Meta's artificial intelligence was able to predict the shape of 600 million proteins

It took him only two weeks.

Scientists of Meta company, which owns Facebook and Instagram, used language artificial intelligence to predict the structure of 600 million proteins from viruses, bacteria and other microbes, Live science reports. And some of these proteins belong to organisms unknown to science.

The program, called ESMFold, used a model that was created to decode human speech. With its help, she was able to predict the “turns” that determine the three-dimensional structure of proteins. The results of AI work can be used to develop new drugs, describe previously unknown microbial functions and track evolutionary relationships between individual related species.

Read also: Scientists have found a protein that can turn an ant from a worker to a queen< /p>

Proteins are the “building blocks” of life, consisting of chains of amino acids – tiny molecular units that connect to each other in countless combinations to form the three-dimensional shape of the protein. Knowing the shape of a protein allows you to determine its functions. But there are many ways in which the same set of amino acids can be combined.

The gold standard for determining protein structure is X-ray crystallography, which allows you to see how high-energy light rays are refracted around proteins, but it is a painstaking method to obtain which can take months or years to produce results, and it works for all types of proteins. After decades of work, more than 100,000 protein structures have been deciphered using X-ray crystallography. To solve this problem, Meta specialists decided to develop their own program that predicts the shape of the protein.

To test the performance of their program, the researchers used a database of metagenomic DNA that was sequenced from soil, water, intestines and from human skin. By entering the data into the program, the scientists managed to predict the shape of 617 million proteins in two weeks.

Related video

This is 400 million more than the AlphaFold program from DeepMind was able to predict.


Based on materials: ZN.ua

Share This Post