Alphabet Inc.’s DeepMind artificial intelligence division has racked up another scientific achievement.

The group this morning revealed that it has built an AI system capable of taking on what is considered to be one of the biggest challenges in biology today: simulating the shape of proteins.

The system, dubbed AlphaFold, took no less than two years to develop. DeepMind decided to announce the project after the software handily beat 97 other algorithms in the CASP simulation contest run by the U.S. National Institute of General Medical Sciences.

Proteins, which play a role in nearly every biological process, are chains of amino acids that twist and fold into various forms. The shape of a protein is one of the main factors that determine its behavior. A better understanding of these configurations could enable scientists to gain new insights into diseases believed to be caused by misfolded proteins and discover molecules that may be useful for drug development.

According to DeepMind, AlphaFold represents “significant progress” toward advancing this goal. The system can calculate what predicts a protein will be taken solely based on what amino acids it includes and with an accuracy that the Alphabet subsidiary described as far higher than existing methods. That’s no small feat, considering how manually simulating every possible configuration of an average-sized protein would take longer than the age of the universe.

AlphaFold breaks down the task into two parts. First, the system uses information about a protein’s constituent amino acids to generate a three-dimensional model of its default (that is to say, unfolded) structure. 

“The properties our networks predict are: (a) the distances between pairs of amino acids and (b) the angles between chemical bonds that connect those amino acids,” DeepMind explained. “We trained a neural network to predict a separate distribution of distances between every pair of residues in a protein. These probabilities were then combined into a score that estimates how accurate a proposed protein structure is.”

Once the 3-D model is ready, AlphaFold figures out the shape the protein most likely to take. It does so by taking advantage of the fact that all proteins are predisposed to fold into the most energy-efficient form available, which helps narrow down the possibilities. The AI tries out different variations of the 3-D model to generate more efficient versions until it finds the optimal form.  

“Our first method built on techniques commonly used in structural biology, and repeatedly replaced pieces of a protein structure with new protein fragments,” DeepMind detailed. “The second method optimized scores through gradient descent — a mathematical technique commonly used in machine learning for making small, incremental improvements — which resulted in highly accurate structures.”

AlphaFold initially took two weeks to generate a prediction, according to The Guardian, but can now complete the task in a couple of hours. It beat the 97 other algorithms tested in the CASP simulation contest by accurately predicting 25 out of the 43 provided proteins. The runner-up correctly predicted just three.

Photo: DeepMind