Researchers have used the protein-structure-prediction tool AlphaFold to identify hundreds of thousands of potential new psychedelic molecules — which could help to develop new kinds of antidepressant. The research shows, for the first time, that AlphaFold predictions — available at the touch of a button — can be just as useful for drug discovery as experimentally derived protein structures, which can take months, or even years, to determine.
The development is a boost for AlphaFold, the artificial-intelligence (AI) tool developed by DeepMind in London that has been a game changer in biology. The public AlphaFold database holds structure predictions for nearly every known protein. Protein structures of molecules implicated in disease are used in the pharmaceutical industry to identify and improve promising medicines. But some scientists had been starting to doubt whether AlphaFold’s predictions could stand in for gold standard experimental models in the hunt for new drugs.
“AlphaFold is an absolute revolution. If we have a good structure, we should be able to use it for drug design,” says Jens Carlsson, a computational chemist at the University of Uppsala in Sweden.
On supporting science journalism
If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.
AlphaFold scepticism
Efforts to apply AlphaFold to finding new drugs have been met with considerable scepticism, says Brian Shoichet, a pharmaceutical chemist at the University of California, San Francisco. “There is a lot of hype. Whenever anybody says ‘such and such is going to revolutionize drug discovery,’ it warrants some scepticism.”
Shoichet counts more than ten studies that have found AlphaFold’s predictions to be less useful than protein structures obtained with experimental methods, such as X-ray crystallography, when used to identify potential drugs in a modelling method called protein–ligand docking.
This approach — common in the early stages of drug discovery — involves modelling how hundreds of millions or billions of chemicals interact with key regions of a target protein, in the hope of identifying compounds that alter the protein’s activity. Previous studies have tended to find that when AlphaFold-predicted structures are used, the models are poor at singling out drugs already known to bind to a particular protein.
Researchers led by Shoichet and Bryan Roth, a structural biologist at the University of North Carolina at Chapel Hill, came to a similar conclusion when they checked AlphaFold structures of two proteins implicated in neuropsychiatric conditions against known drugs. The researchers wondered whether 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 ones that were no less promising.
To test this idea, the team used experimental structures of the two proteins to virtually screen hundreds of millions of potential drugs. One protein, a receptor that senses the neurotransmitter serotonin, was previously determined using cryo-electron microscopy. The structure of the other protein, called the σ-2 receptor, had been mapped using X-ray crystallography.
Drug differences
They ran the same screen with models of the proteins plucked from the AlphaFold database. They then synthesized hundreds of the most promising compounds identified with either the predicted and experimental structures and measured their activity in the lab.
The screens with predicted and experimental structures yielded completely different drug candidates. “There were no two molecules that were the same,” says Shoichet. “They didn’t even resemble each other.”
But to the team’s surprise, the ‘hit rates’ — the proportion of flagged compounds that actually altered protein activity in a meaningful way — were nearly identical for the two groups. And AlphaFold structures identified the drugs that activated the serotonin receptor most potently. The psychedelic drug LSD works partly through this route, and many researchers are looking for non-hallucinogenic compounds that do the same thing, as potential antidepressants. “It’s a genuinely new result,” says Shoichet.
Prediction power
In unpublished work, Carlsson’s team has found that AlphaFold structures are good at identifying drugs for a sought-after class of target called G-protein-coupled receptors, for which their hit rate is around 60%.
Having confidence in predicted protein structures could be game-changing for drug discovery, says Carlsson. Determining structures experimentally isn’t trivial, and many would-be targets might not yield to existing experimental tools. “It would be very convenient if we could push the button and get a structure we can use for ligand discovery,” he says.
The two proteins that Shoichet and Roth’s team picked are good candidates for relying on AlphaFold, says Sriram Subramaniam, a structural biologist at the University of British Columbia in Vancouver, Canada. Experimental models of related proteins — including detailed maps of the regions where drugs bind to them — are readily available. “If you stack the deck, AlphaFold is a paradigm shift. It changes the way we do things,” he adds.
“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. Predicted structures are helpful for some drug targets, but not others, and it’s not always clear which applies. In about 10% of cases, predictions AlphaFold deems highly accurate are substantially different from the experimental structure, a study found.
And even when predicted structures can help to identify leads, more detailed experimental models are often needed to optimize the properties of a particular drug candidate, Akinsanya adds.
Big bet
Shoichet agrees that AlphaFold predictions are not universally useful. “There were a lot of models that we didn’t even try because we thought they were so bad,” he says. But he estimates that in about one-third of cases, an AlphaFold structure could jump-start a project. “Compared to actually going out and getting a new structure, you could advance the project by a couple of years and that’s huge,” he says.
That is the goal of Isomorphic Labs, DeepMind’s drug-discovery spin-off in London. On 7 January, the company announced deals worth a minimum of US$82.5 million — and up to $2.9 billion if business targets are met — to hunt for drugs on behalf of pharmaceutical giants Novartis and Eli Lilly using machine-learning tools such as AlphaFold.
The company says that the work will be aided by a new version of AlphaFold that can predict the structures of proteins when they are bound to drugs and other interacting molecules. DeepMind has not yet said when — or whether — the update will be made available to researchers, as earlier versions of AlphaFold have been. A competing tool called RoseTTAFold All-Atom will be made available soon by its developers.
Such tools won’t fully replace experiments, scientists say, but their potential to help find new drugs shouldn’t be discounted. “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.”
This article is reproduced with permission and was first published on January 18, 2024.