London-based DeepMind lab has developed a deep-learning program known as AlphaFold that bested roughly 100 efforts to predict protein structures in a contest known as Critical Assessment of Structure Prediction (CASP).
Such proteins are found within the cells of organisms ranging from bacteria to humans. Abnormalities in such proteins play a role in numerous diseases.
Predicting the 3D shape of such proteins has been a vexing challenge for decades, although researchers have made breakthroughs in doing so in recent years.
The structure of such proteins dictates how it functions.
While CASP participants were more accurate in the competition than in years’ past, AlphaFold was an outlier. For protein-folding puzzles deemed moderately difficult, the AlphaFold team scored approximately 90 on a 100-point scale for prediction accuracy. The nearest competitors scored roughly 75.
This was the first year that any team came close to accurately predicting protein shapes. The CASP contest was launched in 1994. Scientists have attempted to solve the “protein-folding problem” for some 50 years.
The improved ability to predict protein structures based on their amino-acid sequence will likely be a boon for drug discovery as it will improve scientists’ ability to comprehend the inner workings of cells.
The traditional manner of determining protein structures has relied on lab experiments. Often, figuring out a precise 3D protein shape can take months or years. And in some respects, however, AlphaFold’s predictions replicated those from laboratory techniques such as X-ray crystallography and cryo-transmission electron microscopy.
But the ability to use artificial intelligence to understand protein folding opens up new avenues for studying organisms and disease.
AlphaFold currently does have limitations. It could not accurately predict structures found using nuclear magnetic resonance imaging. It also struggled to predict individual structures in complexes of proteins.