22-Oct-2021
Anastassis Perrakis and Titia Sixma of Instruct-NL explore the role that the new AlphaFold AI system has in the world of structural biology.
The paper, published in EMBO Reports, outlines the problems that have been facing computerised protein prediction models since their inception nearly 40 years ago, and assesses the strengths of the AlphaFold system and its impact now on experimental methods.
Alphafold has been developed by AI company DeepMind. In 2020, it consistently predicted 3D structures very similar to the experimentally determined structures in the 14th round of the biennial CASP (Critical Assessment of protein Structure Prediction) experiment - a regular objective test of protein structure prediction methods. Achieving a median error of 0.96 Å compared to 2.83 Å for the next-best method, it was recognised as the best method for predicting 3D protein structures, significantly outperforming its contemporaries.
The opinion article summarises a number of immediately practical uses for Alphafold database, which was released in collaboration with EMBL-EBI, in the world of structural biology, namely, “To design better protein expression experiments; to solve experimental structures faster; to overcome tedious model building steps; and to interpret lower-resolution cryo-EM maps, where structure solution is limited by dynamic variability.”
The team from The NKI Protein Facility, Netherlands, urge caution for the next generation of structural biologists: “It is always good to keep in mind that AI relies, to some extent, on what has been seen before”. The enthusiasm for solving structures experimentally should not diminish, and the paper outlines a number of reasons in which practical methods are preferred, not least “The danger that people will be less inclined to prepare high-quality protein, which is actually useful for many more experiments than determining structures.”
Read the full paper here, or find out more about the Alphafold system and its potential here.
