Link to paper

The full paper is available here.

You can also find the paper on PapersWithCode here.

Abstract

AlphaFold2 presented at CASP14 in Dec 2020, revolutionizing protein structure predictions
AlphaFold2 code released in summer 2021, can accurately predict structure of most proteins and protein-protein interactions
AlphaFold2 release sparked explosion of development in the field, improving AI-based methods for protein complexes, disordered regions, and protein design

Protein structure prediction is divided into two categories: template-based and de-novo
In the 1980s, methods were developed to predict structure by copying coordinates from an experimentally determined protein structure
Increase in sequence and structure data has enabled more proteins to be modelled accurately
De-novo modelling has blurred over the last two decades by using fragments
In the 1990s, co-evolutionary signals in a multiple sequence alignment were proposed to predict structure
1999 solution proposed to increase accuracy of contact predictions by separating direct and indirect contacts
Ten years ago, the idea of indirect correlation was rediscovered
Combining DCA and machine learning was a way to improve contact prediction
AlphaFold (version 1) introduced at CASP13, deeper architecture and predicted distance probabilities

DeepMind presented AlphaFold2 at CASP14 with impressive results
AlphaFold2 source code was released with an open-source license
AlphaFold2 consists of two main modules: EvoFormer block and structure block
AlphaFold2 input is a “raw” multiple sequence alignment
EvoFormer block uses row and column-wise attention mechanism and triangle updates
Structure module is locally translated and rotational equivariant
Structure module contains information about sidechains and uses Frame Aligned Point Error as a loss function

AlphaFold was released as open source in June 2021
Colabfold was an essential tool for using AlphaFold
AlphaFold was used to predict protein structures and protein-peptide interactions
AlphaFold2 was developed to predict the structure of single protein chains
AlphaFold2 was tricked into predicting the structure of dimers and higher multimers
AlphaFold2 was used to study and design protein-peptide interactions
DeepMind developed a version of AlphaFold aimed at predicting the structure of complexes
AlphaFold2 can accurately predict about half of all complexes up to 6 chains

AlphaFold code and description released, inspiring work to reproduce it
RoseTTAFold published at same time, but initially worse performance than AlphaFold
Later implementations of RoseTTAFold rival AlphaFold in accuracy
AlphaFold uses multiple sequence alignment to predict single protein structure
OmegaFold and ESMfold similar to AlphaFold but without using MSA
AlphaFold used for protein design and machine-learning tools for protein-ligand interactions
Most groups used AlphaFold in some way in CASP15
Two ways to improve over standard AlphaFold: use templates more efficiently or increase sampling

AlphaFold has been released in 2021 and has been used in hundreds of papers.
AlphaFold can produce models close to experimental quality for most proteins and protein complexes.