[ad_1]
Nearly five years ago, the Deepmind company made headlines around the world when its Alphago program defeated Lee Sedol, one of the world’s best players in the Chinese game of Go for over 3,000 years. Go, played with black and white stones on a checkered board, has some simple rules, but the number of possible combinations of stones on the board is one with 170 zeros behind it, or more than all the atoms in the entire universe.
Therefore, the game Go is about intuition and pattern recognition rather than calculations, that is, about subconscious processes that our brain handles well but that are very difficult for computers to handle. Until the match between Sedol and Alphago, such a victory was almost like the holy grail for artificial intelligence developers.
Now Deepmind has solved another more difficult and above all much more useful task with artificial intelligence. His Alphafold 2 program has been able to predict how proteins will fold. Here, the possible variants are even more incredibly numerous: an estimated one with 300 zeros behind it. The advance will also have important practical consequences for medical research and development.
– It is a very important step. Understanding protein folding is the key to understanding proteins. It’s a completely incomprehensible problem for humans, says Matthew Thompson, a researcher at the Enginzyme company in Solna.
Proteins are the building blocks of life. Our DNA contains the genes that are the picture of how the body should function and look, and the proteins put the picture into practice. A protein is a chain of amino acids and a gene is an instruction or recipe for what a particular protein should look like: which of the twenty possible amino acids should be included, in what order they should settle, and how long the chain should be.
Proteins do their work in the body. They control chemical reactions, build body structures, and are neurotransmitters, among other things. Hemoglobin, which binds oxygen to red blood cells and stains the blood, keratin, which forms nails, hair, and the outer layers of the skin, and insulin, which affects metabolism and regulates blood sugar levels are some of the around 20,000 proteins in the human body.
It is fairly easy to figure out what amino acid sequence a protein is made of. Researchers have been able to do this for decades. But that’s only a small part of the truth about protein. Much more important is the shape, or the unique three-dimensional structure to which the protein chain folds when the amino acids it contains interact.
– There are basically an infinite number of possible ways to twist and bend a protein. So predicting the three-dimensional shape based on the amino acid sequence alone is an almost impossible problem to solve, says Matthew Thompson.
That should still work predict the structure of proteins, said the Nobel Prize winner in chemistry Christian Anfinsen in 1972. Since then, researchers have tried it, without success. Instead, they have relied entirely on X-ray crystallography and similar techniques to measure the shape of proteins. It can be expensive and take several years as researchers have to prove it and it is not even always possible. Of the more than 200 million proteins that we know of, we only know the shape of a fraction.
Since 1994, CASP, the Critical Assessment of Protein Structure Prediction, has been held every two years, a competition between computer programs to correct unknown proteins.
– Groups receive protein sequences whose structures have not yet been published and must calculate their appearance. They then get a score between zero and one hundred that shows how well their calculated shape matches the actual structure, says Matthew Thompson.
Deepmind first entered the competition two years ago with the Alphafold program, the predecessor of Alphafold 2. They beat all other participants, but did not reach the score of 90, which is necessary for the program to be useful in practice. It also soon became clear to the Deepmind developers that the program could never go that far.
They started over, creating Alphafold 2, which has been trained on around 170,000 known protein structures. In this year’s competition, Alphafold 2 received an average result of 92.4 in all sub-competitions and 87.0 in the hardest category.
– They basically cleaned the house with the other equipment, and now they work almost as well as all direct measurement methods, says Matthew Thompson.
Deepmind exists since 2015 to Google. Matthew Thompson believes that’s one of the reasons Alphafold 2 is so much more successful than the other competing shows.
– This is probably due to Google’s great experience in machine learning from advertising. Basically, this is how Google makes money. The combination of that experience and a large number of researchers employed gives them both the capacity and the intellectual strength that is difficult to achieve in a regular research group at a university. So it’s a bit of an unfair fight, he says.
Advancement will mean a lot to many different branches of research.
– In the short term, we will be able to model protein structures much better than before. It is important when we develop new drugs or to understand how cells will react to them. It will also be much faster to solve problems.
Matthew Thompson himself is working on reconstructing known proteins so they can be used in the chemical industry.
“Having good models that show what proteins actually look like will allow us to improve their properties much more precisely,” he says.
But perhaps the most important Deepmind’s achievement is that it shows that artificial intelligence and machine learning are useful for solving really complex problems, says Matthew Thompson.
– Being able to transform a string of letters into what looks like a protein and what it does is such a high level of complexity. This indicates that it is really possible to solve these types of problems. And it’s not only possible to solve them, it’s the best available method to solve them, he says.
Deepmind has said that they will make the Alphafold 2 code available to everyone.
– If they keep their promise and publish the code, we will definitely see it. People will be able to use it to solve many other problems and take advantage of it. I guess we’ll see a lot of interesting things in the future, says Matthew Thompson.
Read more:
Nobel Prize: “The pandemic shows how much we need basic research”
Chemistry Laureates Lectured: Genetic Engineering Scissors Sophisticated and Simple Technology
Amina Manzoor: New Data Provides Reassuring News About Covid-19 Vaccine
