Pharmaceutical researchers in the United Kingdom are using the artificial intelligence tool AlphaFold to develop new psychedelic drugs, according to research released last month.
The research, which has not yet been peer-reviewed, shows that using an AI tool is as useful as experimentally produced protein structures, which can take up to years to investigate. The findings give further evidence of the utility of AlphaFold, an artificial intelligence tool developed by DeepMind in London.
“AlphaFold is an absolute revolution,” Jens Carlsson, a computational chemist at the University of Uppsala in Sweden, told Nature. “If we have a good structure, we should be able to use it for drug design.”
AlphaFold, a public database that holds structure predictions for nearly every known protein, has proven to be a significant advancement in biological research. With the tool, pharmaceutical companies can use the protein structure of molecules implicated in disease to identify and improve promising new drugs. But not everyone is sold on the AlphaFold AI technology as a tool for developing new drugs.
“There is a lot of hype,” said Brian Shoichet, a pharmaceutical chemist at the University of California, San Francisco. “Whenever anybody says ‘such and such is going to revolutionize drug discovery’, it warrants some skepticism.”
Shoichet noted that more than 10 studies have shown that predictions made by AlphaFold were not as useful as protein structures obtained with experimental methods such as X-ray crystallography when used to identify potential drugs in a modeling method called protein–ligand docking.
Protein-ligand docking is a method commonly used in the early stages of drug development that uses modeling to determine how hundreds of millions or even billions of compounds interact with key regions of a protein to identify substances that alter the protein’s activity. Previous studies have shown that when structures predicted by AlphaFold are used, the models do not do a good job of identifying drugs that are already known to bind to a particular protein.
Shoichet and Bryan Roth, a structural biologist at the University of North Carolina at Chapel Hill, led a team of researchers who came to a similar conclusion when they tested AlphaFold structures of two proteins implicated in neuropsychiatric conditions against known drugs. The researchers hoped to determine if small differences from experimental structures might cause the predicted structures to miss certain compounds that bind to proteins, but also make them able to identify different potentially useful compounds.
The researchers tested the hypothesis by using experimental structures of two proteins to virtually screen hundreds of millions of potential new drugs. One of the proteins, a receptor that senses the neurotransmitter serotonin, was previously determined through cryo-electron microscopy. The other protein’s structure, called the σ-2 receptor, had been mapped with X-ray crystallography.
The researchers ran the same screen with models taken from the AlphaFold database. They then synthesized hundreds of compounds deemed to be the most promising identified by either the AI-predicted or experimental structures and measured their activity. Significantly, the screens with predicted and experimental structures identified different drug candidates.
“There were no two molecules that were the same,” Shoichet said. “They didn’t even resemble each other.”
The researchers noticed, however, that the proportion of flagged compounds that altered protein activity in a significant manner –known as the hit rate– was almost identical for the two groups. AlphaFold structures identified the drugs that activated the serotonin receptor most strongly.
LSD, the drug commonly known as acid, works in part by activating the serotonin receptor. Many researchers are trying to identify drugs that work similarly without causing hallucinogenic effects so they can be used as antidepressants.
“It’s a genuinely new result,” said Shoichet.
Carlsson and his team of researchers found in research that has yet to be published that AlphaFold structures had about a 60% hit rate class of target called G-protein-coupled receptors. He said that having a tool that can reliably predict protein structures would be revolutionary for the drug development industry.
“It would be very convenient if we could push the button and get a structure we can use for ligand discovery,” he said.
But AlphaFold will not take the place of other methods of discovering new drugs. Predicted structures can be helpful for some but not all drug targets, and which applies is not always readily apparent.
“This is not a panacea,” says Karen Akinsanya, president of research and development for therapeutics at Schrödinger, a drug-software company based in New York City that is using AlphaFold.
AI is not likely to replace experimentation for new drugs, but researchers say the value of AlphaFold and similar tools should be recognized.
“There’s a lot of people that want AlphaFold to do everything, and a lot of structural biologists want to find reasons to say we are still needed,” says Carlsson. “Finding the right balance is difficult.”
