Sunday, January 16, 2022

Method of the Year: Google AlphaFold protein structure prediction

Google collects accolades for its state of the art artificial intelligence research to the benefit of humanity!

"... A pleasant frisson may have set in more recently as you browsed the new and rapidly growing AlphaFold Protein Structure Database or perused papers about a method called AlphaFold and its application to the entire human proteome, or when you dug into the code that drives this inference engine, with its neural network architecture that yields the 3D structure of proteins from a given amino acid sequence. The team behind AlphaFold is DeepMind Technologies, launched as an AI startup in 2010 by Demis Hassabis, Shane Legg and Mustafa Suleyman and now part of Alphabet after being acquired by Google in 2014. DeepMind has presented AlphaFold14 and AlphaFold2 and, more recently, AlphaFold-Multimer5 for predicting the structures of known protein complexes. ...
At CASP14 [Critical Assessment of Protein Structure Prediction] in 2020, AlphaFold2 blew away its competitors. The difference between the DeepMind team results and those of the group in second place “was a bit of a shock,” says University College London researcher David Jones. “I’m still processing that a bit, really.” Only some months later, when DeepMind gave a glimpse of its method and shared the code, were scientists able to begin looking under the hood. No new information was used to transition AlphaFold1 to AlphaFold2; there was no “clever trick,” says Jones. The team used what academics had been doing for years but applied it in a more principled way, he says. ...
DeepMind, in collaboration with EBI [European Bioinformatics Institute], is now filling the AlphaFold Protein Structure Database with hundreds of thousands of computationally generated human protein structures and those from many other organisms, including the ‘classic’ research organisms maize, yeast, rat, mouse, fruit fly and zebrafish.
Every day, the PDB sees around 2.5 million downloads of protein coordinates ...
Compared to other software developed in the academic community ... AlphaFold’s advances include more accurate placement of side chains in the protein models and an improved approach to integrating machine learning with homology modeling, which looks at protein structure in the context of evolutionarily related proteins. The software uses homology modeling at an “ultrafine” level ..."

Method of the Year: protein structure prediction | Nature Methods

In the 14th Critical Assessment of Protein Structure Prediction (CASP14), the performance of AlphaFold2 (first column) was far better than that any of the other participants.


No comments: