Hundreds of thousands of possible new hallucinogenic compounds have been identified by researchers using the protein-structure prediction tool AlphaFold. This could aid in the development of novel antidepressants. For the first time, the study demonstrates that instantaneous AlphaFold predictions can be just as helpful for drug development as empirically generated protein structures, which typically take months or even years to uncover.
The advancement is good news for AlphaFold, the London-based DeepMind artificial intelligence (AI) technology that has revolutionized the field of biology. Nearly all known proteins have structural predictions available in the public AlphaFold database. The pharmaceutical industry uses the protein structures of compounds linked to disease to find and enhance promising treatments. However, a growing number of experts were beginning to question whether AlphaFold’s forecasts could replace industry-standard experimental models when developing new medications.
“AlphaFold represents a complete revolution.” Drug design should be possible if we have a suitable structure, according to Jens Carlsson, a computational chemist at the University of Uppsala in Sweden.
ALPHAFOLD SCEPTICISM
According to Brian Shoichet, a pharmaceutical chemist at the University of California, San Francisco, there has been a great deal of skepticism over efforts to use AlphaFold in the search for new medications. “The excitement is enormous. There should always be some skepticism when someone claims that “such and such is going to revolutionize drug discovery.”
Shoichet cites more than ten studies that show that when protein structures are employed in a modeling technique known as protein–ligand docking to identify possible medications, AlphaFold’s predictions are less helpful than those derived with experimental approaches, such X-ray crystallography.
In an effort to find compounds that change a target protein’s activity, this method, which is frequently used in the early phases of drug discovery, models the interactions between hundreds of millions or billions of molecules and important areas of the protein. Prior research has generally found that the models are not very good at distinguishing medications that are previously known to bind to a certain protein when AlphaFold-predicted structures are employed.
Researchers under the direction of Shoichet and University of North Carolina at Chapel Hill structural biologist Bryan Roth reached a similar conclusion when they compared the AlphaFold structures of two proteins linked to neuropsychiatric disorders with medications that were already on the market. The researchers questioned if minute variations from experimental structures may lead to the predicted structures missing some protein-binding chemicals while allowing them to uncover other, equally promising ones.
The researchers virtually screened hundreds of millions of possible medications using the experimental structures of the two proteins in order to test this theory. Cryo-electron microscopy was used to identify one protein, a receptor that detects the neurotransmitter serotonin. Using X-ray crystallography, the structure of the other protein, known as the σ-2 receptor, had been determined.
DIFFERENCES IN DRUG
Using models of the proteins taken from the AlphaFold database, they ran the same screen. After that, they created hundreds of the most promising compounds that were found to have both experimental and anticipated structures, and then conducted lab tests to see how active they were.
Using both experimental and anticipated structures, the screenings produced entirely different therapeutic candidates. According to Shoichet, “no two molecules were the same.” “They weren’t even similar to one another.”
The “hit rates,” or the percentage of identified chemicals that genuinely changed protein function in a significant way, surprised the team because they were almost the same for both groups. And the medications that most strongly stimulated the serotonin receptor were found using AlphaFold structures. This is one way that the psychedelic drug LSD functions, and many scientists are searching for non-hallucinogenic substances that have a similar effect as possible antidepressants. “It’s a truly novel outcome,” Shoichet declares.
ABILITY TO PREDICT
In previously undisclosed research, Carlsson’s group discovered that AlphaFold structures have a 60% hit probability when it comes to finding medications targeting G-protein-coupled receptors, a popular class of target.
According to Carlsson, having faith in predicted protein structures may revolutionize the drug discovery process. It is not easy to determine structures experimentally, and many potential targets may not yield to current experimental methods. “If we could just press a button and get a structure we can use for ligand discovery, that would be very convenient,” he states.
For relying on AlphaFold, the two proteins selected by Shoichet and Roth’s team make sense, according to Sriram Subramaniam, a structural biologist at the University of British Columbia in Vancouver, Canada. There are many accessible experimental models of related proteins, including thorough diagrams of the locations where medications bind to them. When the cards are stacked, AlphaFold represents a paradigm change. It modifies how we operate,” he continues.
“This is not a magic bullet,” states Karen Akinsanya, president of therapeutics research and development at Schrödinger, a New York City-based provider of medication software that uses AlphaFold. Certain therapeutic targets benefit more from predicted structures than others, and it’s not always obvious which ones. According to a research, roughly 10% of the predictions that AlphaFold considers to be very accurate diverge significantly from the experimental structure.
Furthermore, Akinsanya notes that more thorough experimental models are frequently required to maximize the qualities of a certain drug candidate, even in cases when predicted structures can aid in the identification of leads.
LARGE BET
Shoichet concurs that forecasts made by AlphaFold are not always helpful. He claims, “There were a lot of models that we thought were so bad that we didn’t even try.” However, he believes that an AlphaFold structure could kickstart a project in roughly one-third of cases. “That’s a big advantage—you could move the project forward by a few years compared to actually going out and getting a new structure,” he says.
That is the aim of DeepMind’s London-based drug-discovery spin-off, Isomorphic Labs. On January 7, the company announced agreements to use machine learning techniques like AlphaFold to find medications on behalf of pharmaceutical giants Novartis and Eli Lilly for a minimum of US$82.5 million and up to $2.9 billion if sales targets are accomplished.
According to the business, a new version of AlphaFold that can predict protein structures when bound to pharmaceuticals and other interacting chemicals will help with the task. As with previous AlphaFold iterations, DeepMind has not yet stated whether or when the update would be made available to researchers. The creators of RoseTTAFold All-Atom, a rival tool, want to release it soon.
Scientists assert that while these technologies won’t completely replace tests, they do have the potential to aid in the discovery of new medications. According to Carlsson, “many structural biologists want to find reasons to say we are still needed, and there are many people who want AlphaFold to do everything.” “It’s hard to find the right balance.”