analysis DeepMind’s AlphaFold model has predicted almost all known protein structures discovered to date, although its ability to help scientists discover new drugs remains unproven.
Proteins are complex molecules created by organisms to perform the biological functions necessary for life. These chains generally consist of a chain of 20 amino acids and can be folded in countless ways, with their final shape determining how they function and interact with other things.
Determining how a protein will fold is not a simple process. For example, let’s say you wanted to synthesize a protein or slightly change how it works. You can’t adjust its amino acids or come up with a new set of amino acids and know for sure how they will evolve and function when folded. This is where computers come in.
Advances in AI algorithms and training led to the development of software like AlphaFold that can accurately predict the 3D shapes of proteins based on their amino acid combinations.
AlphaFold is impressive and has now predicted over 200 million proteins from their amino acid chains. Researchers hoped that building such a large database would allow scientists to develop treatments that target specific proteins linked to diseases like cancer or dementia. To develop such drugs, you may need to know the physical structure of the protein, where programs like AlphaFold can be used.
However, a study led by scientists at MIT in America shows how difficult the task is in practice. Essentially, the AI software is useful in one step of the process – structure prediction – but cannot help in other stages, e.g. B. in modeling the physical interaction of drugs and proteins.
“Breakthroughs like AlphaFold expand the possibilities for in silico (computer simulation) of drug discovery, but these developments need to be coupled with additional advances in other aspects of modeling that are part of drug discovery,” James Collins, lead author of the study released in Molecular Systems Biology and Bioengineering Professor at MIT, said in an opinion.
“Our study addresses both the current capabilities and current limitations of computational drug discovery platforms.”
Collins and his colleagues used AlphaFold to simulate interactions between bacterial proteins and antibacterial compounds, a task known as molecular docking. The aim was to use molecular docking to rank the candidate compounds according to how strongly they bind to the target protein. A molecule that binds strongly to a protein is more likely to be an effective drug; it could be more effective in preventing the protein from exerting a pathogenic function, such as tumor growth.
The team tested AlphaFold’s ability to model interactions between 296 essential proteins E. coli Bacteria with 218 antibacterial compounds, including antibiotics such as tetracyclines. AlphaFold was not very effective at accurately modeling molecular docking simulations.
“Using these standard molecular docking simulations, we got an auROC value of about 0.5, which basically says you’re doing no better than if you were randomly guessing,” Collins said.
Not the smartest AI on the block
According to Felix Wong, co-author of the article and a postdoctoral fellow at MIT, other machine learning models were more accurate than AlphaFold on some simulations.
“The machine learning models not only learn the shapes, but also chemical and physical properties of the known interactions, and then use that information to reevaluate the docking predictions,” he said. “We found that if you filtered interactions with these additional models, you would get a higher ratio of true positives to false positives.”
Derek Lowe, a longtime drug discovery chemist and science writer, narrates The registry He wasn’t surprised by the results since AlphaFold wasn’t really trained for molecular docking simulations. ‘Docking small molecules to a specific protein structure is really a different problem than determining that protein structure in the first place,’ he said.
Being able to model these types of chemical interactions is an unsolved problem. No algorithm is perfect. Even when scientists have a good model of the protein, its shape changes when it mysteriously interacts with a potential drug candidate.
“Virtual screening has never reached the ‘always works’ level – sometimes it provides useful information and sometimes it doesn’t, and you’re never sure in advance which of these regimes you’re working in. Add to that the way that different docking software gives you different answers, and for a given goal, one of them could give significantly more useful answers than another – but again, you don’t know in advance which one it will be,” Lowe said.
“Even with perfect protein structures, some of them will lend themselves better to a docking-and-scoring approach than others, and AlphaFold structures, while impressive, aren’t perfect either. But to me, that’s not the case as much about AlphaFold as it is about docking technology.”
AlphaFold may prove useful for other parts of the drug discovery pipeline, where comparing protein structures obtained by different methods with the model’s predictions is valuable.
“The biggest problems in drug discovery are those that contribute to our failure rate of around 85 percent in the clinic. And these are choosing the right targets and early warning of toxicity. Knowing protein structures doesn’t help either of them at all. ‘ Lowe added. ®
https://www.theregister.com/2022/09/08/ai_protein_alphafold/ Limitations of DeepMind’s AlphaFold detailed in MIT study • The Register