While some of the applications for artificial intelligence involve say, winning games of Texas hold’em or recreating pretty paintings, there are areas where the technology could have truly profound consequences. Among those is medical care, and a major breakthrough from Alphabet's DeepMind AI could be a gamechanger in this regard, with the system demonstrating an ability to predict the 3D structures of unique proteins, overcoming a problem that has plagued biologists for half a century.
By understanding the 3D shapes of different proteins, scientists can better understand what they do and how they cause diseases, which in turn paves the way for better drug discovery. Beyond that, as a central component to the chemical processes for all living things, more expedient mapping of 3D protein structures would benefit many fields of biological research, but this process has proven painstaking.
This is because while modern scientific tools such as X-ray crystallography and cryo-electron microscopy allow researchers to study these structures in amazing new detail, they all still hinge on a process of trial and error. Proteins are made up of one-dimensional chains of amino acids that then fold into the final 3D structure, and an ability to use computational methods to take the initial amino acid sequence and predict what the finished 3D product will look like would be a major step forward.
This is known as the “protein folding problem,” and researchers have been searching for solutions to it since the early 1970s. The trouble is that the number of 3D configurations an amino acid chain could fold into is massive, so much so that some experts have said it would take longer than the age of the known universe to determine them all via brute force calculation.
DeepMind’s AlphaFold is an AI system built to tackle this long-standing challenge. In 2018, the initial version of AlphaFold debuted at CASP (Critical Assessment of protein Structure Prediction), a biennial worldwide event for experimenting with state-of-the-art protein structuring technologies. AlphaFold achieved the highest accuracy of the participating technologies at CASP13 in 2018, but has now been developed further into what is being labeled a “stunning advance.”
The system was trained on publicly available data on around 170,000 protein structures and a large database of unknown protein structures ahead of its appearance at CASP14 this week. Technologies are graded from 0-100 for accuracy on what is known as the Global Distance test, which assesses what percentage of beads in the protein chain are within a threshold distance of the correct location. In results released today, AlphaFold scored 92.4 across all targets.
“AlphaFold’s astonishingly accurate models have allowed us to solve a protein structure we were stuck on for close to a decade, relaunching our effort to understand how signals are transmitted across cell membranes,” says Professor Andrei Lupas, Director of the Max Planck Institute for Developmental Biology and a CASP assessor, who also described the technology as a “gamechanger” in an article for Nature.
AlphaFold could help scientists identify malfunctioning proteins and the reasons they lead to certain diseases, opening up entirely new avenues of drug development that could fast-track medical treatments. It could also help develop enzymes to degrade plastic waste, or help with with future pandemics by predicting the protein structures of novel viruses. Beyond that, it could help unravel the hundreds of millions of proteins we currently don’t know the structures for.
“This computational work represents a stunning advance on the protein-folding problem, a 50-year-old grand challenge in biology,” says Professor Venki Ramakrishnan, Nobel laureate and president of the Royal Society. “It has occurred decades before many people in the field would have predicted. It will be exciting to see the many ways in which it will fundamentally change biological research.”
DeepMind says it is currently preparing a paper that will be submitted to a peer-review journal, though a book of abstracts from CASP14, including the company’s contribution, can be found online here. The video below also provides on overview of the breakthrough.
Source: DeepMind
