Artificial intelligence tool cracks code to imagine proteins in 3D

An artificial intelligence network solved a scientific problem that has stumped researchers for half a century, successfully predicting the way proteins fold into three-dimensional shapes, a process that has typically taken expensive and painstaking lab work that could go on for decades.

The way proteins, one of the building blocks of life, fold drives their functionality and behaviour. For instance, SARS-Cov-2 has a protein that folds as a spike. This shape, therefore, is relevant for biologists (including for its ability to find cures for illnesses). It isn’t easy to predict the shape of a protein, though, based on the way amino acids come together to form a protein. That’s because there are countless ways in which a protein can fold into a three-dimensional structure.

DeepMind, owned by Google, created a computer programme called AlphaFold, which predicted to surprising accuracy the 3D shapes of proteins after being fed their constituent parts – data depicting strings of amino acids.

“This computational work represents a stunning advance on the protein-folding problem, a 50-year-old grand challenge in biology. 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,” said professor Venki Ramakrishnan, Nobel laureate and president of the Royal Society, according to a blog post by DeepMind.

“It’s a breakthrough of the first order, certainly one of the most significant scientific results of my lifetime,” a report in Nature quoted Mohammed Al Quraishi, a computational biologist at Columbia University in New York City, as saying. “I think it’s fair to say this will be very disruptive to the protein-structure-prediction field. I suspect many will leave the field as the core problem has arguably been solved,” he added.

Quraishi was part of the Critical Assessment of Structure Prediction (CASP), a competition held every two years to accelerate research into the field, where AlphaFold reached the threshold for what is considered as having “solved” the problem.

DeepMind became a subsidiary of Google after a 2014 acquisition and is best known for its gamer AI, teaching itself to beat Atari video games and defeating world-renowned Go players like Lee Sedol. The company’s ambition has been to develop AI that can be applied to broader problems, and it’s so far created systems to make Google’s data centres more energy-efficient, identify eye disease from scans and generate human-sounding speech.

“These algorithms are now becoming strong enough and powerful enough to be applicable to scientific problems,” DeepMind chief executive officer Demis Hassabis said in a call with reporters, news agency Bloomberg reported. After four years of development “we have a system that’s accurate enough to actually have biological significance and relevance for biological researchers.”

The DeepMind blog post referred to comme nts by eminent scientists on the topic in the past to illustrate the significance of the breakthrough. “In his acceptance speech for the 1972 Nobel Prize in Chemistry, Christian Anfinsen famously postulated that, in theory, a protein’s amino acid sequence should fully determine its structure. This hypothesis sparked a five decade quest to be able to computationally predict a protein’s 3D structure based solely on its 1D amino acid sequence as a complementary alternative to these expensive and time consuming experimental methods,” it said.

“A major challenge, however, is that the number of ways a protein could theoretically fold before settling into its final 3D structure is astronomical. In 1969 Cyrus Levinthal noted that it would take longer than the age of the known universe to enumerate all possible configurations of a typical protein by brute force calculation – Levinthal estimated 10^300 possible conformations for a typical protein,” it added.

CASP scientists analysed the shape of amino acid sequences for about 100 proteins. Competitors were given the sequences, and charged with predicting their shape.

AlphaFold’s assessment lined up almost perfectly with the CASP analysis for two-thirds of the proteins, compared to about 10% from the other teams, and better than what DeepMind’s tool achieved two years ago

Hassabis said his inspiration for AlphaFold came from “citizen science” attempts to find unknown protein structures, like Foldit, which presented amateur volunteers with the problem in the form of a puzzle.

In its first two years, the human gamers proved to be surprisingly good at solving the riddles, and ended up discovering a structure that had baffled scientists and designing a new enzyme that was later confirmed in the lab.