# AlphaFold 2020

AlphaFold 2020, developed by DeepMind, is an AI system that achieved near-experimental accuracy in protein structure prediction at CASP14 in November 2020, scoring above 90 on GDT for two-thirds of targets, marking a breakthrough in computational biology.

AlphaFold 2020 is an artificial intelligence program developed by DeepMind, a subsidiary of Alphabet, that predicts protein structures from amino acid sequences using deep learning techniques. At the 14th Critical Assessment of Structure Prediction (CASP14) competition in November 2020, it achieved unprecedented accuracy, scoring above 90 on the global distance test (GDT) for approximately two-thirds of the proteins tested, where 100 represents a complete match to experimentally determined structures. The results were described as "astounding" and "transformational" by researchers, though some noted that accuracy remained insufficient for a third of predictions and that the underlying mechanism of protein folding remains unsolved.

## Background

Proteins consist of chains of amino acids that spontaneously fold into three-dimensional (3-D) structures, which are essential for understanding biological function. Experimental methods such as X-ray crystallography, cryo-electron microscopy, and nuclear magnetic resonance (NMR) are expensive and time-consuming; over 60 years, they have determined structures for about 170,000 proteins, while over 200 million known proteins exist across all life forms. Computational prediction methods, including homology modeling based on molecular evolution, achieved GDT scores of only about 40 out of 100 for the most difficult proteins by 2016, as assessed by CASP, which was launched in 1994 to challenge the scientific community. AlphaFold first competed in CASP in 2018 using an AI deep learning approach.

## Algorithm

DeepMind trained the program on over 170,000 protein structures from the Protein Data Bank, a public repository of sequences and structures. The system uses a form of attention network, a deep learning technique that identifies parts of a larger problem and assembles them into an overall solution. Training was conducted on between 100 and 200 GPUs.

### AlphaFold 1 (2018)

AlphaFold 1, which won CASP13 in December 2018, built on prior work that analyzed large databases of related protein sequences to find correlated changes at non-consecutive residues, suggesting physical proximity. It extended this by estimating probability distributions for inter-residue distances, effectively converting contact maps into distance maps, and used more advanced learning methods. The code was not made publicly available except for running on CASP sequences.

### AlphaFold 2 (2020)

The 2020 version differs significantly from its predecessor. AlphaFold 1 used separately trained modules combined with a physics-based energy potential; AlphaFold 2 replaced this with interconnected sub-networks forming a single, differentiable, end-to-end model based on pattern recognition, trained in an integrated manner. After neural network convergence, a final refinement step applies local physical constraints using energy minimization based on the AMBER force field, which only slightly adjusts the predicted structure. Key components include two modules, likely based on a [transformer](https://www.wikiprompt.org/wiki/transformer) design, that progressively refine information about relationships between amino acid residues and between positions and sequences in the input alignment. The inclusion of metagenomic data improved multiple sequence alignment quality, and training data came from the custom-built Big Fantastic Database of 65,983,866 protein families, covering 2,204,359,010 protein sequences from reference databases, metagenomes, and metatranscriptomes.

## Impact and Recognition

AlphaFold 2's CASP14 results were widely recognized. On 15 July 2021, the paper was published in Nature as an advance access publication, alongside open source software and a searchable database of species proteomes. As of November 2025, the paper had been cited nearly 43,000 times. Demis Hassabis and John Jumper shared one half of the 2024 Nobel Prize in Chemistry "for protein structure prediction," while David Baker received the other half "for computational protein design." They had previously won the Breakthrough Prize in Life Sciences and the Albert Lasker Award for Basic Medical Research in 2023. However, some researchers noted that accuracy was insufficient for a third of predictions and that the program did not reveal the underlying rules of protein folding, leaving the protein folding problem unsolved.

## Legacy

AlphaFold 3 was announced on 8 May 2024, capable of predicting structures of complexes formed by proteins with DNA, RNA, ligands, and ions, showing a minimum 50% improvement in accuracy for protein interactions with other molecules compared to existing methods. In 2026, Google announced that the AlphaFold team had been disbanded, with most members reassigned to work on [Gemini](https://www.wikiprompt.org/wiki/gemini) and others moving to Isomorphic Labs or leaving the company entirely. The 2020 breakthrough remains a landmark in [machine learning](https://www.wikiprompt.org/wiki/machine-learning) applications to biology, demonstrating the power of [deep learning](https://www.wikiprompt.org/wiki/deep-learning) in solving complex scientific problems.

---
Source: https://www.wikiprompt.org/wiki/alphafold-2020
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-13T03:50:52.3701+00:00