Google’s DeepMind predicts 3D shapes of proteins
AI program’s understanding of proteins could usher in new era of medical progress
Having laid waste to the Atari classics and reached superhuman performance in chess and the Chinese board game, Go, Google’s DeepMind outfit has turned its artificial intelligence on one of the toughest problems in science.
The result, perhaps, was predictable. At an international conference in Cancun on Sunday, organisers announced that DeepMind’s latest AI program, AlphaFold, had beaten all-comers at a particularly fiendish task: predicting the 3D shapes of proteins, the fundamental molecules of life.
The arcane nature of “protein folding”, a mind-boggling form of molecular origami, is rarely discussed outside scientific circles, but it is a problem of profound importance. The machinery of biology is built from proteins and it a protein’s shape defines its function. Understand how proteins fold up and researchers could usher in a new era of scientific and medical progress.
“For us, this is a really key moment,” said Demis Hassabis, co-founder and CEO of DeepMind. “This is a lighthouse project, our first major investment in terms of people and resources into a fundamental, very important, real-world scientific problem.”
DeepMind set its sights on protein folding after its AlphaGo program famously beat Lee Sedol, a champion Go player, in 2016. While games have proved to be a good testing ground for the group’s AI programs, high scores are not their ultimate goal. “It’s never been about cracking Go or Atari, it’s about developing algorithms for problems exactly like protein folding,” Hassabis said.
The human body can make vast numbers of different proteins, with estimates ranging from tens of thousands to billions. Each one is a chain of amino acids, of which there are 20 different types. A protein can twist and bend between each amino acid, so that a protein with hundreds of amino acids has the potential to take on a staggering number of different structures: around a googol cubed, or 1 followed by 300 zeroes.
The 3D form a protein adopts depends on the number and types of amino acids it contains. The shape also determines its role in the body. Heart cells, for example, are dotted with proteins folded in such a way that any adrenaline in the bloodstream sticks to them and ramps up the heart rate. Meanwhile, antibodies in the immune system are proteins that fold into specific shapes which latch onto invading bugs. Nearly every function in the body, from tensing muscles and sensing light to turning food into energy, can be traced back to the shape and movement of proteins.
Normally, proteins take on whatever shape is most energy efficient, but they can become tangled and misfolded, leading to disorders such as diabetes, Parkinson’s and Alzheimer’s disease. If scientists can learn to predict a protein’s shape from its chemical makeup, they can work out what it does, how it might misfold and cause harm, and design new ones to fight diseases or perform other duties, like breaking down plastic pollution in the environment.
DeepMind entered AlphaFold into the Critical Assessment of Structure Prediction (CASP) competition, a biannual protein-folding olympics that attracts research groups from around the world. The aim of the competition is to predict the structures of proteins from lists of their amino acids which are sent to teams every few days over several months. The structures of these proteins have recently been cracked by laborious and costly traditional methods, but not made public. The team that submits the most accurate predictions wins.
On its first foray into the competition, AlphaFold topped a table of 98 entrants, predicting the most accurate structure for 25 out of 43 proteins, compared with three out of 43 for the second placed team in the same category.
To build AlphaFold, DeepMind trained a neural network on thousands of known proteins until it could predict 3D structures from amino acids alone. Given a new protein to work on, AlphaFold uses the neural network to predict the distances between pairs of amino acids, and the angles between the chemical bonds that connect them. In a second step, AlphaFold tweaks the draft structure to find the most energy-efficient arrangement. The program took a fortnight to predict its first protein structures, but now rattles them out in a couple of hours.
Liam McGuffin, a researcher at Reading University, led the highest-scoring UK academic group in the competition. “DeepMind appear to have pushed the bar higher this year and I’m intrigued to find out more about their methods,” he said. “We are not as well resourced, but we can still be very competitive.”
“The ability to predict the shape that any protein will fold in to is a big deal. It has major implications for solving many 21st-century problems, impacting on health, ecology, the environment and basically fixing anything that involves living systems.
“Many groups, including us, have been using machine learning-based methods for several years and improvements in deep learning and AI appear to be having an increasingly important impact. I’m optimistic that as a field we will really nail the problem in the 2020s,” McGuffin said.
Hassabis agrees there is far more to do. “We’ve not solved the protein folding problem, this is just a first step,” he said. “It’s a hugely challenging problem, but we have a good system and we have a tonne of ideas we haven’t implemented yet.”