Andrew W. Senior, Richard Evans, John Jumper, James Kirkpatrick, Laurent Sifre, Tim Green, Chongli Qin, Augustin Žídek, Alexander W. R. Nelson, Alex Bridgland, Hugo Penedones, Stig Petersen, Karen Simonyan, Steve Crossan, Pushmeet Kohli, David T. Jones, David Silver, Koray Kavukcuoglu & Demis Hassabis


This is a paper from DeepMind to delineate using deep learning to better predict three-dimensional protein structure from its amino acid sequence. The authors are able to use several neural networks to improve the accuracy of predictions of the shape of the protein. The first neural network is a convolutional neural network that predicted distance pairs and angles; the second is also a convolutional neural network and it estimated the accuracy of the protein structure; and the last is a variational auto encoder that generated the protein structure. This protein structure prediction system, named AlphaFold, uses simple gradient descent algorithm with a more conventional (and relatively inefficient) fragment assembly/simulated annealing algorithm in order to generate structures without complex sampling processes. AlphaFold performed well and was ranked first (of close to 100 teams) during the Critical Assessment of Protein Structure Prediction (CASP13) biennial event with high accuracy structures in 25 of 43 free modeling domains. This advance in protein-structure prediction has large implications for drug discovery necessary for medications such as anti-viral agents (particularly relevant in light of the present pandemic). It is possible that with protein structure prediction augmented by deep learning (and reinforcement learning in the future) at a very accurate level, drug discovery can be much more precise and expedient as well as more efficacious. DeepMind also deserves much recognition for rendering this project open for collaboration and therefore democratizing computational molecular biology.