Researchers previously doubted the usefulness of the AlphaFold artificial intelligence tool for analyzing protein structure in the drug development process. However, they are now learning how to effectively utilize its advantages.
Using AlphaFold, scientists have been able to predict protein structure and identify hundreds of thousands of potential new psychedelic molecules that could help create new types of antidepressants. This study was the first example that AlphaFold's click-accessible predictions can be as useful for drug discovery as structural data from experimental studies, which can take months or even years.
These advances underscore the significance of AlphaFold, developed by DeepMind in London, which has revolutionized the field of biology. AlphaFold's open database provides structural predictions for almost every known protein. The structures of proteins involved in diseases serve as the basis for the pharmaceutical industry to find and improve promising drugs. However, some scientists have begun to question whether AlphaFold's predictions can replace traditional experimental approaches in the search for new drugs.
Using AlphaFold, scientists have been able to predict protein structure and identify hundreds of thousands of potential new psychedelic molecules that could help create new types of antidepressants. This study was the first example that AlphaFold's click-accessible predictions can be as useful for drug discovery as structural data from experimental studies, which can take months or even years.
These advances underscore the significance of AlphaFold, developed by DeepMind in London, which has revolutionized the field of biology. AlphaFold's open database provides structural predictions for almost every known protein. The structures of proteins involved in diseases serve as the basis for the pharmaceutical industry to find and improve promising drugs. However, some scientists have begun to question whether AlphaFold's predictions can replace traditional experimental approaches in the search for new drugs.
AlphaFold Skepticism
«AlphaFold is a real revolution. When we have a good structure, we can use it to develop drugs» — argues Jens Karlsson, a computational chemist at Uppsala University in Sweden.
Brian Scheuchet, a pharmaceutical chemist at the University of California, San Francisco, noted that attempts to use AlphaFold to find new drugs are often subject to considerable skepticism. «There is a lot of hype. Every time someone claims that this or another tool can transform the drug discovery process, there is justifiable skepticism» — he adds.
Shoichet found more than a dozen studies that showed AlphaFold's predictions were less useful than experimentally derived protein structures, such as X-ray crystallography, for identifying potential drugs using a technique called protein-ligand docking.
«AlphaFold is a real revolution. When we have a good structure, we can use it to develop drugs» — argues Jens Karlsson, a computational chemist at Uppsala University in Sweden.
Brian Scheuchet, a pharmaceutical chemist at the University of California, San Francisco, noted that attempts to use AlphaFold to find new drugs are often subject to considerable skepticism. «There is a lot of hype. Every time someone claims that this or another tool can transform the drug discovery process, there is justifiable skepticism» — he adds.
Shoichet found more than a dozen studies that showed AlphaFold's predictions were less useful than experimentally derived protein structures, such as X-ray crystallography, for identifying potential drugs using a technique called protein-ligand docking.
This method, common in the early stages of drug development, involves modeling the possible interactions of billions of chemicals with key regions of a target protein to identify compounds that help alter the protein's activity. Previous studies have generally shown that using predicted AlphaFold structures, models are weak at recognizing known drugs that bind to a particular protein.
A research team led by Scheuchet and Brian Roth, a structural biologist at the University of North Carolina at Chapel Hill, reached similar conclusions by analyzing the structures of two proteins linked to neuropsychiatric disorders in comparison to known drugs. The scientists wondered whether small differences in the experimentally derived structures could be the reason that the predicted structures missed certain compounds that bind to the proteins, but identified other equally promising ones.
To this end, the team used the experimental structures of two proteins to virtually screen hundreds of millions of potential drugs. One of the proteins under investigation, a receptor sensitive to the neurotransmitter serotonin, had previously been identified using cryo-electron microscopy. The structure of the second protein, known as the σ-2 receptor, was mapped using X-ray crystallography.
Substance differences
The researchers conducted a similar screening based on protein models obtained from the AlphaFold database. As a result, they synthesized hundreds of the most promising compounds, identified from both predicted and experimental structures, and tested their activity in vitro.
Screening based on predictions and experimental data revealed a variety of new drug candidates. «No molecules with similar characteristics were found. They were completely different» — notes Shoichet.
Substance differences
The researchers conducted a similar screening based on protein models obtained from the AlphaFold database. As a result, they synthesized hundreds of the most promising compounds, identified from both predicted and experimental structures, and tested their activity in vitro.
Screening based on predictions and experimental data revealed a variety of new drug candidates. «No molecules with similar characteristics were found. They were completely different» — notes Shoichet.
The team learned that «hit rates» — the proportion of compounds that significantly alter the activity of a protein-were nearly identical for both groups. However, the AlphaFold-based structures identified the drugs that could most effectively activate serotonin receptors.
The psychedelic drug LSD acts in part through this pathway, and many scientists are looking for non-hallucinogenic compounds with similar properties as potential antidepressants. «This result is truly groundbreaking» — comments Shoichet.
According to the researchers, a yet-to-be-published paper from Karlsson's group confirmed that AlphaFold structures are well suited to drug discovery for a popular class of targets - G-protein-coupled receptors, where their hit rate is about 60%.
The psychedelic drug LSD acts in part through this pathway, and many scientists are looking for non-hallucinogenic compounds with similar properties as potential antidepressants. «This result is truly groundbreaking» — comments Shoichet.
According to the researchers, a yet-to-be-published paper from Karlsson's group confirmed that AlphaFold structures are well suited to drug discovery for a popular class of targets - G-protein-coupled receptors, where their hit rate is about 60%.
Prediction
Karlsson emphasizes that confidence in predicted protein structures can dramatically transform the drug discovery process. Determining structures under experimental conditions is challenging, and many potential targets may not be accessible to existing methods. «It would be great if we could just push a button and get a structure that is convenient for finding ligands» — he says.
Sriram Subramaniam, a structural biologist at the University of British Columbia, believes the two proteins chosen by Scheuchet and Roth's team are excellent candidates for the AlphaFold application. Experimental models of the related proteins and detailed maps of their interactions with drug compounds are already available. «AlphaFold is a game changer. It changes approaches to how we do research» — he adds.
However, Karen Akinsanya, President of R&D at Schrödinger, emphasizes that AlphaFold is not a one-size-fits-all solution. The predicted structures may be useful for some purposes but not for all, and it is far from clear which ones are applicable. In about 10% of cases, AlphaFold's predictions, which are considered highly accurate, differ significantly from experimental structures, according to the study.
Karlsson emphasizes that confidence in predicted protein structures can dramatically transform the drug discovery process. Determining structures under experimental conditions is challenging, and many potential targets may not be accessible to existing methods. «It would be great if we could just push a button and get a structure that is convenient for finding ligands» — he says.
Sriram Subramaniam, a structural biologist at the University of British Columbia, believes the two proteins chosen by Scheuchet and Roth's team are excellent candidates for the AlphaFold application. Experimental models of the related proteins and detailed maps of their interactions with drug compounds are already available. «AlphaFold is a game changer. It changes approaches to how we do research» — he adds.
However, Karen Akinsanya, President of R&D at Schrödinger, emphasizes that AlphaFold is not a one-size-fits-all solution. The predicted structures may be useful for some purposes but not for all, and it is far from clear which ones are applicable. In about 10% of cases, AlphaFold's predictions, which are considered highly accurate, differ significantly from experimental structures, according to the study.
Akinsanya also notes that even though predicted structures can be useful for identifying potential compounds, more detailed experimental information is often needed to optimize the properties of specific candidates.
Conclusion
Shoichet agrees that AlphaFold's results are not always positive. «There are quite a few models that we didn't even bother to investigate because of their poor quality» — he states. At the same time, he noted that in about a third of cases, AlphaFold's predictions can significantly speed up a project. «This saves several years, which is a significant advantage» — he adds.
One of the goals of Isomorphic Labs, DeepMind's drug development spin-off, is to use AlphaFold to find new drugs. On Jan. 7, the company announced deals worth at least $82.5 million to find drugs for pharmaceutical companies Novartis and Eli Lilly using machine learning technologies such as AlphaFold.
Conclusion
Shoichet agrees that AlphaFold's results are not always positive. «There are quite a few models that we didn't even bother to investigate because of their poor quality» — he states. At the same time, he noted that in about a third of cases, AlphaFold's predictions can significantly speed up a project. «This saves several years, which is a significant advantage» — he adds.
One of the goals of Isomorphic Labs, DeepMind's drug development spin-off, is to use AlphaFold to find new drugs. On Jan. 7, the company announced deals worth at least $82.5 million to find drugs for pharmaceutical companies Novartis and Eli Lilly using machine learning technologies such as AlphaFold.
The company says it will use an updated version of AlphaFold that will be able to predict protein structures in combination with drugs and other interacting molecules.
However, DeepMind has not yet said when this update will be available to researchers, as it was with previous versions of AlphaFold. The developers plan to introduce a competing tool called RoseTTAFold All-Atom soon. The researchers note that such tools cannot completely replace experimental methods, but their potential for finding new drug compounds should be taken seriously.
«Many people hope that AlphaFold can do everything on its own, while many structural biologists are trying to find reasons why their results are still needed. It is difficult to find the right balance» — concludes Karlsson.
However, DeepMind has not yet said when this update will be available to researchers, as it was with previous versions of AlphaFold. The developers plan to introduce a competing tool called RoseTTAFold All-Atom soon. The researchers note that such tools cannot completely replace experimental methods, but their potential for finding new drug compounds should be taken seriously.
«Many people hope that AlphaFold can do everything on its own, while many structural biologists are trying to find reasons why their results are still needed. It is difficult to find the right balance» — concludes Karlsson.