Ever since the 1970s researchers had tried to predict protein structures from amino acid sequences, but this would prove to be a hard nut to crack. In 2020 though, Demis Hassabis and John Jumper, whose company DeepMind had already been developing famous artificial intelligence (AI) programs that mastered complex board games like chess, were able to solve this 50-year-old problem – and the breakthrough was a fact. 

A few years earlier Demis Hassabis’ company DeepMind (founded in 2010) had constructed a computer program based on a convolutional neural network, which they called AlphaFold (AF1). It was trained on Protein Data Bank structures to produce a map of probability distributions for the distances based on multiple sequence alignments. From this map, a potential of mean force could be constructed and optimized by a gradient descent algorithm to generate structures. However, it did not provide an accuracy competitive with experimental structures for the majority of targets. But the next program, AlphaFold2 (AF2), would. Newly employed John Jumper took on a leading role in developing AF2. His previous experience of protein simulation had given him creative ideas about how to improve AF1 and together with Hassabis he co-led the work that fundamentally reformed the AI model. In this new program the neural network model used for AF1was entirely redesigned, the convolution approach was abandoned, and instead a transformer architecture was used, with the essential attention mechanism for learning which parts of the input are more important for the objective of network. The network is also of the end-to-end type, where atomic coordinates are directly produced as output rather than contact information, which had to be post-processed separately in AF1. 

“The AF2 architecture can been described as an ingenious piece of neural network with a multitude of new inventions, and it can be viewed as the first real scientific breakthrough of AI,” stated the Royal Swedish Academy of Sciences after the announcement.

“AI will make science faster”

Hassabis and Jumper had succeeded in solving the protein structure prediction problem for monomeric proteins to within a backbone accuracy of about 1 Å, and since 2020 more than two million people from 190 countries have used their program. The AF2 source code was also made public, which was a decisive contributor to its impact as it could be extensively tested and validated. 

“It is a key demonstration that AI will make science faster and ultimately help to understand disease and develop therapeutics. This is the work of an exceptional team at Google DeepMind [Google acquired DeepMind in 2014] and this award recognizes their amazing work,” stated John Jumper after receiving the news that he had won the Nobel Prize.

Captivated by chess and the universe

Demis Hassabis, born in 1976 in London, was captivated by chess really early on in life and at the age of 13 he reached Master status. In interviews and on X he has emphasized the importance of chess and playing games for his career and his life. It got him thinking about thinking and also provided him with money to buy his first computer. In his teens he also started working as a programmer and successful games developer. He later found Elixir Studios, producing games for Microsoft and Vivendi Universal, and contributed to many bestselling games. He then studied cognitive neuroscience at University College London, completing his PhD in 2009, and pursued postdoc work at Harvard and MIT. He used what he learned about the brain to develop better neural networks for AI and in 2010 he co-founded DeepMind with the modest goal of “solving the problem of intelligence”.

I’ve always thought if we could build AI in the right way, it could be the ultimate tool to help scientists.

“The reason I’ve worked on AI my whole life is that I’m passionate about science and finding out knowledge, and I’ve always thought if we could build AI in the right way, it could be the ultimate tool to help scientists, help us explore the universe around us,” he said in an interview with Adam Smith right after the announcement.

John Jumper, born in 1985 in Little Rock, Arkansas, USA, is the youngest awardee of the Nobel Prize in Chemistry in 70 years. He received his PhD in theoretical chemistry from the University of Chicago in 2017 and his thesis examined how to apply machine-learning techniques to the study of protein dynamics. He subsequently worked as a postdoctoral researcher before moving to Google DeepMind. John Jumper’s fascination with the universe was what made him start studying physics and mathematics according to the Royal Swedish Academy of Sciences.

You know, we need computation to solve the problems of biology. And I just love that it’s starting to work.

“You know, we need computation to solve the problems of biology. And I just love that it’s starting to work, and I can’t believe we’re getting recognition this fast for it,” he stated to Adam Smith after the announcement.