Editor: Editorial Department
Who would have thought that yesterday's story would repeat itself. The 2024 Nobel Prize in Chemistry was awarded to Demis Hassabis and John M. Jumper of DeepMind and David Baker of the University of Washington. Just in 2023, Hassabis and Jumper won the Lasker Prize, the "Nobel Prize Weather Vane", and today it really worked.
Just now, the 2024 Nobel Prize in Chemistry was announced.
Half of this year's award went to David Baker of United States University of Washington for his contributions to computational protein design.
另一半则授予DeppMind的Demis Hassabis和John M. Jumper,以表彰其在蛋白质结构预测方面的贡献。
All three of them are dedicated to the research of AI proteins.
That's right, everything can be AI, and this year's Nobel Prize is a proper year of AI
According to the commission's official announcement, this year's Nobel Prize in Chemistry focuses on proteins.
David Baker has succeeded in the almost impossible astonishing feat: building a whole new kind of protein.
Demis Hassabis and John Jumper developed AlphaFold 2 in 2020, solving a 50-year-old dream of predicting the complex structure of proteins.
We are now able to predict protein structures and design our own proteins, a breakthrough that has brought tremendous benefits to humanity and opened up endless possibilities.
Unexpectedly, our prophecy came true.
补充阅读：预定诺奖？DeepMind创始人斩获「诺奖风向标」拉斯克奖，AlphaFold成「AI for Science」标杆
The Lasker Awards have once again proven themselves to be the "bellwether of awards".
Hassabis also had a legendary day: wishing his good friend a Nobel Prize during the day and winning the Nobel Prize himself at night.
During his internship at DeepMind, Hassabis mentioned winning multiple Nobel Prizes when asked about the company's goals.
Today, that mission has been partially fulfilled.
Demis Hassabis
Demis Hassabis was born in England in 1976.
From the age of 4, he was a chess master, reaching the grandmaster standard at the age of 13.
He graduated from the University of Cambridge in 1997 with double sums, completed his PhD in Cognitive Neuroscience at University College London, and pursued postdoctoral studies at MIT and Harvard.
In 2010, he co-founded DeepMind with Shane Legg as CEO.
In 2023, following the merger of DeepMind and Google Brain, Hassabis will serve as the CEO of Google's DeepMind team.
There is also a legend about the birth of DeepMind: Hassabis, the world's second-ranked chess player under the age of 14, successfully conquered Peter Thiel at a party, received $2.25 million in venture capital and founded DeepMind.
Additional reading: OpenAI engineers revealed that it only took 8 days to develop ChatGPT! The long article reveals how Google, DeepMind and other top streams in Silicon Valley were born
John M. Jumper
John Jumper is a United States Senior Research Scientist at DeepMind.
He received his Ph.D. from the University of Chicago in 2017.
As a result, he and his colleagues co-founded AlphaFold.
In 2021, he was listed by Nature as one of the top 10 "important people" on its annual list.
David Baker
David Baker was born in 1962 in Seattle, Washington.
He received his B.A. from Harvard in 1984 and his Ph.D. in Biochemistry from UC Berkeley in 1989.
He is the "originator" of the protein design field, and designed RoseTTAFold, a protein structure design algorithm that is earlier than AlphaFold, and earlier than DeepMind.
As a member of the National Academy of Sciences of United States and director of the Institute for Protein Design at the University of Washington, he co-founded more than a dozen biotechnology companies and was named to Time's inaugural list of 100 most influential health people in 2024.
They reveal the secrets of proteins through computing and AI
Proteins play a key role in the vigorous and diverse chemical reactions that support the various activities of living organisms.
Proteins are typically made up of 20 amino acids, which can theoretically be combined in an infinite number of ways. Using the information stored in DNA as a blueprint, these amino acids are linked together to form long chains in our cells.
Next, the wonders of proteins come into play: the chains of amino acids twist and fold into a unique, and sometimes unique, three-dimensional structure that gives proteins their biological functions.
Proteins can be made up of a dozen to a few thousand amino acids, and the chains of amino acids fold into a specific three-dimensional structure that determines the function of the protein
Some proteins become the building blocks of life and can build muscles, horns, or feathers, while others may become hormones or antibodies.
Many of these proteins form enzymes that drive life's various chemical reactions with astonishing precision. Equally important are proteins located on the surface of cells, which act as signaling channels between cells and their surroundings.
The first images of proteins
Chemists have known since the 19th century that proteins are essential to life processes, but it wasn't until the 50s of the 20th century that chemical tools were precise enough that researchers could begin to explore proteins in more detail.
Cambridge researchers John Kendrew and Max Perutz made a breakthrough discovery in the late '50s when they succeeded in presenting the first three-dimensional model of a protein using a method called X-ray crystallography.
For this discovery, they were awarded the Nobel Prize in Chemistry in 1962.
Since then, researchers have been able to map some 200,000 different proteins, mainly using X-ray crystal imaging technology, and with a lot of effort, to lay the foundation for this year's Nobel Prize in Chemistry.
The Mystery of Protein Folding: A 50-Year Challenge
United States scientist Christian Anfinsen made another important early discovery.
Through various chemical techniques, he succeeded in unfolding an existing protein and then folding it again. Interestingly, the proteins take on the exact same shape every time.
In 1961, he concluded that the three-dimensional structure of proteins was determined entirely by amino acid sequences, a discovery that earned him the Nobel Prize in Chemistry in 1972.
However, Anfinsen's logic contains a paradox, which another United States scientist, Cyrus Levinthal, pointed out in 1969.
Levinthal calculated that even if a protein is composed of only 100 amino acids, it can theoretically exhibit at least 10^47 different three-dimensional structures.
If the chains of amino acids were randomly folded, it would take longer to find the correct protein structure than the age of the universe. In cells, however, this process takes only a few milliseconds. So, how exactly are amino acid chains folded?
Anfinsen's findings and Levinthal's skepticism point together to the fact that amino acid folding is a predetermined process. What's more, all the information about how the protein folds must be present in the amino acid sequence.
The new holy grail in biochemistry
This insight led to another decisive realization: if chemists knew the amino acid sequence of a protein, they should be able to predict the three-dimensional structure of the protein.
It's an exciting idea. If successful, they will no longer need to use complex X-ray crystallography techniques and can save significant time; In addition, it is possible to generate structures for all proteins that are not suitable for X-ray crystallography.
These logics and conclusions lead to the new holy grail of biochemistry: the prediction problem.
In order to promote the rapid development of the field, in 1994 researchers launched a project called "Critical Evaluation of Protein Structure Prediction" (CASP), which has since developed into a competition that is held every two years.
In the CASP competition, researchers from all over the world can get their hands on the amino acid sequences of just one batch of proteins. The structure of these proteins has just been determined, but it is kept strictly secret from the contestants. They need to predict protein structure based on known amino acid sequences.
CASP has attracted many researchers, but protein structure prediction has proven to be quite difficult, with progress being slow over the years, and there has been little improvement in heterogeneity between predicted and true structures.
It wasn't until 2018 that a breakthrough finally came when a chess grandmaster, neuroscience expert, and AI pioneer entered the field, bringing a fresh perspective to the long-standing puzzle.
That person is Demis Hassabis, co-founder of DeepMind.
AlphaGo Masters challenge the Protein Olympiad
Demis Hassabis is a great pioneer of today's AI deep learning.
He started playing chess at the age of 4 and reached the level of a grandmaster at the age of 13. As a teenager, he began his career as a programmer and game developer.
Subsequently, Hassabis began to explore AI and dabbled in neuroscience, making several revolutionary discoveries.
He firmly believes that AI neural networks are inseparable from the human brain. Therefore, he decided to use his knowledge of the brain to develop AI neural networks.
2010年,Hassabis与儿时的好友Mustafa Suleyman、以及研究同僚Shane Legg共同在英国创立了DeepMind。
At that time, they created an AI neural network that learned to play board games in a human-like way, and they became famous in one fell swoop.
This neural network can be plugged into an external memory, like a traditional Turing machine, allowing a computer to simulate a human's short-term memory.
In 2014, DeepMind was officially acquired by Google.
In the same year, Hassabis led the team to what is considered by many to be the holy grail of AI.
They developed the AlphaGo algorithm to beat the world champion of Go. Most well-known, AlphaGo won the first place in 2016 against the famous chess player Lee Sedol.
With that, the evolved version of AlphaGo defeated the human player again.
For Hassabis, however, Go is not the ultimate goal, but the means to develop better AI.
After this victory, they are ready to take on a more important challenge to humanity -
So, in 2018, Hassabis and his team signed up for the 13th Protein Structure Prediction Critical Assessment (CASP) competition.
Hassabis' AI model won unexpectedly
In the past few years, CASP researchers have achieved up to 40% accuracy in protein structure prediction.
After the Hassabis team took the AlphaFold to the competition, it achieved a 60% accuracy rate.
As a result, AlphaFold, which made its debut in the CASP competition, won the championship in 2018.
The original AlphaFold created high-precision structures for 24 of the 43 modeling domains. This result is enough to shock many people.
But for them, the potential of AlphaFold is far from being tapped.
To be truly successful, protein structure prediction must be 90% accurate.
As a result, Hassabis and his team continued to work hard.
However, no matter how hard they tried, they could not break through the technical bottleneck.
At that time, the team members were already exhausted.
Surprisingly, a newly hired employee, John Jumper, came up with a groundbreaking idea for improvements to AlphaFold.
John Jumper: Meeting the 'Grand Challenge' of Biochemistry
As the first author of AlphaFold, John Jumper, a senior research scientist at DeepMind, was named one of Nature's Top 10 People of the Year in 2021.
Once, his fascination with the universe prompted him to study physics and mathematics.
But in 2008, when he started working at a company that used supercomputing to model proteins and their kinetics, he began to realize that knowledge of physics could help solve medical problems.
In 2011, while pursuing his PhD in theoretical physics, he began to develop simpler, ingenious ways to model protein dynamics in order to save computing power.
In 2017, he completed his PhD and sent in a job application after hearing that Google's DeepMind was secretly developing technology to predict proteins.
Because of his experience in protein simulation, he had innovative ideas on how to improve AlphaFold, so he was promoted after the team hit a bottleneck.
Subsequently, he co-led the development of AlphaFold 2 with Hassabis.
The revamped AI model delivers amazing results
The new version of AlphaFold2 incorporates Jumper's in-depth knowledge of proteins.
The team also started using Transformer, giving them more flexibility than ever before to find patterns in massive amounts of data, effectively determining what to focus on for a specific goal.
They used a lot of information from databases of all known protein structures and amino acid sequences to train AlphaFold 2, which performed well in the 14th CASP competition.
When the organizers of the CASP evaluated the results in 2020, they realized: the 50-year challenge of biochemistry was over.
For the most part, the AlphaFold2 performs almost as well as X-ray crystallography, which is truly amazing.
How AlphaFold2 works
A textbook on cells changed the course of David Baker's life
When David Baker first entered Harvard, he chose philosophy and social sciences.
然而，在一门进化生物学课程中，他偶然接触到了经典教科书《Molecular Biology of the Cell》的第一版。 正是
This book completely changed the direction of his life.
Since then, he has been exploring cell biology and eventually developing a keen interest in protein structure.
In 1993, Baker joined the University of Washington as a research group leader and began to tackle this "grand challenge" in the field of biochemistry.
Through a series of ingenious experiments, he set out to explore how proteins fold. In the late 90s, he tried to develop a software that could predict the structure of proteins, and Rosetta was born.
In 1998, Bake made his first CASP competition with Rosetta and performed exceptionally well.
This also inspired him to have an innovative idea: to use the software in reverse.
If you can enter the desired protein structure and get a suggestion for the amino acid sequence, you can create a completely new protein, rather than just entering the amino acid sequence into Rosetta to get the protein structure.
Baker: A pioneer in designing proteins from scratch
In the late 90s of the 20th century, the field of protein design began to flourish.
In many cases, researchers have engineered existing proteins to perform new functions, such as degrading environmental pollutants or acting as catalysts in chemical manufacturing.
However, the functional range of natural proteins is limited after all. To break through this limitation and increase the potential of proteins, Baker's team came up with a bold idea: to design entirely new proteins from scratch.
Where does this idea come from? Baker once had a figurative metaphor:
"If you want to build an airplane, you don't start by modifying a bird; Instead, you'll gain a deep understanding of the fundamentals of aerodynamics, and then build entirely new aircraft based on those principles."
Baker's approach to protein design from the ground up ushers in a new era in protein engineering, opening up endless possibilities for future biotechnology and medical applications.
The birth of a unique protein: a breakthrough in design from scratch
The construction of a new protein is known as "de novo design".
Baker's team first mapped a protein with a completely new structure, and then used software called Rosetta to calculate the amino acid sequence that would produce the desired protein.
Rosetta first searched the database for all known protein structures, looking for short protein fragments that were similar to the target structure; Subsequently, the software leverages the basic knowledge of protein energy maps, optimizes these fragments, and proposes the final amino acid sequence.
To verify the software's effectiveness, Baker's team introduced genes corresponding to the designed amino acid sequences into the bacteria to produce the target protein. Subsequently, they used X-ray crystallography to determine the actual structure of the protein.
The results were exciting: Rosetta did indeed succeed in constructing the intended protein. The observed protein structure, named Top7, is almost exactly in line with their design, marking a major breakthrough in the field of protein engineering.
Top7 – the first protein that is completely different from all known existing proteins
Amazing creations from Baker Labs
For researchers in the field of protein design, the emergence of Top7 is undoubtedly a milestone.
Previously, attempts to design proteins from scratch were limited to mimicking structures that already exist in nature. The unique structure of Top7 is unprecedented in nature.
Even more impressively, it is composed of 93 amino acids, which is much larger than any protein previously produced using de novo design methods, which is equivalent to building a miniature "protein edifice" at the molecular scale.
Baker published this groundbreaking discovery in 2003 and generously made Rosetta's source code public, a move that greatly stimulated the continued development and innovative use of the software by the global research community, breathing new life into the field of protein design.
With these breakthroughs emerging, the outline of the 2024 Nobel Prize in Chemistry is already emerging.
What once took years is now just a matter of minutes
When Demis Hassabis and John Jumper confirmed that the AI protein structure prediction tool AlphaFold2 really worked, they began to calculate the structure of all human proteins.
Subsequently, they predicted the structure of almost all of the 200 million proteins that researchers would discover as they explored the Earth's biodiversity.
Not only that, but Google DeepMind has also exposed the code for AlphaFold2, which can be accessed by anyone.
Today, this AI model has become an invaluable resource for researchers. As of October 2024, more than 2 million users from 190 countries have used AlphaFold2.
Previously, it often took years to obtain a protein structure, and it was not always successful. Now it only takes a few minutes to complete.
While this AI model is not perfect, it can estimate the correctness of the resulting structure, so researchers can understand how reliable the predictions are.
Protein structure predicted using AlphaFold2
After the 2020 CASP competition, David Baker realized the potential of Transformer-based AI models.
He then added it to the Rosetta software, which led to the development of protein design.
In recent years, one amazing artificially engineered protein after another has sprung up from Baker's lab.
Artificial proteins designed using Rosetta software
The influence of AI has penetrated into all fields of the Nobel Prize, and I wonder if there will be surprises in the future.