2021: AlphaFold Predicts Protein Structures (2021)

In July 2021, DeepMind published AlphaFold2 in Nature and simultaneously released predictions
for the three-dimensional structures of nearly all proteins encoded by the human proteome—over
350,000 structures in the initial release, subsequently expanded to more than 200 million
structures covering virtually every known protein across the tree of life. AlphaFold2's
accuracy, as measured by Global Distance Test (GDT) scores at the CASP14 protein structure
prediction competition in December 2020, was so dramatically superior to all prior computational
methods that the field described it as a paradigm shift effectively solving a 50-year grand
challenge in biology.

The protein structure prediction problem—determining a protein's three-dimensional folded
conformation from its amino acid sequence—had resisted solution since Christian Anfinsen
demonstrated in 1961 that the sequence contains all information necessary for folding and
Cyrus Levinthal framed the computational paradox (the astronomical number of possible
conformations makes exhaustive search intractable). X-ray crystallography, cryo-electron
microscopy, and NMR spectroscopy could determine structures experimentally, but the process
required months to years per protein and could not be applied at proteome scale.

AlphaFold2 combined a multiple sequence alignment module—leveraging evolutionary co-variation
as a structural constraint—with an attention-based transformer architecture (Evoformer) that
iteratively refined predicted inter-residue distance and angle distributions, followed by a
structure module that assembled atomic coordinates. The system was trained on structures in
the Protein Data Bank.

For drug discovery, the implications are profound: structure-based drug design—previously limited
to the subset of targets with experimentally determined structures—became in principle applicable
to the entire druggable proteome. AlphaFold predictions have since been used to identify
cryptic allosteric pockets, guide fragment screening campaigns, and model protein-protein
interaction interfaces for PPI disruptors.