DeepMind’s Alpha Fold AI Predicts Protein Structure

A network of artificial intelligence (AI) created by a Google AI subsidiary One of biology’s most difficult problems, figuring out a protein’s three-dimensional form from its amino acid composition, has been substantially improved by DeepMind.

In the biannual CASP or Critical Assessment of Structure Prediction, protein-structure prediction challenge, DeepMind’s program, AlphaFold, outperformed almost 100 other teams. At the beginning of the conference, which was held virtually this year and takes stock of the exercise, the results were revealed on November 30.

John Moult, a computational biologist at the University of Maryland in College Park who co-founded CASP in 1994 to advance computer techniques for precisely predicting protein structures, declares that “this is a significant deal.” “The issue has, in some ways, been resolved.”

The fields of life sciences and medicine would greatly benefit from the capacity to precisely anticipate protein structures given their amino-acid sequence. It would greatly speed up efforts to comprehend the cellular building components and facilitate more rapid and sophisticated drug development.

At the most recent CASP, which was held in 2018, London-based DeepMind placed first overall with AlphaFold. But this year, the team’s deep-learning network performed so astoundingly well that, according to scientists, it may have signaled the beginning of a revolution in biology.

The Max Planck Institute for Developmental Biology’s Andrei Lupas, an evolutionary scientist who evaluated the effectiveness of various CASP teams, calls it “a game changer.” He already used AlphaFold to solve the structure of a protein that had plagued his lab for a decade, and he anticipates that it will change the problems he works on and the questions he asks. This will alter medical practice. It’ll modify research. In turn, biotechnology will alter. Everything will change, Lupas continues.

AlphaFold’s structure predictions in some cases were identical to those made by ‘gold standard’ experimental techniques like X-ray crystallography and, more recently, cryo-electron microscopy (cryo-EM). According to scientists, AlphaFold won’t necessarily replace these time-consuming and expensive techniques just yet, but it will enable new approaches to studying living things.

The Structure problem

Proteins are the building blocks of life and are in charge of most cellular functions. A fundamental principle of molecular biology is that “structure is a function”; a protein’s 3D shape determines how and what it does. Proteins frequently assume their structure on their own, relying simply on the principles of physics.

For many years, the primary method for obtaining high-quality protein structures was laboratory experiments. Beginning in the 1950s, scientists used a method that involved firing X-ray beams into crystallized proteins and translating the diffracted light into the atomic coordinates of the protein. The majority of protein structures have been created using X-ray crystallography. However, over the last ten years, cryo-EM has replaced other tools as the go-to option in many structural biology labs.

The multiple folds and twists of a protein’s final shape are mapped out by a string of distinct amino acids, which are the building blocks of proteins. According to academics, early attempts to predict protein structures using computers in the 1980s and 1990s did not go well. When other scientists used the procedures, lofty claims made for them in published studies tended to fall apart.

Moult founded CASP to provide these initiatives with more structure. Teams are required to forecast the structures of proteins that have been solved experimentally but whose structures have not yet been revealed. By putting an end to exaggerated claims, Moult attributes the experiment—which he does not refer to as a competition—to significantly better the field. You’re actually learning what works, what is effective, and what you should ignore, he claims.

Many scientists in the field were taken aback by DeepMind’s 2018 performance at CASP13 because it had previously been the domain of small academic teams. But, as Jinbo Xu, a computational biologist at the University of Chicago in Illinois, points out, it took a general strategy that was similar to that of other teams using AI.

The first version of AlphaFold used genetic and structural data to apply the AI technique known as deep learning to estimate the separation between pairs of amino acids in a protein. According to John Jumper at DeepMind, who is in charge of the project, AlphaFold uses this data in a subsequent, non-AI-based step to create a “consensus” model of how the protein should appear.

The team made an effort to further that strategy but eventually ran across resistance. According to Jumper, the team shifted course and created an AI network that included more details about the geometrical and physical limitations that affect how proteins fold. Additionally, they gave it a more challenging task: the network must predict the final structure of a specific protein sequence rather than only associations between amino acids. It’s a significantly more complicated mechanism, according to Jumper.

Startling accuracy

CASP lasts for a number of months. Teams have many weeks to submit their structure predictions after receiving periodical releases of the target proteins or domains, which number roughly 100 in total. The predictions are then evaluated by a group of impartial scientists using metrics that indicate how similar a predicted protein is to the structure discovered through experimentation. The predictors are anonymous to the assessors.

The forecasts from AlphaFold arrived under the designation “group 427,” but Lupas notes that they stood out due to the astonishing precision of many of their entries. “I assumed it to be AlphaFold. The majority had, he claims.

While some predictions were more accurate than others, over two-thirds were on par with experimental structures in terms of quality. According to Moult, there were certain instances where it was unclear if a mismatch between AlphaFold’s predictions and the experimental outcome was due to a prediction error or an artifact of the experiment.

Although Moult speculates that this may be due to how the raw data is transformed into a model, AlphaFold’s predictions have a poor agreement with experimental structures discovered using the method known as nuclear magnetic resonance spectroscopy. The network also has trouble simulating individual protein complexes or groups, where the geometries of the proteins are altered by interactions with other proteins.

Teams predicted structures more accurately overall this year compared to the previous CASP, but a large portion of the advancement may be ascribed to AlphaFold, according to Moult. According to Moult, other teams’ top performances on fairly challenging protein targets typically scored 75 out of a possible 100 on a prediction accuracy scale, while AlphaFold achieved about 90.

According to Moult, over half of the teams referenced “deep learning” in the abstract that summarised their methodology, indicating that AI is having a significant impact across the board. Although Microsoft and the Chinese technology company Tencent also entered CASP14, academic teams made up the majority of the participants.

When the DeepMind team presents its method on December 1st, computational biologist Mohammed AlQuraishi of Columbia University in New York City, a participant in the CASP competition, is eager to delve into the specifics of AlphaFold’s performance at the competition and learn more about the system’s operation. It’s possible, but improbable, according to him, that the performance was influenced by an easier-than-usual crop of protein targets. AlQuraishi has a good feeling that AlphaFold will have a transformative effect.

“I believe it is safe to state that this will significantly upset the field of protein structure prediction. Since the fundamental issue has arguably been resolved, I predict that many will quit the field, he says. It’s a major development, unquestionably among the most important scientific discoveries of my lifetime.

Faster structures

The structure of a bacterial protein that Lupas’s team had been attempting to decipher for years was discovered thanks to an AlphaFold prediction. Although Lupas’s team has previously gathered raw X-ray diffraction data, it is still necessary to know the protein’s shape in order to convert these Rorschach-like patterns into a structure. Other prediction tools and methods for obtaining this information had been ineffective. After a decade of trying everything, “the model from group 427 provided us our framework in half an hour,” claims Lupas.

AlphaFold will be valuable such that other scientists can utilize it, according to DeepMind co-founder and CEO Demis Hassabis. (It had already disclosed enough information about the original AlphaFold method for other researchers to duplicate the procedure.) The projected structure, which includes assessments of the dependability of various protein sections, can take AlphaFold days to produce. Hassabis continues, “We’re just beginning to understand what biologists would desire,” and suggests that protein design and drug discovery are two possible uses.

The business published predictions of the structures of a few SARS-CoV-2 proteins that hadn’t yet been determined experimentally in the early 2020s. According to molecular neurobiologist Stephen Brohawn of the University of California, Berkeley, whose team published the structure in June, DeepMind’s predictions for a protein called Orf3a turned out to be remarkably close to one later confirmed by cryo-EM. He continues, “What they’ve accomplished is really outstanding.

Real-world impact

Labs that use experimental techniques to solve protein structures, such as Brohawn’s, are unlikely to be shut down by AlphaFold. However, it can imply that all that is required to obtain a sound structure is experimental data that is of lower quality and easier to gather. Some applications, like the investigation of proteins’ evolutionary history, stand to benefit from the tsunami of genetic data that can now be reliably converted into structures. A new generation of molecular biologists will be able to pose more complex queries thanks to this, claims Lupas. There will be less pipetting and more thinking involved.

Janet Thornton, a structural biologist and former CASP assessor at the European Molecular Biology Laboratory-European Bioinformatics Institute in Hinxton, UK, says, “I was beginning to fear that this is a challenge that would not get solved in my lifetime.” She thinks the method will shed light on the purpose of the vast majority of unidentified proteins in the human genome and explain why different people have different disease-causing gene variants.

The success of AlphaFold also represents a turning point for DeepMind. The company is most known for using AI to master games like Go, but its long-term objective is to create systems capable of achieving all-encompassing intellect comparable to that of a human. One of the most crucial uses of its AI, according to Hassabis, is tackling large-scale scientific problems like predicting protein structures. In terms of the impact on the real world, I do believe it to be the most significant thing we have done.

This is all about DeepMind’s AlphaFold AI Predicts Protein Structure, for more informative content visit

About Mark B

Hey Folks! This is Mark and i work as a Content Writer for Techballad. I am flexible to work on different niches. I've got s 3 years experience of Content writing and i aspire to make my future in the same.
View all posts by Mark B →