Google’s DeepMind introduces an AI system, AlphaFold, that can accurately predict how proteins fold into shapes.
Proteins comprise strings of amino acids that are fundamental building parts of the human body. These are encoded into DNA and fold into multiple elaborate shapes that define their function. It is difficult to figure out the protein structure from the genetic sequence. The DNA only contains information about the chain of amino acid residues and not their final form. These structures can assist in finding how proteins cause diseases and producing drugs to treat them. However, predicting the 3D protein form is expensive and time-consuming.
In 2018, DeepMind had won the Critical Assessment of Protein Structure Prediction (CASP13) competition for predicting the 3D structure of proteins. Although the predictions lacked the accuracy required to be biologically useful, DeepMind has revised the system this year. The revised system can now predict protein structures more precisely than before in a matter of days.
The AI system, called AlphaFold, identifies a folded protein structure as a “spatial graph” using a neural network system where residues are comprehended as nodes connected with edges. The team used around 170,000 protein structures from the Protein Data Bank to train the system. This system uses 100-200 GPUs of computing power to run for a few weeks.
This computational work shows a remarkable advance on the protein-folding problem. The recent CASP assessment shows that two-thirds of the system’s predictions match the lab experiment’s accuracy with an average margin of error equivalent to an atom width.
Although the latest version has achieved the SOTA, there is still a long way to employ it in practical applications. The team hopes that the study will accelerate drug discovery and help professionals to understand diseases better.
DeepMind also aims to make AlphaFold available at scale. It is likely to collaborate with partners to explore different edges, such as understanding how multiple proteins form complexes and interact with DNA, RNA, and small molecules. They believe that enhancing the scientific community’s understanding of protein folding can improve the diagnosis and treatment of diseases caused by misfolded proteins.