AlphaFold's Origin Story

DeepMind: the lab that plays games to change science

DeepMind was founded in London in 2010 by Demis Hassabis, Shane Legg, and Mustafa Suleyman. Hassabis is an unusual figure: a former teenage chess prodigy, then a pioneering video game designer (he created the AI for the game Theme Park at age 17), and then a neuroscience researcher at University College London. Google acquired DeepMind in 2014 for a reported £400 million.

DeepMind's philosophy is to use AI to "solve intelligence, and then use that to solve everything else." Their early work famously taught AI to play Atari games and later Go (AlphaGo beat world champion Lee Sedol in 2016). But Hassabis always had biology in mind — he had written his PhD thesis on memory systems and wanted to bring AI to bear on medicine.

In 2016, DeepMind formed a protein structure prediction team. Their hypothesis: the same deep learning techniques that had revolutionised image recognition and game-playing could learn the hidden patterns connecting sequence to structure — if given enough data and the right architectural ideas.

AlphaFold 1 at CASP13 (2018): a shot across the bow

AlphaFold entered CASP13 (held in December 2018) and immediately dominated, winning more target categories than any other group. But "dominating" is relative — AlphaFold 1 was roughly twice as accurate as the next best method on the hardest targets, which sounds impressive until you realise the next best method wasn't very accurate either.

AlphaFold 1's core insight was using evolutionary information — specifically, patterns of correlated mutations across thousands of related sequences — to infer which residue pairs are spatially close in the 3-D structure. (We'll explore this deeply in Part 4.) It used a convolutional neural network to predict inter-residue distances, then used gradient descent to find a 3-D structure consistent with those predicted distances.

CASP13 was a significant result — enough to make headlines in Nature and alarm the protein structure prediction community. But many experts remained sceptical that AI could truly "solve" the problem. They were in for a shock two years later.

AlphaFold 2 at CASP14 (2020): a seismic shift

Between 2018 and 2020, the DeepMind team — led by John Jumper, a computational biophysicist who had joined from the University of Chicago — completely rebuilt AlphaFold from scratch. The new architecture was fundamentally different: rather than predicting distances and then finding structures, AlphaFold 2 directly predicted atomic coordinates using a novel architecture called the Evoformer.

At CASP14 (results revealed in November 2020), AlphaFold 2's performance was shocking. The standard accuracy metric is the GDT_TS score: 100 means perfect, previous state-of-the-art was around 40–50 on the hardest targets. AlphaFold 2 scored 92.4 on average across all targets — comparable to experimental accuracy.

The assessors used words like "transformative" and "game-changing." Some described it as solving the protein folding problem, though this requires qualification — AlphaFold predicts single-chain structures and has limitations (which we'll cover in Part 7). But for the core task of predicting the 3-D structure of globular proteins from sequence alone, the problem was, in the words of CASP founder John Moult, "largely solved."

The aftermath: open science at scale

DeepMind published the AlphaFold 2 paper in Nature in July 2021, releasing both the code (open source on GitHub) and, crucially, a free database of predictions.

The initial AlphaFold Protein Structure Database (hosted by EMBL-EBI) launched with predictions for the entire human proteome (20,000+ proteins) and the proteomes of 20 other key organisms. By 2022 the database had been expanded to cover 214 million protein structures — essentially every protein in UniProt at the time.

Within months of the database launch, papers were appearing showing AlphaFold structures being used to understand antibiotic resistance, design new drugs, study pathogens affecting millions in the developing world, and illuminate the molecular basis of genetic diseases.

AlphaFold 3 (2024): beyond proteins

In May 2024, DeepMind published AlphaFold 3 in Nature, extending the system far beyond single proteins to predict structures of:

• Protein complexes — how multiple proteins interact
• Protein–DNA complexes — transcription factors gripping DNA sequences
• Protein–RNA complexes — ribosomes, spliceosomes, gene regulators
• Protein–small molecule complexes — drugs and their targets, bound together
• Protein–ion interactions — metal cofactors in enzymes

AlphaFold 3 uses a different internal architecture (replacing the Evoformer with a Pairformer, and using a diffusion model for final structure generation — more in Part 4). It is substantially more accurate than AlphaFold 2 on protein–ligand binding, which is directly relevant to drug discovery.

A free web server for AlphaFold 3 predictions is available at alphafoldserver.com, though with some restrictions on commercial use. Concurrently, Isomorphic Labs (a DeepMind spinout co-founded by Hassabis) uses an expanded internal version of AlphaFold 3 for pharmaceutical drug discovery.

The 2024 Nobel Prize in Chemistry

On 9 October 2024, the Royal Swedish Academy of Sciences awarded the Nobel Prize in Chemistry to three scientists:

The Nobel Committee's citation noted that the AlphaFold prize was for "protein structure prediction" and Baker's was for "computational protein design" — two complementary achievements that together are reshaping structural biology.

Full timeline