DeepMind publishes early AlphaFold protein structure work

DeepMind published an account of the AlphaFold system that had placed first among competing methods at CASP13, the biennial Critical Assessment of protein Structure Prediction, in December 2018. The post laid out the approach in more detail than the original competition entry: a deep neural network trained on genomic and structural data predicted the distances between pairs of amino acids and the angles between their chemical bonds, and these predictions were then used either to build up a structure by replacing fragments or, in a second approach, to optimise an entire candidate protein chain directly by gradient descent.

The write-up credited a team spanning structural biology, physics and machine learning, including John Jumper, Andrew Senior and Richard Evans, and DeepMind made the CASP13 code public on GitHub so the result could be checked and built on.

AlphaFold’s CASP13 result was a meaningful jump over the field’s prior methods but still fell well short of experimental accuracy on hard targets, and the system attracted comparatively little attention outside structural biology at the time. That changed the following year: a redesigned version, AlphaFold2, produced predictions at CASP14 that judges compared to experimental accuracy, turning this earlier post into the record of where the project stood before the jump that made it famous.