DeepMind AI's protein folding prediction achieves unprecedented accuracy, opening doors to new disease treatments

The Alphabet subsidiary aced a long-running research competition with its deep-learning system, called AlphaFold.
By Dave Muoio
01:24 pm
Share

Alphabet subsidiary DeepMind has cracked a decades-old protein-folding challenge with an artificial intelligence system that could eventually help identify new treatments for diseases, among other nonmedical uses.

"We have been stuck on this one problem – how do proteins fold up? – for nearly 50 years," John Moult, cofounder and chair of the Critical Assessment of protein Structure Prediction (CASP) competition, said in a statement provided by DeepMind. "To see DeepMind produce a solution for this, having worked personally on this problem for so long, and after so many stops and starts wondering if we'd ever get there, is a very special moment."

Called AlphaFold, the deep-learning system has been declared the winner of the CASP, which launched in 1994 and is run every two years. The contest assigns participants a selection of protein structures that were recently determined using experimental methods, and reviews their ability to blindly predict the proteins' structures from their amino acid sequence.

Following an initial debut in the 2018 CASP competition, in which it led the competition but still fell well short of the goal, DeepMind trained its attention-based neural network system on a public data bank of roughly 170,000 known protein structures and others containing unknown protein structures over the course of a few weeks. Whereas the 2018 iteration of AlphaFold achieved Global Distance Test (GDT) score just short of 60, the most recent version logged a median 92.4 GDT across all targets – surpassing the competition's informal 90 GDT finish line.

Those scores do still leave room for improvement. According to the company, AlphaFold's average error when predicting protein structure is a distance of 1.6 angstroms, which roughly translates to 0.1 nanometers, or the width of an atom. And among a collection of the most difficult protein targets, the system's median score dipped down to 87 GDT.

DeepMind said that it's currently preparing a paper describing its system for publication in a peer-reviewed journal.

"Protein biology is fantastically complex and defies simple characterization," John Jumper, AlphaFold lead at DeepMind, said in a statement. "Our team's work demonstrates that machine learning techniques are finally able to meet the complexity of describing these incredible protein machines, and we are truly excited to see what new breakthroughs in both human health and fundamental biology it will bring."

WHY IT MATTERS

Completion of the CASP challenge is undoubtedly a feather in DeepMind's cap, and continued justification for Alphabet to support the extremely expensive AI business unit.

But more broadly, better understanding of the body via accurate protein-structure prediction could help researchers develop new cures for conditions in which misfolded proteins are believed to play a role, the AlphaFold team wrote in a Nature study published earlier this year.

Examples of such conditions include Alzheimer’s disease, Parkinson’s disease, cystic fibrosis and Huntington’s disease. In yesterday's announcement, the DeepMind team also described its work predicting protein structures within the COVID-19 virus, and noted that the technology could also play a role in environmental sustainability.

"The ultimate vision behind DeepMind has always been to build AI and then use it to help further our knowledge about the world around us by accelerating the pace of scientific discovery," Demis Hassabis, CEO and founder of DeepMind, said in a statement. "For us AlphaFold represents a first proof-point for that thesis. This advance is our first major breakthrough in a long-standing grand challenge in science, which we hope will have a big real-world impact on disease understanding and drug discovery."

THE LARGER TREND

DeepMind's previous demonstrations of its AI have ranged from mastering real-time strategy video games to other healthcare- and biology-focused projects, such as breast cancer detection, acute kidney injury prediction and eye disease diagnosis.

But beyond operating well in the red, the company has been at the heart of multiple patient data privacy concerns that have plagued its parent company, the most notable of which was a class action lawsuit involving a partnership with the University of Chicago. Despite these issues, DeepMind's team was rolled into Google Health just over a year ago.

More regional news

MobiHealthNews:

The latest news in digital health delivered daily to your inbox.

Thank you for subscribing!
Error! Something went wrong!