AlphaFold

AlphaFold is a program capable of determining the 3D shape of proteins from their amino-acid sequence. In 2018, AlphaFold greatly outperformed its competitors in a biennial protein-structure prediction challenge called CASP (Critical Assessment of Structure Prediction). In certain cases, AlphaFold’s predictions were indistinguishable from those determined using standard methods, such as X-ray crystallography and cryo-electron microscopy (cryo-EM).

Nuclear magnetic resonance, X-ray crystallography, and cryo-electron microscopy are some of the standard experimental techniques used to examine and determine the structure of proteins. These techniques depend on laborious and extensive trial and error processes that require the use of multi-million dollar specialized equipment.

AlphaFold combines deep learning with an algorithm that mimics cognitive attention. The process begins by connecting pieces in small clusters of amino acids and then searching for ways to join the clusters in a larger whole. The algorithm had been trained on all 170,000 or so protein structures available in the Protein Data Bank (PDB), powered by a computer network built around 128 machine learning processors.

AlphaFold overview video

Scientists believe that the ability to accurately predict protein structures from their amino-acid sequence would accelerate efforts to understand the building blocks of cells and enable faster and more advanced drug discovery. While AlphaFold may not obviate the necessity for standard protein structure determining techniques, it might open up new research paths. Andrei Lupas, an evolutionary biologist, has stated that AlphaFold's capabilities will have a significant impact on medicine, research, and bioengineering.

At CASP, nearly two-thirds of AlphaFold's predictions were comparable in quality to experimental structures, with a global distance test (GDT) score of approximately 90 on moderately difficult targets, versus the 75 average of competing programs. Scores above 90 on the zero to 100 scale are considered to be on par with experimental methods.

AlphaFold achieved a median GDT score of 92.4 across target proteins and a median of 87 in the most challenging cases, placing it 25 points above the next best predictions. In addition, AlphaFold managed to predict the structures of proteins that sit wedged in cell membranes, which play a significant role in many diseases but are difficult to solve with x-ray crystallography.

One of AlphaFold's limitations appeared to be the modeling of individual structures in protein complexes, or groups, whereby their shape is distorted by interactions with other proteins. Moreover, the program's predictions did not match experimental structures determined by a technique called nuclear magnetic resonance spectroscopy (NMR spectroscopy), which John Moult, a computational biologist, believes might be due to how the raw data is converted into a model.

It was an amalgam of fifty two small repeating segments that distort each others' positions as they assemble. John Jumper, the head of AlphaFold's development at DeepMind, has stated that the AlphaFold team will respond to these shortcomings by training the program to solve the structures it underperformed in during the competition, as well as those of complexes of proteins that co-operate to carry out key functions in the cell.

On August 2020, DeepMind published the results of AlphaFold's protein structure predictions for SARS-CoV-2 membrane protein, protein 3a, Nsp2, Nsp4, Nsp6, and Papain-like proteinase (C terminal domain). The results were released under an open license and without undergoing peer review, due to the urgency of the COVID-19 crisis. The predicted structures have not been experimentally verified.