London, United Kingdom · December 2018
At the 13th Critical Assessment of Protein Structure Prediction (CASP13) competition, DeepMind's AlphaFold system achieved a result that sent shockwaves through the computational biology community. The artificial intelligence system, developed by the London-based subsidiary of Alphabet (then Google), placed first in the overall rankings, significantly outperforming all other competing teams in the most challenging categories of the blind, biennial assessment.
The result was announced in December 2018 at the CASP13 conference held in Cancún, Mexico, and represented the first time a deep learning approach had demonstrated such a decisive advantage in the decades-old problem of predicting protein three-dimensional structures from their amino acid sequences.
The Protein Folding Problem
Proteins are the molecular machines of life. Each protein is a chain of amino acids that folds into a specific three-dimensional shape, and this shape determines the protein's function. The sequence of amino acids — which is encoded by genes and can be read directly from DNA — determines the final folded structure, but understanding exactly how a linear chain of amino acids folds into a complex 3D shape has been one of biology's grand challenges for more than 50 years.
Knowing a protein's structure is critical for understanding disease mechanisms, designing drugs, and engineering enzymes for industrial applications. However, experimental methods for determining protein structures — primarily X-ray crystallography, nuclear magnetic resonance (NMR) spectroscopy, and cryo-electron microscopy — are expensive, time-consuming, and sometimes impossible for certain proteins. As of 2018, the Protein Data Bank contained structures for only a small fraction of known protein sequences.
Computational prediction offered a potential solution, but progress had been slow and incremental for over a decade.
CASP: The Blind Benchmark
The Critical Assessment of Protein Structure Prediction (CASP) is a community-wide experiment held every two years. Organisers collect protein sequences whose structures have been experimentally determined but not yet published. Competing teams are given the amino acid sequences and must predict the 3D structures — without access to the experimental answers. An independent panel of assessors then scores the predictions against the newly solved structures.
CASP is widely regarded as the gold standard for evaluating protein structure prediction methods. Its blind format ensures that teams cannot tune their methods to known answers, making strong performance a genuine indicator of methodological advance.
At CASP13, 98 target proteins were assessed across multiple categories. The most challenging category, Free Modeling (FM), involves proteins for which no similar known structures (templates) exist, requiring methods to predict the fold from first principles rather than by analogy.
AlphaFold's Approach
AlphaFold's success at CASP13 relied on two primary methodological pillars:
1. Co-evolutionary Analysis
During evolution, if two amino acids in a protein are in physical contact (i.e., close together in the 3D structure), mutations in one tend to be compensated by mutations in the other to maintain the contact. By analysing large databases of related protein sequences across species, AlphaFold could identify pairs of positions that co-evolve — a strong signal that those amino acids are in physical proximity. This co-evolutionary information provided a rough "contact map" of which parts of the protein chain are near each other in 3D space.
2. Deep Neural Networks
AlphaFold used deep convolutional neural networks to process the co-evolutionary data and predict the distances and orientations between amino acid pairs. The neural network was trained on known protein structures from the Protein Data Bank, learning to map from sequence-derived features to inter-residue distances and angles. These predicted distance distributions were then used as constraints in a gradient descent optimisation procedure to generate full 3D protein structures.
The Community's Reaction
DeepMind's entry into CASP13 caught the academic community by surprise. While DeepMind was famous for its game-playing AI systems — AlphaGo, which defeated the world champion at Go in 2016 — its application of deep learning to structural biology was unexpected.
The result prompted many research groups worldwide to adopt and extend deep learning approaches for protein structure prediction. The methods pioneered by AlphaFold at CASP13 — combining co-evolutionary analysis with neural network-based distance prediction — became a new baseline for the field, catalysing a wave of innovation that would ultimately transform structural biology.
Implications for Drug Discovery and Disease Research
The ability to predict protein structures computationally has far-reaching implications. Many diseases are caused by misfolded or malfunctioning proteins, and understanding protein structure is essential for rational drug design. Before AlphaFold, researchers often spent months or years trying to determine the structure of a single disease-relevant protein. The prospect of rapid, accurate computational prediction promised to dramatically accelerate the pace of biomedical research.
In the context of global health, protein structure prediction is particularly relevant for understanding pathogens. The spike protein of SARS-CoV-2, for instance, would later be characterised rapidly using structural biology tools, enabling the rapid development of vaccines — a process that computational structure prediction could accelerate further.
From CASP13 to AlphaFold 2
The CASP13 result motivated DeepMind to continue developing the architecture. Two years later, at CASP14 in late 2020, AlphaFold 2 achieved a median GDT score of 92.4 across all targets — a level of accuracy that the CASP organisers described as comparable to experimental methods. This result was widely hailed as having effectively solved the 50-year-old protein folding problem for single-chain proteins.
In July 2021, DeepMind published the AlphaFold 2 code and, in partnership with the European Bioinformatics Institute (EMBL-EBI), released a database of predicted structures for nearly all known proteins across multiple organisms — over 200 million structures by 2022. This resource has been used by hundreds of thousands of researchers worldwide and has been described as one of the most significant contributions of artificial intelligence to science.
Sources
- DeepMind, "AlphaFold: Using AI for scientific discovery," blog post, 2 December 2018 —
- Senior, A.W. et al., "Improved protein structure prediction using potentials from deep learning," Nature 577, 706–710 (2020)
- CASP13 website and results —
- Wikipedia, "AlphaFold" —
- Callaway, E., "'It will change everything': DeepMind's AI makes gigantic leap in solving protein structures," Nature 588, 203–204 (2020)
FIRAT Editorial Board
Institutional Research Desk · Foresight Institute of Research and Translation
The collective editorial and research translation board of FIRAT, synthesising peer-reviewed evidence, policy briefs, and division milestones across our seven foundational research pillars.
