In January 2020 Nature published “Improved protein structure prediction using potentials from deep learning” by Andrew W. Senior, Richard Evans, John Jumper, Demis Hassabis, and colleagues at DeepMind. This is the paper that documents the first version of AlphaFold, the system that had stunned the structural biology community at the CASP13 competition in late 2018.
The core idea was to train a neural network to predict the distances between pairs of amino acid residues in a protein, along with the angles between chemical bonds. From those predictions the team built a “potential of mean force,” a scoring function describing how protein-like a candidate shape is. They could then fold a chain into a plausible structure by simple gradient descent against that potential, without the elaborate sampling procedures earlier methods relied on.
In the blind CASP13 assessment, this approach produced high-accuracy structures (template-modeling scores of 0.7 or higher) for 24 of 43 of the hardest “free modeling” targets, where no similar known structure exists to copy from. The next best method managed only 14 of the same 43. That margin was large enough to mark a genuine shift in how well computers could predict the protein-folding problem, a challenge that had stood for roughly 50 years.
AlphaFold 1 was the proof of concept; AlphaFold 2 a year later turned it into a near-solved problem. For a business or general reader, this paper is the documented starting point of the AI-for-biology wave: it showed that learned statistical patterns across known structures could outperform decades of hand-built physics and homology methods.