Researchers at Stanford Medicine are developing a new artificial intelligence model, SyntheMol, that creates recipes for chemists to synthesize drugs in the lab
Every year, almost five million deaths related to antibiotic resistance occur worldwide. There is therefore an urgent need to find new ways to combat resistant bacterial strains.
Researchers at Stanford Medicine and McMaster University are tackling this problem with generative artificial intelligence. A new model called SyntheMol (for synthesizing molecules) created structures and chemical recipes for six new drugs with the aim of eliminating resistant strains of Acinetobacter baumannii, one of the main pathogens responsible for deaths linked to antibiotic resistance.
“Public health needs to quickly develop new antibiotics,” said Dr. James Zou, associate professor of biomedical data science and co-senior author of the study. “Our hypothesis was that there are many potential molecules that could be effective drugs, but we have not made or tested them yet.” “So we wanted to use AI to design completely new molecules that have never existed before in nature gave.”
Before the advent of generative AI, the same type of artificial intelligence technology that underlies large language models like ChatGPT, researchers had taken different computational approaches to developing antibiotics. They used algorithms to search existing drug libraries and identify compounds most likely to work against a particular pathogen. This technique, which examined 100 million known compounds, produced results but barely scratched the surface in searching for all the chemical compounds that might have antibacterial properties.
“The chemical space is huge,” says Kyle Swanson, a doctoral student in computer science at Stanford and co-author of the study. “It is estimated that there are approximately 1,060 possible drug-like molecules. So 100 million isn’t even close to enough to cover the entire space.
Hallucinate to develop drugs
Generative AI’s tendency to “hallucinate,” or invent answers out of thin air, could be a boon for drug discovery, but previous attempts to use this type of AI to develop new drugs have resulted in compounds that would be impossible to produce in the real world, explains Swanson. The researchers had to set limits on SyntheMol’s activity, that is, ensure that every molecule developed by the model could be synthesized in a laboratory.
“We addressed this problem by trying to bridge the gap between computational work and laboratory validation,” explains Swanson.
The model was trained to develop potential drugs using a library of more than 130,000 molecular blocks and a series of validated chemical reactions. The model generated not only the final compound, but also the steps to follow with those components, giving the researchers a set of recipes for making the drugs.
The researchers also trained the model using existing data on the antibacterial activity of various chemicals against A. baumannii. Using these guidelines and its starting blocks, SyntheMol generated approximately 25,000 possible antibiotics and the recipes for making them in less than nine hours. To prevent bacteria from quickly developing resistance to the new compounds, the researchers filtered the compounds created so that they differed only from the existing ones.
“We now not only have completely new molecules, but also explicit instructions for producing them,” says Zou.
A new chemical space
The researchers selected the 70 compounds with the greatest potential to kill the bacteria and worked with Ukrainian chemical company Enamine to synthesize them. The company was able to effectively produce 58 of these compounds, six of which killed a resistant strain of A. baumannii when researchers tested them in the lab. These new compounds also demonstrated antibacterial activity against other types of infectious bacteria prone to antibiotic resistance, such as E. coli, Klebsiella pneumoniae and MRSA.
The scientists were able to demonstrate the toxicity of two of the six compounds in mice because the other four did not dissolve in water. The two who tried them appeared to be safe; The next step is to test the drugs on mice infected with A. baumannii to see whether they work in a living organism, Zou said.
The six compounds are very different from each other and from existing antibiotics. Researchers don’t know how its antibacterial properties work at the molecular level, but studying these details could provide general principles relevant to the development of other antibiotics.
“This artificial intelligence is designing and teaching us a completely new part of chemical space that humans have not explored before,” says Zou.
REFERENCE
Generative AI for the development and validation of easily synthesized and structurally novel antibiotics