In the new work, DeepMind co-founder David Silver, together with his team, trained a neural network to determine the shape of protein molecules.

In the new work, the authors have created a neural network that determines what form a particular protein molecule will take by the sequence of amino acids that make up it. This will help in the creation of drugs.

Now the shape of a protein is calculated using a particle accelerator, which gives three-dimensional photographs of protein molecules, or using a supercomputer, which calculates their structure in accordance with the laws of chemistry and quantum physics.

The authors have created an evoformer algorithm that tries to determine the structure of individual segments of protein molecules, representing them in the form of a three-dimensional tree of graphs – a mathematical abstraction that consists of a set of objects connected in pairs with each other. Evoformer connects them to each other, relying on already known examples, and gradually changes the structure of connections and the location of nodes, approaching the optimum.

Then they combined such algorithms and created the AlphaFold2 neural network.

Last year, we already presented the first version of our system, AlphaFold, which was able to predict the structure of proteins with near atomic precision in the CASP13 competition. Now we have created a new version of it, which is noticeably superior to all competitors in terms of speed and accuracy. Moreover, its source code is completely open.

Demis Hassabis, CEO of Deepmind

As a result, the neural network reconstructs the three-dimensional shape with atomic precision in about 10 minutes with an error of 0.096 nanometers for each atom inside the protein.