DeepMind creates a ‘transformational’ map of human proteins drawn by artificial intelligence

[ad_1]

The DeepMind AI research lab has created the most comprehensive map of human proteins to date using AI. The company, a subsidiary of Google’s Alphabet subsidiary, is releasing the data for free, with some scientists comparing the work’s potential impact with that of the Human Genome Project, an international effort to map every human gene.

Proteins are long, complex molecules that perform many tasks in the body, from building tissues to fighting disease. Its purpose was determined by its structure, which folds like origami into intricate and irregular shapes. Understanding how a protein folds helps explain its function, which in turn helps scientists with a range of tasks — from pursuing basic research on how the body works, to designing new drugs and treatments.

“The culmination of a full 10 years plus a lifetime of DeepMind”

Previously, protein structure determination relied on expensive and time-consuming experiments. But DeepMind showed last year that it can produce accurate protein structure predictions using an AI program called AlphaFold. Now, the company is releasing hundreds of thousands of predictions that the program has made to the public.

“I see this as the culmination of a full ten years of DeepMind’s life,” Demis Hassabis, CEO and co-founder of The Verge, told The Verge. “From the beginning, this is what we set out to do: to make breakthroughs in AI, to test that in games like Go and Atari, [and] Apply that to real-world problems, to see if we can speed up scientific breakthroughs and use them to benefit humanity.”

Two examples of protein structures predicted by AlphaFold (in blue) compared to experimental results (in green). Photo: DeepMind

There are currently about 180,000 protein structures available in the public domain, each of which is produced by experimental methods and can be accessed through the Protein Data Bank. DeepMind makes predictions about the structure of about 350,000 proteins across 20 different organisms, including animals like mice and fruit flies, and bacteria like Escherichia coli. (There is some overlap between the DeepMind data and pre-existing protein structures, but exactly what is difficult to determine due to the nature of the models.) Importantly, the release includes predictions for 98 percent of all human proteins, about 20,000 different structures, which are collectively known as the human proteome. It is not the first public data set for human proteomics, but it is the most comprehensive and accurate.

If they wish, scientists can download the entire human protein for themselves, says John Jumper, AlphaFold’s technical lead. “The HumanProteome.zip file is effectively there, and I think it’s about 50 GB in size,” Jumper tells The Verge. “You can put it on a flash drive if you like, although it won’t do you much without a computer to analyze it!”

“Anyone can use it for anything”

After releasing this first set of data, DeepMind plans to continue adding proteins to its repository, which will be held by Europe’s leading life sciences laboratory, the European Molecular Biology Laboratory (EMBL). By the end of the year, DeepMind hopes to release predictions for 100 million protein structures, a data set that will be “transformative to our understanding of how life works,” according to Edith Heard, EMBL’s general manager.

Hasbis says the data will be free forever for both scientific and commercial researchers. “Anyone can use it for anything,” the CEO of DeepMind noted at a press conference. “They just need to pay tribute to the people involved in the quote.”

Understanding protein structure is useful to scientists across a range of fields. The information could help design new drugs, synthesize new enzymes that break down waste, and create crops that are resistant to viruses or harsh weather. Already, DeepMind’s protein predictions are being used in medical research, including studying the mode of action of SARS-CoV-2, the virus that causes COVID-19.

“It will certainly have a huge impact on the scientific community.”

New data will speed up these efforts, but the scientists note that turning this information into realistic results will take time. “I don’t think it will be something that will change the way patients are treated during the year, but it will certainly have a significant impact on the scientific community,” said Marcelo C. Souza, a professor in the University of Colorado’s Department of Biochemistry.

Catherine Tuniasovunakul, senior researcher at DeepMind, says scientists will have to get used to having such information at their fingertips. “As a biologist, I can confirm that we don’t have basic evidence to even look at 20,000 structures, so this is [amount of data] Toniasofunakull told The Verge. “Analysing hundreds of thousands of structures – it’s crazy.”

In particular, though, DeepMind produces predictions of protein structures rather than experimentally determined models, which means that in some cases more work will be required to verify the structure. DeepMind says it spent a lot of time building accuracy metrics into its AlphaFold software, which ranks how confident each prediction is.

Example of protein structures predicted by AlphaFold. Photo: DeepMind

However, predictions of protein structures are still very useful. Determining the structure of a protein by experimental methods is expensive, time-consuming and relies on a lot of trial and error. This means that even a low-confidence prediction can save scientists years of work by guiding them in the right direction for research.

Helen Walden, professor of structural biology at the University of Glasgow, tells The Verge that the DeepMind data will “significantly ease” research bottlenecks, but that “the hard work and resource drain of conducting biochemistry and biological assessment, for example, of drug functions” will remain.

Souza, who has previously used data from AlphaFold in his work, says scientists will feel the effect right away. “In our collaboration with DeepMind, we had a data set with a protein sample that we’ve had for 10 years, and we never got to the point of developing a proper model,” he says. “DeepMind agreed to provide us with a chassis, and they were able to solve the problem in 15 minutes after we had been sitting on it for 10 years.”

Why is protein folding so difficult

Proteins are made up of chains of amino acids, which come in 20 different types in the human body. Since any individual protein can consist of hundreds of individual amino acids, each of which can fold and twist in different directions, this means that the final structure of the molecule has an incredibly large number of possible configurations. One estimate is that a typical protein can be folded 10^300 ways – that is, 1 followed by 300 zeros.

Protein folding has been a “big challenge” in biology for decades

Because proteins are too small to be examined with microscopes, scientists have had to indirectly determine their structure using expensive and complex methods such as nuclear magnetic resonance and X-ray crystallography. The idea of ​​determining the structure of a protein simply by reading the list of its constituent amino acids has long been theoretical but difficult to achieve, leading many to describe it as a “big challenge” to biology.

But in recent years, computational methods – particularly those using artificial intelligence – have suggested that such analysis is possible. Using these techniques, AI systems are trained on data sets of known protein structures and use this information to create their own predictions.

DeepMind’s AlphaFold software has significantly increased the accuracy of computational protein folding, as evidenced by its performance in the CASP competition. Photo: DeepMind

Many groups have been working on this problem for years, but DeepMind’s deep platform for AI talent and access to computing resources has allowed it to significantly accelerate progress. Last year, the company competed in an international protein folding competition known as CASP and blew up the competition. His results were so accurate that computational biologist John Molt, one of the founders of CASP, said that “the problem is somewhat [of protein folding] solved.”

DeepMind’s AlphaFold has been upgraded since last year’s CASP competition and is now 16 times faster. “We can fold the average protein in a matter of minutes, and in most cases seconds,” Hasbis says. The company also released the underlying code for AlphaFold last week as open source, allowing others to build on its work in the future.

Liam McGuffin, a professor at the University of Reading who has developed some of the UK’s leading protein folding programs, praised AlphaFold’s technical intelligence, but also noted that the program’s success relied on decades of previous research and public data. “DeepMind has enormous resources to update this database, and is in a better position to do so than any one academic group,” McGuffin told The Verge. “I think the academics would have gotten there eventually, but it could have been slower because we didn’t have enough resources.”

Why does DeepMind care?

Several scientists The Verge spoke to noted DeepMind’s generosity in releasing this data for free. After all, the lab is owned by Google subsidiary Alphabet, which has been pouring massive amounts of resources into commercial healthcare projects. DeepMind itself loses a lot of money every year, and there have been numerous reports of tensions between the company and its parent company over issues such as search independence and commercial viability.

Despite this, accounts tell The Verge that the company has always planned to make this information freely available, and that doing so is true to the founding spirit of DeepMind. He stresses that DeepMind’s work is used in many places at Google – “nearly anything you use, there’s some of our technology that’s part of that under the hood” – but the company’s primary goal has always been basic research.

“There are many ways in which value can be achieved.”

“The agreement when we got it is that we are here primarily to develop state of the art AI and AI technologies and then use that to accelerate scientific achievement,” Hasbis says. “[Alphabet] It has a lot of departments focused on making money,” he adds, noting that DeepMind’s focus on research “brings all kinds of benefits, in terms of prestige and goodwill for the scientific community. There are many ways in which value can be achieved.”

Hassabis anticipates that AlphaFold is a sign of things to come — a project that shows the huge potential of artificial intelligence to tackle problems as messy as human biology.

“I think we’re in a really exciting moment,” he says. “In the next decade, we and others in the field of artificial intelligence hope to make amazing breakthroughs that will really accelerate solutions to the really big problems we face here on Earth.”

[ad_2]