Nobel Prize in Chemistry: Molecular blueprints from the computer

As the complete genetic material of more and more living beings was deciphered, people liked to talk about the blueprint of life. But strictly speaking, that was only part of the blueprint. It is known that three genetic “letters” (bases of DNA) code for one of the 20 amino acids from which all protein molecules in living organisms are made up. But when these amino acids combine to form a protein molecule, something remarkable happens in a fraction of a second: the linked amino acids fold into a very specific shape. This spatial structure is crucial for their biological function.

The US chemist Christian Anfinsen (Nobel Prize 1972) discovered that a certain sequence of amino acids always forms the same spatial structure. He already suspected in his Nobel Prize speech that one day it would be possible to predict the 3D structure of every protein from the sequence of the amino acid building blocks alone. However, it took another 50 years for this vision to come true.

David Baker

dpa/Ian C. Haydon

David Baker, born in 1962 in Seattle, researches at the University of Washington in Seattle.

One half of this year’s Nobel Prize in Chemistry goes to the developers of the AI ​​system “AlphaFold” – the Briton Demis Hassabis and the American John Jumper from the company Google DeepMind – “for predicting protein structures.” The American David Baker gets the other half “for computer-aided protein design.”

Until the beginning of the 21st century, there were only relatively complex experimental methods to clarify the spatial shape of protein molecules: X-ray crystallography (Nobel Prize 1962), nuclear magnetic resonance spectroscopy (Nobel Prize 2002) and cryo-electron microscopy (Nobel Prize 2017). Using these methods, tens of thousands of researchers have elucidated the 3D structure of around 200,000 proteins over the course of several decades. The data from these proteins was now used as learning material for artificial intelligence, which used what it had learned to find many times the number of protein structures in just a few years.

Demis Hassabis

dpa/Toby Melville

Demis Hassabisborn in London in 1976, is managing director of Google DeepMind in London.

This development began in 1994 with the project called “Critical Assessment of Protein Structure Prediction” (CASP), which developed into a competition. Every two years, researchers from around the world were given access to amino acid sequences of proteins whose structure had just been determined. However, the structures were kept secret from participants. The challenge was to predict the protein structures based on the known amino acid sequences.

There were many participants, but over the years the hit rate reached a maximum of 40 percent. The breakthrough only came in 2018, when DeepMind, a company that had previously worked in the field of strategic board games using artificial intelligence (AI) methods, joined. The AlphaGo program, developed under the leadership of Demis Hassabis, defeated a professional player for the first time in 2015. Hassabis, a chess player himself, first studied computer science and then received his doctorate in neuroscience. His AI program used learning processes that use so-called transformers. In machine learning, these program elements ensure that certain elements receive more “attention” than others.

The AI ​​proven in Go also delivered significantly better results for molecule folding than all previous methods for predicting protein structures. At least 60 percent of the structure predicted by AlphaFold was correct. The big jump to over 90 percent came from the physical-chemical knowledge of John Jumper, who joined DeepMind in 2017. With his help, AlphaFold2 was created. AlphaFold2’s program code and training data were disclosed free of charge by the company in 2021. Subsequently, the number of scientific publications on protein structures exploded. Jumper, the youngest Nobel Prize winner in Chemistry in 70 years, was amused in his first interview for the Nobel Prize website that, as a physicist, he was receiving the Chemistry Prize for research using artificial intelligence in the same year as two other physicists for their basic research on AI were honored with the physics prize.

John Jumper

dpa/I am Peter Catchpole

John Jumper, born in 1985 in Little Rock (USA), is a senior scientist at Google DeepMind in London.

But more than 20 years before DeepMind entered structural research, biophysicist David Baker had developed a software tool called Rosetta that modeled protein structures based on physical principles. Before AlphaFold’s success, Rosetta was one of the most successful participants in the annual CASP competition. But early on, Baker was less concerned with the structure of known proteins than with the synthesis of proteins with a desired structure. About novel molecules that can function, for example, as a drug or as a biocatalyst. One of the resulting nanoparticles was the basis for a Covid-19 vaccine.

When AlphaFold2 was announced, it was not yet known whether the program would be free to use and whether the program code would be disclosed. Therefore, Baker’s team, together with the South Korean computational chemist Minkyung Baek, revised the previous Rosetta. They incorporated some of the well-known tricks into their program. The new RoseTTAFold delivered almost as good results as AlphaFold2. A complementary tool helped Baker’s scientists simplify the design of novel proteins.

As revolutionary as the AI ​​approaches in protein research are, one problem remains unsolved, says Lucie Delemotte from the Royal Institute of Technology in Stockholm: the AI ​​algorithms do not explain the folding mechanism. Baker said in an interview after the Nobel Prize was announced: “That’s a big question. I think for the many applications that protein structure prediction has, it’s not that important how you get there.”

AI does not currently make traditional procedures completely superfluous, but it can simplify them considerably. Randy Read from the English University of Cambridge explained to the journal Nature that, for example, to interpret X-ray crystallography images, an initial assumption about the structure is necessary in order to correctly translate the patterns. A historical example of such an approach was the assumption of a double helix for the structure of the genetic molecule DNA by James D. Watson and Francis Crick. With the help of AI, the crystallographers already receive a well-founded basic model.