Protein-structure-predicting neural network goes some distance

Protein structure prediction has benefitted from the analysis of co-variation among homologous sequences, to infer those amino acid residues that are in contact. However, the ability to precisely predict these has been limited to large protein families. Now Andrew Senior and colleagues present AlphaFold, a neural network that is able to predict actual distances between such contacting residues, thus defining a potential of mean force, which can be optimized by a simple gradient descent algorithm, to determine structures without the need for complex sampling procedures. During the latest blind assessment of state-of-the-art algorithms, the CASP-13 contest in December 2018, AlphaFold performed much better than the other competitors, despite being conceptually simpler.