AlphaFold 2
The 2021 DeepMind paper that solved a 50-year-old grand challenge in biology by using AI to predict the 3D structure of a protein from its 1D amino acid sequence.
Paper: Highly Accurate Protein Structure Prediction with AlphaFold
Authors: John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David Reiman, Ellen Clancy, Michal Zielinski, Martin Steinegger, Michalina Pacholska, Tamam Berghammer, Sebastian Bodenstein, David Silver, Oriol Vinyals, Andrew W. Senior, Koray Kavukcuoglu, Pushmeet Kohli, Demis Hassabis · 2021
Read the paper
The Problem
Proteins are the molecular machines of life, and their function is dictated entirely by their 3D shape. Since 1972, biology's grand challenge (the "Protein Folding Problem") was figuring out how to predict this 3D shape using only the 1D string of amino acids that make up the protein. Determining structures experimentally (via X-ray crystallography) took months or years and millions of dollars. Previous computational methods, including AlphaFold 1, were highly inaccurate for novel proteins.
The Idea
DeepMind developed AlphaFold 2, a radically new architecture heavily based on Attention mechanisms. Instead of treating folding as a physics simulation (which requires astronomical compute), it treats it as an information processing problem. It uses Multiple Sequence Alignments (MSA) to find evolutionary patterns (if two amino acids mutate together across different species, they are likely physically touching in 3D space). It uses a novel "Evoformer" architecture to simultaneously update its understanding of the sequence and the physical distance between amino acids, iteratively refining the 3D structure.
How It Works
The AlphaFold 2 pipeline:
- Input: The 1D amino acid sequence.
- Database Search: It searches genomic databases to find similar sequences in other organisms, building a Multiple Sequence Alignment (MSA), and searches structure databases to find known "templates" (similar 3D chunks).
- The Evoformer: This is the core engine. It maintains two representations: a 1D representation of the sequence (MSA) and a 2D representation of the physical distances between every pair of amino acids. Through dozens of layers of attention, information flows back and forth between the 1D and 2D representations, ensuring they agree.
- Structure Module: A 3D transformer that takes the 2D distance map and outputs the exact (x,y,z) coordinates for every atom in the protein, along with a confidence score (pLDDT) for how accurate it believes each part of the prediction is.
Why It Mattered
AlphaFold 2 didn't just advance the field; it effectively solved the protein folding problem. At the CASP14 competition, it achieved a median accuracy of 92.4 GDT (anything above 90 is considered equivalent to experimental accuracy). It is widely regarded as the most significant scientific breakthrough achieved by AI to date, instantly accelerating drug discovery, disease research, and synthetic biology worldwide.
What Came After
DeepMind published the database containing the predicted structures of over 200 million proteins (nearly every protein known to science). They later released AlphaFold 3, which expanded the model's capabilities to predict how proteins interact with DNA, RNA, and drug molecules.