Deep Mind taught your AI to predict protein structure

The “ancestor” of AlphaFold is the AlphaGo algorithm, which began to play go better than any person. Source: DeepMind

The developers from Deep Mind over the past couple of years have become known for their many projects. In particular, they taught artificial intelligence (its weak form) to play Go, classic Atari titles and some other games that are difficult to “understand” by the machine. Now comes the turn of more serious classes - Deep Mind is gradually changing the specialization of AI on molecular biology.

More specifically, artificial intelligence is taught to predict the structure of a protein based on a fragment of the amino acid sequence - these building blocks of protein life. The project in question was named AlphaFold. AI taught to work faster and more accurately than people thanks to training on the basis of sequences collected by geneticists over several years.

In the competition Critical Assessment of Structure Prediction (CASP) , where it was necessary to predict the structure of the protein, artificial intelligence from Deep Mind won first place, becoming the leader among 98 participants. AI was able to correctly predict the structure of 25 of 43 proteins. In second place is the team that managed to correctly predict the structure of 3 out of 43 proteins. During the "competition" each team was sent a certain set of amino acids each month. This happened within a few months. The teams, having received all the elements, had to predict the structure of the protein that these amino acids make up. The structure was previously determined by scientists, so the organizers had the correct answer.

For science, studies of this kind are of paramount importance, since protein is the basis of life. Accordingly, by predicting the structure of a protein, one can learn to understand many biological functions and processes. It is worth noting that in some cases, scientists spend years predicting the structure of a certain protein. The problem is that DNA usually contains data about amino acid sequences, but not the structures that form chains of them.

The human body contains a huge number of varieties of protein. According to different estimates, it can reach several billion. Protein structures and even more - the number describes a number with 300 zeros. The 3D form of the protein depends on many factors - the number of amino acids, the length of the chain, etc. The spatial structure is also determined by the role that a certain protein plays in the human body.

For example, the cells of the heart are built with the help of protein, folded in such a way that the adrenaline molecules going through the human circulatory system are delayed and accelerate the pace of the heart muscle. Almost any of the abilities and capabilities of the organism depends on the form of a certain protein - from muscle contraction to vision.

The more complex the protein structure, the more difficult it is to model it. It is worth noting that some diseases considered to be a problem of the new century are caused by the erroneous folding of protein structures. Such diseases include, in particular, Alzheimer's disease, Parkinson’s disease, cystic fibrosis and Huntington’s disease.


Source: DeepMind

Understanding the structure of proteins of a certain type will allow you to create reagents that can actively act on these proteins. As a yuzkeys can be called the elimination of spilled oil or the creation of an inexpensive, rapidly decomposing plastic.

According toOne of the representatives of DeepMind, their research is a harbinger of a new era. The work is among those that solve the fundamental problems of both science and technology. It is worth noting that DeepMind experts started creating a new AI after their AlphaGo algorithm won the game of Go Lee Sedol, the world champion.

After this, the AI ​​taught computer games to be difficult to play, including “Revenge of Montezuma”. The developers say that their goal was never to get more points in any game to show the power of their AI. The real goal is the development of algorithms that can help a person solve problems of science and technology, such as protein structure and its prediction.

Scientists were able to teach AlphaFold to determine the distance between pairs of amino acids, as well as the chemical bond configuration. The second stage consisted in the search for the most energy-efficient structure of each putative protein. Now the algorithm takes only a few hours to complete the task — while people spend months or even years on the same thing.