- DeepMind has developed a piece of AI software called "AlphaFold" that can accurately predict the structure that proteins will fold into in a matter of days.
- Predicting the shape that a protein will fold into is important because it determines their function and nearly all diseases, including cancer and dementia, are related to how proteins function.
- "DeepMind has jumped ahead," said Professor John Moult, who is the chair of a group called CASP (Critical Assessment for Structure Prediction).
LONDON — Alphabet-owned DeepMind has developed a piece of artificial intelligence software that can accurately predict the structure that proteins will fold into in a matter of days, solving a 50-year-old "grand challenge" that could pave the way for better understanding of diseases and drug discovery.
Every living cell has thousands of different proteins inside that keep it alive and well. Predicting the shape that a protein will fold into is important because it determines their function and nearly all diseases, including cancer and dementia, are related to how proteins function.
"Proteins are the most beautiful, gorgeous structures and the ability to predict exactly how they fold up is really very, very challenging and has occupied many people over many years," Professor Dame Janet Thornton from the European Bioinformatics Institute told journalists on a call.
British research lab DeepMind's "AlphaFold" AI system was entered into a competition organized by a group called CASP (Critical Assessment for Structure Prediction). It's a community experiment organization with the mission of accelerating solutions to one problem: how to compute the 3D structure of protein molecules.
CASP, which has been monitoring progress in the field for 25 years, compares competition submissions with an "experimental gold standard." On Monday, it said DeepMind's AlphaFold system has achieved unparalleled levels of accuracy in protein structure prediction.
"DeepMind has jumped ahead," said Professor John Moult, who is the chair of CASP, on a press call ahead of the announcement. "A 50-year-old grand challenge in computer science has been to a large degree solved."
Moult added that there are "major impacts a little bit down the line for drug design," and in the newly-emerging field of protein design.
With around 1,000 staff and next to no revenue, DeepMind has become an expensive company for Alphabet (Google's parent) to support. However, it has emerged as one of the leaders in the global AI race along with the likes of Facebook AI Research, Microsoft, and OpenAI.
The breakthrough was welcomed by Google Chief Executive Sundar Pichai on Twitter.
DeepMind Co-founder and Chief Executive Demis Hassabis said on the call: "The ultimate vision behind DeepMind has always been to build general AI, and then use it to help us better understand the world around us by greatly accelerating the pace of scientific discovery."
The company, which Google bought for $600 million in 2014, is best-known for creating AI systems that can play games like Space Invaders and the ancient Chinese board game Go. However, it has always said it wants to have more of a scientific impact.
"Games are great proving ground to efficiently develop and test general algorithms that we one day hoped we'd transfer to real world domains like scientific problems," said Hassabis. "We feel AlphaFold is a first proof point for this thesis. These algorithms are now becoming mature enough and powerful enough to be applicable to really challenging scientific problems."
DeepMind also entered a CASP protein folding competition in 2018. While its results at the time were impressive, John Jumper, AlphaFold lead at DeepMind, said the team knew it was some way from producing something with "really strong biological relevance or being competitive with experiment."
This year's competition wasn't plain sailing, however, and Jumper said DeepMind went for three months without making any progress. "We'd sit there and worry have we exhausted the data?" he said.
Even as the competition deadline approached, Jumper and his team were still worried that they may have made mistakes. "There could always be an error that creeps into machine learning systems," he said.
But their efforts seem to have paid off. "We really think that we've built a system that provides correct and actionable information for experimental biologists," he said. "The reason you have a structure is to understand something about the natural world and then ask even more questions. We think we've built a system that will really help people do that."