Smart, smarter, smartest: Folding proteins with AI

In Smart, smarter, smartestthe new Pocket Sciencepart, says science journalist Bennie Mols how artificial intelligence gives humans a turbo boost. In this sneak peek: how does AI help biochemists to unravel the structure of proteins?

Proteins are the biochemical workhorses of your body. They allow your intestines to absorb food, your bones grow, your eyes can see and damaged cells are repaired. Proteins are involved in every bodily function, some ten thousand different types in total.

After a protein is made in a cell, it automatically folds into a complex three-dimensional shape. The folding instructions come from the information stored in the amino acid sequence of the protein molecule. The shape the protein takes is critical to its function. Proteins that fold incorrectly can lead to disease.

In the early 1960s, biologists recognized that the folding of all these different kinds of proteins was one of biology’s great unsolved problems. The question they were so eager to answer was this: suppose you know the sequence of amino acids of a protein, what is the three-dimensional shape that the protein will take? For decades, there was little choice but to try to make a certain protein in a laboratory and try to determine its shape with complicated instruments. It often took a single PhD student years to find out the shape of one protein experimentally.

AlphaFold

Thanks to AI, the problem of protein folding has been largely solved in recent years. Building on the learning techniques developed for the go-playing computer AlphaGo, researchers at AI company DeepMind developed the AI ​​program AlphaFold. This program uses a deep learning neural network trained on thousands of known proteins and their three-dimensional structures to predict protein structures. In 2020, AlphaFold was the first to predict many protein structures at least as accurately as laboratory experiments could map them.

Two years later, in 2022, AlphaFold had already predicted the shape of more than two hundred million proteins from one million biological species. 35 percent of those predictions were very accurate and 45 percent were accurate enough for most practical applications. AlphaFold only takes ten to twenty seconds to make a single protein prediction, which saves biologists an incredible amount of time. “AlphaFold is a unique and momentous advancement in the life sciences that demonstrates the power of AI,” wrote Eric Topoldirector of the US Scripps Research Translational Institute, which focuses on personalizing healthcare through a combination of genetics and digital technologies.

New drugs and proteins

Because the precise shape of a protein largely determines its functions, AlphaFold can help researchers unravel the biochemical causes of diseases and develop new medicines. This is because when developing medicines, use is often made of what a protein looks like in three dimensions, but that shape must first be known.

Researchers from all over the world were quick to use AlphaFold, which was fortunately made available free of charge, to accelerate their scientific work. For example, a biochemist used AlphaFold to determine the structure of a key protein of a malaria parasite. He was then able to figure out where antibodies that could block transmission of the parasite were likely to bind. That insight can now be used to design improved malaria vaccines. Other biochemists used AlphaFold to identify enzymes that occur in nature and that could perhaps be adapted to digest and recycle plastics.

AlphaFold not only accelerates existing research, it also enables entirely new types of research. For example, biologists use the AI ​​instrument to discover new protein families and to study the evolution of proteins.