Google‘s DeepMind: "Protein Folding'' Solved

Google‘s DeepMind: “Protein Folding” Solved

Google‘s DeepMind: “Protein Folding” Solved.

On November 30th, Google’s DeepMind company stated that their development of artificial intelligence software AlphaFold has solved the “protein folding problem” that has plagued the protein science community for 50 years.

The puzzle refers to predicting what shape a protein molecule will fold into, depending on the amino acid sequence that makes up the protein.

Because the shape and function of a protein have a strong correspondence, predicting the folding shape of a protein is very important for biological research.

Google‘s DeepMind: "Protein Folding'' Solved — Source: DeepMind

This problem has plagued computational biologists since the 1960s.

In 1969, American biologist Cyrus Levinthal pointed out that if brute force calculations are used to enumerate all possible configurations of a protein, at least 10³⁰⁰ possibilities must be exhausted for each protein molecule.

In the biannual critical assessment of protein structure prediction (CASP), scientists believe that AlphaFold has solved this problem. The training of AlphaFold lasted only a few weeks.

The data set contains about 170,000 protein structures and a large number of protein sequences, and uses 128 TPUv3 cores (computing power approximately equal to 100 to 200 GPUs).

After training, AlphaFold achieved an average of 92.4 GDT at the 14th CASP (it can be roughly regarded as the similarity between the predicted protein structure and the actual protein structure, with a full score of 100 GDT).

Google‘s DeepMind: “Protein Folding” Solved

(Source: internet/Google‘s DeepMind: Protein Folding Solved)