How Deepmind’s AI decodes millions of proteins

Why do researchers study proteins?

Dieticians typically advise us to eat meat, eggs and fruits for a high-protein diet to help us stay healthy. But virtually every body part or tissue — be it our muscles, bones, skin or hair — comprises thousands of different proteins (the human body is estimated to have 20,000 to over 100,000 unique types of proteins within a cell) with each having a specific function. These proteins are made up of long chains of 20-22 different types of amino acids linked by peptide (shorter chain of amino acids) bonds. Their order determines how the protein chain will fold upon itself into a 3D structure.

What’s the role of AlphaFold in this?

Folding lets proteins adopt a functional shape or conformation. If researchers can predict how proteins fold, they can better learn how cells function and how mis-folded proteins can cause diseases. AlphaFold uses AI to predict a protein’s 3D structure from its amino acid chain. It released AlphaFold2 and the AlphaFold Protein Structure Database which it likens to a ‘Google Search’ for protein structures) in 2021. On 28 July, 2022, Deepmind and EMBL’s European Bioinformatics Institute released predicted structures for over 200 million proteins, covering almost every catalogued protein.

Where is AlphaFold being used?

Deepmind says AlphaFold has been used by over half a million researchers for work on real-world problems. The Centre for Enzyme Innovation at the University of Portsmouth is using it to develop faster enzymes to recycle single-use plastics, while the University of California San Francisco has used it to better understand the biology of the SARS-CoV-2 virus.

How is AI speeding up the process?

Proteins can fold in seconds or even milliseconds, but it would take longer than the age of the known universe (about 13.8 billion years) to find all possible configurations of a typical protein using brute force. Systems like AlphaFold and RoseTTAFold leverage advances in AI to transform how drugs are discovered and developed. AlphaFold can do “in seconds”, “what used to take many months or years”, says Eric Topol, founder and director of the Scripps Research Translational Institute.

But are the predictions always reliable?

AlphaFold can predict the structure of a single protein chain with high accuracy, but can’t do the same for multi-chain protein complexes for which it’s training another model called Multimimer. It also cannot predict the effect of disease-causing mutations. Also, when proteins take on multiple conformations, AlphaFold usually only produces one of them. Still, Deepmind believes “AI might turn out to be just the right technique to cope with the dynamic complexity of biology” even as it works to reduce its limitations.

First article