In December 2020, DeepMind, Google’s artificial intelligence arm, took the world of biology by surprise by solving a massive 50-year challenge with AlphaFold — an AI tool that can predict the structure of proteins. Last week, the London-based company published all the details of the tool and released its source code.
3D folded protein image.Source: DeepMind
Now, the company has announced that it has used its AI to predict the shapes of nearly every protein in the human body, in addition to the shapes of hundreds of thousands of other proteins found in 20 of science’s most widely studied organisms, including yeast, fly flies. fruits and mice. The discovery could allow biologists around the world to better understand diseases and develop new drugs.
So far, the database consists of 350,000 new protein structures. DeepMind says it will predict and release the structures of more than 100 million of them in the coming months — more or less every protein known to science.
“Protein folding is a problem I’ve been looking at for more than 20 years,” DeepMind co-founder and CEO Demis Hassabis said in an interview with MIT Technology Review. “It’s been a big project for us. I would say this is the biggest thing we’ve done so far. And it’s the most exciting in a way, because it’s got to have the biggest impact on the world outside of artificial intelligence,” added Hassabis.
Protein structure revealed
Proteins are made of long strands of amino acids, which twist into intricate knots. Knowing the knot shape of a protein can reveal what it does. Access to this twist is crucial to understanding how diseases work and developing new drugs — or even finding new ways to help the planet, for example by identifying organisms that can help fight pollution and climate change.
But figuring out the shape of a protein takes weeks or months in the lab. With AlphaFold, you can predict these shapes in a day or two. And the new database should make life even easier for scientists.
Repercussion in the scientific community
When AlphaFold was announced, scientists told the journal Nature that artificial intelligence “would change everything.” Now, the impacts can already be felt. David Baker of the University of Washington’s Institute of Protein Design has built his own tool to predict the structure of proteins — called RoseTTAFold, based on the AlphaFold approach.
For the past few months, his team has been working with biologists who were trying to figure out — unsuccessfully — the shape of the proteins they were studying. “There’s a lot of really cool biological research that’s been really accelerated,” he told MIT Technology Review. A public database, containing hundreds of thousands of forms of ready-made proteins, should be an even bigger accelerator.
“It looks surprisingly impressive,” said Tom Ellis, a biologist at Imperial College London who studies the yeast genome. He told the site that he was excited to try the database, but warned that most of the forms provided by the AI have not yet been verified in the lab.
The new version of AlphaFold features predictions with a confidence score — signaling how close each predicted shape is to the real one. So far, the AI has predicted shapes for 36 percent of human proteins with an accuracy that’s correct down to the level of individual atoms — good enough for drug development, Hassabis said.
Before, even after decades of work, only 17% of the human body’s proteins had their structures identified in the laboratory. If AlphaFold’s predictions are as accurate as DeepMind claims, the tool has more than doubled the number of proteins identified in just a few weeks.
For now, DeepMind is releasing the tools and forecasts for free and has so far not revealed whether it has any plans to make money from them in the future. But the possibility is not ruled out.