Global experimental contest of protein folding calculations has been won by the machine learning newcomer – AlphaFold from DeepMind startup.

Prediction of protein folding always posed a serious bioinformatic problem. A sequence in the gene codes an order of amino acids in a protein molecule. However, those amino acids are folded to a three-dimensional structure which performs its functions on the basis of that shape. Information about protein structure is not present in DNA, and amino acid sequence and researchers are trying to predict it using sophisticated algorithms.

Competition relies on amino acid sequences with known but not publicly available structures. Participating 97 teams sent their predictions of structures and received the results on a conference in December.

Out of 43 amino acid sequences, AlphaFold predicted 25 most accurately in comparison to other competitors. Second most successful participant predicted three protein structures.

Although the result seems outstanding, Robert F. Service noted that predictions were mostly only slightly better than others. This is a consequence of machine learning methods use, which are based on previously known data about how amino acid sequences translate to folded protein structure.

In-depth analysis: