LAST UPDATED: Dec. 2, 2020, 5:06 p.m.
New Delhi (Anish Yande): DeepMind, Google’s artificial intelligence unit, has made a giant leap in resolving one of the most persistent problems in Biology. DeepMind, through its AlphaFold program, has determined a 3D shape of a protein from its amino-acid sequence. The first variant of AlphaFold has used deep learning to predict the distance between amino acid pairs in a protein
A challenge named the Critical Assessment of Structure Prediction is held every two years for research on protein structure prediction. This challenge commenced in 1994 and is held after every two years. The conference for the challenge was held on 30 November, and the results were announced at the start of the conference.
In biology, analysis of the protein structure could reveal clues about how the protein reacts with different molecules. The different folds in a protein interact differently with other molecules. The analysis of the structure leads to a better understanding of how diseases such as Covid-19 invade cells. The analysis through AI could also be used for improving agricultural produce and deconstructing pollutants.
For the challenge, competitors were given protein sequences and challenged with predicting the shapes.
How AlphaFold Achieved Results:
The AlphaFold then uses this information to create a ‘consensus’ model of the protein structure.
The DeepMind team changed its approach after encountering an obstacle. They started to build a network that predicts the final structure of a target protein sequence. In the results, other teams usually scored 75out of 100 for accuracy of predictions. AlphaFold scored 90 for accuracy of predictions.
Scientists from Critical Assessment of Structure Prediction sent out a set of 100 proteins for the shape of amino acid sequences. The results from AlphaFold matched up with the protein sequences from CASP, almost perfectly. The results were better than the DeepMind results from two years ago.
DeepMind has won the challenge earlier in 2018, after accurately predicting 25 out of 43 protein sequences given by CASP.
Image Credit: Google