Since Google acquired it in six years, DeepMind has been touting a long list of artificial intelligence milestones. He has competed in the Go Championship, competed with StarCraft professional players, and has turned his attention to chess and shogi.
Except for her work in healthcare – which became part of Google Health in September 2019 – something that DeepMind has not been particularly sincere about is applying its practical form to more practical problems. There are some exceptions – DeepMind’s AI has already helped make Google’s data centers more energy-efficient and improve the firm’s text-to-speech system – but most of the titles that work on the game are O Focused on using the system as a basis to prove it.
Google Deepmind Tackle
But now Deep-Mind is starting to tackle one of science’s toughest problems: protein folding. An article published in the journal Nature describes how DeepMind’s AI system managed to defeat all its competitors, where the algorithm predicted its genetic makeup-based protein structure. Is. Being able to predict the structure of proteins can make it easier for us to develop new medicines, to understand how genetic mutations cause disease and produce artificial proteins.
“They blew the field,” says colleague Paul Bates, a leader of the Biomacultural Modeling Laboratory at the Francis Kirk Institute in London and critical critic of protein structure prediction (CASP) competition techniques. ” “We were all surprised that they did it as well because it was their first attempt at the field.”
Although CASP results were announced in December 2018 (the process of publishing results in scientific journals can be lengthy and cumbersome), DeepMind’s work on protein folding began two years ago, when some of the team members began researching the topic during a two-day hackathon. General Chat Chat Lounge Part of this focus, explains DeepMind’s CEO and co-founder Dames Matisse, is that protein folding is the kind of problem that AI is particularly well placed to help solve. Is.
“It seems that this issue is applicable to some kind of human awareness.” “If you think about turning a protein backbone, it’s something like moving a sport.”
The protein folding field is also well-established for training artificial intelligence agents. It has a huge dataset – Protein Data Bank, a repository of 3D structures and a genetic make-up of 150,000 proteins used to train DeepMind’s protein prediction system, Called alpha fold. There are even simulators that helped guide whether alpha fold was accurately predicting protein structure, as well as a useful test in the case of CASP competition.
Protein folding is a problem that – if solved – can have a big impact in many areas, including drug discovery, disease research, and synthetic protein preparation. “We try and find root node problems. If you solve them, it will open up new fields of research for us and other people,” says Mathews.
The structure of a protein signifies both its function and its pharmaceuticals or other molecules that are likely to interact with that protein. But right now we only have reliable structures for half the protein in the human body. This is why knowing the exact protein structure can help researchers to develop drugs that specifically target this protein only.
Protein Folding Prediction
Understanding protein folding will also help us design our own proteins, which can be used as drugs or even digest waste plastics. Butts says, “If you can get [Protein Folding Prediction] then you can start designing beyond nature, then you can design products that can be more beneficial as a drug. More specifically targeted. “
We already know how to determine the structure of a protein, but this involves the use of hard-working techniques such as X-ray crystallography and cryo-electron microscopy. Determination of the structure of a single protein can be masterly by one’s Ph.D. research. Being able to calculate the structure of a protein can speed up and reduce the cost of research. It can also help researchers develop new proteins without the need for protein. “It benefits from the kind of computational approach you can give,” says Pashmit Kohli, head of DeepMind’s science, robustness and reliability research.
Google DeepMind’s AI
But how accurate must it be before the protein modeling algorithm is useful? Although DeepMind’s AI was better than the competition, Alpha Food’s highly-predicted Scratch-derived proteins predicted 25 out of 43 proteins (managed by its closest competitors), There is still a long way to go in the real world. In order to be sufficiently accurate for real-world applications, AlphaFold will need to score a World Distance Score (GDT) – a measurement that measures CASP test accuracy – from 85 to 90. Is between. As of summer 2018, Alpha Fold’s overall average GDT was 63. (Google DeepMind)
“It’s too long to say before we say that we’ve solved [protein folding] in any meaningful way,” says Mathews. To get there, Deep Mind is thinking about enrolling better Alpha Fold in this year’s CASP Test. If it ever gets to the desired level of accuracy, then the next challenge will be to turn it into a product that people can use. “We’re still focusing on resolving the issue, and then we’ll know how to split it.”
“If you think about it, you realize that maybe (Google DeepMind) should do well because they are a big company, they are machine learning specialists, they know the tools of TensorFlow and they Extensive computing resources are available for calling. ” Bits say. He says the real test for Alpha Fold will be how it performs in this year’s test. If it manages to significantly improve its 2018 performance, then solving the problem of protein folding may not be too far off.