AlphaFold won its first CASP contest in 2018

AlphaFold is best known for its 2020 result at CASP14, but DeepMind’s first version had already won the previous contest. At CASP13 in December 2018, the original AlphaFold achieved the highest accuracy among participants, putting a large margin between itself and the second-placed group in the hardest free-modeling category. DeepMind published that earlier system in Nature in 2019 as “Improved protein structure prediction using potentials from deep learning.” The 2018 win signaled the approach was working; the 2020 CASP14 result, with a median score of 92.4 GDT, is what was hailed as solving the 50-year problem.