From Move 37 to Nobel Prize: How AI Transformed Games and Revolutionized Science
DeepMind's AI evolved from AlphaGo, which stunned the world by defeating Go champion Lee Sedol, to AlphaFold, which solved protein folding and earned its creators a Nobel Prize.
In early March 2016, an extraordinary tournament took place in Seoul, South Korea, attracting millions of viewers worldwide. In a series of five matches of Go—one of the oldest and most complex strategy games—the world's top player at the time, Lee Sedol, competed against AlphaGo, a computer program developed by the company DeepMind. Before the tournament, experts confidently predicted an easy victory for the South Korean grandmaster, but it quickly became apparent they had severely underestimated AlphaGo.
The most dramatic moment of the tournament occurred on the second day. Lee Sedol had already lost the first match but remained determined to demonstrate his mastery in subsequent games. During the second match, he briefly stepped away from the playing hall for a short break to clear his mind. Meanwhile, AlphaGo played a move that forever changed the game of Go and is now considered one of the key milestones in artificial intelligence development.
On move 37, the program placed a stone unexpectedly on the fifth line, close to the edge of the board. Commentators fell momentarily silent before questioning if there had been an error. The move was so unconventional that no one anticipated it, as such moves simply were not part of the established repertoire of human masters.
Beyond Human Strategy: How AlphaGo Developed Novel Approaches
When Lee Sedol returned to the board, he stared at his computer opponent's unusual move in astonishment. He studied the board extensively, trying to grasp its significance. Initially, he could not understand AlphaGo's strategy, but soon he realized it was not a mistake but rather a brilliantly conceived plan. He recognized the strength of artificial intelligence, surpassing previous human comprehension of the game. Soon thereafter, he lost the second match as well. By the end of the tournament, he was greatly relieved to have managed at least one victory against the computer.
Move 37 became a symbol of a turning point in the relationship between artificial intelligence and humanity. AlphaGo did not base this remarkable move on analysis of historical human games. Instead, it developed the move independently through extensive self-play scenarios. The artificial intelligence played millions of matches against itself, experimenting with various approaches, and identified strategies through feedback that proved most successful. Thus, AlphaGo created entirely novel methods previously unknown to human players.
DeepMind and Demis Hassabis: From Games to Science
Demis Hassabis, the founder and head of DeepMind (now part of Google), had emphasized during the development of AlphaGo that programs like this represent just the first steps toward advanced artificial intelligence. As a youth, Hassabis was ranked second in the world in chess within his age group, providing him with deep insights into the significance of games as testing grounds for evaluating new artificial intelligence algorithms.
After returning from Korea, the DeepMind team embarked on a new, ambitious project. They decided to direct their AI expertise toward solving one of biology's greatest unsolved problems, which had puzzled scientists for decades. Similar to AlphaGo predicting moves in Go, they aimed to create AlphaFold—a program capable of predicting protein structures from amino acid sequences.
The Protein Folding Challenge: Biology's Grand Puzzle
Proteins can be thought of as miniature biological machines performing diverse, vital tasks within living organisms. For example, hemoglobin in blood binds and transports oxygen, insulin regulates glucose levels, and antibodies recognize and neutralize harmful invaders such as viruses and bacteria. Each protein's precise instructions are stored in the cell's genetic code within DNA molecules. These instructions determine the sequence of amino acids, linked together during synthesis much like beads on a necklace. The resulting amino acid chain spontaneously folds into a complex three-dimensional structure. This spatial form dictates the protein's function and its interactions with other molecules in biological processes.
For decades, predicting a protein’s three-dimensional structure from its amino acid sequence remained one of biochemistry's biggest mysteries. Solving this puzzle would allow scientists to precisely influence protein structure and function by strategically altering DNA. Such insight would open new avenues in medicine, biotechnology, and related sciences—from customized drug design to the creation of artificial enzymes with enhanced capabilities.
Revolutionary Results: AlphaFold Stuns the Scientific Community
In 2020, DeepMind presented an enhanced version of AlphaFold to the public. They tested their program in a new competition—this time not against human champions but in a contest of algorithms. The challenge was to determine the three-dimensional structures of proteins solely based on their amino acid sequences, using proteins whose structures were known to the competition organizers but not publicly disclosed.
AlphaFold's remarkable performance astonished the scientific community, similarly to AlphaGo’s achievement four years earlier. Its protein structure predictions matched experimental accuracy, representing a significant leap forward in molecular biology. Shortly thereafter, DeepMind used their new tool to determine structures for over 200 million proteins, making the entire dataset freely available to researchers worldwide.
In 2024, Demis Hassabis and John Jumper were awarded the Nobel Prize in Chemistry for developing this groundbreaking method of predicting protein structures from amino acid sequences. The research paper describing AlphaFold quickly became one of the most cited scientific publications of all time.