AlphaFold Predicts the Shapes of Proteins

Proteins perform vital biological tasks determined by the complex three-dimensional arrangements into which their amino acid chains fold. For six decades, mapping these shapes required painstaking laboratory techniques such as X-ray crystallography, nuclear magnetic resonance, and cryo-electron microscopy. While these methods successfully resolved roughly 170,000 structures, they left the overwhelming majority of known proteins uncharacterized. Computational methods attempted to bridge this gap, yet for decades struggled with accuracy on difficult targets. Founded in 1994, the Critical Assessment of Techniques for Protein Structure Prediction, a biennial blind competition, tracked this slow progress by scoring predictions on the global distance test, which runs from 0 to 100.

In 2018, the research company DeepMind entered the thirteenth cycle of the competition with AlphaFold, an artificial intelligence system trained on publicly cataloged protein structures. That initial iteration placed first overall by predicting contact distance distributions rather than simple contact maps. However, its software pipeline still relied on separate modules that paired machine learning inferences with traditional physics potentials. Although AlphaFold 1 demonstrated superior results on targets lacking homologous structural templates, it attained high accuracy on only a tiny fraction of its predictions, leaving open the broader question of how reliably computers could model complex macromolecules.