Google’s DeepMind bests experts at predicting 3D protein shapes

Google’s AI, DeepMind, has just come top in a competition for predicting the 3D shapes of proteins in Cancun.

The AI’s latest project, AlphaFold, was able to best all 98 rivals at the international conference for Critical Assessment of Structure Prediction (CASP). AlphaFold’s goal was to tackle one of the hardest problems in biology: predicting protein structures.  

On the first day of the competition, DeepMind came a clear first out of a possible 98 different teams, predicting the most accurate structure for 25 out of 43 proteins. Second place was only able to predict three.

READ NEXT: How Google’s DeepMind is learning like a child

Proteins play a vital role in the functioning of all living systems, interwoven into every physiological process. To have a healthy body, you need a good balance of proteins. In cancerous patients, there’s an overproduction of proteins; in those with Parkinson’s disease, proteins aren’t developing properly.

The problem is that protein structures are humbling pieces of data, to say the least. The body makes anything between tens of thousands to billions of the things. Each is made up of hundreds of amino acids, which bend and shape differently depending on the bonds kind of bonds the acids make. In short, a single protein can be made up of structures numbering into a googol cubed, or 1 followed by 300 zeroes.

Fortunately, AlphaFold has come up with some notable developments. DeepMind taught AlphaFold to use neural networking to interpret huge datasets on protein structures and their respective amino acids. Neural networking is when AI thinks like a human brain, considering multiple pieces of data at once and building upon them to find a conclusion.

For the case of 3D protein structures, AlphaFold would calculate the distances in the protein as a result of the different amino acid pairings. It would then calculate the angles between each chemical bond to give it a 3D shape. Finally, AlphaFold would reiterate this initial structure to find the most energy-efficient arrangement, a natural arrangement for a healthy body.

Being able to map proteins accurately means scientists can develop either replicas of current proteins, or new ones, in order to tackle an issue. This means being able to tackle masses of health problems that arise from problematic proteins – the scope of this reaches as far as overcoming ageing.

READ NEXT: What is AI?

The DeepMind team isn’t stopping at humanity, though. Their aim also to utilise the findings to help counteract our species’ impact on the planet. They hope to take AlphaFold’s research and utilise it to design proteins for biodegradable enzymes. Such proteins would be able to break down things like plastic or oil.

AlphaFold’s success in making such accurate predictions marks a great step for the DeepMind program. Demis Hassabis, co-founder and CEO of DeepMind, describes the AI’s achievement as its “first major investment in terms of people and resources into a fundamental, very important, real-world scientific problem.”