Google’s AlphaFold software, which uses artificial intelligence to forecast the structure and form of molecules inside the human body, received a much-needed update last week.
Because the structure and activity of any given molecule may be inferred from its form, biologists have spent a great deal of time studying the folding patterns of amino acid chains, which are the building blocks of proteins. This process can be sped up and streamlined by the AI tool, creating new opportunities for innovation, particularly in the fields of medication and vaccine research.
Building on its previous two incarnations, AlphaFold 3, the latest upgrade from Google DeepMind was published in the journal Nature last week. The program may be able to predict the three-dimensional structure of proteins with accuracy, according to its 2018 teaser; nevertheless, its 2020 upgrade, AlphaFold 2, brought about several notable advancements. Google revealed the 3D structures of almost all known proteins in the human body, coupled with an open-source version of AlphaFold in 2021. Two million predicted protein structures were shared the next year.
However, despite these advancements—which allowed scientists to better understand the eggs of ancient birds and map the human heart—AlphaFold 2 was only capable of simulating proteins.
According to Mohammed AlQuraishi, a systems biologist at Columbia University who is not connected to Google DeepMind, “the AlphaFold 2 system only knew about amino acids, so it was of very limited utility for biopharma,” as James O’Donnell of the MIT Technology Review reports.
The most recent iteration of the program can forecast not just the conformation of proteins but also that of DNA, RNA, and other compounds, including ligands. Most importantly, this improvement will enable researchers to more accurately investigate and forecast the geometric interactions between various molecules in the human body, including where a medicine may bind to a protein.
According to the MIT Technology Review, AlphaFold 3 provides researchers with a degree of confidence with each prediction it models, often ranging from 40 percent to 80 percent. High confidence portions of a structure are represented in blue, whereas less certain areas are represented in red. Its imprecision is a drawback in several situations—for instance, the system isn’t very accurate when modeling RNA-protein interactions.
The model’s propensity to generate incorrect information or “hallucinate” is another possible disadvantage. The developers of AlphaFold 3 took inspiration for their molecular library and function from other A.I. models that produce images and videos, like OpenAI’s DALL-E 2 and Sora. Although this enhanced AlphaFold 3’s ability to generate massive 3D pictures of molecular forms, it still has the potential to cause hallucinations. More training data, the researchers thinks, would help address this problem. They remark in the paper that hallucinated structures would normally be classified as low confidence.
In contrast to its predecessor, only a restricted version—the AlphaFold Server—will be made available to the general public, and the AlphaFold 3 code will not be published open-source.
Google DeepMind CEO Demis Hassabis tells WIRED’s Will Knight, “This is a big advance for us.” “Determining how a small molecule will bind to a drug, how strongly, and what else it might bind to is exactly what you need for drug discovery.”
