures have validated its antibiotic nature. And the good news is that this molecule is not expected to be toxic to humans. In any case, this is what a second prediction model created by the researchers provides. This new antibiotic has since been called “halicin”, in homage to the HAL computer in the 2001 film : Space Odyssey .
An antibiotic that should stay effective for a long time
If this molecule is so effective, it is because it attacks the electrochemical gradient of the membrane of the bacteria, a property which allows it to store energy and therefore to stay alive. According to the researchers, bacteria should have trouble developing resistance against this type of product, because the mutations in the membranes are long and complex.
Following this discovery, the researchers decided to scan a set of 100 million molecules from the ZINC15 library, which contains 1.5 billion. In the space of three days, the software has selected 23 candidates, eight of whom have antibiotic properties. And among these, two appear to be particularly powerful. In the future, researchers believe they can use deep learning to create antibiotics tailored to the needs of patients. For example to target only certain bacteria, but not others. In short, pharmaceutical engineering is about to enter a new era.