AlphaFold can predict the shapes of proteins up to an atom-wide scale. This revolutionary method will help scientists produce new drugs and better analyze the causes of diseases.
Now the level reached by artificial intelligence today is becoming more and more dazzling to our eyes. DeepMind is an important place in this regard. He’s a great artificial intelligence who can play many different games with super-human features. Go has successfully learned and finished all kinds of StarCraft and arcade games.
But Demis Hassabis, one of the founders of DeepMind and the publicly known face of it, is driven only by the awareness and stress that these achievements are milestones leading to a much bigger goal: Artificial Intelligence (AI) will actually help us understand the whole world.
Another success came with DeepMind’s artificial intelligence, which showed that Hassabis had reached the capacity to fulfill the role he had assigned him. AlphaFold, the latest version of DeepMind and its deep learning system, has overcome a major obstacle to the science of biology, as announced by the organizers of the long-running competition critical evaluation on protein structure predictions (CASP).
It can accurately predict protein structures up to an atom wide. “This is the first time an AI has been used to solve a serious problem.”the founder of the CASP team,” says John Moult of the University of Maryland.
What Is Protein?
A protein is a chain of amino acids that fold by itself in many complex ways. This structure determines what it will function, and learning what proteins do by experimenting in conditions where they work allows us to understand the simple mechanisms of life. Studies in the development of the covid-19 vaccine have focused on the virus’s spear (spike) protein, for example.
The adhesion of coronavirus to human cells depends on the structure of this protein and the protein structures in the cells to which it binds. Spear protein is only one of the billions of varieties among all living things, and only in human cells there are tens of thousands of different types of it.
At this year’s CASP competition, AlphaFold correctly predicted the structure of a dozen proteins with a margin of error of just 1.6 anstrom –that is, 0.16 nanometers, or atom width. This result exceeded the sensitivity of all other computational methods and provided information with the accuracy obtained by techniques used in the laboratory for the first time.
Cryogenic electron microscopy, nuclear magnetic resonance and X-ray crystallography methods, such as olcusen AlphaFold sized; these techniques are very expensive, slow years of work with the fact that the sums for each protein in the structure of a protein with high throw and artificial intelligence spending a few days in the great revolution in bulabilme more is revealed.
Making new drugs and understanding the causes of diseases
This revolution will help researchers a lot in making new drugs and understanding the causes of diseases. In the long term, thanks to this development, synthetic proteins and enzymes can be produced and more efficient work can be achieved on issues such as waste disposal and biofuel production. Researchers are making new discoveries through synthetic proteins to make agricultural products and plants more nutritious.
“It’s a great achievement that comes from a very fundamental level. It’s a development I never expected to be made so quickly, and it’s very surprising.”that’s what we’re talking about,” says Mohammed AlQuaraishi, a systems biologist at Columbia University.
“It’s a really big achievement. Like they did with Go.”that’s what David Baker, the head of the Institute for Protein Design at the University of Washington and the team behind Rosetta, a family of proetin analysis tools, says.
Defining the structure of a protein is very difficult. For most proteins, researchers have an amino acid sequence in their hands, but they do not know the structure in which they fold. And to estimate this structure, indeed, when you take into account each sequence, there are possibilities in very large astronomical numbers. Scientists have been dealing with this challenge since the 1970s, when Christian Anfinsen won the Nobel Prize for showing the sequences that determine the structure.
Accelerated by the creation of the CASP organization in 1994, decades of efforts were made to make progress by overcoming the great difficulties and slowness of laboratory work. Dozens of research groups from around the world competed with each other to find the right protein folding through software. The software developed through this CASP was already used by medical researchers, but the process was very slow.
Things started to change when DeepMind entered the CASP competition in 2018 with the first version of Alphafold. Although he could not produce results with the accuracy obtained in a laboratory, he dusted off other computational methods.
In this year’s competition, more than half of the applicants participated in the competition using various types of deep learning. Overall accuracy in the results was high. Baker’s new system, called trRosetta, was using some of DeepMind’s ideas in 2018. But he still finished second by a large margin.
After an evaluation using a global scaling Test between 0-100 in the competition, DeepMind AlphaFold scored 90 and above in two-thirds of the proteins given to it, and according to Team Leader John jumper, it was 25 points ahead of its nearest competitor in the prediction of the most difficult protein.
A result above 90 tells us that the margin of error between the predicted structure and the main structure can be reduced to the margin of error in laboratory conditions. A software error is almost non-existent. In addition, the predicted structure provides a valid alternative configuration that remains within the limit of the natural variations of the structure defined in the laboratory.
Jumper, the leader of the team, says that the incomplete structure of 4 proteins that the independent jury worked on in the lab in the competition can also be accurately predicted thanks to AlphaFold.
According to alquaraishi, the researchers could only reach the point achieved by this development of Alphafold within 10 years under normal circumstances. It’s a situation close to the physical limits of how accurate we can be, he says.
AlphaFold builds step by step on the work of researchers around the world. DeepMind also brings together a wide range of expertise, including biologists, physicists and computer scientists. Details of how it works will be shared in the CASP this week and will be published as an article in a special issue of the Journal Proteins next year. But we know for sure that this artificial intelligence uses some kind of attention network and develops a deep-learning technique that allows it to work on the problem from a wide range of angles. Jumper compares this approach to the formation of a puzzle: first connecting the regional pieces together and then ensuring their harmony with the whole.
DeepMind trained Alphafold with around 170,000 proteins from the public protein database. Multiple sequences in the bank were compared and amino acid pairs that converged together in folded structures were searched. This data was then used to estimate the distances of amino acid pairs within the three-dimensional structure, which is not yet known how they are. The training of artificial intelligence, which can also assess how accurate these estimates are, took only “a few weeks” with computers with computing power between 100 and 200 GPUs.
Dame Janet Thornton, of the European Institute of Bioinformatics in Cambridge, England, has been working on proteins for 50 years. His words at last week’s press conference explain his feelings: “I have been working since the emergence of this problem, and I have begun to think that it cannot be solved in my lifetime.”
Many drugs are studied using 3-D molecular structure simulations and designed to place molecules in target proteins like a key-lock. This process, of course, succeeds if the structure of this protein is known. “AlphaFold will open a whole new field in these studies, ” says Thornton, who says that this method is used for studies of only a quarter of the estimated 20,000 human proteins, leaving 15,000 drug targets that have not yet been touched.”
DeepMind plans to focus on all parasite-borne tropical diseases, particularly those associated with unknown proteins, including leishmania, sleeping sickness and malaria.
The only disadvantage of alphafold is that it is slower than other competing techniques. Alquraishi’s system uses an algorithm called repetitive geometric network (RGN), which finds protein structures millions of times faster, reducing them from a few days to a few seconds. “For those working on enzymes or bacteria, even a result with less accuracy would be very good news, and even what we have is better: a quick transition to pharmaceutical applications.”
Researchers are currently working on understanding and explaining exactly how Alphafold works. According to Baker: “once they can fully describe to the world how this system works, thousands of flowers will bloom. People will be able to use this information for a wide range of things, things we haven’t even dreamed of yet..”
Source: MIT Technology Review