AlphaFold 2 subverts protein structure prediction and changes the history of science

Oxford University professor Matthew Higgins is grappling with a classic head-scratching question: What does protein actually look like?

Since 2005, his laboratory has been focusing on malaria-related issues.

Traditional techniques could only produce vague outlines of protein structures, which confused Higgins.

But by using a new artificial intelligence technology called AlphaFold 2, he deciphered the structure of a key protein used by the parasite that causes malaria.

This breakthrough helped him develop an experimental malaria vaccine that is now being tested in humans.

Malaria kills more than 6 million people each year, and these vaccines may hold the key to fighting the disease. Without AlphaFold, he said, we might still be hitting a brick wall.

It is easy to see from Higgins’ achievements that AlphaFold 2 is rapidly subverting science and medicine.

In just a few years, Alphabet’s artificial intelligence startup DeepMind has grown from winning the game of Go to solving biology’s grand challenges, and now it has been used by more than 100 Thousands of researchers use it, including researchers from universities and researchers from large pharmaceutical companies.

DeepMind CEO Demis Hassabis said in a podcast last year: "AlphaFold is surprising, but it is just the beginning."

From winning Go to changing the history of science

Nowadays, proteins are the main target of almost all drugs, so understanding the protein structure is the key to solving the problem of how to use specific methods to Key to intervening in disease phenotypes.

Before AlphaFold, finding the structure of a protein was a difficult task.

The traditional approach is for researchers to crystallize the protein, turning it into a salt form that the protein is resistant to. If this worked, they bombarded each crystal with X-rays and watched how electrons bounced off it to create the image.

By repeating this process, scientists can learn the 3D structure of a protein.

Higgins said it might take a PhD student a year or two to discover a new structure, but the results are often vague and uncertain.

DeepMind CEO Demis Hassabis is a chess prodigy and evangelist for artificial intelligence. He founded DeepMind in 2010 with the goal of building artificial intelligence systems that can perform certain tasks as well as or better than humans.

In 2016, DeepMind’s artificial intelligence system AlphaGo defeated world-class players in the Go game.

After his victory at Go, Hassabis and David Silver, a top scientist at DeepMind, decided it was time to move from competing at Go to solving problems. Real world problems now.

So they began to focus on protein problems, and decades of work by biologist John Moult paved the way for DeepMind to enter biology.

In 1994, he founded the CASP Protein Structure Prediction Competition (Critical Assessment of Protein Structure Prediction).

Contestants will be assigned the amino acid sequences of about 100 unknown proteins. The three structures of these proteins have been determined but have not been published.

Teams will have several months to develop and use mathematical models to solve these unknown structures. Moult rates their predictions for accuracy. On a scale of 100, a score above 90 indicates that the structure prediction is close to perfect.

DeepMind made its first public attempt at the 2018 CASP conference. The first version of AlphaFold won competitions and beat world standards. In competitions, winner prediction accuracy is typically around 40%, while AlphaFold's result was 60%.

Although this result is impressive, AlphaFold’s predictions have many errors and are not perfect yet. Hassabis wants to do better.

Months before the CASP results were released, John Jumper, one of the top scientists behind AlphaFold, was working with his team Planning together, we want to make incremental improvements to the technology.

Hassabis unexpectedly stopped them, probably meaning "Is it too difficult to solve this problem with the current model? Should we make separate models?"

After that conversation, Jumper abandoned the first version of AlphaFold and started from scratch. Jumper said, "AlphaFold 2 is built on the basis of more biological and physical knowledge of proteins."

At the CASP at the end of 2020, AlphaFold 2 handed in the answer sheet , the accuracy of predicting protein structure reached nearly 90%, which was much higher than other contestants. Experts believe it effectively solves the problem.

"At that moment, I knew we had changed the history of science," Jumper said.

Explosive growth in life sciences

In the months following CASP, DeepMind moves fast.

The team predicted all 20,000 proteins in the human body around Christmas 2020. The results were published in July 2021, along with the code for the software, in a seminal paper in Nature, which has been cited more than 8,800 times, or about 15 times a day.

Hassabis said that the decision to release AlphaFold 2 for free was to maximize the benefit of mankind.

According to CNBC, DeepMind, as a subsidiary of Alphabet, makes money by selling software and services to other Alphabet companies, such as YouTube and Google.

Then, Hassabis established the biotechnology start-up company Isomorphic Labs in 2021 to concentrate on researching drugs. Meanwhile, AlphaFold 2 has been running smoothly, releasing 200 million protein structure predictions last summer.

The pace of research is accelerating rapidly.

According to data from the biomedical research directory PubMed, only 4 papers referenced AlphaFold in 2020. This number will grow to 92 articles in 2021 and 546 articles in 2022. There will be more than 1,000 papers in 2023.

Accelerator for Drug Research

Several biotech companies are now using AlphaFold 2 to develop drugs.

“AlphaFold sparked a wave of innovation by showing people what was possible,” said Chris Bahl, chief scientist at AI Proteins, a Boston startup that also uses AlphaFold to help develop drugs. .

In 2019, Raphael Townshend worked at AlphaFold as a DeepMind intern while he was completing his computer science degree at Stanford University Doctorate of Science.

Now, he runs a startup called Atomic AI in San Francisco, hoping to develop what he calls the "AlphaFold of RNA."

RNA reads the instructions in our genetics (DNA) to create proteins in the body.

His company wants to predict the structure of RNA molecules and hopes to use these studies to develop drugs. Other biotech companies are also using AlphaFold in conjunction with other AI technologies to quickly and cheaply discover potential new drugs.

For example, startup Insilico Medicine is using its artificial intelligence system with AlphaFold to design molecules that block proteins associated with liver cancer. It created one of the molecules and used lab tests to confirm it worked. The company published the study in January.

The company’s CEO, Alex Zhavoronkov, claims that it took his team only about 50 days, less than 100, to go from finding the drug target to designing the drug and testing it in the lab. million, which he believes is a record for drug development.

Zhavoronkov has a photo of Hassabis in his office. “AlphaFold is a wonderful discovery, but it is part of a huge Lego puzzle that you need to have to successfully succeed. Bringing drugs to the market."

However, although this artificial intelligence technology makes drug development faster and easier, the company does not plan to use it due to the cost of clinical trials. Drugs advance into human studies because the process of testing them on animals and humans still takes many years and hundreds of millions of dollars.

What’s next

The potential of artificial intelligence in biotechnology is limited.

AlphaFold's predictions are not always perfect. The prediction model is very accurate in solving a small group of unknown proteins, but this does not guarantee that all predicted structures are correct.

Oxford University’s Higgins said he himself would use laboratory experiments to double-check AI predictions, so he is wary of research papers that rely entirely on AlphaFold’s predictions because There is a lack of experimental verification.

Despite these limitations, AlphaFold 2 is already a major breakthrough that has even sparked talk of a Nobel Prize, especially if it wins $3 million in 2022 After the award.

Pedro Domingos, professor of computer science at the University of Washington, said that the AlphaFold team’s research is deeper, such as how proteins interact with other proteins or small molecules. Such questions are very meaningful. .

Their research will become increasingly difficult in the future, and it is unclear whether AI will be able to handle the next research. But Domingos believes that DeepMind’s team is very good, so he is very optimistic about its future development.

DeepMind has done some research in genetics and predicting more complex protein interactions, but the big biological questions they will target next remain mysterious. It was revealed that the application of its technology by other institutions and companies in the future will be "increasingly difficult to grasp."

DeepMind’s Jumper said his AlphaFold team is focused on clearing the next big hurdle in biological research. But it remains a secret.

"I have my theories about where this might go, what kind of technology it is, and what the future might look like, and I won't reveal it."

The above is the detailed content of AlphaFold 2 subverts protein structure prediction and changes the history of science. For more information, please follow other related articles on the PHP Chinese website!